Artificial

Patrick Krauss

Intelligence and
Brain Research
Neural Networks, Deep Learning
and the Future of Cognition
Artificial Intelligence and Brain Research
 Patrick Krauss

Artificial Intelligence
and Brain Research
Neural Networks, Deep Learning and
the Future of Cognition
Patrick Krauss
University of Erlangen-Nuremberg
Erlangen, Germany

ISBN 978-3-662-68979-0 ISBN 978-3-662-68980-6 (eBook)

https://doi.org/10.1007/978-3-662-68980-6

Translation from the German language edition: “Künstliche Intelligenz und Hirnforschung” by Patrick
Krauss, © Der/die Herausgeber bzw. der/die Autor(en), exklusiv lizenziert an Springer-Verlag GmbH, DE, ein
Teil von Springer Nature 2023. Published by Springer Berlin Heidelberg. All Rights Reserved.

This book is a translation of the original German edition “Künstliche Intelligenz und Hirnforschung” by
Patrick Krauss, published by Springer-Verlag GmbH, DE in 2023. The translation was done with the help of
an artificial intelligence machine translation tool. A subsequent human revision was done primarily in terms
of content, so that the book will read stylistically differently from a conventional translation. Springer Nature
works continuously to further the development of tools for the production of books and on the related tech-
nologies to support the authors.

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer-Verlag GmbH, DE,
part of Springer Nature 2024

This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the
whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustra-
tions, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar meth-
odology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does
not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors
or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer-Verlag GmbH, DE, part of Springer
Nature.
The registered company address is: Heidelberger Platz 3, 14197 Berlin, Germany

If disposing of this product, please recycle the paper.

For Sofie and Hannes
 Preface

How does Artificial Intelligence work? How does the brain function? What
are the similarities between natural and artificial intelligence, and what are
the differences? Is the brain a computer? What are neural networks? What is
Deep Learning? Should we attempt to recreate the brain to create real gen-
eral Artificial Intelligence, and if so, how should we best proceed?
We are in an extremely exciting phase of cultural and technological devel-
opment of humanity. Recently, Artificial Intelligence (AI) and Machine
Learning have been making their way into more and more areas, such as
medicine, science, education, finance, engineering, entertainment, and even
art and music, and are becoming ubiquitous in twenty-first-century life.
Particularly in the field of so-called Deep Learning, the progress is extraor-
dinary in every respect, and deep artificial neural networks show impressive
performance in a variety of applications such as processing, recognition, and
generation of images or natural language. Especially in combination with a
method called Reinforcement Learning, the networks are becoming increas-
ingly powerful, for example when it comes to playing video games, or they
even achieve superhuman abilities in complex board games like Go, when
they are trained by playing millions of games against themselves.
Many of the algorithms, design principles, and concepts used in AI today,
such as neural networks or the aforementioned reinforcement learning, have
their origins in biology and psychology. Therefore, neuroscience lectures
are becoming an integral part of courses such as computer science or artifi-
cial intelligence at more and more universities. But it is also worthwhile for
brain researchers to engage with artificial intelligence, as it not only provides
important tools for data evaluation, but also serves as a model for natural

vii
viii      Preface

intelligence and has the potential to revolutionize our understanding of the

brain.
When considering the goals of AI and neuroscience, it becomes appar-
ent that they are complementary to each other. The goal of AI is to achieve
cognition and behavior at a human level, and the goal of neuroscience is
to understand human cognition and behavior. One could therefore say that
artificial intelligence and brain research are two sides of the same coin. The
convergence of both research fields promises profound synergies, and it is
already certain that the insights gained from this will shape our future in a
sustainable way.
In recent years, I have given many lectures on these and related topics.
From the subsequent discussions and numerous follow-up questions, I
learned that the deep connection between AI and brain research is imme-
diately apparent to most, but was not really conscious before. Although
this is gradually beginning to change, most people associate AI exclusively
with degree programs such as computer science or data science, and less so
with cognitive science or computational neuroscience, even though these
branches of science can contribute a lot to basic research in AI. Conversely,
artificial intelligence has become indispensable in modern brain research.
To understand how the human brain works, research teams are increasingly
using models based on artificial intelligence methods, gaining not only neu-
roscientific insights, but also learning something about artificial intelligence.
There are already many excellent textbooks and non-fiction books in
which the various disciplines are each presented in isolation. However, an
integrated presentation of AI and brain research has not yet existed. With
this book, I want to close this gap. Based on exciting and current research
results, the basic ideas and concepts, open questions, and future devel-
opments at the intersection of AI and brain research are clearly presented.
You will learn how the human brain is structured, what fundamental mech-
anisms perception, thinking, and action are based on, how AI works, and
what is behind the spectacular achievements of AlphaGo, ChatGPT, and
Co. Please note that I am not aiming for a comprehensive introduction to
AI or brain research. You should only be equipped with what I consider to
be the theoretical minimum, so that you can understand the challenges,
unsolved problems, and ultimately the integration of both disciplines.
The book is divided into four parts, some of which build on each other,
but can also be read independently of each other. So there are different ways
to approach the content of this book. Of course, I would prefer if you read
the book as a whole, preferably twice: once to get an overview, and a second
 Preface     ix

time to delve into the details. If you want to get an overview of how the
brain works, then start with Part I. However, if you are more interested in
getting an overview of the state of research in Artificial Intelligence, then
I recommend you start with Part II. The open questions and challenges of
both disciplines are presented in Part III. If you are already familiar with the
basics and open questions of AI and brain research and are primarily inter-
ested in the integration of both research branches, then read Part IV.
I have tried to clarify complex issues through illustrative diagrams wher-
ever possible. My children have actively supported me in creating these dia-
grams. English quotes have been translated by me, unless otherwise noted.
Colleagues, friends, and relatives have greatly helped in correcting errors
and improving the clarity and readability of the text. I would like to thank
Konstantin Tziridis, Claus Metzner, Holger Schulze, Nathaniel Melling,
Tobias Olschewski, Peter Krauß, and Katrin Krauß for this.
My special thanks go to Sarah Koch, Ramkumar Padmanaban, and Ken
Kissinger from Springer Publishing, who have supported me in the realiza-
tion of this book project.
My research work has been and continues to be supported by the German
Research Foundation. I am grateful to those in charge. Without the inspir-
ing working atmosphere at the Friedrich-Alexander University Erlangen-
Nuremberg and the University Hospital Erlangen, many of my ideas and
research projects would not have been possible. My special thanks go to
Holger Schulze, Andreas Maier, and Thomas Herbst for their support, as
well as Claus Metzner and Achim Schilling for the countless inspiring con-
versations. I sincerely thank my father for the many discussions on the var-
ious topics of this book. My greatest thanks go to my wife, who has always
supported everything over the years and continues to do so. What I owe her,
I cannot put into words. I dedicate this book to my children.

Großenseebach Patrick Krauss

in April 2023
 Contents

1 Introduction 1

Part I Brain Research

2 The Most Complex System in the Universe 15
3 Building Blocks of the Nervous System 19
4 Organization of the Nervous System 27
5 Organization of the Cortex 41
6 Methods of Brain Research 53
7 Memory 59
8 Language 69
9 Consciousness 77
10 Free Will 97

Part II Artificial Intelligence

11 What is Artificial Intelligence? 107
12 How Does Artificial Intelligence Learn? 113
13 Game-playing Artificial Intelligence 125
14 Recurrent Neural Networks 131

xi
xii      Contents

15 Creativity: Generative Artificial Intelligence 139

16 Talking AI: ChatGPT and Co. 147
17 What are AI Developers Researching Today? 161

Part III Challenges

18 Challenges of AI 171
19 Challenges of Brain Research 179

Part IV Integration
20 AI as a Tool in Brain Research 191
21 AI as a Model for the Brain 197
22 Understanding AI Better with Brain Research 203
23 The Brain as a Template for AI 209
24 Outlook 219

Glossary 231
 1
Introduction

An elephant is like a fan!

The fifth blind man

ChatGPT Passes the Turing Test

In the field of Artificial Intelligence (AI), there have been a number of
spectacular breakthroughs in the last approximately 10–15 years—from
AlphaGo to DALL-E 2to ChatGPT –, which until recently were completely
unthinkable.
The most recent event in this series is certainly also the most spectacular:
It is already clear that November 30, 2022 will go down in history. On this
day, the company OpenAI made the Artificial Intelligence ChatGPT freely
accessible to the public. This so-called Large Language Model can gener-
ate any type of text in seconds, answers questions on any topic, gives inter-
views and conducts conversations, remembering the course of which and
thus responding adequately even in longer conversations. Since then, mil-
lions of people have been able to convince themselves daily of the amazing
capabilities of this system. The responses and texts generated by ChatGPT
are indistinguishable from those produced by humans. ChatGPT thus
becomes the first in the history of Artificial Intelligence to pass the Turing
Test, a procedure devised to determine whether a machine has the ability
to think (Turing, 1950). An artificial system that passes the Turing Test has

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 1
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_1
2    
P. Krauss

long been considered the Holy Grail of research in the field of Artificial
Intelligence. Even though passing the Turing Test does not necessarily mean
that ChatGPT actually thinks, you should still remember November 30,
2022 well. It not only represents the most important milestone in the his-
tory of Artificial Intelligence to date, but its significance is certainly com-
parable to the invention of the loom, the steam engine, the automobile, the
telephone, the internet, and the smartphone, which often only turned out to
be game-changers and decisive turning points in development in retrospect.

The Next Affront

In addition to the much-discussed consequences that ChatGPT and simi-
lar AI systems will have on almost all levels of our societal life, the aston-
ishing achievements of these new systems also challenge our explanations of
what fundamental concepts such as cognition, intelligence andconsciousness
mean at all. In particular, the influence that this new type of AI will have on
our understanding of the human brain is already immense and its impacts
are not yet fully foreseeable.
Some are already talking about the next great affront to humanity. These
are fundamental events or insights that have profoundly shaken man’s
self-understanding and his relationship to the world throughout history.
The Copernican affront, named after the astronomer Nicolaus
Copernicus, refers to the discovery that the Earth is not the center of the
universe, but revolves around the sun. This realization in the sixteenth cen-
tury fundamentally changed the world view and led to a loss of self-cen-
teredness and self-assurance. With the discovery of thousands of exoplanets,
i.e., planets outside our solar system, in recent decades, this affront has even
been intensified. This has shown that planetary systems are very common in
our galaxy and that there may even be many planets that orbit in the habita-
ble zone around their stars and thus represent possible places for life.
Another insight that affected man’s self-understanding was the Darwinian
affront. Charles Darwin’s theory of evolution in the nineteenth century
showed that man is not a species created by God, but has evolved like all
other species through evolution. This discovery questioned man’s self-under-
standing as a unique species separated from the rest of nature.
Another affront, which Sigmund Freud modestly named after the theory
he developed, the psychoanalytic affront, refers to the discovery that human
behavior and thinking are not always consciously and rationally controlled,
but are also influenced by unconscious and irrational drives. This realization
 1 Introduction    
3

shook man’s confidence in his ability for self-control and rationality. The
Libet experiments, which even question the existence of free will, further
intensified the impact of this affront.
AI can be considered as a newly added fourth major affront to human
self-understanding. Until now, our highly developed language was consid-
ered the decisive distinguishing feature between humans and other spe-
cies. However, the development of large language models like ChatGPT
has shown that machines are in principle capable of dealing with natural
language in a similar way to humans. This fact challenges our concept of
uniqueness and incomparability as a species again and forces us to at least
partially rethink our definition of being human.
This “AI affront” affects not only our linguistic abilities, but our cognitive
abilities in general. AI systems are already capable of solving complex prob-
lems, recognizing patterns, and achieving human-like or even superhuman
performance in certain areas (Mnih et al., 2015; Silver et al., 2016, 2017a,
b; Schrittwieser et al., 2020; Perolat et al., 2022). This forces us to reinter-
pret human intelligence and creativity, where we have to ask ourselves what
role humans play in a world where machines can take over many of our pre-
vious tasks. It also forces us to think about the ethical, social, and philo-
sophical questions that arise from the introduction of AI into our lives. For
example, the question arises as to how we should deal with the responsibility
for decisions made by AI systems, and what limits we should set on the use
of AI to ensure that it serves the good of humanity (Anderson & Anderson,
2011; Goodall, 2014; Vinuesa et al., 2020 ).
Less than half a year after the publication of ChatGPT, its successor GPT-
4 was released in March 2023, which significantly surpasses the performance
of its predecessor. This prompted some of the most influential thinkers in
this field to call for a temporary pause in the further development of AI sys-
tems, which are even more powerful than GPT-4, in a widely noticed open
letter1 to prevent a potentially impending loss of control.

Artificial Intelligence and Brain Research

The remarkable achievements of ChatGPT and GPT-4 also have direct
implications for our understanding of the human brain and how it func-
tions. They therefore not only challenge brain research, but even have the

1 https://futureoflife.org/open-letter/pause-giant-ai-experiments/
4    
P. Krauss

potential to revolutionize it. Indeed, AI and brain research have always been
closely intertwined in their history. The so-called cognitive revolution in
the middle of the last century can also be seen as the birth of research in
the field of AI, where it developed as an integral part of the newly emerged
research agenda of cognitive sciences as an independent discipline. In fact,
AI research was never just about developing systems to take over tedious
work. From the beginning, it was also about developing and testing theories
of natural intelligence. As we will see, some astonishing parallels between
AI systems and brains have been uncovered recently. Therefore, AI plays an
increasingly important role in brain research, not only as a pure tool for data
analysis, but especially as a model for the function of the brain.
Conversely, neuroscience has also played a key role in the history of arti-
ficial intelligence and has repeatedly inspired the development of new AI
methods. The transfer of design and processing principles from biology to
computer science has the potential to provide new solutions for current
challenges in the field of AI. Here too, brain research not only plays the role
of providing the brain as a model for new AI systems. Rather, a variety of
methods for deciphering the representation and calculation principles of
natural intelligence have been developed in neuroscience, which can now
in turn be used as a tool for understanding artificial intelligence and thus
contribute to solving the so-called black box problem. An endeavor occa-
sionally referred to as Neuroscience 2.0. It is becoming apparent that both
disciplines will increasingly merge in the future (Marblestone et al., 2016;
Kriegeskorte & Douglas, 2018; Rahwan et al., 2019; Zador et al., 2023).

Too Blind to See the Elephant

The realization that different disciplines must work together to understand
something as complex as human-level cognition is of course not new and is
vividly illustrated in the well-known metaphor of the six blind men and the
elephant (Friedenberg et al., 2021):

Once upon a time, there were six blind scientists who had never seen an ele-
phant and wanted to research what an elephant is and what it looks like. Each
examined a different part of the body and accordingly came to a different
conclusion.
The first blind approached the elephant and touched its side. “Ah, an ele-
phant is like a wall,” she said.
 1 Introduction    
5

The second blind touched the elephant’s tusk and exclaimed: “No, an ele-
phant is like a spear!”
The third blind touched the elephant’s trunk and said: “You are both
wrong! An elephant is like a snake!”
The fourth blind touched a leg of the elephant and said: “You are all wrong.
An elephant is like a tree trunk.”
The fifth blind touched the elephant’s ear and said: “None of you know
what you’re talking about. An elephant is like a fan.”
Finally, the sixth blind approached the elephant and touched its tail: “You
are all wrong,” he said. “An elephant is like a rope.”

If the six scientists had combined their findings, they would have come
much closer to the true nature of the elephant. In this story, the elephant
represents the human mind, and the six blind people represent the various
scientific disciplines that try to understand its functioning from different
perspectives (Fig. 1.1). The punchline of the story is that while each indi-
vidual’s perspective is valuable, a comprehensive understanding of cognition

Fig. 1.1 The Blind Men and the Elephant. Each examines a different part of the body
and accordingly comes to a different conclusion. The elephant represents the mind
and brain, and the six blind represent different sciences. The perspective of each indi-
vidual discipline is valuable, but a comprehensive understanding can only be achieved
through collaboration and interdisciplinary exchange
6    
P. Krauss

can only be achieved when the different sciences work together and
exchange ideas.
This is the founding idea of cognitive science, which began in the 1950s
as an intellectual movement referred to as the cognitive revolution (Sperry,
1993; Miller, 2003). During this time, there were significant changes in
the way psychologists and linguists worked and new disciplines such as
computer science and neuroscience emerged. The cognitive revolution was
driven by a number of factors, including the rapid development of personal
computers and new imaging techniques for brain research. These technolog-
ical advances allowed researchers to better understand how the brain works
and how information is processed, stored, and retrieved. As a result of these
developments, an interdisciplinary field emerged in the 1960s that brought
together researchers from a wide range of disciplines. This field went by vari-
ous names, including information processing psychology, cognition research,
and indeed cognitive science.
The cognitive revolution marked a significant turning point in the his-
tory of psychology and related disciplines. It fundamentally changed the way
researchers approach questions of human cognition and behavior, paving the
way for numerous breakthroughs in areas such as artificial intelligence, cog-
nitive psychology, and neuroscience.
Today, cognitive science is understood as an interdisciplinary scientific
endeavor to explore the different aspects of cognition. These include lan-
guage, perception, memory, attention, logical thinking, intelligence, behav-
ior and emotions. The focus is primarily on the way natural or artificial
systems represent, process, and transform information (Bermúdez, 2014;
Friedenberg et al., 2021).
The key questions are: How does the human mind work? How does cog-
nition work? How is cognition implemented in the brain? And how can cog-
nition be implemented in machines?
Thus, cognitive science addresses some of the most difficult scientific
problems, as the brain is incredibly difficult to observe, measure, and manip-
ulate. Many scientists even consider the brain to be the most complex sys-
tem in the known universe.
The disciplines involved in cognitive science today include linguistics,
psychology, philosophy, computer science, artificial intelligence, neuro-
science, biology, anthropology, and physics (Bermúdez, 2014). For a time,
cognitive science fell somewhat out of fashion, particularly the idea of inte-
grative collaboration between different disciplines was somewhat forgotten.
Specifically, AI and neuroscience developed independently and thus also
away from each other. Fortunately, the idea that AI and brain research are
 1 Introduction    
7

complementary and can benefit greatly from each other is currently expe-
riencing a real renaissance, with the term “cognitive science” apparently
being interpreted differently in some communities today or considered too
old-fashioned, which is why terms like Cognitive Computational Neuroscience
(Kriegeskorte & Douglas, 2018) or NeuroAI (Zador et al., 2023) have been
suggested instead.
The legacy of the cognitive revolution is evident in the many innovative
and interdisciplinary approaches that continue to shape our understanding
of the human mind and its functioning. Whether through state-of-the-art
brain imaging techniques, sophisticated computer models, or new theoreti-
cal frameworks—researchers are constantly pushing the boundaries of what
we know about the human brain and its complex processes.

Brain-Computer Analogy
Many researchers believe that computer models of the mind can help us
understand how the brain processes information, and that they can lead to
the development of more intelligent machines. This assumption is based
on the brain-computer analogy (Von Neumann & Kurzweil, 2012). It
is assumed that mental processes such as perception, memory, and logical
thinking involve the manipulation of mental representations that correspond
to the symbols and data structures used in computer programs (Fig. 1.2).
Like a computer, the brain is capable of receiving, storing, processing, and
outputting information.2
However, this analogy does not mean that the brain is actually a com-
puter, but that it performs similar functions. By considering the brain as a
computer, one can abstract from biological details and focus on the way it
processes information to develop mathematical models for learning, mem-
ory, and other cognitive functions.
The brain-computer analogy is based on two central assumptions that
underlie cognitive science. These are computationalism and functionalism.

2A fundamental difference is that a computer processes information with different components

than those with which it stores the information. In the brain, both are done by the—sometimes
same—neurons.
8    
P. Krauss

Input Representation Transformation Output

Linguistic
utterances are
a form of
behaviour

Fig. 1.2 Brain-Computer Analogy. Information processing includes the input, rep-
resentation, transformation, and output of information. For a computer, the input
may come from the keyboard, for a biological organism from the sensory organs.
This input must then be represented: by storing it on a hard drive or in the comput-
er’s RAM, or in the brain as momentary neuronal activity in short-term memory or
in long-term memory in the interconnection of neurons. Then a transformation or
processing takes place, i.e., mental processes or algorithms must act on the stored
information and change it to generate new information. For a computer, this could
be text processing, for humans, for example, logical reasoning. Finally, the result of
information processing is output. The output can be, for example, via a printer for
a computer. In living beings, the output corresponds to observable behavior or, as a
special case of behavior, to human linguistic utterances

Computationalism
In computationalism, it is assumed that cognition is synonymous with
information processing, i.e., that mental processes can be understood as
calculations and that the brain is essentially an information processing sys-
tem (Dietrich, 1994; Shapiro, 1995; Piccinini, 2004, 2009). Like any such
system, the brain must therefore represent information and then transform
these represented information, i.e., there must be mental representations of
information and there must be mental processes that can act on these rep-
resentations and manipulate them. Computationalism has greatly influenced
the way cognitive scientists and researchers in the field of artificial intelli-
gence think about intelligence and cognition.
However, there is also criticism of this view, as evidenced by numerous
ongoing debates in philosophy and cognitive science. Some critics argue, for
example, that the computer model of the mind is too simple and cannot
 1 Introduction    
9

fully capture the complexity and richness of human cognition. Others argue
that it is unclear whether mental processes can really be understood as calcu-
lations or whether they fundamentally differ from the way processes occur in
computers.

Functionalism
Is cognition only possible in a (human) brain? Functionalism clearly answers
this question with a no. Accordingly, mental states and processes are defined
exclusively by their functions or their relationship to behavior, not by their
physical or biochemical properties (Shoemaker, 1981; Jackson & Pettit,
1988; Piccinini, 2004). What does this mean in concrete terms?
Please imagine a car in your mind’s eye. And now remember the last sit-
uation in which you ate chocolate, and try to recall the taste as accurately
as possible. Did you succeed? I assume you did. As I write these lines, I
have brought to mind the same two mental states “seeing a car” and “tast-
ing chocolate”. Obviously, each of us can activate the corresponding mental
representations in our brains, even though you, I, and every other reader of
these lines have completely different brains. All human brains are of course
similar in their basic structure. But they are certainly not identical down to
the smallest detail, especially not in the exact wiring of the neurons, if only
because every person has had completely different, individual experiences in
their life, which affect the wiring pattern of the brain. In computer science
terminology, one would say that each person has a different, individual hard-
ware. Yet we can all bring to mind the same mental state.
While in the previous example the systems were somehow very similar—
they were always human brains—the following example may illustrate how
much the different physical implementations of the same algorithm can dif-
fer from each other. Consider the addition of two numbers. The representa-
tion of these numbers, as well as the associated process or algorithm to add
them, can be implemented in your brain when you “calculate in your head”,
or for example also in a laptop with spreadsheet program, a slide rule, a cal-
culator or a calculator app on your smartphone. Each time, the same num-
bers are represented and added, while the information processing systems are
completely different. This is the concept of multiple realizability.
Accordingly, the same mental state or process can in principle be real-
ized by completely different natural or artificial systems. Put simply, this
means that cognition and presumably also consciousness can in principle
be implemented in any physical system capable of supporting the required
10    
P. Krauss

Fig. 1.3 Functionalism. Human-level cognition is not limited to a human brain,

but could in principle also be implemented in any other system that supports the
required calculations, such as correspondingly advanced robots or aliens

calculations. If many different human brains are already capable of this, why
should this ability be limited to humans or biological systems? From the per-
spective of functionalism, it is therefore quite possible that the ability for
human-like cognition can also be implemented in correspondingly highly
developed machines or alien brains (Fig. 1.3).

Conclusion
In recent years, spectacular advances in artificial intelligence have turned
our understanding of cognition, intelligence, and consciousness upside
down and will have profound impacts on society and our understanding of
the human brain. Cognitive science is the key to a deeper understanding of
brain and mind, and computer models of the mind can help us understand
how the brain processes information and contribute to the development of
smarter machines. These models are based on the central assumptions of
computationalism and functionalism, which emphasize the equivalence of
cognition and information processing as well as the independence of cogni-
tive processes from their physical implementation.
The advances in artificial intelligence have also led to the fields of neu-
roscience and computer science becoming increasingly intertwined. The
transfer of construction and processing principles from biology to computer
science promises new solutions for current challenges in artificial intelli-
gence. Conversely, the close collaboration of these disciplines will become
increasingly important in the future to understand complex systems like the
human brain.
 1 Introduction    
11

The recent advances in artificial intelligence and their applications have

opened the door in neuroscience to new insights and technologies far
beyond what was previously possible. We are only at the beginning of a new
era of research and innovation, and it remains to be seen what fascinating
discoveries and developments await us in the future.

References
Anderson, M., & Anderson, S. L. (Hrsg.). (2011). Machine ethics. Cambridge
University Press.
Bermúdez, J. L. (2014). Cognitive science: An introduction to the science of the mind.
Cambridge University Press.
Dietrich, E. (1994). Computationalism. In thinking computers and virtual persons
(pp. 109–136). Academic.
Friedenberg, J., Silverman, G., & Spivey, M. J. (2021). Cognitive science: An intro-
duction to the study of mind. Sage.
Goodall, N. J. (2014). Machine Ethics and Automated Vehicles. In G. Meyer & S.
Beiker (Eds.), Road vehicle automation. Lecture notes in mobility. Springer. https://
doi.org/10.1007/978-3-319-05990-7_9.
Jackson, F., & Pettit, P. (1988). Functionalism and broad content. Mind,97(387),
381–400.
Kriegeskorte, N., & Douglas, P. K. (2018). Cognitive computational neuroscience.
Nature neuroscience,21(9), 1148–1160.
Marblestone, A. H., Wayne, G., & Kording, K. P. (2016). Toward an integration of
deep learning and neuroscience. Frontiers in computational neuroscience,10, 94.
Miller, G. A. (2003). The cognitive revolution: A historical perspective. Trends in
cognitive sciences,7(3), 141–144.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G.,
… & Hassabis, D. (2015). Human-level control through deep reinforcement
learning. Nature,518(7540), 529–533.
Perolat, J., De Vylder, B., Hennes, D., Tarassov, E., Strub, F., de Boer, V., … &
Tuyls, K. (2022). Mastering the game of Stratego with model-free multiagent
reinforcement learning. Science,378(6623), 990–996.
Piccinini, G. (2004). Functionalism, computationalism, and mental contents.
Canadian Journal of Philosophy,34(3), 375–410.
Piccinini, G. (2009). Computationalism in the philosophy of mind. Philosophy
Compass,4(3), 515–532.
Rahwan, I., Cebrian, M., Obradovich, N., et al. (2019). Machine behaviour.
Nature,568, 477–486.
12    
P. Krauss

Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S.,
… & Silver, D. (2020). Mastering Atari, Go, Chess and Shogi by planning with
a learned model. Nature,588(7839), 604–609.
Shapiro, S. C. (1995). Computationalism. Minds and Machines,5, 517–524.
Shoemaker, S. (1981). Some varieties of functionalism. Philosophical topics,12(1),
93–119.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G.,
… & Hassabis, D. (2016). Mastering the game of Go with deep neural networks
and tree search. Nature,529(7587), 484–489.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., …
& Hassabis, D. (2017a). Mastering the game of Go without human knowledge.
Nature,550(7676), 354–359.
Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., … &
Hassabis, D. (2017b). Mastering Chess and Shogi by self-play with a general rein-
forcement learning algorithm. arXiv preprint arXiv:1712.01815.
Sperry, R. W. (1993). The impact and promise of the cognitive revolution.
American Psychologist,48(8), 878.
Turing, A. M. (1950). Computing machinery and intelligence. Mind,59(236),
433–460.
Vinuesa, R., Azizpour, H., Leite, I., Balaam, M., Dignum, V., Domisch, S., … &
Fuso Nerini, F. (2020). The role of artificial intelligence in achieving the sustain-
able development goals. Nature Communications,11(1), 233.
Von Neumann, J., & Kurzweil, R. (2012). The computer and the brain. Yale
University Press.
Zador, A., Escola, S., Richards, B., Ölveczky, B., Bengio, Y., Boahen, K., … &
Tsao, D. (2023). Catalyzing next-generation artificial intelligence through
NeuroAI. Nature Communications,14(1), 1597.
 Part I
Brain Research

In the first part of the book, the aim is to familiarize you with the most
important aspects of the structure and function of the brain. In doing so,
a detailed and systematic description of many molecular biological, physio-
logical, and anatomical details is deliberately avoided. The presentation also
makes no claim to completeness. Interested readers may deepen their knowl-
edge with one of the many excellent textbooks available on psychology and
neuroscience. Rather, these first chapters are intended to convey the basics
necessary from the author’s point of view, on the basis of which we want
to show the numerous cross-connections to Artificial Intelligence in later
chapters.
 2
The Most Complex System in the Universe

There is always a bigger fish.

Qui-Gon Jinn

The Brain in Numbers

The human brain consists of approximately 86 billion nerve cells, known as
neurons (Herculano-Houzel, 2009). These are the fundamental processing
units responsible for the reception, processing, and transmission of informa-
tion throughout the body. The neurons are connected via so-called synapses
and form a gigantic neural network. On average, each neuron receives its
input from about 10,000 other neurons and sends, also on average, its out-
put to about 10,000 subsequent neurons (Kandel et al., 2000; Herculano-
Houzel, 2009). The actual number of connections per neuron can vary
significantly, over several orders of magnitude, which is why we also speak
of a broad distribution of connections per neuron. Some neurons, such as
those in the spinal cord, are only connected to a single other neuron, while
others, for example in the cerebellum, can be connected to up to a million
other neurons.
Based on the total number of neurons and the average number of con-
nections per neuron, the total number of synaptic connections in the brain
can be roughly estimated at one quadrillion. This is a number with 15 zeros
and can also be written as 1015. In recent years, we have become somewhat

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 15
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_2
16    
P. Krauss

accustomed to amounts beyond a thousand billion, i.e., in the trillion range,

in the context of politics and economics. The approximate number of syn-
apses in the brain is a thousand times larger!
This may sound like a lot, but let’s consider how many synapses would
theoretically be possible in the brain. Each of the approx. 1011 neu-
rons could in principle be connected to every other, with the information
between any two neurons potentially running in two directions: either from
neuron A to neuron B or vice versa. Additionally, each neuron can indeed
be connected to itself. These special types of connections are called autapses.
Purely combinatorially, this results in 1011 times 1011, or 1022, as the possi-
ble number of synapses. A comparison with the number of actually existing
synapses shows that only about one in 10 million theoretically possible con-
nections is actually realized. The network that the neurons form in the brain
is therefore anything but dense (dense), but on the contrary extremely sparse
(sparse) (Hagmann, 2008).

How Many Different Brains Can There Be?

In reference to the genome, which refers to the entirety of all genes of an
organism, the connectome is the entirety of all connections in the nervous
system of a living being (Sporns et al., 2005). To answer the question of how
many different brains there can be, one must estimate how many different
connectomes are combinatorially possible. At this point, it should be noted
that not every theoretically possible connectome must result in a functioning
viable nervous system. It turns out that it is quite complicated to calculate
the exact number, which is why we want to settle here with an estimate for
the lower limit of the actual number based on some simplifications. Let’s
assume for simplicity that each of the 1022 theoretically possible connections
can either be present or not present. So we assume binary connections, with
one representing an existing and zero a non-existing connection. As we will
see later, reality is even more complicated. But even under this strong simpli-
fication, an absurdly high number of 210 22 (read “two to the power of ten to
the power of 22”) results. This corresponds to a number with a trillion zeros.
Of course, not every one of these connectomes leads to a powerful and via-
ble nervous system, so the realizable number should only correspond to a
tiny subset of all theoretically possible connectomes. On the other hand, the
synaptic weights are not binary numbers, but can take any continuous value,
which significantly increases the number of possibilities again.
 2 The Most Complex System in the Universe    
17

How Many Different Mental States are

Possible?
Let’s now consider how many different brain or mental states there can be.
Why is this important, you ask? Well, one of the central assumptions of mod-
ern cognitive and neuroscience is that every mental state (mental state) is based
on a neuronal state (brain state) or a whole sequence of brain states. In other
words, everything we can think, feel and experience has a specific neuronal
correlate, i.e., an activation or even a temporal sequence of activations of neu-
rons in our brain. This is also referred to as spatiotemporal activation patterns.
In this estimation, we again assume simplification. At any given moment,
a specific neuron can either be active, i.e., send an action potential, which
corresponds to one, or it does not do this and is thus inactive, which cor-
responds to zero. So here too, we end up with the binary system. As a fur-
ther simplification, we divide the continuous flow of time into the smallest
meaningful units. The typical duration of an action potential is one milli-
second, i.e., one thousandth of a second. This is the characteristic time scale
of the brain. Therefore, in each millisecond, each of the 1011 neurons in
the brain can either be active or inactive. This results in a purely combina-
torial 210 11 different activation patterns per millisecond. Of course, we do
not have a new thought, perception or feeling impression every millisecond
or a thousand times per second. It is assumed that what we experience as
now, the present or the moment, lasts about three seconds (Pöppel, 1997),
which corresponds to 3000 milliseconds. Each mental state would therefore
correspond to a sequence of 3000 different activation patterns within our
11
simplifications. Since there can be (210 )3000 different such sequences, our
estimate results in just as many different brain states.

Conclusion
If you feel like the author, then you can’t imagine anything meaningful
under these orders of magnitude, as these numbers are essentially approach-
ing infinity. While we were able to rewrite the number of possible brains
or connectomes with a number with a trillion zeros, the number of men-
tal states can no longer be put into words, as we would already have to use
exponential notation for the number of digits of this number.
We can try to at least somewhat classify these absurd orders of magnitude.
Apart from infinity, what is the largest number that still has a meaningful
18    
P. Krauss

meaning in the natural sciences? It is 1082. That’s about how many atoms
there are in the observable universe (Eddington, 1931). Or in other words:
there are significantly fewer of the smallest in the largest than there are possi-
ble brains and mental states.
We can rightly claim that the human brain is probably the most com-
plex system in the universe—that we know of. Of course, it is by no means
excluded that there are far more complex systems somewhere in the vast-
ness of the universe, such as the natural or possibly also artificial information
processing systems of a highly developed species or artificial intelligence.

References
Eddington, A. S. (1931). Preliminary note on the masses of the electron, the pro-
ton, and the universe. In Mathematical Proceedings of the Cambridge Philosophical
Society (Vol. 27, No. 1, pp. 15–19). Cambridge University Press.
Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Van Wedeen,
J., et al. (2008). Mapping the structural core of human cerebral cortex. PLoS
Biology, 6(7), 1479–1493.
Herculano-Houzel, S. (2009). The human brain in numbers: A linearly scaled-up
primate brain. Frontiers in Human Neuroscience, 3, 31.
Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S., Hudspeth, A. J., &
Mack, S. (Eds.). (2000). Principles of neural science (Vol. 4, pp. 1227–1246).
McGraw-Hill.
Sporns, O., Tononi, G., & Kötter, R. (2005). The human connectome: A structural
description of the human brain. PLoS Computational Biology, 1(4), 0245–0251.
Pöppel, E. (1997). A hierarchical model of temporal perception. Trends in Cognitive
Sciences, 1(2), 56–61.
 3
Building Blocks of the Nervous System

There is not the slightest reason to doubt that brains are nothing more than machines
with a huge number of parts, functioning in perfect accordance with the laws of physics.
Marvin Minsky

Neurons and Synapses

Neurons (nerve cells) are the fundamental processing units of every known
nervous system (Kandel et al., 2000). The basic structure and function are
the same in all neurons (Fig. 3.1). Like most other cells, they consist of a cell
body with a nucleus, which contains the genetic information. What makes
them unique, however, are the dendrites. These are many small, highly
branched, antenna-like protrusions on the cell surface, through which the
nerve cell receives signals from other nerve cells. Another characteristic is the
axon, which represents the output channel of each neuron, through which
it transmits signals to subsequent nerve cells. The axon is also usually highly
branched at its end and forms the so-called telodendron (Kandel et al.,
2000).
The coupling between two nerve cells, i.e., the place where a branch
of the telodendron or axon of the predecessor neuron meets a dendrite of
the successor neuron, is called a synapse. Characteristic of the structure of
the synapse is a small gap with a width of about 10 to 50 nanometers, the
so-called synaptic gap, which separates the presynaptic neuron, which sends
signals, from the postsynaptic neuron or target cell, which receives signals.

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 19
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_3
20    
P. Krauss

Cell body
Telodendron

Dendrites

Axon

Cell nucleus
Myelin

Fig. 3.1 Biological neuron. Characteristic of the neuron are two types of extensions
of the cell membrane. The dendrites receive signals from other neurons and pass
them on to the cell body. Action potentials are transmitted to other neurons via the
axon. In many neurons, the axon is surrounded by a myelin sheath, which acts like a
kind of insulation and increases the signal conduction speed

When an electrical signal, a so-called action potential, reaches the end of the
presynaptic neuron, it triggers the release of chemical messengers, so-called
neurotransmitters, into the synaptic gap. These neurotransmitters diffuse
through the synaptic gap and then bind to receptors on the postsynaptic
neuron or target cell. So, there is a chemical transmission. Depending on
the number of receptors and the amount of released neurotransmitters, the
synapses can be of varying strength, i.e., they can transmit a signal more
or less well and thus have a varying influence on the activity of the succes-
sor neuron. In addition, there are basically two different types of synapses,
which can either stimulate or inhibit the activity of the postsynaptic cell:
excitatory synapses, which favor the activation of the successor neuron,
and inhibitory synapses, which rather prevent the activation of the succes-
sor neuron. Formally, a real number can be assigned to each synapse, where
the absolute value corresponds to the signal transmission strength and the
sign corresponds to the type of synapse: positive for excitatory and negative
for inhibitory. This number is called synaptic weight or short weight, often
abbreviated as w. In the entirety of the synapses of the nervous system, more
precisely: in the resulting network and the resulting directed information
flow, all the information within a nervous system is stored, i.e., all knowl-
edge as well as memories, learned skills, innate behaviors and reflexes, and
even character traits (Kandel et al., 2000).
All signals arriving at the dendrites are transmitted at the respective syn-
apses and weighted with the value of the synapse, which mathematically
 3 Building Blocks of the Nervous System    
21

corresponds to a multiplication of the signal with the continuous synaptic

weight. The information transfer from the synapse to the cell body thus
uses an analog encoding. Finally, all weighted signals in the cell body of the
neuron are spatially and temporally integrated into the so-called membrane
potential. If this exceeds a certain threshold, the neuron becomes active and
fires an action potential or spike at the axon hillock (Koch, 1999). At the
same time, the membrane potential, after a short refractory period of 2 to
4 milliseconds during which no further action potential can be generated,
is reset to the resting membrane potential. The action potential corresponds
to a brief change in the electrical potential, triggered by ions entering the
cell. Since each action potential always has about the same time course, the
same strength, and the same duration of about one millisecond, the out-
put sequence of action potentials of a neuron corresponds to a quasi-digi-
tal code. In each millisecond, either an action potential can be sent out or
not. The information is represented in the temporal rate (frequency) and the
exact temporal sequence of the action potentials. The emitted action poten-
tial runs along the axon until it is finally transmitted again via synapses to
further subsequent nerve cells. In the nervous system, there is thus a con-
stant change from analog encoding through chemical transmission and qua-
si-digital encoding through electrical transmission. It is not yet fully clarified
whether the simultaneous use of both types of encoding is necessary to ena-
ble cognitive functions in the brain, or whether purely analog or purely dig-
ital encodings can also lead to comparable performances (Kolb & Whishaw,
1989, 1998; Kandel et al., 2000).
The synaptic weights of a neuron to all its successors have—although this
would of course be mathematically possible and is usually the case in artifi-
cial neural networks—not different, but all the same sign. This means that a
neuron only releases one type of neurotransmitter at its synapses. In contrast
to artificial neural networks, there are in biological neural networks (with
few exceptions) exclusively purely excitatory or purely inhibitory neurons—a
property known as Dale’s principle1 (Dale, 1935; Strata & Harvey, 1999).

1 No rule without exception: Actually, the effect depends on the receptor at the postsynapse. In the
dopaminergic system, for example, the same transmitter can have opposite effects (Missale et al., 1998).
Neurons have also been discovered that, contrary to Dale’s Principle, release two different transmitters
at their synapses (Vaaga et al., 2014).
22    
P. Krauss

Neuroplasticity
Learning is synonymous with the experience-dependent change in the con-
nection structure of the neural networks in the brain, which is referred to
as neuroplasticity. Various types of changes can be distinguished (Kolb &
Whishaw, 1989, 1998; Kandel et al., 2000).
Pruning refers to the process of eliminating unused or weak synap-
tic connections between neurons in the brain. This thinning is an impor-
tant mechanism for the development and refinement of neuronal circuits
in the brain. During the development of the brain, neurons form synaptic
connections on a large scale to enable a multitude of synaptic connections.
However, through the process of pruning, only the most frequently used
connections between neurons are reinforced and maintained, while unused
or weak connections are eliminated. This leads to neural networks with more
efficient and specific connections. Pruning does not only occur during the
development of the brain, but can also occur in adulthood. It is believed
that pruning can help to relieve the brain by removing no longer needed or
redundant synaptic connections to free up resources for important connec-
tions. Moreover, pruning plays a crucial role in imprinting learning during
the development of the brain (Kandel et al., 2000).
Synaptic plasticity refers to the process of strengthening or weakening
the synaptic connection between neurons by enhancing or reducing activity
(Kandel et al., 2000). Various mechanisms are discussed that could under-
lie this type of plasticity. The Hebb’s rule states that the synaptic connec-
tion between two neurons is strengthened when the presynaptic neuron is
active and simultaneously activates the postsynaptic neuron. In other words:
“Cells that fire together wire together” (Hebb, 2005). So, if a neuron repeat-
edly fires and another neuron fires at the same time, then the synapse and
thus the connection between the two neurons is strengthened. This is also
referred to as long-term potentiation. Conversely, synaptic connections are
weakened if the neurons are rarely or never active at the same time, which is
also referred to as long-term depression.
The so-called Spike Timing Dependent Plasticity (STDP), on the other
hand, describes how the synaptic connection between neurons changes
due to the exact temporal sequence of their activity (Gerstner et al., 1996).
Accordingly, the synapse is only strengthened when the presynaptic neuron
is active shortly before the postsynaptic neuron, and weakened when the
presynaptic neuron is active shortly after the postsynaptic neuron. In a way,
STDP can be considered an extension of Hebb’s rule. The dependence on
 3 Building Blocks of the Nervous System    
23

the temporal sequence of neuronal activity corresponds to the consideration

of causality. A neuron that already sends an action potential before the pre-
decessor neuron did is either randomly active or was excited by another neu-
ron. Therefore, in this case, it makes sense to weaken the synapse rather than
strengthen it.
While long-term potentiation and depression act activity-dependent and
thus locally on certain synapses, there is another type of synaptic plasticity,
which works rather globally and is referred to as synaptic scaling (Tononi &
Cirelli, 2003; De Vivo, 2017). This means that all input synapses of a neu-
ron are weakened proportionally to their respective strength. The purpose of
this process, which mainly takes place during sleep, is to prevent individual
synapses from becoming increasingly stronger and the sum of all inputs of a
neuron from becoming too large, causing the neuron to fire too frequently.
A similar purpose is served by intrinsic plasticity, through which the
activity of a neuron can be changed (Daoudal & Debanne, 2003; Zhang &
Linden, 2003). The goal is to maintain the long-term average activity level
of a neuron at a medium level. From an information-theoretical point of
view, a neuron that never fires is just as pointless as a neuron that constantly
fires. Both processes, synaptic scaling and intrinsic plasticity, are collectively
referred to as homeostatic plasticity and ensure to maintain a balance in the
neural network, which is optimal for information processing (Desai, 2003;
Marder & Goaillard, 2006).
Finally, neurogenesis and myelin plasticity also count as neuroplasticity.
Neurogenesis refers to the brain’s ability to form new neurons, which in
the adult brain only occurs in the hippocampus. Myelin plasticity refers to
changes in the myelin sheath, which surrounds the axons as electrical insula-
tion and influences the speed of signal transmission.

Glia Cells
Not to be left unmentioned are the glia cells, which make up more than half
of the volume of neural tissue. Although this important type of cell cannot
generate action potentials itself, it still plays a significant role in the func-
tion of the nervous system. Their main functions include the myelination of
neurons to accelerate signal transmission and fix the position of neurons. In
addition, glia cells provide neurons with nutrients and oxygen, fight path-
ogens, remove dead neurons, and recycle neurotransmitters (Kandel et al.,
2000).
24    
P. Krauss

Recent research suggests that glia cells may also play an active role in neu-
ral processing. For example, it has been found that astrocytes are involved in
the regulation of neurotransmitter levels, in maintaining the correct chemi-
cal environment for neural signal transmission, and even in modulating syn-
aptic transmission (Clarke & Barres, 2013; Sasaki et al., 2014). Microglia,
on the other hand, have been associated with pruning and neural develop-
ment (Schafer et al., 2012).
In some theoretical works, computer models have been simulated in
which glia cells, especially astrocytes, play a role in the generation and mod-
ulation of neuronal oscillations. In these models, it is assumed that glia cells
act as a kind of buffer for extracellular potassium ions (K+) released by neu-
rons during action potentials. By absorbing excess K+ ions and modulating
extracellular ion concentrations, glia cells could potentially regulate neuronal
excitability and synchronization, which could contribute to the emergence
of neuronal oscillations (Wang et al., 2012).

Conclusion
The brain has a remarkable ability to flexibly adapt to new circumstances
and change its structure. The potential implications of the various types of
neuroplasticity for the development of new learning algorithms in Artificial
Intelligence are considerable. By studying the learning and adaptability of
the brain, researchers can develop new algorithms that can learn and adapt
in a similar way to the human brain. The exploration of synaptic plasticity,
i.e., the strengthening or weakening of connections between neurons, has
for example led to the development of artificial neural networks that mimic
the structure and function of the human brain and are capable of learning
from examples and adapting their behavior over time.
The exploration of structural plasticity, which involves changes in the
physical structure of neurons and their connections, can also lead to the
development of algorithms that can restructure in response to new data and
experiences. This could lead to more flexible and adaptable AI systems that
can learn from new data and adjust their behavior accordingly.

References
Clarke, L. E., & Barres, B. A. (2013). Emerging roles of astrocytes in neural circuit
development. Nature Reviews Neuroscience, 14(5), 311–321.
 3 Building Blocks of the Nervous System    
25

Dale, H. H. (1935). Pharmacology and nerve-endings. Proceedings of the Royal

Society of Medicine, 28, 319–332.
Daoudal, G., & Debanne, D. (2003). Long-term plasticity of intrinsic excitability:
Learning rules and mechanisms. Learning & Memory, 10(6), 456–465.
De Vivo, L., Bellesi, M., Marshall, W., Bushong, E. A., Ellisman, M. H., Tononi,
G., & Cirelli, C. (2017). Ultrastructural evidence for synaptic scaling across the
wake/sleep cycle. Science, 355(6324), 507–510.
Desai, N. S. (2003). Homeostatic plasticity in the CNS: Synaptic and intrinsic
forms. Journal of Physiology – Paris, 97(4–6), 391–402.
Gerstner, W., Kempter, R., Van Hemmen, J. L., & Wagner, H. (1996). A neuronal
learning rule for sub-millisecond temporal coding. Nature, 383(6595), 76–78.
Hebb, D. O. (2005). The organization of behavior: A neuropsychological theory.
Psychology press.
Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S., Hudspeth, A. J., &
Mack, S. (Eds.). (2000). Principles of Neural Science (Vol. 4, pp. 1227–1246).
McGraw-Hill.
Koch, C. (1999). Biophysics of computation: Information processing in single neurons.
Oxford University Press.
Kolb, B., & Whishaw, I. Q. (1989). Plasticity in the neocortex: Mechanisms
underlying recovery from early brain damage. Progress in Neurobiology, 32(4),
235–276.
Kolb, B., & Whishaw, I. Q. (1998). Brain plasticity and behavior. Annual Review of
Psychology, 49(1), 43–64.
Marder, E., & Goaillard, J. M. (2006). Variability, compensation and homeostasis
in neuron and network function. Nature Reviews Neuroscience, 7(7), 563–574.
Missale, C., Nash, S. R., Robinson, S. W., Jaber, M., & Caron, M. G. (1998).
Dopamine receptors: From structure to function. Physiological Reviews, 78(1),
189–225.
Sasaki, T., Ishikawa, T., Abe, R., Nakayama, R., Asada, A., Matsuki, N., & Ikegaya,
Y. (2014). Astrocyte calcium signalling orchestrates neuronal synchronization in
organotypic hippocampal slices. The Journal of Physiology, 592(13), 2771–2783.
Schafer, D. P., Lehrman, E. K., Kautzman, A. G., Koyama, R., Mardinly, A. R.,
Yamasaki, R., & Stevens, B. (2012). Microglia sculpt postnatal neural circuits in
an activity and complement-dependent manner. Neuron, 74(4), 691–705.
Strata, P., & Harvey, R. (1999). Dale’s principle. Brain Research Bulletin, 50(5),
349–350.
Tononi, G., & Cirelli, C. (2003). Sleep and synaptic homeostasis: A hypothesis.
Brain Research Bulletin, 62(2), 143–150.
Vaaga, C. E., Borisovska, M., & Westbrook, G. L. (2014). Dual-transmitter neu-
rons: Functional implications of co-release and co-transmission. Current Opinion
in Neurobiology, 29, 25–32.
26    
P. Krauss

Wang, F., Smith, N. A., Xu, Q., Fujita, T., Baba, A., Matsuda, T., & Nedergaard,
M. (2012). Astrocytes modulate neural network activity by Ca2+-dependent
uptake of extracellular K+. Science Signaling, 5(218), ra26–ra26.
Zhang, W., & Linden, D. J. (2003). The other side of the engram: Experience-
driven changes in neuronal intrinsic excitability. Nature Reviews Neuroscience,
4(11), 885–900.
 4
Organization of the Nervous System

Never trust something that can think for itself if you can’t see where it keeps its brain!
Arthur Weasley

Modularity of the Nervous System

A fundamental organizational principle of the nervous system of more
advanced species is its strong modularity. It is therefore not a gigantic homo-
geneous neural network, but rather consists of various functional units,
which are not only anatomically distinguishable from each other, but can
also operate largely independently of each other.
Another characteristic that all vertebrate nervous systems have in com-
mon is their bipolar mirror symmetry. Each anatomically distinguishable
module exists twice, once in the left and once in the right half of the body.
There are indeed examples in the animal kingdom of other architectures.
Cephalopods, e.g. octopuses, have a nervous system with circular symmetry.
At the highest level, a distinction can be made between the central and
peripheral nervous system. The central nervous system (CNS) is the main
control unit of the body and consists of the brain, the brainstem, which con-
nects the brain with the spinal cord and plays an important role in regulat-
ing vital functions such as respiration, cardiovascular and blood pressure, as
well as the spinal cord. The spinal cord in turn acts as a connection between
the CNS and the peripheral nervous system, which includes the autonomic

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 27
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_4
28    
P. Krauss

and somatic nervous system and can be considered as an interface between the
organism and the world. The autonomic nervous system controls a number of
involuntary body processes such as heart rate, respiration, digestion, and pupil
contraction. These processes run automatically and are not subject to con-
scious control. The autonomic nervous system consists of two complementary
systems: Sympathetic and Parasympathetic. The sympathetic nervous system
prepares the organism for activity by activating organs and body functions,
thus putting the body into fight-or-flight mode. The parasympathetic, on the
other hand, is active during rest and relaxation phases and contributes to the
calming and recovery of the body. The autonomic nervous system responds
to a variety of stimuli, including stress, physical activity, and changes in the
environment. It works with other body systems to maintain homeostasis and
allow the body to function optimally. While the autonomic functions cannot
be consciously controlled, they can be influenced in various ways, e.g. through
relaxation techniques, movement and breathing exercises. The somatic nerv-
ous system, finally, is responsible for transmitting sensory information from
the sensory organs to the CNS and for forwarding motor commands from the
CNS to the muscles to control voluntary movements (Fig. 4.1).

The nervous system

Central nervous system Peripheral nervous system

Brain
Autonomic Somatic
nervous system nervous system
Brain
stem

Spinal Sympathetic Parasympathetic

cord nervous system nervous system

Fig. 4.1 Organization of the nervous system. The central nervous system controls
the body and consists of the brain, the brainstem, and the spinal cord. The peripheral
nervous system includes the autonomic and the somatic nervous system. The auto-
nomic nervous system regulates involuntary body processes and includes the sympa-
thetic and the parasympathetic nervous system. The somatic nervous system transmits
sensory information to the CNS and forwards motor commands from the CNS to the
muscles
 4 Organization of the Nervous System    
29

Spinal Cord and Brainstem

The spinal cord is part of the central nervous system and runs from the brain
through the vertebral canal to the lower back. It is a long, narrow strand of
nerve tissue, consisting of nerve cells and nerve fibers. It plays a crucial role
in transmitting motor control signals between the brain and skeletal muscles
and in transmitting sensations such as pain, temperature, touch, and pres-
sure from the sensory organs to the brain. However, the spinal cord should
not be imagined as a large cable harness that simply transmits information
in both directions. Rather, it is an independent control unit for the coor-
dination of movements, which contains the circuits for controlling many
reflexes as well as the programs for complex movement sequences such as
walking. Reflexes are differentiated into self-reflexes and foreign reflexes
(Kandel et al., 2000).
A self-reflex is an involuntary and automatic reaction of the body to
a specific stimulus, which occurs without the involvement of the brain. It
is a local reaction, in which sensory nerve endings in muscles or tendons
transmit information to the spinal cord, which is directly connected to cor-
responding motor neurons, thereby triggering a motor reaction in the same
muscle. Since the information from input to output is guided via only one
synapse, these are also referred to as monosynaptic reflexes. A typical exam-
ple is the patellar reflex, where a light blow to the tendon below the knee-
cap leads to a stretching of the quadriceps muscle, which in turn triggers
a contraction of this muscle as a counter-reaction. Self-reflexes are impor-
tant for controlling muscle tension and help to prevent muscle injuries and
unwanted movements.
In contrast, a foreign reflex, also known as a polysynaptic reflex, is a reflex
in which the reflex response does not originate in the organ receiving the
stimulus. The reflex arc is transmitted here via several synapses from the sen-
sory nerve endings in the body to the spinal cord, where the motor reac-
tion is then triggered. The information is also forwarded to higher centers in
the brain. However, this only serves to inform the brain retrospectively that
something has just happened. The control of the foreign reflex takes place
in the spinal cord, as the reaction would otherwise be much too slow. An
example of a foreign reflex is the pain reflex, in which a pain stimulus on the
skin (such as touching a hot stove or stepping on a sharp object) triggers an
immediate withdrawal movement. Foreign reflexes serve to protect the body
from possible damage and to enable a quick reaction to dangers.
30    
P. Krauss

The advantages of this principle of subsidiary movement control in the

spinal cord are obvious. On the one hand, disturbances in the movement
sequence, such as stumbling, can be reacted to very quickly, and on the other
hand, it is very efficient, as the brain is relieved of recurring, trivial control
tasks.

The Thalamus
The thalamus is a small but important structure deep in the center of the
brain. It acts as a relay station for sensory information by processing signals
from the sensory organs and forwarding them to the corresponding areas of
the cerebral cortex (cortex). Only it alone is responsible for the perception and
conscious processing of all sensory stimuli and the (motor) response to them.
The thalamus is often referred to as the “gatekeeper” of the cerebral cortex, as
it filters and regulates the flow of information to the cerebral cortex, so that
only the most relevant and important sensory impressions reach conscious-
ness. This is achieved through a network of thalamocortical loops, i.e., recip-
rocal connections between the thalamus and various regions of the cerebral
cortex. These loops enable the activity of the cortex to be modulated and regu-
lated by filtering out irrelevant information and amplifying important signals.
The thalamus also plays a role in regulating sleep and wakefulness as well
as certain cognitive functions such as memory, attention, and language. It
is connected to various other parts of the brain, including the basal ganglia,
the hypothalamus, and the amygdala (parts of the so-called limbic system,
which is also responsible for processing emotions), contributing to a variety
of functions and behaviors.
The thalamus is not a homogeneous structure, but consists of several
nuclei, each with its own function. Specific nuclei, also referred to as relay
nuclei, are responsible for forwarding specific sensory information from
the periphery (such as eyes, ears, skin, or taste buds) to the corresponding
regions of the cerebral cortex. These nuclei act as filters, amplifying or atten-
uating the sensory input depending on the significance of the information.
The non-specific nuclei, on the other hand, are involved in regulating the
overall activity of the cortex. These nuclei receive input from various regions
of the brain, including the basal ganglia, the limbic system, and the brain-
stem, and send projections to the cortex to modulate its activity. Non-
specific thalamic nuclei play a crucial role in regulating attention, arousal,
and sleep-wake cycles (Kandel et al., 2000).
 4 Organization of the Nervous System    
31

The Cortex
The cerebrum, the largest part of the human brain and most higher mam-
mals, consists of two hemispheres, which are connected by the so-called
corpus callosum. This powerful bundle of fibers contains about 200 million
nerve fibers, the so-called commissures. For comparison: The optic nerve,
which transmits all information from the retina of the eye to the brain, con-
sists only of 1 million nerve fibers.
The nerve cells are organized in six layers parallel to the surface of the
cortex. To accommodate as large a brain surface as possible in a limited skull
volume, the cerebral cortex (Cortex) is strongly folded by numerous convo-
lutions (gyri) and furrows (sulci). This folding effect increases the number
of nerve cells that can be accommodated along the surface of the cerebrum.
Since the skull is subject to anatomical constraints, it is as if one were stuff-
ing a large towel into a small cooking pot. To solve the problem, one must
crumple the towel. The outer areas of the cortex contain the gray matter,
which consists of the cell bodies of the approximately 16 billion neurons of
the cerebrum. The rest of the cerebrum consists of the axons of the neu-
rons, the “connecting cables”, which are referred to as white matter due to
their myelin sheath. With an area of about 2500 cm2, which corresponds to
about four DIN A4 pages, the cortex is only about 3 mm thick. The cortex
is therefore a flat, approximately two-dimensional structure. All higher cog-
nitive performances such as conscious perception, language, thinking, mem-
ory, movement and feelings are located in the cortex (Kandel et al., 2000),
which is why the organization of the cortex will be presented in detail in the
next chapter.

The Hippocampus
The Hippocampus is located in the medial temporal lobe. It owes its name
to the fact that it looks like a seahorse. Together with the amygdala, it
belongs to the limbic system and forms the so-called hippocampus forma-
tion with the entorhinal cortex. It receives input from virtually all regions of
the cortex and is essential for the formation of new declarative and episodic
memory contents, especially in the consolidation and retrieval of episodic
and spatial memories, as well as for spatial navigation.
The hippocampus is considered the highest level of the cortical hierarchy,
as it integrates and processes information from various brain regions before
32    
P. Krauss

forwarding it for storage and further processing to other regions (Hawkins

& Blakeslee, 2004).
Hippocampal replay is a phenomenon that occurs in the formation, con-
solidation, and retrieval of memories (Wilson & McNaughton, 1994). It
involves the reactivation of neuronal activity patterns that were originally
generated during a specific experience or event. It is believed that this rep-
etition strengthens and stabilizes the memory trace. There are two main
types of replay: awake and sleep. Awake replay occurs during rest or inactiv-
ity phases when a person is still conscious, while sleep replay occurs during
non-REM sleep (deep sleep). Both types of replay are crucial for memory
consolidation and the integration of new information into existing memory
networks.
It is believed that the fast-learning hippocampus stores new memory con-
tents during the day and then transfers them as a trainer to the slow-learning
cortex during sleep (replay). The hippocampus thus serves as a kind of buffer
or RAM (Rapid Access Memory). This is the brain’s solution to the so-called
stability-plasticity dilemma (Grossberg, 1982).
This dilemma is a fundamental challenge for neuronal systems, espe-
cially when it comes to learning and memory. It illustrates the need to find
a balance between retaining already learned information and adapting to
new experiences. Every neuronal system must be able to store and retrieve
long-term memories, and at the same time be flexible enough to take in new
information and adapt to changed circumstances, without forgetting what
has already been learned.
A possible solution to the stability-plasticity dilemma is therefore the
existence of separate memory systems in the brain, such as the hippocam-
pus and the cortex (McClelland, 1995). The hippocampus is involved in the
rapid encoding and initial consolidation of new memories, while the cor-
tex is responsible for the slow, gradual process of memory consolidation and
storage. This separation of memory systems can help ensure that new learn-
ing does not collide with already stored information, thus creating a balance
between stability and plasticity.
The hippocampus also plays a key role in pattern separation (O’Reilly &
McClelland, 1994) and pattern completion (McClelland, 1995). Pattern
separation is the process by which similar experiences or events are repre-
sented as different memories, enabling precise retrieval and minimizing
interference between memories. Pattern completion, on the other hand, is
the process by which incomplete or disturbed or noisy signals can trigger the
retrieval of a complete memory.
 4 Organization of the Nervous System    
33

Recent studies also suggest that the hippocampus, in addition to spatial

navigation (Morris et al., 1982), also enables navigation in abstract, mental
spaces and is thus involved in the organization of thoughts across domains
by constructing cognitive maps of any content and flexibly transferring them
to new situations and contexts (Bellmund et al., 2018). If accurately repli-
cated, this functionality could significantly improve the capabilities of our
machine learning systems (Bermudez-Contreras et al., 2020).

The Basal Ganglia

The basal ganglia are a group of nuclei deep in the brain that are responsible
for many important functions of motor control and learning. One of their
main tasks is to act as a kind of “servomechanism” for the cortex, i.e., they
assist in the regulation and fine-tuning of movements and actions initiated
by the cerebral cortex (Kolb & Whishaw, 2009).
An important function of the basal ganglia is the learning and storing of
complex input-output relationships. To this end, they receive input from all
sensory cortex areas and send their output via the thalamus to the motor
control areas of the cortex. These direct connections act like shortcuts in
information transmission and enable more efficient and effective, auto-
mated movement control, as the sensory input can be converted faster and
more precisely into the corresponding motor output, without having to go
through the comparatively slow and complex hierarchy of the cortex areas.
Finally, the basal ganglia are also involved in a form of learning known as
model-free reinforcement learning (Bar-Gad et al., 2003). This means that
they learn from the results of actions and can adjust their behavior accord-
ingly, without explicitly having to create a model of the environment. This
type of learning is essential for adaptive behavior and allows us to continu-
ously improve our actions and decisions over time. The basal ganglia are also
involved in cognitive processes such as decision-making and motivation.
It is assumed that most of our movement control takes place uncon-
sciously and automatically. For example, when learning to drive, this is a
very complex and strenuous task. Among other things, you have to operate
three pedals with only two feet and at the same time keep an eye on all road
users and the road signs. Practically all areas of the cortex are strongly acti-
vated. But with increasing driving practice, something interesting happens.
The cerebral cortex becomes less and less active. An experienced driver can
drive almost unconsciously, completely automated. Especially on routine
34    
P. Krauss

routes like the daily commute to work or on long motorway journeys, the
cortex is in the so-called default mode, a resting state that the cortex always
assumes when there is “nothing to do”. Therefore, driving can often even be
meditative for experienced drivers, as the cortex is essentially idling and one
can indulge in daydreams or think about something completely different
than driving. However, there is also the danger that it becomes too monot-
onous and boring and one falls asleep. This certainly would not have hap-
pened during the first driving lesson.1
The cortex becomes fully active again when a critical situation occurs. In
such situations, the cortex takes over control again, and the driver focuses
his full attention on the road because he has to react quickly to the unex-
pected. Surely you have experienced this or a similar situation yourself,
whether driving or riding along. You are having a conversation during the
drive, thanks to their basal ganglia the driver can use her cortex for some-
thing other than steering the car. When suddenly a child runs onto the road
from behind a parked car, the driver immediately stops talking and focuses
her full attention on the road.

The Cerebellum
The cerebellum (Cerebellum) is one of the oldest brain regions in terms of
evolutionary history and makes up about 10% of the total volume of the
brain. The cerebellum consists of two hemispheres and is divided into var-
ious lobes and nuclei. Of the total 86 billion neurons in the brain, the cer-
ebellum contains the most neurons of all brain parts with about 69 billion
neurons. In evolution, the volume of the cerebellum has increased parallel
to the surface enlargement of the cortex and is most pronounced in humans
(Barton & Venditti, 2014; Sereno et al., 2020).
The cerebellum is responsible for the coordination and fine-tuning of
movements by integrating information from sensory systems and other brain
regions and interacting with motor systems. It also supports the planning
and execution of movements as well as adaptation to changed conditions
and the correction of errors. In addition, the cerebellum also plays a role in

1 The sequence of different phases of neural activity described here corresponds to the development of

competence levels, a model from developmental psychology, starting with unconscious incompetence,
conscious incompetence and conscious competence, to unconscious competence (Adams, 2011).
 4 Organization of the Nervous System    
35

cognition and emotion, although its exact function in these areas is not yet
fully understood (Kandel et al., 2000; Kolb & Whishaw, 2009).
One of the most important functions of the cerebellum is the temporal
coordination of movements, which enables the smooth execution of com-
plex movements such as walking, running, or playing a musical instrument.
The cerebellum receives information from the spinal cord, the sense of bal-
ance, and practically all sensory cortex areas and uses this to coordinate and
adjust movements. In addition to coordinating movements, the cerebellum
is also responsible for correcting movement errors. If there is a discrepancy
between the intended and the actual movement, the cerebellum recog-
nizes this error and makes the necessary adjustments to correct the move-
ment. The cerebellum also plays a crucial role in learning new motor skills.
Through constant adjustment and refinement of movements, the cerebellum
enables humans to improve their motor skills over time (Kandel et al., 2000;
Kolb & Whishaw, 2009).
The cerebellum is divided into three main parts. The vestibulocerebellum
receives information about the position and movement of the body from the
vestibular system (balance organs in the inner ear), which it uses to regulate
body posture and balance. It is also responsible for the precise coordination
of almost all eye movements that originate from various oculomotor centers
in the brainstem (Kandel et al., 2000; Kolb & Whishaw, 2009).
The spinocerebellum receives sensory signals from the spinal cord, which
provide information about the position of joints and muscles, as well as con-
tinuous feedback about the movement signals sent to the spinal cord and the
periphery. It also ensures fine-tuning of the movement signals and that the
movement is executed as intended. This includes the complex coordination
of the facial and laryngeal muscles required for speaking.
The pontocerebellum is functionally connected to the cerebral cortex. It
receives input from various areas, particularly from the premotor centers in
the frontal lobe (premotor cortex and supplementary motor cortex), where
motor plans are formed. In the cerebellum, these are temporally precisely
modulated and the planned activity of the involved muscles is coordinated.
The results of these calculations are forwarded to the thalamus, from where
they are finally forwarded to the motor cortex as input (Kandel et al., 2000).
There is some evidence that the cerebellum plays an important role not
only for motor functions but also for cognitive processes. There are various
arguments for this thesis. As already mentioned, the hemispheres of the cere-
bellum in humans are particularly pronounced, which in evolutionary terms
goes hand in hand with the growth of the cerebral cortex and the devel-
opment of human cognitive abilities. On the other hand, the cerebellum
36    
P. Krauss

also receives a tremendous amount of information from virtually all areas

of the cerebral cortex. Moreover, it has been shown that the output of the
cerebellum via the thalamus reaches not only motor cortex areas but also
other areas of the cortex. Interestingly, it has also been observed that cer-
tain lesions, i.e., injuries, of the cerebellum have no effect on movement
coordination, suggesting that the corresponding areas must be responsible
for other, non-motor functions. Finally, functional studies using modern
imaging techniques have shown that the cerebellum is also activated in many
cognitive tasks.
One of the most remarkable features of the cerebellum is its extremely
regular strictly geometric circuit diagram (Fig. 4.2). The parallel fibers are
the axons of the granule cells, which represent the most common type of
neurons in the cerebellar cortex. These fibers first rise vertically, then branch
into two opposite branches and finally run horizontally through the molec-
ular layer of the cerebellum, where they form synapses with the fan-like
dendrites of Purkinje cells. The parallel fibers are generally considered to be
non- or very thinly myelinated, which gives them a very slow conduction
speed. It is believed that this property allows the cerebellum to detect the

Purkinje cells
Parallel fibers

Output

Granule cells

Input

Fig. 4.2 Circuit diagram of the cerebellum. The granule cells represent the input
layer of the cerebellum, their axons rise and then branch into parallel fibers. There
they form synapses with the fan-like dendrites of the Purkinje cells, which function
like a kind of coincidence detectors. Only when enough action potentials arrive at
the same time via the parallel fibers, they are activated
 4 Organization of the Nervous System    
37

smallest time differences between incoming signals. This ability is important

for the precise synchronization and coordination of neuronal activity, which
in turn is crucial for the role of the cerebellum in motor control, learning,
and coordination (Kandel et al., 2000; Kolb & Whishaw, 2009).
Interestingly, there is a structure in the nervous system that has a very
similar structure to the cerebellum, the Dorsal Cochlear Nucleus (DCN)
(Oertel & Young, 2004; Bell et al., 2008; Singla et al., 2017). The DCN is
one of the first processing stages of the auditory system. It is believed that
here the information content of the auditory input from the cochlea (hear-
ing snail) is determined and monitored by calculating the autocorrelation of
the signal (Licklider, 1951; Krauss et al., 2016, 2017). Due to the anatom-
ical similarity, it is obvious to assume that the cerebellum performs similar
calculations and thus at least partially has a similar function to the DCN,
namely to quantify and regulate the information content.

Conclusion
The brain is not a single large neural network, but on the contrary, it exhib-
its a very high degree of modularity. The study of the brain’s modularity can
provide valuable insights for the development of new AI systems, architec-
tures, and algorithms. An important concept of machine learning is the idea
of modular learning, where a complex problem is broken down into smaller,
more manageable sub-problems. This is comparable to the way the brain is
divided into different modules that perform specific functions.
Furthermore, studying the modularity of the brain also offers the oppor-
tunity to develop more efficient and specialized AI algorithms that can per-
form specific tasks with high accuracy. By breaking down a complex task
into smaller subtasks and developing separate modules for each subtask, an
AI system can be optimized for this specific task. This can lead to faster and
more accurate performance than an AI system that tries to handle all tasks in
a single module.
The modular organization of the brain allows for flexibility and adapt-
ability, as different modules can be reconfigured and used for different tasks.
Similarly, AI systems with modular architectures can be developed, allow-
ing for easy integration of new modules and adaptation to new tasks and
environments. For example, the hippocampus-cortex system could serve as a
model for future AI systems to solve the stability-plasticity dilemma.
38    
P. Krauss

References
Adams, L. (2011). Learning a new skill is easier said than done. Gordon Training
International.
Bar-Gad, I., Morris, G., & Bergman, H. (2003). Information processing, dimen-
sionality reduction and reinforcement learning in the basal ganglia. Progress in
Neurobiology, 71(6), 439–473.
Barton, R. A., & Venditti, C. (2014). Rapid evolution of the cerebellum in humans
and other great apes. Current Biology, 24(20), 2440–2444.
Bell, C. C., Han, V., & Sawtell, N. B. (2008). Cerebellum-like structures and their
implications for cerebellar function. Annual Review of Neuroscience, 31, 1–24.
Bellmund, J. L., Gärdenfors, P., Moser, E. I., & Doeller, C. F. (2018). Navigating
cognition: Spatial codes for human thinking. Science, 362(6415), eaat6.
Bermudez-Contreras, E., Clark, B. J., & Wilber, A. (2020). The neuroscience of
spatial navigation and the relationship to artificial intelligence. Frontiers in
Computational Neuroscience, 14, 63.
Grossberg, S. (1982). How does a brain build a cognitive code? Studies of Mind
and Brain. Boston Studies inthe Philosophy of Science, 70. Springer. https://doi.
org/10.1007/978-94-009-7758-7_1.
Hawkins, J., & Blakeslee, S. (2004). On intelligence. Macmillan.
Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S., Hudspeth, A. J., &
Mack, S. (Eds.). (2000). Principles of neural science (Vol. 4, pp. 1227–1246).
McGraw-Hill.
Kolb, B., & Whishaw, I. Q. (2009). Fundamentals of human neuropsychology.
Macmillan.
Krauss, P., Tziridis, K., Metzner, C., Schilling, A., Hoppe, U., & Schulze, H.
(2016). Stochastic resonance controlled upregulation of internal noise after hear-
ing loss as a putative cause of tinnitus-related neuronal hyperactivity. Frontiers in
Neuroscience, 10, 597.
Krauss, P., Metzner, C., Schilling, A., Schütz, C., Tziridis, K., Fabry, B., & Schulze,
H. (2017). Adaptive stochastic resonance for unknown and variable input sig-
nals. Scientific Reports, 7(1), 2450.
Licklider, J. C. R. (1951). A duplex theory of pitch perception. The Journal of the
Acoustical Society of America, 23(1), 147–147.
McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are
complementary learning systems in the hippocampus and neocortex: Insights
from the successes and failures of connectionist models of learning and memory.
Psychological Review, 102(3), 419.
Morris, R. G., Garrud, P., Rawlins, J. A., & O’Keefe, J. (1982). Place navigation
impaired in rats with hippocampal lesions. Nature, 297(5868), 681–683.
Oertel, D., & Young, E. D. (2004). What’s a cerebellar circuit doing in the auditory
system? Trends in Neurosciences, 27(2), 104–110.
 4 Organization of the Nervous System    
39

O’Reilly, R. C., & McClelland, J. L. (1994). Hippocampal conjunctive encoding,

storage, and recall: Avoiding a trade-off. Hippocampus, 4(6), 661–682.
Sereno, M. I., Diedrichsen, J., Tachrount, M., Testa-Silva, G., d’Arceuil, H., &
De Zeeuw, C. (2020). The human cerebellum has almost 80% of the surface
area of the neocortex. Proceedings of the National Academy of Sciences, 117(32),
19538–19543.
Singla, S., Dempsey, C., Warren, R., Enikolopov, A. G., & Sawtell, N. B. (2017). A
cerebellum-like circuit in the auditory system cancels responses to self-generated
sounds. Nature Neuroscience, 20(7), 943–950.
Wilson, M. A., & McNaughton, B. L. (1994). Reactivation of hippocampal ensem-
ble memories during sleep. Science, 265(5172), 676–679.
 5
Organization of the Cortex

A person consists of two parts—his brain and his body. But the body has more fun.
Woody Allen

Division of the Cortex into Hemispheres

and Lobes
Like practically every part of the nervous system, the cerebral cortex, or
Cortex, also exists in duplicate. The two halves of the Cortex are referred
to as the left and right hemisphere. Anatomically, each hemisphere can be
roughly divided into five so-called lobes, which are separated from each
other by deeper fissures. The division of the Cortex into lobes is anatomi-
cally and functionally significant, as each lobe can be assigned a specific pri-
mary processing area as well as a whole group of functions. On the surface
of the brain are the frontal or frontal lobe, the parietal or parietal lobe, the
occipital or occipital lobe, and the temporal or temporal lobe. The insular
lobe or insular cortex is not visible from the outside, as it is covered by the
frontal, parietal and temporal lobes (Fig. 5.1).
This anatomical division also correlates with the major functional sys-
tems of the Cortex. In the frontal lobe, which is particularly pronounced
in humans, are the motor areas responsible for movement control and plan-
ning. In addition, fundamental personality traits seem to be located here at
the foremost pole, the so-called orbitofrontal cortex, which lies above the

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 41
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_5
42    
P. Krauss

Parietal lobe

Frontal lobe
Occipital lobe

Insular cortex
Temporal lobe

Fig. 5.1 Division of the Cortex into Lobes. Each hemisphere of the Cortex consists of
five lobes: frontal lobe, parietal lobe, temporal lobe, and occipital lobe. Not visible
from the outside, as it is covered by the frontal, parietal, and temporal lobes, is the
fifth, the insular lobe (dotted line)

eye sockets. The parietal lobe, which contains the primary somatosensory
center and areas of spatial perception, follows to the rear. The areas respon-
sible for visual perception are located in the occipital lobe. The temporal
lobe contains the auditory center as well as higher visual areas, e.g., for face
recognition, and areas of long-term memory, so-called multimodal associa-
tion areas. The insular cortex is the least explored and houses, among other
things, the primary taste area and other areas that represent vegetative body
states. Here, the primary center for basic viscerosensitivity, i.e., for informa-
tion from the viscera, is also suspected.

Columns: The Radial Organization of the Cortex

In addition to the vertical organization in layers, there is a radial organiza-
tion into so-called pillars or columns.
Microcolumns (Mountcastle, 1997) are smaller structures consisting of
groups of about 80 neurons with similar response characteristics, which are
strongly interconnected through vertical connections. Each microcolumn is
selectively responsible for a specific feature of a sensory stimulus or a cogni-
tive process. In the visual cortex, for example, a microcolumn may contain
neurons that all respond to a specific orientation (angle) or a specific color
of a visual stimulus at a specific location in the visual field. It is believed that
microcolumns could represent the basic units of information processing in
 5 Organization of the Cortex    
43

the cortex. According to this view, information is encoded in the activity of

the neurons within the microcolumns and integrated across the microcol-
umns to form the respective output.
Due to the number of neurons in the cortex (approx. 16 billion) and the
number of neurons per microcolumn (approx. 80), there is an approximate
number of about 200 million microcolumns in the cortex. Interestingly, this
corresponds to both the number of commissures (nerve fibers) in the corpus
callosum, which connect both hemispheres, and the number of nerve fib-
ers that connect the cortex to the cerebellum via the pons. This could sug-
gest that each microcolumn sends its information on average exactly via one
nerve fiber both into the opposite hemisphere and into the cerebellum, sup-
porting the hypothesis that the microcolumns are the fundamental process-
ing units of the cortex.
Macrocolumns, on the other hand, are groups of about a hundred micr-
ocolumns, which work together to process information from a specific sen-
sory modality or cognitive function (Mountcastle, 1997). A macrocolumn
consists of the representations of all possible manifestations of a certain
feature at a certain location; in visual processing, for example, all possible
angles or colors of a stimulus at a certain location in the visual field.
Of course, this does not mean that we can only distinguish between about
a hundred different colors or angles at a given location in the visual field.
Through so-called population coding, the actual number of representable
feature manifestations is virtually unlimited. Population coding is a neu-
ral coding scheme that represents sensory or motor information through
the activity of an entire population of neurons or microcolumns, not just
the activity of individual neurons or columns. In other words, the infor-
mation is represented by the combined activity of a group of neurons or
microcolumns.
This coding scheme is used in the brain to represent a variety of informa-
tion such as visual stimuli, auditory stimuli, and motor commands. Each
neuron in the population has a preferred feature manifestation or stimulus,
and its firing rate is set to respond best to this manifestation of the feature.
When a stimulus is presented to the population, each neuron contributes to
the overall response in proportion to its sensitivity to the stimulus.
Population coding has advantages over other coding schemes such as rate
coding or temporal coding, as it can represent complex stimuli with greater
accuracy and tolerates noise or fluctuations in the response of individual
neurons. It also allows for the integration of information from different sen-
sory modalities or sources, which is crucial for many cognitive functions
(Averbeck et al., 2006).
44    
P. Krauss

Layers: The Vertical Organization of the Cortex

The cortex is remarkably uniformly structured, which is why it is also
referred to as isocortex (uniform cortex). If you cut it crosswise to the sur-
face and look at the sections under different stains in the microscope, you
can recognize a vertical layer structure. The different layers are packed at dif-
ferent densities and consist of different types of neurons. In addition, there
are various vertical and horizontal connections between the cells of the indi-
vidual layers. Overall, six different layers can be distinguished, which are
arranged and interconnected in a highly organized manner. Although this
structure is in principle the same at any point in the cortex, there are also
differences. Depending on where you examine, the relative thickness and cell
density of the individual layers vary in the cross-section.

Brodmann Areas
In the early twentieth century, the psychiatrist and neuroanatomist
Korbinian Brodmann systematically examined many sections from all
regions of the cortex and found that the cortex can be divided into different
areas, each characterized by a unique combination of cell types, densities,
and layers (Brodmann, 1910). Brodmann published a system for mapping
the human brain based on its cytoarchitecture, i.e., the organization of cells
in the various brain regions. The Brodmann areas named after him are num-
bered according to their location in the brain, from the primary sensory and
motor areas in the upper central part of the brain to the more complex asso-
ciation areas in the front part of the brain. The numbering system ranges
from 1 to 52, with some numbers omitted because it was later found that
the areas were duplicates or combinations of other areas (Fig. 5.2).
Interestingly, these structural anatomical differences correlate with func-
tional differences. This means that each Brodmann area, which externally
differs from its neighbors under the microscope, actually performs a differ-
ent function. For example, it has been found that visual information is pro-
cessed in Brodmann areas 17, 18, and 19 in the back of the brain, while
auditory information is processed in areas 41 and 42 at the top edge of the
temporal lobe. Area 4 corresponds to the primary motor cortex, which sends
control commands to the skeletal muscles, and areas 1, 2, and 3 correspond
to the primary sensory cortex, which is responsible for processing tactile
stimuli from the body surface.
 5 Organization of the Cortex    
45

Fig. 5.2 Brodmann Areas. Each Brodmann area is defined by a unique combination
of cell types, cell densities, and layers and is associated with different functions and
behaviors. For example, Brodmann area 4 in the primary motor cortex is responsible
for controlling voluntary movements, while Brodmann area 17 in the primary visual
cortex is responsible for processing visual information

However, the actual number of cortex areas is significantly larger than the
division into 52 Brodmann areas suggests (Amunts & Zilles, 2015; Coalson
et al., 2018; Gordon et al., 2016). Using multimodal MRI data from the
Human Connectome Project (Elam et al., 2021) and an objective, semi-au-
tomatic neuroanatomical approach, 180 areas per hemisphere have recently
been identified that differ in their architecture, function, connectivity, or
topography (Glasser et al., 2016).
But even this is likely to be only a preliminary, approximate number. The
practical difficulty of dividing the cortex into meaningful maps or areas,
and the fundamental difficulty of assigning specific functions to these maps
that are neither too narrowly defined nor too general, is described by David
Poeppel as the Maps Problem (division into maps) and Mapping Problem
(assignment of functions) (Poeppel, 2012).

Maps in the Head

Another special feature is the map-like organization of these areas. If we
look at the primary somatosensory area, i.e., the cortex area where informa-
tion from the touch and touch receptors of the skin surface is processed, we
notice that there is a system. Input from neighboring areas of the skin is also
46    
P. Krauss

processed in neighboring areas of this cortex area. This organizational princi-

ple is referred to as Somatotopy (Fig. 5.3).
What is also noticeable is that the relative sizes are quite different from
what we know from the body surface. Hand, tongue, and lips, for example,
take up significantly more space than the foot and even more space than the
entire back. The explanation for this is that the density of touch receptors on
the body surface varies greatly depending on the region. We find the highest
receptor density in the fingertips, on the tongue and lips, and the lowest on
the legs and back. This makes immediate sense, as we can feel very high-res-
olution and precise with our fingers, which is also the basis for reading
Braille. We can’t do that with our back. The stimulus discrimination thresh-
old or tactile acuity at fingertips, lips, and tongue is about one millimeter:
If we are pricked with two needles without seeing it, we can only distin-
guish whether we were pricked with one or two needles from a distance of
less than one millimeter. In contrast, tactile acuity on legs and back is even

Fig. 5.3 Somatosensory Homunculus. Projection of the body surface onto the pri-
mary sensorimotor cortex area. The relative size of the body parts is greatly distorted
and is based on the respective density of touch and tactile receptors. The representa-
tions of hand, tongue, and lips take up significantly more space than those for foot
and back
 5 Organization of the Cortex    
47

several centimeters. High tactile acuity means that a lot of information from
the corresponding body region reaches the cortex, which in turn requires a
lot of processing capacity, i.e., a relatively large area of the corresponding
area. In contrast, only a relatively small area is needed for processing for the
back, where the receptor density is lowest.
The primary motor area, which sends control commands to the muscles,
behaves analogously. We can move our fingers very finely tuned and pre-
cisely, but not so much our foot. Accordingly, the areas for controlling hands
and face are relatively large compared to the areas that control the muscles of
the feet and toes. Try playing the piano with your feet!
The primary visual area (visual cortex) also follows this topographic
organizational principle. Here we speak of retinotopy, as the input comes
from the retina. The retina is structured in such a way that the spatial rela-
tionships between the different parts of the visual field are preserved. This
means that adjacent parts of the retina correspond to adjacent parts of the
visual field. When the information from the optic nerve reaches the visual
cortex in the brain, it is organized in a similar way. Adjacent regions of the
visual cortex correspond to adjacent regions of the visual field, creating a
“map” of the visual world. Again, this map is not a perfect one-to-one rep-
resentation of the visual field. Rather, there are distortions and irregularities
in the way different parts of the visual field are represented in the brain. The
fovea, the point of sharpest vision in the relatively small central area of the
retina, is greatly overrepresented in the primary visual cortex. Although the
fovea only makes up about 1 percent of the total area of the retina, about
half of the primary visual cortex is responsible for processing information
from the fovea.
The reason for this strong overrepresentation lies again in the high sen-
sitivity and visual acuity of the fovea. Since the fovea contains a high den-
sity of photoreceptors and is responsible for our sharpest vision, it provides
the brain with the most detailed and accurate information about the visual
world. Therefore, the brain assigns a larger proportion of its processing
resources to the fovea to ensure that this information is processed with the
greatest possible accuracy. In contrast, the peripheral regions of the visual
field are represented in the visual cortex in a highly compressed form.
The last example refers to the auditory system, which is responsible for
the perception of sounds and speech. Here too, there is a map-like struc-
ture, the tonotopy, where adjacent frequencies are processed at adjacent loca-
tions of the primary auditory area. The acoustic input is broken down into
48    
P. Krauss

its individual frequency components at the basilar membrane of the cochlea

(hearing snail), with similar frequencies represented at adjacent locations on
the basilar membrane. This tonotopy continues in all further processing sta-
tions of the auditory pathway up to the primary auditory cortex area.
As is well known, most bats use ultrasonic echo location to locate their
prey and obstacles in their environment even in darkness. They can then
derive information about the position and movement of objects in their
environment based on the runtime and frequency shift of the echo. The fre-
quency range of the ultrasonic waves emitted by the bats is therefore par-
ticularly important for them and is represented in their auditory system with
high resolution, i.e., they can recognize the smallest frequency differences
in this range and it therefore takes up a disproportionately large space in
their primary auditory cortex area, which is why, by analogy with the human
visual system, we also speak of an “auditory fovea”.
In fact, the map-like organization can be found in all sensory modalities
and in the motor system, where it is most obvious in the primary areas, as
these areas are closest to the outside world, so to speak. The primary sensory
areas represent the input areas of the cortex, i.e., they receive their inputs
from the sensory organs, while the primary motor area is the output area of
the cortex, which sends its output to the muscles and thus quasi to the out-
side world. In contrast, all other areas receive their input from other cortex
areas and also send their output back to other cortex areas. Since the structure
of the cortex is fundamentally the same everywhere, it can be assumed that
the fundamental functional principles are also the same in all cortex areas.
This means that each area creates a map of its input, with similar input repre-
sented at adjacent locations and more frequent and therefore more important
input taking up more space on the map. However, as the respective area gets
further away from the outside world, the representations become increasingly
abstract and the map-like organization becomes harder to understand.

The Canonical Circuit of the Cortex

The six layers of the cortex are interconnected in a specific way. Although
fundamentally all possible combinations of connections exist, some are
much more pronounced than others. From this fact, the following pattern
for the flow of information within a cortical area emerges: The input from
the sensory organs (in primary sensory areas) or from other, lower cortical
areas first reaches layer 4. From there, the information is forwarded to layers
 5 Organization of the Cortex    
49

2/3 and finally to layer 5/6, which serves as the output layer of the cortex,
from where the information is transmitted back to the thalamus or to lower
cortical areas as feedback. In addition, fibers from layers 2/3 project into
further cortical areas. Depending on which layer is the source and target,
three different types of connections can be distinguished between two corti-
cal areas: ascending (bottom-up, feed-forward), descending (top-down, feed-
back), and horizontal (lateral) connections (Kolb & Whishaw, 2009; Kandel
et al., 2000; Imam & Finlay, 2020) (Fig. 5.4).

Hierarchy of Cortical Areas

By analyzing the type of connections between the cortical areas, a kind of
circuit diagram of the information flow through the areas can be recon-
structed. It turns out that the areas of the cortex are hierarchically-parallel
organized, i.e., at each hierarchy level there is usually not just one, but sev-
eral areas, with the number of areas per level also increasing with increasing
hierarchy level. Interestingly, it turns out that the hippocampus is at the top
of the hierarchy of sensory areas (Kolb & Whishaw, 2009; Felleman & Van
Essen, 1991; Van Essen et al., 1992) (Fig. 5.5).

Layer 5/6
Feedback to
thalamus or
lower areas
Feedback from
higher areas
Layer 2/3
Output to
higher areas

Input from Layer 4

Thalamus or
lower areas

Fig. 5.4 Information flow in the cortical layers. The six layers of the cortex are spe-
cifically interconnected. Input from sensory organs or hierarchically lower cortical
areas reaches layer 4, is forwarded to layers 2/3, and finally reaches the output layers
5/6. From there, feedback is sent back to the thalamus or hierarchically lower corti-
cal areas. In addition, fibers from layers 2/3 project into hierarchically higher cortical
areas or across the corpus callosum into homologous areas on the other hemisphere.
Feedback from higher areas ends in layers 2/3 and 5/6
50    
P. Krauss

Fig. 5.5 Hierarchical-parallel organization of the cortex. The analysis of the con-
nections between cortical areas allows the reconstruction of a circuit diagram of the
information flow. The areas are hierarchically and parallel arranged, with several
areas present at each level and the number of areas per level increasing with higher
hierarchy level

Conclusion
The cerebral cortex has an impressively uniform structure that allows a hier-
archical-parallel processing, with the cortical areas being map-like organized.
They represent their respective input topologically preserving similarity and
importance. Similar input is processed in adjacent areas, with the relative
size of the area correlating with the amount of information and thus the
importance of the input. Understanding the organization and function of
the cortex could have significant implications for the development of new AI
architectures.
Insights into the vertical and horizontal organization of the cortex could
contribute to the development of new architectures for deep neural networks
to make information processing in different layers and between different
areas of the network more efficient. In addition, the study of the function-
ing of micro- and macro-columns as well as population coding in the cortex
could contribute to developing new approaches for information processing
and representation in artificial neural networks that may be more robust
and efficient. The analysis of canonical circuits and the flow of information
 5 Organization of the Cortex    
51

in the cortex could contribute to the development of new algorithms and

learning strategies for artificial neural networks that consider both local and
global information and respond better to feedback.
Finally, understanding the role of the hippocampus in the hierarchy of
sensory areas and its involvement in information processing and integration
could contribute to developing new models of artificial intelligence that are
better able to process spatio-temporal information and build a long-term
memory.

References
Amunts, K., & Zilles, K. (2015). Architectonic mapping of the human brain
beyond Brodmann. Neuron, 88(6), 1086–1107.
Averbeck, B. B., Latham, P. E., & Pouget, A. (2006). Neural correlations, popula-
tion coding and computation.Nature Reviews Neuroscience, 7(5), 358–366.
Brodmann, K. (1910). Feinere Anatomie des Grosshirns (pp. 206–307). Springer.
Coalson, T. S., Van Essen, D. C., & Glasser, M. F. (2018). The impact of tradi-
tional neuroimaging methods on the spatial localization of cortical areas.
Proceedings of the National Academy of Sciences, 115(27), E6356–E6365.
Elam, J. S., Glasser, M. F., Harms, M. P., Sotiropoulos, S. N., Andersson, J. L.,
Burgess, G. C., … & Van Essen, D. C. (2021). The human connectome project:
A retrospective. NeuroImage, 244, 118543.
Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in
the primate cerebral cortex. Cerebral Cortex, 1(1), 1–47.
Glasser, M. F., Coalson, T. S., Robinson, E. C., Hacker, C. D., Harwell, J., Yacoub,
E., … & Van Essen, D. C. (2016). A multi-modal parcellation of human cere-
bral cortex. Nature, 536(7615), 171–178.
Gordon, E. M., Laumann, T. O., Adeyemo, B., Huckins, J. F., Kelley, W. M., &
Petersen, S. E. (2016). Generation and evaluation of a cortical area parcellation
from resting-state correlations. Cerebral Cortex, 26(1), 288–303.
Imam, N., & Finlay, B. L. (2020). Self-organization of cortical areas in the devel-
opment and evolution of neocortexs. Proceedings of the National Academy of
Sciences, 117(46), 29212–29220.
Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S., Hudspeth, A. J., &
Mack, S. (Eds.). (2000). Principles of neural science (Vol. 4, pp. 1227–1246).
McGraw-Hill.
Kolb, B., & Whishaw, I. Q. (2009). Fundamentals of human neuropsychology.
Macmillan.
Mountcastle, V. B. (1997). The columnar organization of the neocortex. Brain: A
Journal of Neurology, 120(4), 701–722.
52    
P. Krauss

Poeppel, D. (2012). The maps problem and the mapping problem: Two challenges
for a cognitive neuroscience of speech and language. Cognitive Neuropsychology,
29(1–2), 34–55.
Van Essen, D. C., Anderson, C. H., & Felleman, D. J. (1992). Information pro-
cessing in the primate visual system: An integrated systems perspective. Science,
255(5043), 419–423.
 6
Methods of Brain Research

One only sees what one already knows and understands.

Johann Wolfgang von Goethe

Imaging Techniques: Watching the Brain Think

Imaging techniques are methods used to create images of the brain or other
body structures. Imaging techniques play a crucial role in brain research as
they allow researchers to collect and analyze detailed information about the
brain. With these techniques, brain activity can be measured during vari-
ous tasks, and it can be traced how different brain regions interact with each
other. Imaging techniques can also help detect and investigate brain diseases
and disorders such as stroke, dementia, and epilepsy. Overall, imaging tech-
niques are an essential part of brain research and contribute to expanding
our understanding of the brain and how it works.

CT
The Computed Tomography (CT) uses X-rays to create detailed cross-sec-
tional images of the brain. X-ray images of the head are taken from various
angles and reconstructed into a three-dimensional image of the brain on a
computer. The resulting image shows the structures of the brain, including

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 53
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_6
54    
P. Krauss

the ventricles, the skull, and the blood vessels. CT imaging is a common
method for diagnosing neurological diseases such as stroke, traumatic brain
injury, and brain tumors. However, CT imaging does not provide informa-
tion about the functional activity of the brain, which can be obtained with
other imaging techniques such as PET, fMRI, EEG, and MEG (Ward,
2015; De Groot & Hagoort, 2017).

PET
The Positron Emission Tomography (PET) visualizes the metabolic activity
of cells and tissues in the body. A small amount of a radioactive substance,
a so-called radiotracer, is injected into the body. The radiotracer emits pos-
itrons, positively charged particles that interact with electrons in the body.
When a positron encounters an electron, they annihilate each other and pro-
duce gamma rays, which are detected by the PET scanner. The PET scan-
ner records these gamma rays and creates a three-dimensional image of the
brain’s metabolic activity from them. However, since the brain is constantly
active, meaningful data with PET can only be generated by subtracting two
images. Typically, an image is taken during a specific cognitive task or stim-
ulus, and another image of the brain’s background activity is taken, and
then the difference image is calculated (Ward, 2015; De Groot & Hagoort,
2017).

MRI
The Magnetic Resonance Imaging (MRI), also known as nuclear magnetic
resonance imaging, uses the property of hydrogen nuclei to align along mag-
netic field lines to generate detailed images. Short magnetic pulses stimulate
the aligned nuclei to emit electromagnetic waves, which are detected and
used to determine the distribution of hydrogen, mainly in the form of water.
Functional Magnetic Resonance Imaging (fMRI) uses the fact that the red
blood pigment affects the magnetic signal of the hydrogen nuclei differently,
depending on whether it has bound oxygen or not. Comparing measure-
ments under different stimulation conditions allows changes in blood oxy-
gen saturation (Blood Oxygenation Level Difference, BOLD) to be derived,
which serve as an indirect measure of the change in blood flow to a spe-
cific brain area. The assumption behind this is that active brain areas require
more oxygen and are therefore more heavily perfused. The BOLD signal
 6 Methods of Brain Research    
55

builds up slowly and reaches a maximum about six to ten seconds after the
stimulus begins, before it slowly decreases again. Compared to MEG and
EEG, nuclear magnetic resonance imaging has a lower temporal resolution
of about one recording per second (1 Hz), but a much higher spatial reso-
lution in the range of about one cubic millimeter (Menon & Kim, 1999),
i.e., the activity of the brain can be measured in about one million so-called
voxels1 simultaneously (Ward, 2015; De Groot & Hagoort, 2017).

EEG and MEG

The Electroencephalography (EEG) is a non-invasive method for measuring
the electrical activity of the brain using electrodes placed on the scalp. The
various patterns of the brain’s electrical activity are used to diagnose neu-
rological diseases such as epilepsy, sleep disorders, brain tumors, and head
injuries, and are also important in research for studying brain functions and
behavior. The EEG is known for its extremely high temporal resolution, as it
can perform up to 100,000 measurements per second (100 kHz). However,
the spatial resolution is limited, and the electrical fields are strongly atten-
uated by the brain tissue. With so-called high-density EEG systems, brain
activity can be measured simultaneously in 128 channels distributed over
the scalp. The EEG is best suited for measuring the activity of the cerebral
cortex, especially the gyri, as they are in close proximity to the skull bone.
For measuring the activity of deeper brain regions, the MEG is better suited
(Ward, 2015; De Groot & Hagoort, 2017).
The Magnetoencephalography (MEG) is also a non-invasive method
and is used to measure the magnetic activity of the brain using supercon-
ducting magnetometers (Hämäläinen et al., 1993), so-called SQUIDs
(Super Conducting Quantum Interference Devices). Similar to the EEG,
the MEG can be used to diagnose neurological diseases such as epilepsy or
brain tumors, but also in research for studying brain functions and behavior.
The MEG also has an extremely high temporal resolution, as brain activ-
ity can be recorded at up to 100,000 measurements per second (100 kHz).
The spatial resolution is slightly better than that of the EEG (approx. 250
magnetometers distributed over the scalp). Since magnetic fields are prac-
tically not attenuated or distorted by brain tissue, the MEG is particularly
well suited for measuring activity in the sulci of the cortex and also in deeper

1A voxel, from volume pixel, is the three-dimensional analogue to a pixel (picture element).
56    
P. Krauss

brain regions below the cerebral cortex (subcortical regions). MEG and
EEG therefore complement each other well and can be used simultaneously
in combined M/EEG measurements (Ward, 2015; De Groot & Hagoort,
2017).

Event-related Potentials and Fields

Event-related potentials (ERPs) (engl. Event-Related Potentials, ERP) are a

measure of brain activity that can be measured with EEG while the test per-
son performs a certain task or is confronted with certain stimuli (Handy,
2005; Sur & Sinha, 2009). ERPs are temporally aligned with the presenta-
tion of the stimulus or event and are only visible when averaged over many
(at least 50) measurement repetitions. This is because the noise on the actual
ERPs is ten to a hundred times greater than the actual signal. ERPs repre-
sent the neuronal activity associated with the cognitive or sensory process-
ing of an event. ERPs are usually characterized by their polarity, latency, and
amplitude. The polarity refers to whether the electrical potential recorded at
the scalp is positive or negative relative to a reference electrode. The latency
is the time interval between the stimulus presentation and the occurrence of
the peak of the ERP waveform. The amplitude reflects the strength or size of
the electrical potential recorded at the scalp. ERPs are often used in cogni-
tive neuroscience to investigate cognitive processes such as attention, mem-
ory, language, perception, and decision-making. They can provide insights
into the neuronal mechanisms underlying these processes and can also be
used as biomarkers for various neurological and psychiatric disorders. Event-
related fields (Event-Related Fields, ERF) correspond to the magnetic ana-
logue of the ERPs measured with MEG (Ward, 2015; De Groot & Hagoort,
2017).

Intracranial EEG
Intracranial Electroencephalography (iEEG) is an invasive technique for
measuring brain activity (Parvizi & Kastner, 2018). Electrodes are placed
directly on or in the brain, usually during surgery, to record electrical activ-
ity. This allows for very high spatial resolution, as the electrodes are placed
in specific regions of the brain, often in close proximity to the area of inter-
est. iEEG also has a very high temporal resolution, allowing electrical signals
to be recorded at a rate of up to several thousand times per second (Ward,
 6 Methods of Brain Research    
57

2015; De Groot & Hagoort, 2017). Patients being treated for drug-resist-
ant epilepsy have iEEG electrodes implanted for diagnostic purposes a few
weeks before surgery to resect their epilepsy focus, to determine which brain
regions need to be spared during surgery to preserve important functions
such as language. These patients often participate in neuropsychological
studies, as the data obtained in this way are extremely rare and valuable.

Conclusion
Today, various techniques are available to measure the structure and activity
of the brain. Each method has specific advantages and disadvantages, and no
method is perfect. Ideally, complementary methods such as MEG and EEG
or fMRI and EEG are used simultaneously to get a more complete picture
of brain activity. But even then, we are still far from being able to read the
brain in the spatiotemporal resolution that would be necessary to capture
the exact temporal activity course of each neuron and each synapse. This is
one reason why computer models of brain function are absolutely necessary.
In contrast to the brain, simulated models offer the crucial advantage that
all internal parameters and variables can be read out at any time with any
desired accuracy.

References
De Groot, A. M., & Hagoort, P. (Eds.). (2017). Research methods in psycholinguistics
and the neurobiology of language: A practical guide. Wiley.
Hämäläinen, M., Hari, R., Ilmoniemi, R. J., Knuutila, J., & Lounasmaa, O. V.
(1993). Magnetoencephalography – Theory, instrumentation, and applications
to noninvasive studies of the working human brain. Reviews of Modern Physics,
65(2), 413.
Handy, T. C. (Eds.). (2005). Event-related potentials: A methods handbook. MIT
press.
Menon, R. S., & Kim, S. G. (1999). Spatial and temporal limits in cognitive neuro-
imaging with fMRI. Trends in Cognitive Sciences, 3(6), 207–216.
Parvizi, J., & Kastner, S. (2018). Promises and limitations of human intracranial
electroencephalography. Nature Neuroscience, 21(4), 474–483.
Sur, S., & Sinha, V. K. (2009). Event-related potential: An overview. Industrial
Psychiatry Journal, 18(1), 70.
Ward, J. (2015). The student’s guide to cognitive neuroscience. Psychology Press.
 7
Memory

Long-term memory is the result of permanent changes brought about

by the growth of new synaptic connections.
Eric Kandel

Memory as an Information Processing System

Memory is a crucial aspect of human cognition, as it allows the brain to
encode, store, and retrieve information over an extended period (Milner
et al., 1998; Kandel, 2007). The recollection of past events forms the basis
for the development of language, social relationships, and personal identity.
Without memory, human experience would be severely limited.
Memory is organized as a multi-stage information processing system. In
the first stage, the sensory organs absorb information from the environment.
The subsequent modality-specific neural processing acts as ultra-short-term
or sensory memory. The preprocessed information from the sensory sys-
tems is then transferred to short-term and possibly working memory. In
this phase, the incoming sensory information is temporarily stored and pos-
sibly integrated and processed with information from other modalities and
cognitive systems. From there, the information may be consolidated and
transferred to long-term memory (Atkinson & Shiffrin, 1968). Long-term
memory is responsible for storing information over longer periods up to the
entire lifespan. In this phase, memories become part of the personal biogra-
phy and contribute to the development of personality and self-image.

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 59
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_7
60    
P. Krauss

The ability to store and retrieve information over an extended period is

also crucial for controlling future actions. The information stored in mem-
ory forms the basis for decisions and allows the individual to draw on past
experiences to influence current actions. Thus, memory plays a decisive
role in controlling an organism’s behavior and responses to its environment
(Squire, 1987, 1993; Mandler, 1967).

Sensory Memory
The sensory memory is the first stage in the memory’s information process-
ing system. It stores information received through the sensory organs only
for a very short period, usually less than a second. Sensory memory is crucial
for us to perceive and process the world around us. Depending on the sen-
sory modality, different types of sensory memory are distinguished, includ-
ing iconic, echoic, and haptic memory.
Iconic memory refers to visual stimuli and allows an image to be briefly
retained in the mind’s eye (Sperling, 1963). This type of memory plays a
crucial role in tasks such as reading, where the eyes need to move quickly
to process the text. Echoic memory, on the other hand, refers to auditory
stimuli and allows sounds or words to be retained for a short time, which
forms the basis for the ability to process language and communication.
Finally, haptic memory refers to tactile stimuli and allows sensations of phys-
ical touch to be stored for a short time. This type of memory is important
for people to navigate their environment and interact with objects in their
surroundings.
Sensory memory is believed to correspond to short-term neural activity in
specific sensory brain regions such as the visual or auditory cortex. The exact
neural correlates of sensory memory vary depending on the type of sensory
information being stored, and research continues into how this information
is processed and maintained in the brain (Gazzaniga et al., 2006).
The partial-report paradigm, introduced by George Sperling in 1963
(Sperling, 1963), is a method for investigating sensory memory. In the clas-
sic version, subjects are presented with a matrix of letters or numbers for
a very short time, usually about 50 milliseconds. Immediately afterwards, a
tone or cue is given indicating which row of the grid the participants should
name. The subjects are then asked to repeat the letters or numbers from the
indicated row. Sperling found that participants were able to name almost
all the letters in the indicated row if the tone or cue was given immediately
after the grid disappeared, even though the participants did not know which
 7 Memory    
61

row was relevant before the matrix was presented. This suggests that sensory
memory must have a relatively large storage capacity, as apparently all rows
of the matrix can be stored for a short period of time. However, if the cue
to the row was delayed by a few hundred milliseconds, participants usually
could only remember very few letters from the corresponding row. This, in
turn, suggests that sensory memory fades quickly.
Sperling also introduced another variant of the paradigm, the full-report
paradigm, in which the test subjects are asked to remember all the letters or
numbers of the grid, not just those in a particular row. In general, however,
participants are then only able to remember a few letters or numbers from
the entire matrix, regardless of whether the cue is given immediately or later.
This suggests that conscious access to sensory memory has a limited capacity,
which is smaller than the actual storage capacity of sensory memory.
The partial-report paradigm is widely used in research on sensory memory
and attention and has provided important insights into the nature of these
processes. It has also been used to investigate visual and auditory perception
as well as the encoding and retrieval of memories (Fig. 7.1).

Short-term Memory
Short-term memory refers to the ability to temporarily store a small amount
of information for a few seconds to a minute without altering it. It is based
on short-term neural activity in multimodal brain regions, particularly in
the prefrontal, parietal, and temporal cortex (Fuster & Alexander, 1971;
Funahashi et al., 1989). The involvement of the thalamus, basal ganglia, and
cerebellum is also discussed.
Short-term memory is essential for daily life as it allows us, for example,
to remember important details such as phone numbers, addresses, or direc-
tions. However, short-term memory has a very limited capacity. This is often
associated with the “magic number” seven. This goes back to a publication
by George Miller from the 1950s (Miller, 1956). There he suggested that
the average number of elements or “chunks” that can be stored in short-
term memory is seven plus or minus two. Accordingly, most people should
be able to store five to nine units of information at a given time. However,
recent research suggests that the actual number may be somewhat lower,
with an average capacity of only about four chunks (Cowan, 2001; Luck &
Vogel, 1997; Rouder et al., 2008). This means that if we try to remember
more than these four to five objects or contents, we do it very poorly and we
will probably forget some of them.
62    
P. Krauss

Fig. 7.1 Partial-Report Paradigm. Top: Sensory memory has a relatively large stor-
age capacity, as almost all letters in the indicated row can be named when the cue is
given immediately. Middle: However, it fades quickly, as shown when the cue is given
with a delay. Bottom: The full-report paradigm shows that conscious access to sensory
memory has a limited capacity, which is smaller than the storage capacity of sensory
memory itself

Fortunately, there are methods to increase the capacity of short-term

memory. An effective strategy is so-called chunking, where individual pieces
of information are grouped into larger blocks. For example, if we try to
remember a phone number, we can divide it into groups of three or four
digits to make it easier to remember.
Give it a try. Please remember this number:
3471892341
Did you manage? If you’re like most people, including the author, you’ll find
it difficult or even impossible to remember the number this way.
Now try to memorize this number:
347 189 2341
 7 Memory    
63

It should have been much easier for you now. The difference is that in the
first case the number consists of ten digits that you have to remember, while
in the second case only three chunks, each consisting of three to four digits,
need to be remembered.
As we have just seen, this chunking interestingly works even when the
individual chunks have no special meaning. However, if the chunks have a
meaningful significance, the storage capacity of short-term memory can be
significantly further increased, as it can then be applied recursively, which is
also referred to as hierarchical chunking.
Try to remember these two characters:

你好

Unless you happen to speak and especially read Mandarin, it is virtually

impossible to quickly memorize these two characters, which by the way
mean “You good” or “Hello”. This is because in reality you would have to
remember not just two, but actually—depending on the count—form, size,
position, and orientation of about twenty individual objects, the strokes.
If you had learned the characters before because you learned Mandarin,
then instead of the twenty strokes, you would now only have to remember
two individual objects, the characters for “you” and “good” in short-term
memory, or if you know that both together mean as much as “Hello”, even
only a single object.

Working Memory
Working memory is a theoretical concept that plays a central role in cog-
nitive psychology, neuropsychology, and cognitive neuroscience. It is an
important cognitive system that allows us to temporarily store and manip-
ulate information. This system is essential for many higher cognitive func-
tions such as logical reasoning, problem-solving, decision-making, and
behavioral control.
Working memory consists of several subcomponents (Baddeley & Hitch,
1974), each of which has its own function (Fig. 7.2). The central executive
component is responsible for controlling attention and coordinating infor-
mation processing. Another component is the phonological loop, which is
responsible for the temporary storage of verbal information. It can be further
divided into the articulatory loop, which is responsible for the subvocal rep-
etition of verbal information, and the acoustic buffer, which is responsible
64    
P. Krauss

Central executive

phonological episodic visuo-spatial

loop buffer scratchpad

articulatory auditory
loop buffer

Fig. 7.2 Multicomponent model of working memory. Working memory consists of

several subcomponents. The central executive controls attention and coordinates infor-
mation processing. The phonological loop stores verbal information, while the episodic
buffer integrates information from different modalities and the visuospatial sketchpad
enables the temporary storage and manipulation of visuospatial information

for the temporary storage of linguistic information. The episodic buffer

is another component of working memory, serving as a temporary storage
system for the integration of information from different modalities, such as
visual and auditory information. Finally, the visuospatial sketchpad is respon-
sible for the temporary storage and manipulation of visuospatial information.
We use this system, for example, when mentally rotating an object.
The neural correlates of short-term and working memory are the result
of numerous complex brain activities involving a multitude of brain regions
(Goldman-Rakic, 1995). Studies have shown that the central executive com-
ponent of working memory is associated with activities in the prefrontal
cortex and the basal ganglia (Hikosaka et al., 1989; Graybiel, 1995). The
phonological loop is associated with activities in the left hemisphere, par-
ticularly in the parietal and temporal lobes. The visuospatial sketchpad is
associated with activity in the posterior parietal lobe and the prefrontal cor-
tex region (Milner et al., 1985). The episodic buffer, on the other hand, is
associated with activity in the medial temporal region and the hippocampus.
The thalamus is also believed to play an important role in maintaining work-
ing memory through the control and modulation of feedback loops with the
cortex (Watanabe & Funahashi, 2012). Finally, the cerebellum is also attrib-
uted a (yet not fully understood) function in the control and regulation of
short-term memory (Ivry, 1997; Desmond & Fiez, 1998).
 7 Memory    
65

Long-term Memory
Long-term memory is essentially based on synaptic plasticity (Bailey &
Kandel, 1993) and allows us to store and retrieve information over longer
periods—from hours to decades. Two categories of this type of memory are
distinguished: declarative (explicit) memory and non-declarative, procedural
(implicit) memory (Eichenbaum & Cohen, 2004).
Declarative memory refers to the conscious retrieval of information and
can be further subdivided into semantic memory, which contains general
factual knowledge, and episodic memory, which contains specific events and
their sequence (Tulving, 1972, 1983). These do not necessarily have to be
events or episodes experienced by oneself, but can also be movies watched or
the plot of books read. A special case of episodic memory is autobiograph-
ical memory, which contains memories of specific events from one’s own
life (Conway & Pleydell-Pearce, 2000). The medial temporal lobe and the
hippocampus are important brain regions involved in the formation and
retrieval of declarative long-term memories (Milner, 1970; Zola-Morgan
& Squire, 1986); Squire, 1987, 1992; Gazzaniga et al., 2006). As we have
already seen, it is particularly the case that the hippocampus acts as a kind
of intermediate storage for declarative memories (Scoville & Milner, 1957;
Penfield & Milner, 1958) and transfers its information to the cortex during
deep sleep. This process is also referred to as memory consolidation (Squire,
1992; Conway & Pleydell-Pearce, 2000).
Non-declarative, procedural memory, on the other hand, refers to our
unconscious memory for skills, procedures, and motor abilities such as play-
ing the piano, swimming, or cycling. This type of memory often manifests
itself in our behavior and does not need to be consciously retrieved. The cer-
ebellum and the basal ganglia are key regions of the brain involved in the
formation and retrieval of procedural long-term memory (Graybiel, 1995;
Desmond & Fiez, 1998).

Conclusion
In summary, memory is an integral part of human cognition that allows
us to encode, manipulate, store, and retrieve information over time.
Understanding how memory functions as an information processing system
is crucial to understanding its role in our experience and behavior.
66    
P. Krauss

The study of human memory has significant implications for the develop-
ment of Artificial Intelligence. If we understand how human memory works,
we can develop AI systems that learn and process information more effi-
ciently than previous systems. One of the most important areas of research
in this context is the development of working memory models for AI sys-
tems. Working memory is crucial for many cognitive tasks, including prob-
lem-solving and decision-making. By integrating a working memory into AI
systems, their abilities to handle complex tasks and make more accurate pre-
dictions could be significantly improved.
In addition, the study of long-term memory has implications for the
development of Artificial Intelligence that is supposed to learn and adapt
over time. Just as humans store and retrieve information over long periods,
these systems could be designed to store and retrieve large amounts of data
and use this information to improve their performance over time. This is
particularly important in areas such as natural language processing or real
image processing, where the systems constantly need to adapt to new data
and contexts.
Since human memory is a complex, multi-layered system that is not yet
fully understood, the development of AI systems that simulate human mem-
ory still poses a major challenge.

References
Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: A proposed sys-
tem and its control processes. In Psychology of learning and motivation (Vol. 2,
pp. 89–195). Academic: Elsevier.
Baddeley, A. D., & Hitch, G. (1974). Working memory. In Psychology of learning
and motivation (Vol. 8, pp. 47–89). Academic, Elsevier.
Bailey, C. H., & Kandel, E. R. (1993). Structural changes accompanying memory
storage. Annual Review of Physiology, 55(1), 397–426.
Conway, M. A., & Pleydell-Pearce, C. W. (2000). The construction of autobiograph-
ical memories in the self-memory system. Psychological Review, 107(2), 261.
Cowan, N. (2001). The magical number 4 in short-term memory: A reconsidera-
tion of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.
Desmond, J. E., & Fiez, J. A. (1998). Neuroimaging studies of the cerebellum:
Language, learning and memory. Trends in Cognitive Sciences, 2(9), 355–362.
Eichenbaum, H., & Cohen, N. J. (2004). From conditioning to conscious recollection:
Memory systems of the brain (No. 35). Oxford University Press on Demand.
 7 Memory    
67

Funahashi, S., Bruce, C. J., & Goldman-Rakic, P. S. (1989). Mnemonic cod-

ing of visual space in the monkey’s dorsolateral prefrontal cortex. Journal of
Neurophysiology, 61(2), 331–349.
Fuster, J. M., & Alexander, G. E. (1971). Neuron activity related to short-term
memory. Science, 173(3997), 652–654.
Gazzaniga, M. S., Ivry, R. B., & Mangun, G. R. (2006). Cognitive Neuroscience. The
biology of the mind. Norton.
Goldman-Rakic, P. S. (1995). Cellular basis of working memory. Neuron, 14(3),
477–485.
Graybiel, A. M. (1995). Building action repertoires: Memory and learning func-
tions of the basal ganglia. Current Opinion in Neurobiology, 5(6), 733–741.
Hikosaka, O., Sakamoto, M., & Usui, S. (1989). Functional properties of mon-
key caudate neurons. III. Activities related to expectation of target and reward.
Journal of Neurophysiology, 61(4), 814–832.
Ivry, R. (1997). Cerebellar timing systems. International Review of Neurobiology, 41,
555–573. PMID: 9378608.
Kandel, E. R. (2007). In search of memory: The emergence of a new science of mind.
Norton.
Luck, S. J., & Vogel, E. K. (1997). The capacity of visual working memory for fea-
tures and conjunctions. Nature, 390(6657), 279–281.
Mandler, G. (1967). Organization and memory. In Psychology of learning and moti-
vation (Vol. 1, pp. 327–372). Academic, Elsevier.
Milner, B., Petrides, M., & Smith, M. L. (1985). Frontal lobes and the temporal
organization of memory. Human Neurobiology, 4(3), 137–142.
Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits
on our capacity for processing information. Psychological Review, 63(2), 81.
Milner, B. (1970). Memory and the medial temporal regions of the brain. Biology of
Memory, 23, 31–59.
Milner, B., Squire, L. R., & Kandel, E. R. (1998). Cognitive neuroscience and the
study of memory. Neuron, 20(3), 445–468.
Penfield, W., & Milner, B. (1958). Memory deficit produced by bilateral lesions in
the hippocampal zone. AMA Archives of Neurology & Psychiatry, 79(5), 475–497.
Rouder, J. N., Morey, R. D., Cowan, N., Zwilling, C. E., Morey, C. C., & Pratte,
M. S. (2008). An assessment of fixed-capacity models of visual working memory.
Proceedings of the National Academy of Sciences, 105(16), 5975–5979.
Scoville, W. B., & Milner, B. (1957). Loss of recent memory after bilateral hip-
pocampal lesions. Journal of Neurology, Neurosurgery, and Psychiatry, 20(1), 11.
Sperling, G. (1963). A model for visual memory tasks. Human Factors, 5(1), 19–31.
Squire, L. R. (1987). Memory and brain. Oxford University Press.
Squire, L. R. (1992). Memory and the hippocampus: A synthesis from findings
with rats, monkeys, and humans. Psychological Review, 99(2), 195.
68    
P. Krauss

Squire, L. R., Knowlton, B., & Musen, G. (1993). The structure and organization
of memory. Annual Review of Psychology, 44(1), 453–495.
Tulving, E. (1972). Episodic and semantic memory. In Organization of memory
(pp. 381–402). Academic, Elsevier.
Tulving, E. (1983). Elements of episodic memory. Oxford University Press.
Watanabe, Y., & Funahashi, S. (2012). Thalamic mediodorsal nucleus and working
memory. Neuroscience & Biobehavioral Reviews, 36(1), 134–142.
Zola-Morgan, S., & Squire, L. R. (1986). Memory impairment in monkeys follow-
ing lesions limited to the hippocampus. Behavioral Neuroscience, 100(2), 155.
 8
Language

Language is, by its nature, a communal thing, that is, it never expresses the exact,
but a compromise—that which is common to you, me, and everyone else.
Thomas Ernest Hulme

How does Man Come to Language?

The question of how man comes to language is one of the oldest scien-
tific puzzles. According to the Greek historian Herodotus, the Egyptian
Pharaoh Psammetichus tried to fathom the origin of language 2500 years
ago. To this end, he conducted an experiment with two children, whom
he handed over to a shepherd as newborns to feed and care for them, but
who was instructed not to speak to them. Psammetichus hypothesized that
the first word of the infants would be spoken in the original language of
all humans. When one of the children cried “bekos”, which was the sound
of the Phrygian word for “bread”, Psammetichus concluded that Phrygian
was the original language of all humans (Rawlinson & Wilkinson, 1861).
Behind this cruel experiment was obviously the assumption that humans are
born with innate words and their meanings and that this original language is
somehow deformed or overwritten during individual development and first
language acquisition.
Nowadays, of course, it is clear that words and meanings are not
innate, but are learned during language acquisition, and that there is no
causal relationship between the sound pattern and the meaning of a word

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 69
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_8
70    
P. Krauss

(de Saussure, 1916). However, it is still very controversial to what extent lan-
guage skills are innate or need to be learned (Goodluck, 1991).

Noam Chomsky’s Universal Grammar

According to the theory of universal grammar, every human being has an
innate, genetically determined language ability, which, for example, dis-
tinguishes between different word classes such as nouns and verbs, which
should not only facilitate, but even make it possible for children to learn to
speak (Chomsky, 2012, 2014).
The theory states that there are certain universal principles and rules that
underlie all human languages, and that these principles are hard-wired in
our brains from birth. The concept of universal grammar was first proposed
by the linguist Noam Chomsky in the 1950s. Chomsky held the view that
children are born with an innate knowledge of the basic grammatical struc-
ture of language, which enables them to quickly learn the specific rules of
their mother tongue when confronted with it.
One of the basic assumptions of universal grammar is that all languages
have a set of basic structural features, such as the use of verbs, nouns, and
adjectives, and the ability to form sentences with subject, verb, and object.
It is believed that these features reflect the innate cognitive abilities of the
human brain and are present in all languages regardless of their specific lin-
guistic features.

Cognitive Linguistics: Usage-Based Approaches

In contrast to Universal Grammar and other approaches in linguistics that
focus on formal rules and structures, Cognitive Linguistics investigates how
language is processed and represented in our brain. It also explores how lan-
guage and cognition interact and how they are influenced by factors such
as culture and social interaction. In so-called usage-based approaches (usage-
based approaches) to language acquisition, a profound relationship between
language structure and language use is assumed (Goldberg, 1995, 2003,
2019; Tomasello, 2005; Langacker, 2008). In particular, it is assumed that
contextual mental processing and mental representations have the cognitive
capacity to capture the complexity of actual language use at all levels (Bybee
et al., 1994; Hopper & Bybee, 2001; Bybee, 2013; Diessel et al., 2019;
Schmid, 2020).
 8 Language    
71

In the usage-oriented approaches, the importance of a child’s experience

with language for the language acquisition process is emphasized. These the-
ories assume that children learn language by observing and imitating the
language usage patterns in their linguistic environment. In particular, the
idea that there is a fixed, innate language knowledge that children possess
from birth, as postulated by the theory of Universal Grammar, is rejected.
Instead, it is argued that children build their linguistic knowledge through
the accumulation of language usage patterns.
An important concept in usage-based approaches is the idea of construc-
tion learning. According to this theory, children learn to use language by
acquiring certain language usage patterns or constructions that they hear in
the language around them. Constructions are form-meaning pairs and can
range from individual words to simple phrases and sentences to more com-
plex structures and are learned through repeated contact and use. Cognitive
Linguistics views language as a complex system of constructions that are
formed on the basis of experience and perception and shape our linguistic
abilities and our understanding of language. This implies that, contrary to
the assumptions of classical linguistics, there is no strict separation between
words (lexicon) and grammar rules (syntax), but rather a lexicon-syntax con-
tinuum, i.e., there can also be all possible transitional forms from words on
the one hand and rules on the other. An example of this would be so-called
collocations.
A collocation is a combination of words that often occur together and are
perceived as a natural unit in language use. These combinations are often
idiomatic and cannot always be explained by the meaning of their individual
words. The use of a collocation can contribute to making a sentence sound
more fluent and natural. A collocation is particularly present when there are
several possibilities to express a fact through different word combinations,
but speakers or writers tend to prefer a certain combination, otherwise the
statement loses its naturalness. For example, “strong soup” and “strong
coffee” sound much more natural than “strong soup” and “strong coffee”,
although grammatically and semantically all combinations would be pos-
sible. With a strict separation of words and grammar rules, such phenom-
ena can only be poorly explained, as then all alternative word combinations
should occur approximately equally frequently.
Usage-based approaches also emphasize the role of social interaction and
communication in language acquisition. They argue that language is learned
through social interaction with other speakers and that children’s under-
standing of language is shaped by the communicative functions that lan-
guage fulfills in social situations.
72    
P. Krauss

Language in the Brain

Language is probably the most complex cognitive ability of humans, involv-
ing all cognitive subsystems: attention, movement planning and control,
memory, perception, etc. Accordingly, language processing in the brain con-
sists of complex, interconnected neural networks that are responsible for
various aspects of language, including language comprehension, language
production, and language acquisition.
The primary areas of the brain responsible for language processing are
located in the left hemisphere, particularly in the regions of the frontal, tem-
poral, and parietal lobes (Pulvermüller, 2002). These areas work together
to process various aspects of language, e.g., phonology (the sounds of lan-
guage), syntax (the rules for combining words into sentences), semantics
(the meaning of words and sentences), and pragmatics, i.e., the use of lan-
guage in social contexts (Herbst, 2010).
The Broca area is a brain region in the left frontal lobe and is considered
the motor speech center. It is named after the French neurologist Paul Broca,
who was the first to recognize the importance of this area for language pro-
cessing. The Broca area is primarily involved in language production, par-
ticularly in the ability to produce language and form grammatically correct
sentences. Damage to this area can lead to a condition known as Broca’s
aphasia, characterized by difficulties in forming coherent sentences, although
the affected person still has a relatively good understanding of language.
Recent studies also suggest that the Broca area may play a role in other
cognitive functions beyond language processing, such as working memory,
attention, and decision-making (Kemmerer, 2014).
The second important area in language processing is the Wernicke area,
which is also considered the sensory language center and is located in the
posterior part of the left temporal lobe near the auditory areas (Wernicke,
1874). It is named after the German neurologist Carl Wernicke, who was
the first to recognize that the Wernicke area is primarily involved in lan-
guage comprehension, particularly in the ability to understand and interpret
spoken and written language. Damage to this area can lead to a condition
known as Wernicke’s aphasia, characterized by difficulties in language com-
prehension, but also in the production of coherent language. However,
recent studies also suggest that the Wernicke area may play a role in other
cognitive functions beyond language processing, such as visual perception
and memory (Kemmerer, 2014). These recent findings on the function of
the Broca and Wernicke areas point to two problems that neuroscientist
 8 Language    
73

David Poeppel refers to as the Maps Problem and the Mapping Problem (see
also Chap. 5).
One of the most important processes in language comprehension is the
so-called parsing, in which the brain breaks down a sentence into its com-
ponents and assigns a meaning to each part. This requires working memory,
attention, and the ability to quickly process and integrate information from
various sources.
Language production involves a similar series of processes, in which the
brain first generates a message and then encodes it into a sequence of words
and grammatical structures. This process also requires the integration of vari-
ous information sources, including semantic knowledge, syntactic rules, and
social context (Kemmerer, 2014).
Language acquisition is a complex process in which various cognitive and
neural mechanisms such as attention, memory, and social perception inter-
act. It is assumed that children are particularly sensitive to language during
critical developmental phases and that contact with language during these
phases is crucial for the acquisition of language competence.
According to a new model proposed for understanding the functional
anatomy of language, the early stages of language processing take place in
the auditory areas located in the temporal lobe of the cortex on both sides of
the brain. Analogous to the visual system, which has a ventral “where” and
a dorsal “what” path of processing, language processing in the cortex also
divides into a ventral and a dorsal path (Hickok & Poeppel, 2004, 2007,
2016). While the “what” path is responsible for assigning sounds to mean-
ings, the “where” path maps sounds onto motor-articulatory representations
(where is the sound produced).
Of course, not only the cortex is involved in the processing of language.
The thalamus, the basal ganglia, and the cerebellum also play an important
role in language perception and production.
The thalamus, as a sensory relay station in the brain that receives and
processes incoming sensory information before it is forwarded to the cor-
responding cortical areas for further processing, is also involved in the pre-
processing of auditory information, including speech sounds.
The basal ganglia are involved in the planning and execution of speech
movements. They also play a role in the formation of grammatical structures
and in the selection of appropriate words during language production.
And finally, the cerebellum, traditionally associated with motor coordi-
nation and balance, is also involved in cognitive processes such as language
processing. It is responsible for the coordination of speech movements as
74    
P. Krauss

well as for the timing of language production and plays an important role in
learning (foreign) languages.

Conclusion
The question of how humans acquire language has occupied science for
millennia. The study of human language and language development has
deepened our understanding of how language is acquired, processed, and
represented in the brain. Theories such as Chomsky’s Universal Grammar
and usage-based approaches offer different perspectives on language acquisi-
tion, while neuroscience research expands our knowledge of the underlying
brain mechanisms.
In the context of Artificial Intelligence and large language models like
GPT-4, these insights provide important knowledge for the development
and improvement of language processing systems. The study of human lan-
guage abilities and language development can help optimize the architecture
and learning mechanisms of such models and improve their ability to under-
stand and generate natural language.
Future research in linguistics and neuroscience could help further close
the gap between human and artificial language processing. The integration
of insights from various disciplines such as cognitive linguistics, neurosci-
ence, and AI research could lead to even more powerful and human-like lan-
guage models. Furthermore, such models could deepen our understanding
of human language and cognition by serving as tools for investigating lin-
guistic phenomena and cognitive processes.

References
Bybee, J. L., Perkins, R. D., & Pagliuca, W. (1994). The evolution of grammar:
Tense, aspect, and modality in the languages of the world (Vol. 196). University of
Chicago Press.
Bybee, J. L. (2013). Usage-based theory and exemplar representations of con-
structions. In Th. Hoffmann & G. Trousdale (Eds.), The Oxford Handbook of
Construction Grammar (online edn, 16 Dec. 2013), Oxford Academic. https://
doi.org/10.1093/oxfordhb/9780195396683.013.0004. Accessed 9 Aug 2023.
Chomsky, N. (2012). On the nature, use and acquisition of language. In Language
and meaning in cognitive science (pp. 13–32). Taylor and Francis.
Chomsky, N. (2014). Aspects of the theory of syntax (Vol. 11). MIT press.
 8 Language    
75

De Saussure, F. (1916). Nature of the linguistic sign. Course in General Linguistics,

1, 65–70.
Diessel, H., Dabrowska, E., & Divjak, D. (2019). Usage-based construction gram-
mar. Cognitive Linguistics, 2, 50–80.
Goldberg, A. E. (1995). Constructions: A construction grammar approach to argument
structure. University of Chicago Press.
Goldberg, A. E. (2003). Constructions: A new theoretical approach to language.
Trends in Cognitive Sciences, 7(5), 219–224.
Goldberg, A. E. (2019). Explain me this: Creativity, competition, and the partial pro-
ductivity of constructions. Princeton University Press.
Goodluck, H. (1991). Language acquisition: A linguistic introduction. Blackwell.
Herbst, T. (2010). English Linguistics. De Gruyter Mouton.
Hickok, G., & Poeppel, D. (2004). Dorsal and ventral streams: A framework for
understanding aspects of the functional anatomy of language. Cognition, 92(1–
2), 67–99.
Hickok, G., & Poeppel, D. (2007). The cortical organization of speech processing.
Nature Reviews Neuroscience, 8(5), 393–402.
Hickok, G., & Poeppel, D. (2016). Neural basis of speech perception. In G.
Hickok & S. L. Small (Eds.), Neurobiology of Language (pp. 299–310).
Academic.
Hopper, P. J., & Bybee, J. L. (2001). Frequency and the emergence of linguis-
tic structure. In Typological Studies in Language (pp. 1–502). John Benjamins
Publishing Company. http://digital.casalini.it/9789027298034.
Kemmerer, D. (2014). Cognitive neuroscience of language. Psychology Press.
Langacker, R. (2008). Cognitive grammar as a Basis for Language Instruction. In
BookHandbook of Cognitive Linguistics and Second Language Acquisition (Vol. 1).
Imprint Routledge.
Pulvermüller, F. (2002). The neuroscience of language: On brain circuits of words and
serial order. Cambridge University Press.
Rawlinson, H. C., & Wilkinson, J. G. (1861). The history of Herodotus (Vol. 1). D.
Appleton & Co. New York.
Schmid, H. J. (2020). The dynamics of the linguistic system: Usage, conventionaliza-
tion, and entrenchment. Oxford University Press.
Tomasello, M. (2005). Constructing a language: A usage-based theory of language
acquisition. Harvard University Press.
Wernicke, C. (1874). Der aphasische Symptomencomplex: Eine psychologische Studie
auf anatomischer Basis. Cohn & Weigert.
 9
Consciousness

Consciousness is not a journey upward, but a journey inward.

Bernard Lowe

An age-old mystery
For more than 2000 years, the question of understanding consciousness has
been at the center of interest for many philosophers and scientists. Modern
philosophy distinguishes between an easy and a hard problem (Chalmers,
1995). While the easy problem consists of explaining the function, dynam-
ics, and structure of consciousness, the hard problem is to explain whether
and why any physical system, be it a human, an animal, a fetus, a cell orga-
noid, or an AI (Bayne et al., 2020), is conscious and not unconscious at all.
Throughout history, many different perspectives have been proposed, rang-
ing from the pessimistic view of “Ignorabimus”—which means as much as
“We will never know”1—to more optimistic mechanistic ideas that even aim
at the construction of an artificial consciousness. These different views have
led to ongoing debates and discussions about the nature of consciousness
and whether it can ultimately be understood or not.

1 Emildu Bois-Reymond made this statement at the 45th annual meeting of German natural scientists
and doctors in 1872 in his lecture on the limits of scientific knowledge, referring to the relationship
between brain processes and subjective experience.

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 77
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_9
78    
P. Krauss

At the core, it is about the mind-body problem or, more modernly

expressed, the problem of the relationship between the brain and the mind.
The core question is: How do mental, mental states and processes relate to
physical states and processes?
On the one hand, we have the brain, which is a material, physical object
that can be measured and studied. On the other hand, we have the mind,
which consists of subjective, conscious experiences that cannot be observed
or measured in the same way as physical properties. This raises two fun-
damental questions. The first is whether the mind is physical or some-
thing completely different. Some philosophers believe that the mind is
a non-physical entity that exists independently of the body, while others
believe that the mind is merely a byproduct of the physical processes in the
brain. The second question concerns the causal relationship between mind
and brain: How do physical processes in the brain lead to subjective con-
scious experiences? Some philosophers believe that there is a one-to-one
correspondence between the physical processes in the brain and the men-
tal processes in the mind. Others believe that mental processes cannot be
reduced to physical processes and that the mind is more than just the activ-
ity of the brain.
Closely related to this is the concept of qualia. These are subjective
experiences from the first-person perspective that we have when we per-
ceive or interact with the world. These experiences include sensations such
as color, taste, and sound, but also more complex experiences like feelings
and thoughts. Qualia are often considered ineffable, which means that they
cannot be fully captured or conveyed by language or other forms of rep-
resentation. They cannot, therefore, be studied from the “outside”, i.e., from
the third-person perspective. This has led some philosophers to claim that
qualia represent a special kind of phenomena that cannot be reduced to or
explained by physical or objective properties of the world.

Monism and Dualism

Over the course of history of philosophy, an innumerable variety of ideas
and concepts have been developed to solve the mind-body problem, all of
which can essentially be divided into one of two basic views: monism and
dualism.
Monism is the idea that there is only one kind of state or substance in the
universe. This concept goes back to the ancient Greek philosopher Aristotle,
who believed that mind and body relate to each other as form and matter.
 9 Consciousness    
79

For Aristotle, the various forms in the physical world were simply different
physical states, and there was no non-physical or mental substance beyond
the physical world.
Dualism, on the other hand, assumes that both mental and physical sub-
stances are possible. This thought was first developed by Plato, who believed
that mind and body exist in two separate worlds. According to Plato’s idea,
the mind was part of the ideal world of forms, it was immaterial, non-ex-
tended, and eternal. The idea of an ideal circle, for example, exists in the
mind as a perfect concept, although no physical circle in the material world
can ever truly correspond to it. In contrast, the body belongs to the mate-
rial world, it is extended and perishable. Concrete physical circles could be
found in the material world, but they would always be imperfect and change.

What is it Like to be a Bat?

The main problem in analyzing consciousness is its subjectivity. Our mind
is capable of perceiving and processing our own states of consciousness.
Through induction, we are also able to attribute conscious processes to other
people. However, as soon as we try to imagine being another species, as phi-
losopher Thomas Nagel describes in his groundbreaking essay “What is it
like to be a bat?” (Nagel, 1974), we immediately fail to consciously pursue
this experience.
The basic idea is to imagine what it would be like to experience the world
as a bat, from the bat’s own subjective perspective.

“…imagine that one has webbing on one’s arms,which enables one to fly
around at dusk and dawn catching insects in one’smouth; that one has very
poor vision, and perceives the surrounding world by asystem of reflected
high-frequency sound signals; and that one spends the dayhanging upside
down by one’s feet in an attic. In so far as I can imagine this(which is not very
far), it tells me only what it would be like for me tobehave as a bat behaves.
But that is not the question. I want to know what itis like for a bat to be a
bat.” (Nagel, 1974).

Nagel argues that it is impossible for us to fully understand the subjective

experience of a bat or any other creature, as we are fundamentally limited
by our own human perspective. We can study the behavior and physiology
of bats, but we will never truly know what it is like to perceive the world
through echolocation or to experience the unique sensory and cognitive
80    
P. Krauss

processes that a bat uses to navigate and interact with its environment in
total darkness.
According to Nagel’s view, consciousness is not just about objective facts
or physical processes, but also about subjective experiences, i.e., what it is
like to be a particular organism. He suggests that we need to develop a new
kind of science that takes into account subjective experience if we hope to
fully understand the nature of consciousness.
He further argues that the subjective nature of consciousness undermines
any attempt to decipher it with objective and reductionist means, i.e., the
means of natural sciences. He believes that the subjective character of experi-
ence cannot be explained by a system of functional or intentional states.
He concludes that consciousness cannot be fully understood if its subjec-
tivity is ignored, as it cannot be explained reductionistically, as it is a mental
phenomenon that cannot be reduced to materialism. Nagel concludes with
the assertion that physicalism, while not wrong, is also only incompletely
understood, as it lacks the characterization of subjective experience. This, in
turn, is a necessary prerequisite for understanding the mind-body problem.
At its core, he thus represents a dualistic perspective, much like Plato two
and a half millennia before him.

Consciousness as a Useful Illusion?

Philosopher Daniel Dennett, on the other hand, holds a monistic view and
contradicts Nagel’s claim that the consciousness of the bat is fundamentally
inaccessible to us. He justifies his view by stating that all important features
of the bat’s consciousness can be observed from the outside. For example, it
is clear that bats cannot recognize objects that are more than a few meters
away, as their biological echolocation system has a limited range. Dennett
believes that, analogously, all aspects of a bat’s subjective experience can be
discovered through further scientific investigations (Dennet, 1991).
Furthermore, Dennett holds the view that consciousness is not a unified
phenomenon, but rather a collection of mental processes that constantly
change and interact with each other. He describes consciousness as a kind
of useful illusion that arises from the functioning of the brain (Cohen &
Dennet, 2011). Similarly, neuroscientist Anil Seth equates our subjective
experience with a constantly newly generated hallucination, which is con-
tinuously compared with the input from the sensory organs from the world
(Seth, 2021). For Dennett, the subjective experience of consciousness is an
emergent property of brain activity and not a separate thing that exists in
 9 Consciousness    
81

itself. He describes consciousness as a “virtual machine” that the brain creates

to help us navigate the world. Dennett also holds the view that consciousness
does not take place in a single central location in the brain, but rather arises
from a series of parallel, distributed processes that constantly generate and
update multiple different drafts of our experience (Dennet, 1991).
Dennett’s perspective on consciousness emphasizes the importance of
understanding the underlying mechanisms of the brain in order to com-
prehend the nature of subjective experience. He is skeptical of traditional
dualistic views of consciousness that postulate a separate immaterial soul or
mind, and instead emphasizes the importance of studying consciousness as a
product of the physical brain.

Limits of the Philosophy of Mind

A disadvantage of exploring consciousness with philosophical means is that
we will never be able to investigate the underlying physical, biological, or
information-theoretical processes. By thinking alone, we will not be able to
open the black box.
While philosophy plays an important role in expanding our knowledge
and understanding of the nature of consciousness, its ability to ask more
comprehensive questions, to look at the bigger picture, and to explore the
relationships between different areas, rightly makes it an essential part of
cognitive science. However, philosophy can only help us gain new insights
and a more comprehensive understanding of consciousness through collabo-
ration with empirical sciences.

Neural Correlates of Consciousness

In neuroscience, it is currently assumed that consciousness functions as a
kind of user interface (Seth & Bayne, 2022). It comes into play whenever
we have to find our way in new situations, anticipate possible future events,
plan actions, consider different scenarios, and choose between them (Graves
et al., 2011). While the philosophical tradition from Aristotle to Descartes
considered consciousness to be an exclusively human phenomenon, modern
neuroscience tends to view consciousness as a gradual phenomenon that can
fundamentally also occur in animals (Boly et al., 2013).
The neural correlates of consciousness (NCC) refer to specific patterns of
neural activity that are believed to be associated with conscious experience
82    
P. Krauss

(Crick & Koch, 2003; Koch, 2004). In other words, NCCs are the brain
processes necessary for the occurrence of consciousness. The concept of NCC
is based on the idea that there is a close relationship between neural activity
and conscious experience. When we perceive something consciously, there
are certain patterns of neural activity that are always associated with this
experience. By investigating these patterns of neural activity, we hope to gain
a better understanding of the mechanisms underlying conscious experience.
There are various approaches to identifying NCCs. One approach is to com-
pare brain activity during conscious and unconscious states such as sleep or
anesthesia. Another approach is to investigate the changes in neural activity
that occur when a person perceives a stimulus, e.g., a visual or auditory cue.
Francis Crick2 and Christof Koch proposed that brain waves with a fre-
quency range between 30 and 100 cycles per second, so-called gamma oscil-
lations, play a crucial role in the emergence of consciousness (Crick & Koch,
1990). Koch further developed this concept and investigated the neural
correlates of consciousness in humans (Tononi & Koch, 2008; Koch et al.,
2016). Accordingly, activity in the primary visual cortex is essential for con-
scious perception, but not sufficient, as activity in the hierarchically higher
cortical areas of the visual system is more closely correlated with the various
aspects of visual perception and damage to these areas can selectively impair
the ability to perceive certain features of stimuli (Rees et al., 2002)—a phe-
nomenon referred to as agnosia. We will return to this later in this chapter.
Furthermore, Koch suggests that the precise timing or synchronization of
neural activity may be much more important for conscious perception than
simply the extent of neural activity. Recent studies using imaging techniques
on visually triggered activity in parietal and prefrontal cortex regions seem to
confirm these hypotheses (Boly et al., 2017).
Based on the idea of the neural correlates of consciousness, a kind of
measurement procedure was even developed, which aims to determine the
degree of consciousness (Perturbational Complexity Index) e.g., in comatose
patients, independently of the activity associated with sensory and motor
processing (Seth et al., 2008; Casali et al., 2013; Casarotto et al., 2016).
While this approach may be suitable for quantifying the complexity of neu-
ral activity (Demertzi et al., 2019), it does not provide information about
the underlying neural circuits and the “algorithms” implemented in them.

2 This is, by the way, the same Francis Crick who deciphered the double helix structure of DNA in

1953 together with James Watson.

 9 Consciousness    
83

Integrated Information Theory

Building on observations of the neural correlates of consciousness, Giulio
Tononi proposed the Integrated Information Theory (IIT) (Tononi et al.,
2016). This theoretical framework attempts to explain how conscious experi-
ence arises from the physical activity of the brain (Tononi, 2004). According
to the theory, consciousness is not simply the result of the activity of indi-
vidual neurons or brain regions, but arises from the integrated activity of the
brain as a whole. Tononi argues that an information-processing system can
only be conscious if the information is integrated into a unified whole. In
other words, it must be impossible to decompose the system into quasi-in-
dependent parts, as these parts would otherwise appear as two separate con-
scious entities (Koch, 2013).
The theory further assumes that the key to consciousness is the degree of
information integration in the brain. This degree of integration is supposed
to be quantifiable by a mathematical quantity called “phi”, which repre-
sents the amount of causal information generated by interactions between
different parts of the brain. The higher the phi value, the more integrated
the information processing in the brain and the more conscious the system
should be. The Integrated Information Theory assumes that consciousness is
a fundamental property of certain physical systems that are capable of influ-
encing themselves. Accordingly, any system with a phi value above a certain
threshold would be conscious to some degree.
Theoretical physicist Max Tegmark even generalized Tononi’s framework
further from a consciousness based on neural networks to any physical
quantum systems. He proposed that consciousness can be understood as a
state of matter with pronounced information processing capabilities—thus
quasi an additional state of aggregation besides solid, liquid and gaseous.
He suggested the name “perceptronium” for this type of conscious matter
(Tegmark, 2014).

Is the Brain a Quantum Computer?

Where we are already at physics, we want to turn to another question—at
least for physicists—that is obvious.
Although there is a large consensus that consciousness can be under-
stood as a form of information processing (Cleeremans, 2005; Seth,
2009; Reggia et al., 2016; Grossberg, 2017; Dehaene et al., 2014, 2017),
84    
P. Krauss

there is a disagreement about what the appropriate level of description is

(Kriegeskorte & Douglas, 2018). Physicist Roger Penrose and physician
Stuart Hameroff hypothesized that the brain is a kind of quantum com-
puter. In their view, quantum calculations take place in the so-called micro-
tubules of the cell skeleton of neurons (Penrose, 1989, 1994; Hameroff &
Penrose, 1996a, b; Hameroff, 2001; Hameroff et al., 2002). However, this
view is highly controversial.
Max Tegmark and Christof Koch argue that the brain can be understood
within a purely neurobiological framework, without having to resort to
quantum mechanical properties: Quantum computations, which are based
on the phenomenon of entanglement of (sub-)atomic particles, require on
the one hand that the qubits (quantum mechanical analogue to the bit)
are perfectly isolated from the rest of the system, while on the other hand
a coupling of the system with the outside world is necessary for the input,
control and output of the computations (Nielsen & Chuang, 2002). Due to
the wet and warm nature of the brain, all these operations introduce distur-
bances in the form of noise into the computations, leading to decoherence
of the quantum states and thus making quantum computations impossible.
Furthermore, they argue that the molecular machines of the nervous system,
such as the pre- and postsynaptic receptors, are so large that they can be
treated as classical physical systems and not as quantum systems. They con-
clude that cognition can be fully understood within the theoretical frame-
work of neural networks, without having to consider quantum phenomena.
(Tegmark, 2000; Koch & Hepp, 2006, 2007; Koch, 2013).

Global Workspace Theory

As an alternative and competing theory to the Integrated Information
Theory, Baars introduced the concept of a virtual global workspace (Global
Workspace) to describe consciousness in the 1990s, which arises from the
networking of various brain areas (Newman & Baars, 1993; Baars, 1994;
Baars & Newman, 1994; Baars, 2017). This global workspace allows the
integration and exchange of information between different specialized brain
regions, and it is assumed that it is responsible for the emergence of con-
scious experiences by enabling the selective transmission of information rele-
vant to the individual’s current goals and needs.
In the Global Workspace Theory (GWT), conscious perception is com-
pared to a spotlight that illuminates the contents of the global workspace
and makes it consciously perceptible (Mashour et al., 2020). The contents
 9 Consciousness    
85

of the workspace can be sensory information, thoughts, feelings, and memo-

ries. The theory states that the more comprehensively and intensively infor-
mation is processed in the global workspace, the more likely it is to reach
the threshold of consciousness. This means that stimuli that are particu-
larly noticeable, novel, or emotionally significant reach consciousness more
quickly.
The idea of the global workspace was later taken up and further developed
by Stanislas Dehaene, and neural correlates were proposed that underlie
it, especially distributed networks of prefrontal, parietal, and sensory cor-
tex areas (Dehaene et al., 1998; Dehaene & Naccache, 2001; Dehaene &
Changeux, 2004; Sergent & Dehaene, 2004; Dehaene et al., 2011, 2014)
(Fig. 9.1).
Starting from the implications of this theory, namely that “consciousness
arises from specific types of information processing processes that are physically
realized by the hardware of the brain” (Dehaene et al., 2017), Dehaene argues
that

“a machine equipped with these processing capacities would behave as if it had

consciousness; it would, for example, know that it sees something, express its
confidence in what it has seen, communicate it to others, have hallucinations
when its control mechanisms fail, and even experience the same perceptual
illusions as humans” (Dehaene et al., 2017) .

Attention system
Long-term memory

Limbic system

Global
workspace

Motor system

Perception systems

Fig. 9.1 Global workspace. Various neural processes interact with each other and
exchange information. The global workspace acts as a central “stage” where con-
sciously perceived information is processed and coordinated
86    
P. Krauss

The feeling of what happens!—Damasio’s

Model of Consciousness
The neuroscientist and psychologist Antonio Damasio has created the
world’s largest database of brain lesions. It contains information about the
location of the lesion and the associated cognitive deficits of tens of thou-
sands of patients, which he has collected over decades. Based on this data-
base and his observations, he has developed his own theory of consciousness.
His view of consciousness can be summarized in the idea that conscious-
ness arises from the interaction between our biological self and the environ-
ment with which we interact (Damasio, 1999, 2014; Damasio & Meyer,
2009). For Damasio, consciousness is not a single thing or a single pro-
cess, but a complex and dynamic phenomenon that arises from the inter-
action between different brain regions and other body systems. In his view,
consciousness is closely associated with emotions, feelings, and subjective
experiences.
Damasio assumes that the neural systems underlying consciousness are
hierarchically organized, with basic sensory and motor processing taking
place at lower levels and more complex cognitive and emotional processing
at higher levels. Consciousness arises when information from these different
levels is coherently and uniformly integrated.
One of Damasio’s key ideas is the concept of the body loop, which refers
to the constant flow of information between the brain and the rest of the
body. According to Damasio, this loop is essential for the generation of con-
scious experiences, as it allows the brain to constantly monitor and respond
to the state of the body and the environment.
Damasio also emphasizes the importance of emotions in the generation of
conscious experiences. He argues that emotions are not simply mental states,
but complex physical reactions that play a crucial role in how we perceive
the world and control our behavior.

Emotions and Feelings

In Damasio’s theory, feelings and emotions are different but interconnected

phenomena (Damasio, 2001). He defines emotions as direct signals from
the body, indicating a positive or negative state. These complex patterns of
physiological reactions are triggered by stimuli from the environment. In
contrast, feelings are quasi second-order emotions, which correspond to the
conscious perception of these physiological reactions.
 9 Consciousness    
87

Emotions are body states caused by so-called somatic markers, i.e., any-
thing that any receptors inside the body can measure, such as blood pressure,
heart rate, blood pH, blood sugar level, concentration of certain hormones,
and so on and so forth. The totality of all somatic markers, i.e., all measure-
ments from inside the body at a certain point in time, corresponds to the
body state. This is represented and controlled by a multitude of brain regions,
including the amygdala, the insula, and the prefrontal cortex. These brain
regions work together to represent the state of the organism and to evoke and
control the coordinated physiological reactions to environmental stimuli in a
feedback loop, e.g., changes in heart rate, breathing, and hormone levels.
Feelings, on the other hand, arise from the conscious perception and
interpretation of these physiological reactions. For example, the feeling of
fear can be evoked by the physiological reactions associated with the emo-
tion of fear, such as an increased heart rate and sweating. In Damasio’s view,
feelings give us a conscious representation of our bodily states and enable us
to understand and respond to the world around us.
Damasio also emphasizes the role of emotions and feelings in deci-
sion-making and behavior. He assumes that emotions provide us with
important information about the environment and contribute to steer-
ing our actions in an adaptive way. For example, the feeling of disgust can
help us avoid potentially harmful substances, while the feeling of love can
encourage us to seek social contacts.

Hierarchies of Consciousness

In Damasio’s theory of consciousness, a hierarchy of neuronal and psycho-

logical levels is assumed, which contribute to the generation of conscious
experience. He distinguishes between proto-self, core-self, and extended or
autobiographical self.
At the lowest level of this hierarchy is the proto-self, which arises from
the constant monitoring of the internal body state and external stimuli. In
humans and other mammals, it is continuously generated by neuronal activ-
ity in the brainstem and thalamus, integrating sensory and motor signals
from the entire body. The proto-self does not have the ability to recognize
itself. It is a pure processing chain that reacts to inputs and stimuli like a
machine, completely unconscious. According to this definition, every animal
with a nervous system, from the snail to the dolphin, also has a proto-self.
The next level in the hierarchy is the core-self, which arises from the inte-
gration of information from the proto-self with cognitive and emotional
88    
P. Krauss

processing at a higher level. Here, higher common representations of feelings

and objects from the environment are created, along with one’s own actions
and the emotional changes they cause. Thus, the core-self contributes to
the experience of agency, self-responsibility, and self-awareness, and enables
a more complex and differentiated understanding of ourselves as individu-
als. It can anticipate immediate reactions in its environment and adapt to
them. It is also able to recognize itself and its parts in its own image of the
world. This allows it to anticipate the world and react appropriately to it.
The core-self is also fleeting and unable to linger for hours and forge com-
plex plans. It is limited to the here and now. It remains constant through-
out the organism’s lifetime and is continuously recreated by mental patterns
that arise from the interaction with internal body states and external stim-
uli (objects). These interactions lead to a representation referred to as the
core-self, which is limited to the present moment. This can be imagined as
a continuous stream of consciousness, which is constantly recreated by these
interactions and gives our conscious experience a basic sense of coherence
and continuity. Unlike many philosophical approaches, the core-self does
not rely on representing the world in the form of language. Damasio even
believes that the fixation on language has hindered progress in understand-
ing the nature of consciousness. Damasio holds the view that every animal
capable of adapting to its environment also has a core-self.
At the top of the hierarchy is the extended or autobiographical self, which
encompasses our personal memories, beliefs, and narratives about ourselves
and our lives by accessing extensive memory systems. It enables a human-
like interaction with the environment and builds on the core-self. The ability
to process language also falls into the category of the extended self and can
be interpreted as a form of serialization of parallel consciousness contents. It
is assumed that this level of self arises from the integration of information
from the core-self with higher-level cognitive processes such as thinking and
introspection. According to Damasio, the extended self is closely linked to
our sense of identity and our understanding of ourselves as unique individ-
uals with a past, present, and future. Higher mammals such as cats, dogs,
chimpanzees, and dolphins also have an extended self, with this being most
developed and thus most pronounced in humans.
While the proto-self must be hardwired from birth to ensure the survival
of the organism, the other two levels of consciousness can be altered by
experience. In particular, the extended self, which depends on various types
of memory, develops gradually over the course of individual development.
From birth, children have a proto- and core-self. The extended self, however,
only comes into play during child development and becomes increasingly
 9 Consciousness    
89

pronounced and differentiated. New functions are added through the pro-
gressive myelination and the associated functional integration of ever new
cortex areas into the existing nervous system.

Explanatory Power of Damasio’s Model

Due to its strongly mechanistic character, Damasio’s model is able to explain

some of the most important neurological consciousness disorders compara-
tively simply.
The vegetative state, also known as a coma, is a state in which a person
appears to be awake but shows no conscious actions or reactions to their
environment. People in a vegetative state usually have their eyes open and
can maintain sleep-wake cycles, but are unable to respond to verbal, tactile,
or painful stimuli. There are also no signs of conscious perception, thoughts,
or feelings. The vegetative state can occur after a severe brain injury, stroke,
or other brain disease. According to Damasio’s model, the emergence of
the vegetative state can be traced back to two different causes: The brain no
longer receives information about the body’s state, or the parts of the brain
responsible for generating the core self are impaired.
The flow of sensory information from the body to the brain is crucial for
the formation and maintenance of the core self. In a vegetative state, this
flow can be interrupted or altered, which can prevent the brain from gener-
ating the core self. This interruption can have various causes, such as a trau-
matic brain injury, lack of oxygen in the brain, or another severe nervous
system disease. Without the continuous feedback loop from the body’s inte-
rior, the brain cannot build the core self and consciousness is lost. In some
cases, the neural structures responsible for the formation of the core self
can also be damaged, leading to a collapse of this basic level of conscious-
ness. This can occur as a result of a traumatic brain injury or other neuro-
logical disorders. In both cases, the core self collapses—in one case because
it lacks input (software problem), and in the other case because it can no
longer process input (hardware problem). Since the core self is the basis
for the autobiographical or extended self, this too is lost and consciousness
disappears.
The Locked-in Syndrome is a condition in which a person is unable to
move or speak due to damage to the brainstem, but is usually fully con-
scious. People with this syndrome can usually still move their eyelids and eye
muscles to communicate. However, they are unable to control their mus-
cles to actively move or speak, and are therefore proverbially trapped in their
90    
P. Krauss

own body. Here too, we can try to interpret the symptoms in the context of
Damasio’s consciousness model.
Since both the flow of sensory information from the body to the brain and
the neural structures responsible for generating the core self are not impaired
in Locked-in Syndrome, the core self remains intact. Despite the almost
complete lack of motor output, the brain still receives sensory input and
can thus continue to generate the core self. This allows the person to main-
tain a basic level of self-perception and consciousness. However, Locked-in
Syndrome interrupts the person’s ability to interact with the world through
motor actions, which can have long-term effects on the extended self (Kübler
& Birbaumer, 2008), as the extended self is based on autobiographical mem-
ories, personal identity, and the ability to plan and think about the future.
The inability to communicate and interact with the environment can lead
to difficulties in maintaining and updating the extended self. However, the
extended self does not completely disappear, as the individual is still able to
process and create new memories based on sensory experiences.
Agnosia is a neurological phenomenon in which people have difficulty
recognizing sensory information, even though the corresponding sensory
organs are intact. There are a wide variety of different forms of agnosia in
all sensory modalities. In visual agnosia, the recognition of visual stimuli is
disturbed, while vision is intact. Examples are prosopagnosia, where sufferers
can no longer recognize faces, object agnosia, which impairs object recogni-
tion, or color agnosia, where sufferers perceive the world in black and white
only. Particularly bizarre is akinetopsia, which impairs motion perception.
Patients suffering from it perceive the world only as a stroboscopic sequence
of single images. Depending on the form, aphasias impair various aspects
of language comprehension and speech. And there are many more of these
impairments of the extended self.
Of course, in all these cases, the affected individuals are fully conscious.
Their proto- and core self is completely intact. Only very specific functions
of their extended self are isolatedly impaired.

Conclusion
The mind-body problem is far from solved and remains the subject of
intense debates among philosophers, neuroscientists, and psychologists.
However, it is clear that our understanding of the relationship between mind
and body is essential for understanding the nature of human consciousness
and subjective experience.
 9 Consciousness    
91

The debate between monism and dualism continues to shape philosophi-

cal discourse on the nature of the mind. While monism postulates a unified
and interconnected worldview, dualism emphasizes the uniqueness of mind
and body and suggests that there might be more behind our experiences
than can be explained by physical phenomena alone.
Philosophy allows us to ask much broader questions than many other sci-
entific disciplines (Laplane et al., 2019). It is capable of looking at the big-
ger picture and providing important insights into the relationships between
different areas of knowledge. Philosophy is particularly important for the
interdisciplinary efforts of cognitive science, where it helps to bridge gaps
between different disciplines and pave new ways for research. Unlike scien-
tific methods, philosophizing is a non-empirical approach that attempts to
validate concepts through logical thinking and argumentation. Philosophers
tend to ask questions rather than provide definitive answers, and their con-
tributions often consist of challenging established assumptions and pro-
posing new research approaches. However, for a more comprehensive
understanding of the nature of consciousness, close collaboration between
philosophy and neuroscience is required (Lamme, 2010). This means that
while philosophy can provide valuable insights into theoretical concepts and
broader ethical questions, it needs to be supplemented by empirical findings
and experiments to reach a more comprehensive understanding.
The Integrated Information Theory (IIT) and the Global Workspace
Theory (GWT) are the two most prominent neuroscientific theories
attempting to explain the nature of consciousness. There are some similar-
ities between these theories, but also some significant differences. One of
the main differences between IIT and GWT is that they focus on different
aspects of consciousness. IIT emphasizes the subjective experience of con-
sciousness and attempts to quantify the amount of integrated information
present in a particular system, while GWT emphasizes the cognitive and
computational aspects of consciousness and the role of attention in shap-
ing conscious experience. Another significant difference between the two
theories is their approach to explaining the neural basis of consciousness.
IIT assumes that consciousness arises from the integration of information
in complex neural networks, while GWT assumes that consciousness arises
from the global activation of a distributed network of brain regions that serve
as a “workspace” for the integration and processing of information. In terms
of the nature of conscious experience, IIT suggests that consciousness is a
fundamental property of certain complex systems to influence themselves,
while GWT assumes that conscious experience is a byproduct of cognitive
processes that include attention and the global activation of neural networks.
92    
P. Krauss

Prominent representatives of both competing theories proposed an ambi-

tious and large-scale experiment, distributed across many labs worldwide,
in 2021 to decide between the two theories (Melloni et al., 2021). GWT
predicts that consciousness arises from the global transmission and amplifi-
cation of information through interconnected networks of prefrontal, pari-
etal, and sensory cortex areas, while IIT suggests that consciousness is based
on the intrinsic ability of a neural network to influence itself by generating
maximally integrated information, with the posterior, parietal cortex being
the ideal location for this. Various imaging techniques such as fMRI, EEG,
and MEG, as well as sophisticated experimental paradigms, will now be
used to test which prediction and thus which of the two theories is correct.
We can look forward to the results. The author would not be surprised if the
truth—as often in the history of science—lay somewhere in the middle and
the researchers found activation in all mentioned cortex regions and thus
evidence for both theories.
From the author’s perspective, the processes described by both theories are
probably necessary but not sufficient to actually generate consciousness. The
only theory that offers a mechanistic explanation is Damasio’s theory. While
according to Damasio’s framework the neural structures and processes of
GWT and IIT are fully accounted for in the extended self or consciousness,
his theory of somatic markers, the body loop, the proto-self and core self
provide a basis on which the other higher-level processes can build. In the
terminology of computer science, while GWT, IIT, and extended self cor-
respond to a user interface with various “apps” (language, visual perception,
etc.), the core self would be a kind of operating system and the proto-self a
kind of BIOS.

References
Baars, B. J. (1994). A global workspace theory of conscious experience. In
Consciousness in Philosophy and Cognitive Neuroscience, (pp. 149–171). Erlbaum.
Baars, B. J. (2017). The global workspace theory of consciousness. In The Blackwell
Companion to Consciousness, (pp. 236–246). Wiley.
Baars, B. J., & Newman, J. (1994). A neurobiological interpretation of global work-
space theory. Consciousness in Philosophy and Cognitive Neuroscience, 211–226.
Bayne, T., Seth, A. K., & Massimini, M. (2020). Are there islands of awareness?
Trends in Neurosciences, 43(1), 6–16.
Boly, M., Massimini, M., Tsuchiya, N., Postle, B. R., Koch, C., & Tononi, G.
(2017). Are the neural correlates of consciousness in the front or in the back of
 9 Consciousness    
93

the cerebral cortex? Clinical and neuroimaging evidence. Journal of Neuroscience,

37(40), 9603–9613.
Boly, M., Seth, A. K., Wilke, M., Ingmundson, P., Baars, B., Laureys, S., … &
Tsuchiya, N. (2013). Consciousness in humans and non-human animals: Recent
advances and future directions. Frontiers in Psychology, 4, 625.
Casali, A. G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S., Casali, K. R., …
& Massimini, M. (2013). A theoretically based index of consciousness independ-
ent of sensory processing and behavior. Science Translational Medicine, 5(198),
198ra105–198ra105.
Casarotto, S., Comanducci, A., Rosanova, M., Sarasso, S., Fecchio, M., Napolitani,
M., … & Massimini, M. (2016). Stratification of unresponsive patients by an
independently validated index of brain complexity. Annals of Neurology, 80(5),
718–729.
Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of
Consciousness Studies, 2(3), 200–219.
Cleeremans, A. (2005). Computational correlates of consciousness. Progress in Brain
Research, 150, 81–98.
Cohen, M. A., & Dennett, D. C. (2011). Consciousness cannot be separated from
function. Trends in Cognitive Sciences, 15(8), 358–364.
Crick, F., & Koch, C. (1990). Towards a neurobiological theory of conscious-
ness. In Seminars in the neurosciences (Vol. 2, pp. 263–275). Saunders Scientific
Publications.
Crick, F., & Koch, C. (2003). A framework for consciousness. Nature Neuroscience,
6(2), 119–126.
Damasio, A. R. (1999). The feeling of what happens: Body and emotion in the making
of consciousness. Houghton Mifflin Harcourt.
Damasio, A. (2001). Fundamental feelings. Nature, 413(6858), 781–781.
Damasio, A., & Meyer, K. (2009). Consciousness: An overview of the phenomenon
and of its possible neural basis. In S. Laureys & T. Giulio (Eds.), The Neurology
of Consciousness: Cognitive Neuroscience and Neuropathology (Vol. 1, 10. Okt.
2008) (pp. 3–14). Academic.
Damasio, A. R. (2014). Ich fühle, also bin ich: Die Entschlüsselung des Bewusstseins.
Ullstein eBooks.
Dehaene, S., & Changeux, J. P. (2004). Neural mechanisms for access to conscious-
ness. The Cognitive Neurosciences, 3, 1145–1158.
Dehaene, S., Changeux, J. P., & Naccache, L. (2011). The Global Neuronal
Workspace Model of Conscious Access: From Neuronal Architectures to Clinical
Applications. In S. Dehaene & Y. Christen (Eds.), Characterizing Consciousness:
From Cognition to the Clinic?. Research and Perspectives in Neurosciences Springer.
https://doi.org/10.1007/978-3-642-18015-6_4.
Dehaene, S., Charles, L., King, J. R., & Marti, S. (2014). Toward a computational
theory of conscious processing. Current Opinion in Neurobiology, 25, 76–84.
94    
P. Krauss

Dehaene, S., Kerszberg, M., & Changeux, J. P. (1998). A neuronal model of a

global workspace in effortful cognitive tasks. Proceedings of the National Academy
of Sciences, 95(24), 14529–14534.
Dehaene, S., Lau, H., & Kouider, S. (2017). What is consciousness, and could
machines have it? Science, 358(6362), 486–492.
Dehaene, S., & Naccache, L. (2001). Towards a cognitive neuroscience of con-
sciousness: Basic evidence and a workspace framework. Cognition, 79(1–2),
1–37.
Demertzi, A., Tagliazucchi, E., Dehaene, S., Deco, G., Barttfeld, P., Raimondo, F.,
… & Sitt, J. D. (2019). Human consciousness is supported by dynamic complex
patterns of brain signal coordination. Science Advances, 5(2), eaat7603.
Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company.
Graves, T. L., Maniscalco, B., & Lau, H. (2011). Volition and the function of con-
sciousness. In Conscious Will and Responsibility (pp. 109–121) Graves.
Grossberg, S. (2017). Towards solving the hard problem of consciousness: The vari-
eties of brain resonances and the conscious experiences that they support. Neural
Networks, 87, 38–95.
Hameroff, S. (2001). Consciousness, the brain, and spacetime geometry. Annals of
the New York Academy of Sciences, 929(1), 74–104.
Hameroff, S., Nip, A., Porter, M., & Tuszynski, J. (2002). Conduction pathways in
microtubules, biological quantum computation, and consciousness. Bio Systems,
64(1–3), 149–168.
Hameroff, S., & Penrose, R. (1996a). Orchestrated reduction of quantum coher-
ence in brain microtubules: A model for consciousness. Mathematics and
Computers in Simulation, 40(3–4), 453–480.
Hameroff, S. R., & Penrose, R. (1996b). Conscious events as orchestrated space-
time selections. Journal of Consciousness Studies, 3(1), 36–53.
Koch, C. (2004). The quest for consciousness. Engineering and Science, 67(2),
28–34.
Koch, C. (2013). Bewusstsein: Bekenntnisse eines Hirnforschers. Springer.
Koch, C., & Hepp, K. (2006). Quantum mechanics in the brain. Nature,
440(7084), 611–611.
Koch, C., & Hepp, K. (2007). The relation between quantum mechanics and
higher brain functions: Lessons from quantum computation and neurobiology.
Citeseer. https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.652.712&
rep=rep1&type=pdf.
Koch, C., Massimini, M., Boly, M., & Tononi, G. (2016). Neural correlates of con-
sciousness: Progress and problems. Nature Reviews Neuroscience, 17(5), 307–321.
Kriegeskorte, N., & Douglas, P. K. (2018). Cognitive computational neuroscience.
Nature Neuroscience, 21(9), 1148–1160.
Kübler, A., & Birbaumer, N. (2008). Brain – computer interfaces and communi-
cation in paralysis: Extinction of goal directed thinking in completely paralysed
patients? Clinical Neurophysiology, 119(11), 2658–2666.
 9 Consciousness    
95

Lamme, V. A. (2010). How neuroscience will change our view on consciousness.

Cognitive Neuroscience, 1(3), 204–220.
Laplane, L., Mantovani, P., Adolphs, R., Chang, H., Mantovani, A., McFall-Ngai,
M., … & Pradeu, T. (2019). Why science needs philosophy. Proceedings of the
National Academy of Sciences, 116(10), 3948–3952.
Mashour, G. A., Roelfsema, P., Changeux, J. P., & Dehaene, S. (2020). Conscious
processing and the global neuronal workspace hypothesis. Neuron, 105(5),
776–798.
Melloni, L., Mudrik, L., Pitts, M., & Koch, C. (2021). Making the hard problem
of consciousness easier. Science, 372(6545), 911–912.
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4),
435–450.
Newman, J. B., & Baars, B. J. (1993). A neural attentional model for access to con-
sciousness: A global workspace perspective. In Concepts in Neuroscience, 4(2),
109–121. Graves.
Nielsen, M., & Chuang, I. (2002). Quantum computation and quantum informa-
tion. Cambridge University Press.
Penrose, R. (1989). The emperor’s new mind. Oxford University Press.
Penrose, R. (1994). Mechanisms, microtubules and the mind. Journal of
Consciousness Studies, 1(2), 241–249.
Rees, G., Kreiman, G., & Koch, C. (2002). Neural correlates of consciousness in
humans. Nature Reviews Neuroscience, 3(4), 261–270.
Reggia, J. A., Katz, G., & Huang, D. W. (2016). What are the computational
correlates of consciousness? Biologically Inspired Cognitive Architectures, 17,
101–113.
Sergent, C., & Dehaene, S. (2004). Neural processes underlying conscious percep-
tion: Experimental findings and a global neuronal workspace framework. Journal
of Physiology-Paris, 98(4–6), 374–384.
Seth, A. (2009). Explanatory correlates of consciousness: Theoretical and computa-
tional challenges. Cognitive Computation, 1(1), 50–63.
Seth, A. (2021). Being you: A new science of consciousness. Penguin.
Seth, A. K., & Bayne, T. (2022). Theories of consciousness. Nature Reviews
Neuroscience, 23(7), 439–452.
Seth, A. K., Dienes, Z., Cleeremans, A., Overgaard, M., & Pessoa, L. (2008).
Measuring consciousness: Relating behavioural and neurophysiological
approaches. Trends in Cognitive Sciences, 12(8), 314–321.
Tegmark, M. (2000). Importance of quantum decoherence in brain processes.
Physical Review E, 61(4), 4194.
Tegmark, M. (2014). Consciousness is a state of matter, like a solid or gas. New
Scientist, 222(2964), 28–31.
Tononi, G. (2004). An information integration theory of consciousness. BMC
Neuroscience, 5, 1–22.
96    
P. Krauss

Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information
theory: From consciousness to its physical substrate. Nature Reviews Neuroscience,
17(7), 450–461.
Tononi, G., & Koch, C. (2008). The neural correlates of consciousness: An update.
Annals of the New York Academy of Sciences, 1124(1), 239–261.
 10
Free Will

Life is like a card game. The hand you receive is deterministic;

how you play it is free will.
Jawaharlal Nehru

Is Free will Just a Pious Wish?

The question about the nature of consciousness is closely linked to the ques-
tion of the existence of free will, that is, the ability to choose freely between
different possible courses of action, free from external influences or desires
(O’Connor, 1972; Kane, 2001; Watson, 2003; Harris, 2012; Ekstrom,
2018). The concept of free will is closely linked to notions of moral respon-
sibility, praise, guilt, sin, and other ethical and legal concepts (Roth, 2004,
2016; Roth et al., 2006; Lampe et al., 2008).
In principle, free will also means the ability to make decisions that are
not determined by past events or so-called deterministic causal chains, i.e.,
that the actor can choose between different possible outcomes and that the
result of his decision is not predetermined by previous events (James, 1884;
Van Inwagen, 1975). In classical determinism, however, it is assumed that
all events are determined by past causes and thus only one course is possible
(Earman, 1986).

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 97
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_10
98    
P. Krauss

The Laplace’s Demon, named after the French mathematician Pierre-Simon

Laplace, is a philosophical and physical thought experiment that examines
the concept of determinism (Van Strien, 2014; Kožnjak, 2015). The thought
experiment involves a hypothetical demon who knows the position and speed
of all particles in the universe at a certain point in time. With this knowledge,
the demon would be able to predict the future of the universe with absolute
certainty, as he would know the exact outcome of each event based on the
physical laws. The Laplace’s Demon is often used to illustrate the idea that if
determinism is true, the future of the universe is already predetermined and
all events are necessary and inevitable. This notion has significant implications
for the concept of free will, because if the future is predetermined, there seems
to be no room for real decisions or options for action. Incompatibilists there-
fore argue that if determinism is true, free will is not possible.
However, Laplace’s Demon is also the subject of various criticisms and
objections. One objection is that the demon’s complete knowledge of the
universe could be impossible, as Heisenberg’s uncertainty principle implies
that there are limits to our ability to simultaneously measure the position
and speed of particles with arbitrary accuracy (Robertson, 1929; Busch
et al., 2007). Moreover, it is argued that even if the demon’s knowledge were
possible, it would not necessarily imply complete predictability, as quan-
tum physics gives the universe an inherent randomness and unpredictability
(Dirac, 1925, 1926; Schrödinger, 1926).
And finally, determinism and predictability are not necessarily identical
(Van Kampen, 1991; Loewer, 2001), as anyone who has ever been annoyed
by the wrong weather forecast knows. The dynamics of complex systems are
sensitive to initial conditions, which in principle would have to be known
with infinite accuracy in order to predict the further course exactly (Strogatz,
2018). Otherwise, the slightest deviations over time can lead to exponen-
tially growing differences in behavior (Murphy, 2010). This is the infamous
butterfly effect (Lorenz, 2000). The weather is certainly a deterministic sys-
tem and follows the causal laws of cause and effect, but it is far from being
exactly predictable into the distant future (Koch, 2009).
Some philosophers therefore advocate a compatibilist view of free will,
which states that free will is compatible with determinism (Vihvelin, 2013).
Although the universe is largely deterministic, there is still enough random-
ness and chaos in the world to save free will. Others argue that the apparent
contradiction between free will and determinism is a category mistake. The
opposite of determinism is not free will, but randomness and chaos. And the
opposite of free will is not determinism, but compulsion (Mittelstraß et al.,
2004; Hallett, 2009; 2011) (Fig. 10.1).
 10 Free Will    
99

Fig. 10.1 Free will. The existence of free will is disputed in philosophy and neurosci-
ence. However, there is no doubt about the existence of the village of the same name
in the district of Schleswig-Flensburg in the north of Germany.

The Libet Experiment

The Libet Experiment, named after Benjamin Libet, is a famous and con-
troversial study in neuroscience and psychology, intended to investigate the
nature of free will and the timing of conscious decisions (Libet et al., 1983;
Libet, 1985). The experiment was conducted in the early 1980s to deter-
mine whether the conscious decision to perform a voluntary action occurs
before or after the brain triggers the action.
The basic experimental setup consisted of four components. First, volun-
teers were asked to perform simple voluntary movements, such as bending
the wrist or pressing a button. Meanwhile, their brain activity was moni-
tored using electroencephalography (EEG), focusing on the readiness poten-
tial. This refers to the slow increase in electrical activity in the brain that
precedes voluntary movements and can thus be interpreted as an indication
of the emergence of a conscious intention. Second, participants were shown
a clock-like device that served as a timer. A rotating dot allowed them to
report the exact time of their conscious decision to perform the voluntary
movement. Third, the participants were asked to indicate the position of the
rotating dot on the clock at the time of their conscious decision. And finally,
100    
P. Krauss

EEG amplitude

Readiness potential

Time (ms)

Fig. 10.2 The Libet Experiment. A subject performs a voluntary movement while his
brain activity is monitored using EEG. A clock-like device allows him to record the
time of his conscious decision. He then indicates the position of the rotating pointer
on the clock at the decision time. Finally, brainwave measurements and decision
reports are compared. Results suggest that unconscious processes already initiate
actions (−500 ms) before the conscious decision to do so is made (−200 ms) and the
action is then finally executed (0 ms)

fourth, the results of the brainwave measurement were compared with the
decision reports to determine whether the participants’ conscious decision
was indeed the cause of the voluntary movement (Fig. 10.2).
The most important result of Libet’s experiment was that the readiness
potential in the brain already increased significantly (500−300 milliseconds)
before the time when the participants reported that they had become aware
of their decision to perform the movement. Libet concluded from this that
unconscious processes in the brain initiate voluntary actions before con-
scious decisions are made.

Conclusion
The results of the Libet experiment seem to argue against the existence of
free will. They have therefore been widely discussed and have led to heated
controversies. Critics claim that the methodology of the experiment is
 10 Free Will    
101

flawed and the conclusions about free will are too deterministic. It is also
argued that the decision to perform a certain movement at a certain time
oversimplifies the nature of free will. For example, this simple action is not
comparable to freely deciding for or against a certain course of study, a cer-
tain place of residence, or a life partner.
From the author’s perspective, it would be highly strange if our (free)
decisions were completely independent of our genetic, neural, and autobi-
ographical history. Of course, important life decisions like the ones men-
tioned above are made against the backdrop of our personal experience and
life history, which in turn are stored in the brain as knowledge and memo-
ries (Hallet, 2007; Lau, 2009).
Imagine if Libet’s results had been exactly the opposite: The test subjects
report their decision to perform the movement before the corresponding neu-
ral activity is measurable. This result would have been much more disturbing,
as it would have implied that there are biology-independent mental processes
that are capable of causing changes in the brain. This would have been de
facto evidence of dualistic views on the relationship between brain and mind.
Some even went so far as to question the practice of punishment with
the argument that if the will is not free, the person cannot be held respon-
sible for their actions (Roth, 2010, 2012; Singer, 2020). This of course
overlooks that punishments can still be meaningful as they serve different
purposes and work on different levels (Feinberg, 1965; Van den Haag, 1975;
Jakobs, 2011; Kühl, 2017). First, victims of crimes, even if the existence of
free will is questioned, still have a need for justice, atonement, and repara-
tion. Second, resocialization, regardless of the question of free will, can con-
tribute to rehabilitating the offender and imparting the skills and knowledge
to become a constructive member of society. In this context, resocialization
focuses on changing behaviors and thought patterns that have led to crimes,
not on whether the offender consciously made these decisions. Third, the
punishment of offenders can serve as a deterrent for potential imitators by
clarifying the consequences of crimes. Even if people do not have free will,
they can still respond to incentives and deterrence. The sanction system can
thus help to reduce the frequency of crimes by deterring potential offend-
ers from criminal behavior. And finally, the imprisonment of an offender
can protect society from further crimes by this person. Even if the offender
does not have free will, he could still pose a danger to society if left free.
Thus, punishment and imprisonment of an offender can remain meaningful
and justified even if the existence of free will is questioned (Viney, 1982;
Stroessner & Green, 1990; Hallett, 2007; Mobbs et al., 2009; Hodgson,
2009; Vincent et al., 2011; Focquaert et al., 2013).
102    
P. Krauss

In addition to the impacts on human decision-making, the Libet

experiments also raise interesting questions about artificial intelligence.
Although AI is fundamentally deterministic, these experiments can influ-
ence our understanding of AI decisions, responsibility, and moral action.
Furthermore, AI can serve as a tool to explore and understand the concept
of free will in humans.

References
Busch, P., Heinonen, T., & Lahti, P. (2007). Heisenberg’s uncertainty principle.
Physics Reports, 452(6), 155–176.
Dirac, P. A. M. (1925). The fundamental equations of quantum mechanics.
Proceedings of the Royal Society of London. Series A, Containing Papers of a
Mathematical and Physical Character, 109(752), 642–653.
Dirac, P. A. M. (1926). On the theory of quantum mechanics. Proceedings of the
Royal Society of London. Series A, Containing Papers of a Mathematical and
Physical Character, 112(762), 661–677.
Earman, J. (1986). A primer on determinism (Vol. 37). Springer Science & Business
Media.
Ekstrom, L. (2018). Free will. Routledge.
Feinberg, J. (1965). The expressive function of punishment. The Monist, 49(3),
397–423.
Focquaert, F., Glenn, A. L., & Raine, A. (2013). Free will, responsibility, and the
punishment of criminals. The future of punishment, 247–274.
Hallett, M. (2007). Volitional control of movement: The physiology of free will.
Clinical Neurophysiology, 118(6), 1179–1192.
Hallett, M. (2009). Physiology of volition. Downward causation and the neurobiol-
ogy of free will, 127–143.
Hallett, M. (2011). Volition: How physiology speaks to the issue of responsibility.
Conscious will and responsibility, 61–69.
Harris, S. (2012). Free will. Simon and Schuster.
Hodgson, D. (2009). Criminal responsibility, free will, and neuroscience.
Downward causation and the neurobiology of free will, 227–241.
Jakobs, G. (2011). Strafrecht: Allgemeiner Teil. de Gruyter.
James, W. (1884). The dilemma of determinism (pp. 1878–1899). Kessinger
Publishing.
Kane, R. (Hrsg.). (2001). Free will. Wiley.
Koch, C. (2009). Free will, physics, biology, and the brain. Downward causation
and the neurobiology of free will, 31–52.
Kožnjak, B. (2015). Who let the demon out? Laplace and Boscovich on determin-
ism. Studies in History and Philosophy of Science Part A, 51, 42–52.
 10 Free Will    
103

Kühl, K. (2017). Strafrecht: Allgemeiner Teil. Vahlen.

Lampe, E. J., Pauen, M., Roth, G., & Verlag, S. (Eds.). (2008). Willensfreiheit und
rechtliche Ordnung. Suhrkamp.
Lau, H. C. (2009). Volition and the function of consciousness. Downward causation
and the neurobiology of free will, 153–169.
Libet, B., Wright, E. W., Jr., & Gleason, C. A. (1983). Preparation-or inten-
tion-to-act, in relation to pre-event potentials recorded at the vertex.
Electroencephalography and Clinical Neurophysiology, 56(4), 367–372.
Libet, B. (1985). Unconscious cerebral initiative and the role of conscious will in
voluntary action. Behavioral and brain sciences, 8(4), 529–539.
Loewer, B. (2001). Determinism and chance. Studies in History and Philosophy
of Science Part B: Studies in History and Philosophy of Modern Physics, 32(4),
609–620.
Lorenz, E. (2000). The butterfly effect. World Scientific Series on Nonlinear Science
Series A, 39, 91–94.
Mittelstraß, J., Menzel, R., Singer, W., & Nida-Rümelin, J. Zur Freiheit des
Willens II: Streitgespräch in der Wissenschaftlichen Sitzung der Versammlung
der Berlin-Brandenburgischen Akademie der Wissenschaften am 2. Juli 2004.
Mobbs, D., Lau, H. C., Jones, O. D., & Frith, C. D. (2009). Law, responsibility,
and the brain (pp. 243–260). Springer.
Murphy, R. P. (2010). Chaos theory. Ludwig von Mises Institute.
O’Connor, D. J. (1972). Free will. Springer.
Robertson, H. P. (1929). The uncertainty principle. Physical Review, 34(1), 163.
Roth, G. (2004). Freier Wille, Verantwortlichkeit und Schuld. Zur Freiheit des
Willens, 63–70.
Roth, G., Lück, M., & Strüber, D. (2006). „Freier Wille“ und Schuld von
Gewaltstraftätern aus Sicht der Hirnforschung und Neuropsychologie. Neue
Kriminalpolitik, 18(2), 55–59.
Roth, G. (2010). Free will: Insights from neurobiology. Homo novus–a human with-
out illusions, 231–245.
Roth, G. (2012). Über objektive und subjektive Willensfreiheit. Theory of Mind:
Neurobiologie und Psychologie sozialen Verhaltens, 213–223.
Roth, G. (2016). Schuld und Verantwortung: Die Perspektive der Hirnforschung.
Biologie in unserer Zeit, 46(3), 177–183.
Schrödinger, E. (1926). SCHRÖDINGER 1926C. Annalen der Physik, 79, 734.
Singer, W. (2020). Ein neues Menschenbild?: Gespräche über Hirnforschung.
Suhrkamp.
Strogatz, S. H. (2018). Nonlinear dynamics and chaos with student solutions manual:
With applications to physics, biology, chemistry, and engineering. CRC Press.
Stroessner, S. J., & Green, C. W. (1990). Effects of belief in free will or determin-
ism on attitudes toward punishment and locus of control. The Journal of Social
Psychology, 130(6), 789–799.
Van den Haag, E. (1975). Punishing criminals (Vol. 10). Basic Books.
104    
P. Krauss

Van Inwagen, P. (1975). The incompatibility of free will and determinism.

Philosophical Studies, 27(3), 185–199.
Van Kampen, N. G. (1991). Determinism and predictability. Synthese, 89,
273–281.
Van Strien, M. (2014). On the origins and foundations of Laplacian determinism.
Studies in History and Philosophy of Science Part A, 45, 24–31.
Vihvelin, K. (2013). Causes, laws, and free will: Why determinism doesn’t matter.
Oxford University Press.
Vincent, N. A., Van de Poel, I., & Van Den Hoven, J. (Eds.). (2011). Moral respon-
sibility: Beyond free will and determinism (Vol. 27). Springer Science & Business
Media.
Viney, W., Waldman, D. A., & Barchilon, J. (1982). Attitudes toward punishment
in relation to beliefs in free will and determinism. Human Relations, 35(11),
939–949.
Watson, G. (Eds.). (2003). Free will. Oxford readings in philosophy.
 Part II
Artificial
Intelligence
In the second part of the book, the focus will be on the most important
ideas and concepts of Machine Learning and Artificial Intelligence. Since
mathematical details can be as daunting as neuroanatomical and molecular
biological ones, this part will also avoid an in-depth mathematical presenta-
tion. The focus is on a general understanding and an overview of the big
picture and the most important ideas and concepts, in order to grasp the
essence and potentials of AI and Machine Learning. This approach helps us
to build a bridge between the two disciplines of AI and brain research and to
work out their similarities and differences without getting lost in technical
details.
 11
What is Artificial Intelligence?

Despite all the hype and excitement about AI, it is still very limited
compared to human intelligence.
Andrew Ng

History of AI
The idea that human intelligence or the cognitive processes that humans use
could be mechanized or automated is very old. One of the first mentions
of this idea can be found in Julien Offray de La Mettrie’s work L’Homme
Machine, published in 1748. Another theoretical precursor of AI is the
Laplace’s Demon, which we have already encountered in the chapter on free
will. It is named after the French mathematician, physicist, and astronomer
Pierre-Simon Laplace. The concept is based on the idea that the entire uni-
verse functions like a mechanical machine, similar to a clock, and that the
human mind and intelligence function in the same way.
Artificial Intelligence (AI) is based on the belief that human thinking can
be structured and systematized. The roots of this concept go back to the 1st
millennium BC, when Chinese, Indian, and ancient Greek philosophers
developed structured techniques for formal reasoning. Later, in the seven-
teenth century, philosophers like René Descartes and Gottfried Wilhelm
Leibniz tried to formalize rational thinking and make it as precise as algebra
or geometry. They considered thinking as equivalent to the manipulation

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 107
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_11
108    
P. Krauss

of symbols. This model later served as the basis for research in the field of
Artificial Intelligence.
The study of mathematical logic played a crucial role in the development
of AI in the twentieth century. One of the most important contributions
came from George Boole and Gottlob Frege, who laid the foundations for
the formal manipulation of symbols. These works established a set of rules
and principles for the manipulation of symbols in a logical system, which
proved essential for the development of intelligent machines.
Another important contribution was the Church-Turing thesis, a funda-
mental concept in computer science. It states that any mathematical prob-
lem that a human can solve with paper and pencil can also be solved by a
machine, specifically a mechanical machine that can process zeros and ones.
This concept laid the foundation for the development of digital computers
and the theory of computing.
The Turing Machine, a theoretical computing model proposed by
Alan Turing, was a crucial breakthrough in the development of Artificial
Intelligence. It is a simple and abstract model that captures the essential
features of any mechanical machine capable of performing abstract sym-
bol manipulations. The Turing Machine allowed researchers to formalize
the concept of computing and develop algorithms that can be executed by
machines.
The Dartmouth Conference, which took place in the summer of 1956
at Dartmouth College in Hanover, New Hampshire, is considered the
founding event of the academic field of Artificial Intelligence (AI). The
conference was a six-week workshop titled “Dartmouth Summer Research
Project on Artificial Intelligence,” organized by John McCarthy as part of a
research project funded by the Rockefeller Foundation. The term “Artificial
Intelligence” was used for the first time in the announcement of this con-
ference. Some of the brightest minds in computer science and related fields
attended the conference, including McCarthy himself, Marvin Minsky,
Nathaniel Rochester, and Claude Elwood Shannon, the founder of informa-
tion theory. During the conference, the participants discussed the possibility
of creating machines that can think and argue like humans. They exam-
ined various AI approaches such as logic-based systems, neural networks,
and heuristic search algorithms. They also discussed the ethical implications
of developing intelligent machines and the question of whether it is even
possible to develop truly intelligent machines. The Dartmouth Conference
established AI as an academic discipline and provided a roadmap for future
research in this field.
 11 What is Artificial Intelligence?    
109

There is a close connection between the research direction of Artificial

Life and AI. The long-term goal of AI is to develop an intelligent agent capa-
ble of understanding or learning any intellectual task that can be accom-
plished by a human or other living being. This is also referred to as strong AI
or Artificial General Intelligence (Artificial General Intelligence, AGI).

Concept Clarification
Artificial Intelligence generally refers to machines capable of intelligent
behavior (Russell, 2010). Unlike the natural intelligence of animals and
humans, which is based on biological processes, AI is based on algorithms
and software developed by humans. AI systems are programmed to perform
tasks that normally require human intelligence, such as pattern recognition,
learning from experience, decision-making, and problem-solving. An impor-
tant aspect of AI systems is their ability to be autonomous and adaptable.
They can learn from experiences and feedback and improve over time. This
process is also referred to as Machine Learning and includes various tech-
niques such as supervised learning, unsupervised learning and reinforcement
learning. There are many different types of AI systems, each specialized in
their own way. Natural language processing, for example, allows machines to
understand and respond to human language (Görz & Schneeberger, 2010;
Russell, 2010).
Expert systems use knowledge in the form of symbols and rules to solve
problems and make decisions. Autonomous robots are another example of
AI systems specialized in physical actions. They can perform tasks such as
navigation, object manipulation, and interaction with the environment.
Multi-agent systems, on the other hand, are a type of swarm of autonomous
agents that work together to achieve a common goal (Görz & Schneeberger,
2010; Russell, 2010).
When artificial intelligence is discussed in the media and politics today, it
usually refers to only a relatively small part of these approaches. All the spec-
tacular successes and breakthroughs of the last one or two decades—whether
it’s composing music, playing Go, generating text, or classifying images—are
based on machine learning and pattern recognition, a subfield of AI, and
here in particular on deep learning, which is only a part of machine learn-
ing. In public discourse, the terms artificial intelligence, machine learning,
and deep learning are often used synonymously today (Fig. 11.1). Often in
the context of AI, “the algorithms” are mentioned. An algorithm is similar to
110    
P. Krauss

Artificial intelligence

Machine learning /
pattern recognition

Deep
learning

Fig. 11.1 Subareas of Artificial Intelligence. The terms Artificial Intelligence,

Machine Learning, and Deep Learning are often seen as interchangeable today and
are used synonymously in public discourse, although they actually stand in a hierar-
chical relationship to each other

a cooking recipe, a step-by-step guide to solving a problem or performing a

specific task. It consists of an ordered sequence of instructions formulated in
such a way that they can be executed by a machine, a computer, or a human.

Machine Learning and Pattern Recognition

Machine Learning deals with the development of algorithms and statisti-
cal models that enable computers to learn from data and make predictions
or decisions without being explicitly programmed (Bishop & Nasrabadi,
2006). The focus is on analyzing large amounts of data to recognize patterns
and draw conclusions. Machine learning is an important part of modern
data analysis and is used in many areas, including image and speech recog-
nition, autonomous driving, and personalized advertising. An important
concept of machine learning is pattern recognition. Patterns in data are rec-
ognized to make predictions or decisions. To do this, algorithms and models
must be developed that can learn from data and recognize patterns.
 11 What is Artificial Intelligence?    
111

The human brain also uses pattern recognition to process and interpret
sensory impressions. It has specialized regions responsible for certain types
of pattern recognition. The brain’s ability to learn patterns and adapt to new
patterns is crucial for intelligent, goal-directed behavior and our ability to
interact with the world.

Deep Learning
Deep Learning is a branch of Machine Learning that focuses on artificial
neural networks with multiple layers of interconnected neurons. The more
layers a neural network has, the “deeper” it is. Modern architectures can
have hundreds of layers and are capable of processing and learning from
large amounts of data by recognizing complex patterns and establishing
abstract connections. One of the greatest achievements of Deep Learning
is the ability to train neural networks to process large amounts of unstruc-
tured data, such as images or voice recordings, and recognize complex cor-
relations and patterns. This has led to breakthroughs in many application
areas, including language processing, image recognition and segmentation,
and has enabled advances in medical research, natural sciences, and other
fields (Schmidhuber, 2015; LeCun et al., 2015; Goodfellow et al., 2016).
The main advantage of Deep Learning is the ability to recognize and learn
meaningful patterns in complex and unstructured data without the need for
human expertise. Deep Learning models can even work with data that is dif-
ficult or impossible for the human mind to interpret, such as digital X-rays.
This makes Deep Learning an enormously powerful tool in many fields.
Despite the incredible advances in recent years, there are still some chal-
lenges in the development and use of Deep Learning models. One of the
biggest problems is the difficulty of interpreting and explaining the work-
ings of deep neural networks. It is often hard to understand why a particular
model made a certain decision or how it recognized certain patterns.

Conclusion
The successes of AI have been remarkable in recent years and cover a wide
range of applications. One of the most impressive advances is the ability of
AI systems to understand human language and process natural language.
This has led to significant advances in speech recognition and translation.
AI has also made progress in image processing by being able to recognize
112    
P. Krauss

and analyze images. This has contributed to the development of face recog-
nition, object detection, and medical imaging. Many practical applications
of Artificial Intelligence (AI) are often integrated so quickly into everyday
products that they are no longer perceived as AI. An example of this is text
recognition, which is now standard in many smartphones and is taken for
granted.
Because of this phenomenon known as the AI effect, it may seem that AI
research is only struggling with difficult problems that it has not yet solved,
such as the ability to make complex decisions in dynamic environments, or
the problem that it is difficult to interpret and explain the workings of deep
neural networks.
A famous quote by computer scientist Larry Tesler, “Intelligence is what
machines have not yet done,” expresses this.

References
Bishop, C. M., & Nasrabadi, N. M. (2006). Pattern recognition and machine learn-
ing (Vol. 4, No. 4, p. 738). Springer.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.
Görz, G., & Schneeberger, J. (Eds.). (2010). Handbuch der künstlichen Intelligenz.
De Gruyter.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. nature, 521(7553),
436–444.
Russell, S. J. (2010). Artificial intelligence a modern approach. Pearson Education,
Inc.
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural
networks, 61, 85–117.
 12
How Does Artificial Intelligence Learn?

After my training as a computer scientist in the 90s, everyone knew that AI did not
work. People tried. They tried with neural networks, but nothing worked.
Sergey Brin

Artificial Neuron
As in the brain, the neuron is also the fundamental processing unit in many
areas of AI. Artificial neurons are simplified mathematical models of their
biological counterparts. They receive their input in the form of real num-
bers through several input channels. As in the brain, this input can either
come “from outside” or from other neurons. Each input channel has a
weight, which corresponds to the synapse in biology, with which the respec-
tive input is multiplied. Subsequently, all weighted inputs of the neuron are
summed up. Since all inputs of a neuron can be summarized into an input
vector and analogously all corresponding weights can be summarized into
a weight vector, the weighted sum of the inputs corresponds to the scalar
product of the input and weight vector. This then goes into an activation
function to determine the output of the neuron. In the McCulloch-Pitts
neuron (McCulloch & Pitts, 1943), which represents the earliest and sim-
plest artificial neuron, the activation function corresponds to a threshold
function, i.e., it is compared whether the weighted sum of the inputs is
greater or smaller than a certain threshold. If it is greater, the neuron sends
an output (a one), otherwise not (which corresponds to a zero). Thus, the

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 113
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_12
114    
P. Krauss

Sum Activation function

Output
Artificial
neuron

Input vector Weight vector

Biological neuron

Fig. 12.1 Artificial Neuron. The basic processing unit in Artificial Intelligence.
Analogous to its biological model, it takes input x through several channels and
multiplies it with associated weights w, similar to the synapses in the brain. These
weighted inputs are summed up and processed through an activation function to
determine the output y of the neuron. In simple models like the McCulloch-Pitts neu-
ron, the activation function is a threshold function, leading to binary outputs (0 or 1).
More complex models use advanced activation functions for continuous outputs

perceptron is a binary classifier. However, more complicated activation func-

tions are usually used, which lead to continuous outputs (Schmidhuber,
2015; LeCun et al., 2015; Goodfellow et al., 2016) (Fig. 12.1).

Artificial Neural Networks

From these artificial neurons, artificial neural networks can now be con-
structed by connecting several of them together, with a weight for each
connection. Each neuron receives input signals from other neurons or from
external sources, then processes this input and calculates an output sig-
nal, which is then transmitted further to other neurons or as output to the
outside.
Depending on the architecture of the network, the neurons can be
arranged in different so-called layers, with several neurons in one layer. As
a rule, all neurons of a layer receive their input from the same source and
send their output to the same target. In a neural network, one distinguishes
between input layer, output layer and (usually several) intermediate layers or
also hidden layers.
The weights of a neural network are summarized in the weight matrix.
The columns of the weight matrix correspond to the weight vectors of the
 12 How Does Artificial Intelligence Learn?    
115

neurons. This matrix can either contain all pairwise weights of all neurons
of the network including all self-connections (Autapses). In this case, one
speaks of a complete weight matrix. However, a weight matrix can also only
contain the forward-directed weights between the neurons of two successive
layers of a network.
In the simplest case, the network consists only of an input and an output
layer. These networks, also known as Perceptron (Rosenblatt, 1958), calcu-
late a weighted sum of the inputs and then apply an activation function to
generate the output. Due to this simple architecture, these neural networks
can only learn linear classifications, i.e., the data in the input space must be
linearly separable (Figs. 12.2 and 12.3).
Due to their linear character, two-layer networks are only limitedly able
to solve classification tasks. A famous example that illustrates the limits of
two-layer neural networks is the XOR problem. The XOR function (eXclu-
sive OR) is a binary operation that outputs 1 when the two input values are
different and 0 when they are the same. It turns out that this problem is not
linearly solvable and thus cannot be learned by two-layer networks.
To overcome this limitation and solve problems with more complex deci-
sion boundaries, multi-layer neural networks with one or more hidden layers
can be used. These hidden layers allow the network to learn nonlinear trans-
formations of the input data, so they can capture more complex relation-
ships and decision boundaries, making them ultimately more versatile and
powerful for various tasks.
In fact, the so-called universal approximation theorem states that in prin-
ciple, a neural network with a single hidden layer and a finite number of

Input layer Output layer

Weight vectors Weight matrix

Fig. 12.2 Two-layer network. The perceptron consists only of an input and an out-
put layer. The weight vectors are summarized as a weight matrix
116    
P. Krauss

Fig. 12.3 The XOR problem. A two-layer neural network cannot solve the XOR prob-
lem. There is no suitable set of weights and threshold, as the two solutions (0, 1)
are not linearly separable, i.e., the decision boundary between the classes is not a
straight line. However, the logical AND can be solved with a simple two-layer net-
work, as the solutions here are linearly separable

neurons can approximate any continuous function with arbitrary accuracy

(Cybenko, 1989). However, usually neither the number of neurons nor the
specific architecture of the network is clear. The theorem only guarantees the
existence of such a network. Moreover, the theorem does not state that a
network with a single hidden layer is always the best choice for a particular
approximation problem or that it converges quickly during training.
In practice, therefore, neural networks with multiple hidden layers are
often used to approximate complex functions, as they often achieve higher
accuracy with fewer neurons than a network with a single hidden layer.
Deep learning refers to the use of deep neural networks, which consist of a
large number of hidden or intermediate layers. The more layers such a net-
work has, the deeper it is (Fig. 12.4).
Four fundamentally different methods of learning in artificial neural net-
works can be distinguished: supervised, unsupervised, and self-supervised
learning as well as reinforcement learning.

Supervised Learning
For supervised learning, labeled data, i.e., label-data pairs, are required. An
application example is image classification. Here, there must be a label for
each image with information about what is seen in the image or to which
 12 How Does Artificial Intelligence Learn?    
117

Input layer Hidden layers Output layer

Fig. 12.4 Multi-layer neural network. Artificial neural networks are complex struc-
tures built from artificial neurons and interconnected. In such a structure, several
neurons can be present in one layer, connected to other neurons via weights. Hidden
layers or intermediate layers are located between the input and the output layer.
Deep learning refers to the use of deep neural networks, which consist of a large
number of hidden layers. The more layers a neural network has, the deeper it is. All
these networks are feedforward architectures, i.e., forward-directed networks, as the
input flows in one direction without feedback loops, namely from the input to the
output layer

category the image belongs. In an image dataset, “cat”, “apple” or “car” can
be possible labels. The labels correspond to the desired output of the model
in supervised learning.
For the training of the neural network, the images are first broken down
into individual pixels, with the color or grayscale values coded as numerical
values serving as input channels for the first layer of the neural network. In a
fully connected network, each input channel is connected to each neuron of
the next layer. Activation is achieved by applying the weights to the inputs.
This process can be repeated as often as desired in the various layers. At the
end of the neural network is the output layer, which makes a decision, e.g.,
the classification of the image. If the network is already trained and has
learned to classify a certain image correctly, the task is completed. However,
if the network is not yet sufficiently trained, an error may occur. For exam-
ple, in the output layer, the “apple” neuron could be most strongly activated,
even though the input layer received an image of a “banana” as input.
During training, the desired output is compared with the actual output
and an error is calculated using the loss or cost function. Since it is similar
to a teacher monitoring the performance of the network and returning cor-
responding error messages, this learning paradigm is referred to as supervised
learning. The desired output for a given input, such as the appropriate label
(apple, banana) for an image, is usually provided by humans. You are prob-
ably familiar with these CAPTCHAs (completely automated public Turing
test to tell computers and humans apart) on certain websites, where you are
asked to confirm that you are not a robot. For example, you are asked to
118    
P. Krauss

click on all images showing a traffic light. In this way, large amounts of labe-
led data for supervised learning are generated.
The goal of supervised learning is to minimize the cost function, i.e., to
reduce the sum of all errors. To do this, the individual errors are propagated
backwards from the output layer to the input layer through all layers. This
process is therefore referred to as Error-Backpropagation and allows for a
readjustment of the synaptic weights. This means that the errors are used to
change the weights between two layers in such a way that the correct output
is produced with a slightly higher probability the next time. This is done
by calculating so-called gradients, which indicate the directions in which
the weights need to be changed. This basic technique of machine learning is
therefore also referred to as Gradient Descent (Schmidhuber, 2015; LeCun
et al., 2015; Goodfellow et al., 2016).
In a figurative sense, the error landscape corresponds to a mountain range
with many peaks and valleys. The higher you are, the greater the total error
of the network. The gradients indicate at each point the direction in which
it goes down steepest. By taking a small step in the direction of the steepest
descent at each time step, you eventually end up in a valley, which corre-
sponds to a lower total error of the network.
At the end of the training, the test accuracy of the neural network is usu-
ally determined. This is done by calculating the ratio of correctly predicted
or classified objects to the total number of objects in the dataset. For exam-
ple, if a model trained to classify images correctly classifies 90 out of 100
images, then the test accuracy of the model is 90%.
Usually, the entire available dataset is randomly divided into a training
and a test dataset, usually in a ratio of 80 to 20, a practice referred to as
dataset splitting. The idea behind this is that you want to test how well the
neural network can generalize, i.e., how well it can handle previously unseen
data. The network is therefore trained exclusively with the training dataset.
At the end of the training, the test dataset is used to determine the test accu-
racy or accuracy of the neural network.

Unsupervised Learning
In contrast to supervised learning, no labeled data is required for unsuper-
vised or self-supervised learning.
Unsupervised learning is about extracting patterns and structures from
data without the need for labeled data. This type of learning typically
 12 How Does Artificial Intelligence Learn?    
119

includes tasks such as clustering, where similar data points are grouped, and
dimensionality reduction, where high-dimensional data is represented in a
low-dimensional space while preserving important information. Examples of
unsupervised learning algorithms include K-Means clustering, hierarchical
clustering, and Principal Component Analysis (PCA). Since they belong to
the field of machine learning but do not represent artificial neural networks
and thus are not the focus of this book, we do not want to delve further into
the presentation of these methods (MacKay, 2003; Bishop & Nasrabadi,
2006).

Self-supervised Learning
Self-supervised Learning is on the one hand a special case of unsupervised
learning and on the other hand a category of learning methods in its own
right (Liu et al., 2021). In this case, the algorithm generates its own super-
vision signal (label) from the input data. For example, the model is trained
to predict or reconstruct certain aspects of the data, e.g., the next image in
a video or the context of a specific word in a sentence. In self-supervised
learning, the data itself provides the labels or the supervision signal, with
which the parameters of the model are updated during training. In this way,
it is possible to learn from large amounts of unlabeled data, which are often
much more extensive than labeled data. By learning to extract useful rep-
resentations from the data, self-supervised learning can be used to improve
the performance of a wide range of downstream tasks, including classifica-
tion, object detection, and language understanding.
Often, the neural network is trained on an auxiliary task using the
self-generated labels, which should help it, for example, in learning useful
representations. After learning the auxiliary task, the learned representations
can be used for downstream tasks.
An example of self-supervised learning is the task of predicting the next
word in a sentence (or a masked word) in the context of natural language
processing. In this case, the model generates its own supervision signal by
using the surrounding words as context and thus learns useful language rep-
resentations. In this way, so-called word vectors are generated, which form
the basis of ChatGPT and Co. More on this in the chapter on language-tal-
ented AI.
Another example of self-supervised learning are autoencoders, which are
also referred to as encoder-decoder networks. This type of neural networks
consists of two parts. In the encoder, the layers from the input layer to the
120    
P. Krauss

so-called bottleneck (Bottleneck Layer) become narrower with each further

intermediate layer, thus containing fewer neurons. In the decoder, starting
from the bottleneck, the layers successively become wider again up to the
output layer, which has the same width (number of neurons) as the input
layer (Fig. 12.5).
The idea behind this architecture is that the input is increasingly com-
pressed by the encoder, while the decoder should be able to reconstruct the
original input signal as accurately as possible from this compression. The
compression must therefore be as lossless as possible. In the bottleneck layer,
an abstract, reduced to the essentials representation of the input is created,
which is also called embedding.
Autoencoders can be used to reduce noise in data or to complete incom-
plete data. The embeddings can also be read directly from the bottleneck
layer and used for visualization or as input for further processing steps. A
variant are autoencoders that are not trained end-to-end, i.e., completely
over all layers, but instead layer by layer, with each layer having the task
of compressing the input as well as possible. Once the training of a layer
is completed, its weights are frozen, i.e., no longer changed, and its output
serves as input for the next layer to be trained. This is an elegant way to solve
the problem of vanishing gradients (Bengio et al., 1994), as only a shallow
network is trained at a time. In this way, increasingly abstract feature and
object representations are created with each additional layer.

Fig. 12.5 Autoencoder. Schematic representation of an autoencoder, consisting of

an encoder and a decoder part. The encoder compresses the input data into a nar-
rower bottleneck, while the decoder reconstructs the data from this compressed rep-
resentation. Compression and decompression should be as lossless as possible. The
representation generated in the bottleneck, the so-called embedding, then repre-
sents a reduced to the essentials representation of the original input data
 12 How Does Artificial Intelligence Learn?    
121

Neurons for Cats and Jennifer Aniston

In 2012, a Google team made an interesting discovery when it trained a

deep autoencoder with a large dataset of images from YouTube videos.
They found that one of the neurons in the network had learned to respond
strongly to images of cats, even though the model had not been explic-
itly trained to recognize cats. Instead, it had learned this independently by
finding patterns in the data (Le et al., 2012). The researchers subsequently
named this neuron the “cat neuron”.1 In the human brain, too, there are
neurons that react very selectively to certain concepts or objects. The
so-called “Jennifer Aniston” neuron in the brain of an epilepsy patient, who
had electrodes implanted in his brain before his operation to remove the
epilepsy focus, gained fame in this context. This neuron reacted selectively
to images of actress Jennifer Aniston, but not to images of other people
(Quiroga et al., 2005).

Reinforcement Learning
In contrast to all previous learning methods, reinforcement learning is a
type of machine learning where a model or agent is trained to learn useful
input-output functions, i.e., to make a series of decisions in an uncertain
environment that maximize a cumulative reward (Dayan & Niv, 2008; Li,
2017; Sutton & Barto, 2018; Botvinick et al., 2019; Silver et al., 2021).
In contrast to supervised learning, no outputs or labels are provided to the
model. Instead, the agent receives feedback in the form of rewards or penal-
ties after each action. The goal of the algorithm is to learn a strategy (Policy)
that maps states of the environment to actions in such a way that it leads to
a maximum reward in the long term.
The agent uses trial and error to learn from its experiences in the environ-
ment and randomly tries out different actions to find out which actions lead
to the highest rewards in certain situations, by increasingly using actions
that have already proven to be successful. Over time, the agent’s strategy is
continuously refined and optimized, so that it can make better decisions and
achieve higher rewards.

1 https://www.nytimes.com/2012/06/26/technology/in-A-big-network-of-computers-evidence-of-ma-

chine-learning.html
122    
P. Krauss

There are two types of reinforcement learning. In model-based rein-

forcement learning, a model of the environment is also learned, which can
predict the feedback of the environment (rewards or penalties) to certain
actions. In model-free reinforcement learning, on the other hand, the agent
limits itself to learning the best action for a given state (Kurdi et al., 2019).
Reinforcement learning is used in a variety of applications, including
games, robotics, autonomous driving, and recommendation systems.
The origins of reinforcement learning lie in psychology, particularly in the
study of animal behavior and learning. The concept of reinforcement was
first introduced in the 1930s and 1940s by B.F. Skinner, who developed
the theory of operant conditioning (Skinner, 1963). Skinner’s theory states
that behavior is influenced by subsequent consequences such as reward or
punishment.

Conclusion
Supervised learning requires labeled data to train models that can correctly
classify input data such as images. In contrast, both unsupervised and
self-supervised learning can process data without labels. In unsupervised
learning, the focus is on discovering patterns and structures in the data,
while self-supervised learning essentially generates its own labels from the
data and thus learns useful representations by solving auxiliary tasks.
In the case of the autoencoder, each input pattern is simultaneously its
own label, as it should be restored as accurately as possible from the com-
pression of the bottleneck layer. The input and desired output are therefore
identical. If the image of an apple is input, the desired output is this image
of the apple. In predicting, for example, the next element to a certain input
sequence (video or text), the next image or word is essentially the label of
this sequence. Training is then again, as in the supervised case, with error
backpropagation and gradient descent.
In reinforcement learning, training is focused on making decisions in an
uncertain environment that maximize a cumulative reward, without using
predefined outputs or labels. Instead, the agent uses trial and error to learn
which actions yield the highest rewards, and refines its strategy over time.
The concept originally comes from psychology and the study of animal
behavior.
While we have made considerable progress in artificial intelligence and
machine learning, it is important to emphasize that we are still at the begin-
ning. Neuroscience has the potential to inspire new learning paradigms by
 12 How Does Artificial Intelligence Learn?    
123

giving us deep insights into the workings of the human brain. For exam-
ple, insights from studies on brain plasticity could lead to new approaches
for learning and adapting neural networks. Similarly, investigations into
the mechanisms the brain uses to process complex sensory data could lead
to new architectures for deep neural networks that are more efficient and
robust. Furthermore, understanding the neural processes involved in deci-
sion-making and problem-solving could provide new approaches for rein-
forcement learning and optimization.
Finally, deeper investigations into the coordination and communication
between different brain regions could contribute to the development of
improved methods for training and coordinating larger ensembles of sev-
eral different neural networks develop. Autoencoder with an encoder and a
decoder part or Generative Adversarial Networks, which we will get to know
more closely in the chapter on generative AI, are examples of ensemble net-
works made up of two subnetworks. In principle, however, ensembles of
many subnetworks specialized in various tasks are also conceivable.

References
Bengio, Y., Simard, P., & Frasconi, P. (1994). Learning long-term dependencies
with gradient descent is difficult. IEEE transactions on neural networks, 5(2),
157–166.
Bishop, C. M., & Nasrabadi, N. M. (2006). Pattern recognition and machine learn-
ing (Vol. 4, No. 4, p. 738). Springer.
Botvinick, M., Ritter, S., Wang, J. X., Kurth-Nelson, Z., Blundell, C., & Hassabis,
D. (2019). Reinforcement learning, fast and slow. Trends in Cognitive Sciences,
23(5), 408–422.
Cybenko, G. (1989). Approximation by superpositions of a sigmoidal function.
Mathematics of Control, Signals and Systems, 2(4), 303–314.
Dayan, P., & Niv, Y. (2008). Reinforcement learning: The good, the bad and the
ugly. Current Opinion in Neurobiology, 18(2), 185–196.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.
Kurdi, B., Gershman, S. J., & Banaji, M. R. (2019). Model-free and mod-
el-based learning processes in the updating of explicit and implicit evaluations.
Proceedings of the National Academy of Sciences, 116(13), 6035–6044.
Li, Y. (2017). Deep reinforcement learning: An overview. arXiv preprint
arXiv:1701.07274.
Liu, X., Zhang, F., Hou, Z., Mian, L., Wang, Z., Zhang, J., & Tang, J. (2021). Self-
supervised learning: Generative or contrastive. IEEE Transactions on Knowledge
and Data Engineering, 35(1), 857–876.
124    
P. Krauss

Le, Q. V., Monga, R., Devin, M., Corrado, G., Chen, K., Ranzato, M. A., … &
Ng, A. Y. (2012). Building high-level features using large scale unsupervised learn-
ing. arXiv preprint. arXiv:1112.6209.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553),
436–444.
MacKay, D. J. (2003). Information theory, inference and learning algorithms.
Cambridge University Press.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in
nervous activity. The Bulletin of Mathematical Biophysics, 5, 115–133.
Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C., & Fried, I. (2005). Invariant
visual representation by single neurons in the human brain. Nature, 435(7045),
1102–1107.
Rosenblatt, F. (1958). The perceptron: A probabilistic model for information stor-
age and organization in the brain. Psychological Review, 65(6), 386.
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural
Networks, 61, 85–117.
Silver, D., Singh, S., Precup, D., & Sutton, R. S. (2021). Reward is enough.
Artificial Intelligence, 299, 103535.
Skinner, B. F. (1963). Operant behavior. American Psychologist, 18(8), 503.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT
press.
 13
Game-playing Artificial Intelligence

If there are intelligent beings on other planets, then they play Go.
Emanuel Lasker

Video Games
The development of reinforcement learning was a major breakthrough in
artificial intelligence. In a groundbreaking work, the authors proposed com-
bining reinforcement learning with so-called convolutional networks for
image recognition, thereby learning directly from pixel values as input the
actions to control a video game (Mnih et al., 2015). Specifically, the neural
network received the last five images of the ongoing game and the score as
input. The task was to generate the control commands for the controller or
joystick, such as “left”, “right”, “up”, “down” or “fire”. The goal of the neural
network was, as usual in reinforcement learning, to maximize the reward,
i.e., the score.
The researchers trained their system on various Atari games and showed
that a single algorithm without task-specific knowledge can achieve human-
level performance. The AI system surpassed all previous methods of rein-
forcement learning and in some cases even the performance of human
experts.
However, it turned out that the system was particularly good at action
games that do not require a forward-looking strategy and where success

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 125
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_13
126    
P. Krauss

does not depend on the previous course of the game, such as Video Pinball,
Boxing or Breakout. All these games are pure reaction games, which can be
won by purely tactical action (pressing the right button at the right time). In
this type of game, the neural network reached human level or above. In con-
trast, the system completely failed at comparatively simple strategy games
like Montezuma’s Revenge or Ms. Pac-Man, where you have to navigate
through a manageable world and solve certain tasks.
Nevertheless, the impact of the work on the development of AI was con-
siderable. It marked a milestone in the research on reinforcement learning
and showed that deep learning techniques can be combined with reinforce-
ment learning algorithms to achieve unprecedented performance in a variety
of tasks.

Go and Chess
The Asian board game Go, whose origins date back to ancient China, is
considered the most complex strategy game ever. Its basic rules are relatively
simple—much simpler than in chess—and can be learned within a few min-
utes. The complexity of the game arises from the sheer number of possible
positions, which is estimated at 10170 and thus exceeds the number of possi-
ble positions in chess (approx. 1040) by many orders of magnitude.
In recent decades, chess computers have achieved impressive perfor-
mances and are now able to beat even the best human players. In contrast,
Go computers until recently still had difficulties winning against even medi-
ocre human players. This is because the brute-force method used in chess in
Go is not applicable. In brute force, all possible moves for a certain number
of moves in the future are simulated and the resulting positions are evalu-
ated. The chess computer then selects the best move. Since there is a rela-
tively limited number of moves and possible positions in chess, a computer
with sufficient computing power can successfully apply this technique.
In contrast, brute force is ruled out in Go due to the astronomically high
number of possibilities. There are more possible positions on the Go board
than atoms in the universe, which means that even the most powerful com-
puters currently would not be able to calculate all possible moves. Instead,
Go computers must rely on a combination of heuristic techniques and
machine learning to determine their moves.
Another difference between chess and Go is that chess can produce rel-
atively simple positions after just a few moves, while Go can still be very
complex even after many moves. After only two or four moves in chess,
 13 Game-playing Artificial Intelligence    
127

Fig. 13.1 The game Go. The board game, whose origins date back to ancient China,
is considered the most complex strategy game ever, although the basic rules are rel-
atively simple, for example compared to chess. This is mainly due to the sheer num-
ber of possible positions, which exceeds that of chess by many orders of magnitude.
Developing an AI that can master the game at an advanced human level was long
considered unattainable

there are usually only a few possible positions, while in Go there are already
thousands of possible positions after two or four moves. This makes Go an
even greater challenge for computers, as they not only have to calculate the
best moves, but also have to analyze and evaluate the complexity of the posi-
tions (Fig. 13.1).

AlphaGo Achieves the Breakthrough

While the supercomputer Deep Blue managed to defeat the then world
chess champion Garry Kasparov for the first time in 1996, it was long con-
sidered unattainable to develop an AI that could master the game of Go at
an advanced human level. However, this changed in 2016 with the AI sys-
tem AlphaGo from the company DeepMind (Silver et al., 2016).
Initially, AlphaGo, a neural network, was trained on millions of histori-
cal Go games played by masters and grandmasters in tournaments, and on
games against other computer programs. The network was supposed to pre-
dict the probabilities for different moves in a given situation. In addition,
the network was used to analyze the game and evaluate possible moves and
128    
P. Krauss

positions for both players. In a sensational match in March 2016, AlphaGo

defeated the then best Go player in the world, Lee Sedol, in four out of
five games, thus making history. For the first time ever, an AI had beaten a
human at grandmaster level in Go. This spectacular success marked a mile-
stone in the development of AI.
Just one year later, DeepMind presented AlphaGo Zero (Silver et al.,
2017), a new approach to developing AI in the game of Go. Instead of train-
ing the network on historical games, DeepMind had AlphaGo Zero play bil-
lions of games against itself using reinforcement learning. At the beginning,
the network was only given the rules, i.e., the allowed moves, and the goal
of the network was to maximize the reward, i.e., the number of games won.
At the beginning, AlphaGo Zero simply tried random moves, but it quickly
improved.
After just four days of training, the new AlphaGo Zero was as good as
AlphaGo and reached a level of play after another 40 days of training that
was so superior that it could defeat its predecessor version AlphaGo—which
had already beaten the human world champion—in 100 games just as often.
The worldwide Go elite set about analyzing AlphaGo Zero’s style of play.
Some grandmasters said that watching this AI play was like watching a
hyper-intelligent alien play. AlphaGo Zero developed strategies, tactics, and
moves that were completely unknown in the more than 4000-year history of
the game and were significantly superior to the known ones in some cases.
This AI has forever changed the game of Go. In the meantime, the analysis
of human players’ behavior against self-learning Go programs even resulted
in a significant improvement in the abilities of the playing humans (Choi
et al., 2022).
Now there is AlphaZero (Zhang & Yu, 2020), a further generalization of
AlphaGo Zero. This system can teach itself any strategy game, such as chess,
shogi (Japanese chess), checkers, and of course Go. The AlphaZero variant
AlphaStar is even capable of playing the massive parallel online player strat-
egy game StarCraft II at a human level against a variety of professional gam-
ers (Jaderberg et al., 2019).

Conclusion
The success of AlphaGo and AlphaGo Zero demonstrates how far AI tech-
nology has come and what possibilities are opening up for the future. The
complex game of Go serves as a vivid example of the impressive capabilities
of artificial intelligence and the progress that has been made in this area.
 13 Game-playing Artificial Intelligence    
129

However, cognitive scientist Gary Marcus points out that a large part of
human knowledge has flowed into the development of AlphaZero. And he
suggests that human intelligence seems to include some innate abilities, such
as the intuitive ability to develop language. He advocates considering these
innate abilities in the development of future AI systems, i.e., using a priori
knowledge, instead of always starting training from scratch (Marcus, 2018).
Josh Tenenbaum, a professor at the Massachusetts Institute of Technology
who also deals with human intelligence, argues similarly and says that we
should study human flexibility and creativity if we want to develop real arti-
ficial intelligence at a human level. He highlighted, among other things,
the intelligence of Demis Hassabis and his colleagues at the company
DeepMind, who conceived, designed, and created AlphaGo in the first place
(Lake et al., 2017).

References
Choi, S., Kim, N., Kim, J., & Kang, H. (2022). How does AI improve human
decision-making? Evidence from the AI-powered Go program. Evidence from the
AI-powered Go program (April 2022). USC Marshall School of Business Research
Paper Sponsored by iORB, No. Forthcoming.
Jaderberg, M., Czarnecki, W. M., Dunning, I., Marris, L., Lever, G., Castaneda,
A. G., … & Graepel, T. (2019). Human-level performance in 3D multiplayer
games with population-based reinforcement learning. Science, 364(6443),
859–865.
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building
machines that learn and think like people. Behavioral and Brain Sciences, 40,
e253.
Marcus, G. (2018). Innateness, alphazero, and artificial intelligence. arXiv preprint.
arXiv:1801.05667.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G.,
… & Hassabis, D. (2015). Human-level control through deep reinforcement
learning. Nature, 518(7540), 529–533.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G.,
… & Hassabis, D. (2016). Mastering the game of Go with deep neural networks
and tree search. Nature, 529(7587), 484–489.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., …
& Hassabis, D. (2017). Mastering the game of Go without human knowledge.
Nature, 550(7676), 354–359.
Zhang, H., & Yu, T. (2020). AlphaZero. Deep reinforcement learning: Fundamentals,
research and applications (pp. 391–415). Springer.
 14
Recurrent Neural Networks

Recurrent neural networks are universal parallel-sequential computers.

Jürgen Schmidhuber

Recurrence in the Brain

Recurrence is a fundamental aspect of neural processing and integration
of information in biological neural networks, especially in the brain. As
we have already seen in the chapter on the structure of the nervous sys-
tem, numerous hierarchically nested feedback loops exist in the brain. Due
to the total number of neurons in the brain (1011) and the average num-
ber of connections per neuron (104) to its successors, an estimate suggests
that each signal on average returns to its source after only three synapses
(Braitenberg & Schüz, 1991): Since each neuron passes the signal to
10,000 other neurons, the signal after the first synapse is at 104, after the
second synapse at 104 × 104 = 108 and after the third synapse theoretically
at 104 × 104 × 104 = 1012 neurons, which however exceeds the actual num-
ber of existing neurons tenfold. Thus, the signal must now have returned to
the original neuron among others. This highly recurrent structure allows the
brain to process information extremely efficiently and dynamically.
The recurrent nature of biological neural networks plays an impor-
tant role in many cognitive processes. For example, recurrent connections
are crucial for the formation and retrieval of memories. When information

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 131
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_14
132    
P. Krauss

flows through the network and returns to its source, it can strengthen the
connections between the involved neurons, leading to the consolidation of
memories. This process is facilitated by synaptic plasticity, i.e., the ability of
synapses to change their strength over time depending on their activity.
In addition, the recurrent connectivity in the brain contributes to the
integration of sensory information from different modalities, which is essen-
tial for perception and decision-making. For example, when we simulta-
neously see and hear something, the brain’s recurrent circuits allow these
separate inputs to be merged into a coherent representation of the external
environment. This multisensory integration enables us to respond appropri-
ately to our environment and make informed decisions.
In addition to memory formation and sensory integration, recurrent con-
nections in the brain also play a crucial role for attention and consciousness.
By selectively amplifying or inhibiting certain signals in feedback loops, the
brain can modulate its attention to specific stimuli and allocate cognitive
resources accordingly. This mechanism is important to filter out irrelevant
information and maintain cognitive flexibility.

Recurrence in Artificial Neural Networks

In contrast to pure feedforward architectures, as we have learned in the previ-
ous chapter, artificial recurrent neural networks (RNN) also offer several advan-
tages that make them suitable for a variety of tasks. One of the main advantages
of RNNs is their ability to process temporal dependencies and sequences in
the data, which is essential for tasks with time series data, natural language
processing, and other problems where the order of the input is important.
In addition, RNNs can process or generate input and output sequences of
variable length, while pure feedforward networks require input and output
sequences of fixed size. Due to this flexibility, RNNs are better suited for
tasks where the length of the input and output sequences varies, such as in
language translation or speech recognition.
The recurrence can be expressed in different ways. In the simplest case,
the network as a whole is a feedforward network, with the neurons of some
or all layers having additional self-connections, so that in each time step
of processing they receive their own output from the previous time step as
input in addition to the input from the preceding layer. However, there are
also network architectures where entire layers are fed back, up to fully recur-
rent neural networks, which consist almost entirely of a single layer of recur-
rently connected neurons.
 14 Recurrent Neural Networks    
133

LSTMs
Long-Short-Term Memories (LSTMs) are recurrent neural networks that
are specifically designed to store information over long periods of time and
selectively forget (Hochreiter & Schmidhuber, 1997). The core idea behind
LSTMs is the use of a memory cell that can store information over a longer
period of time. The memory cell is updated by a series of gate mechanisms
that control the flow of information into and out of the cell. The gates are
trained to learn which information to store, forget, or update at each time
step. LSTMs have been successfully used for a variety of tasks, including
speech recognition, machine translation, and automatic caption genera-
tion. They have also been used in combination with other neural network
architectures such as convolutional networks to create more powerful mod-
els for tasks like object recognition and classification in images and video
sequences.

Elman Networks
Elman Networks are another type of recurrent neural networks, proposed by
Jeffrey Elman (Elman, 1990). In the simplest case, these are three-layer net-
works with an input, intermediate, and output layer, with the intermediate
layer being extended by a so-called context layer. This context layer stores
the state of the intermediate layer from the previous time step and passes it
on to the intermediate layer. Thus, the intermediate layer receives the new
input from the input layer and additionally its own activation state from the
previous time step. Therefore, Elman Networks are also capable of process-
ing input sequences and generating output sequences (Fig. 14.1).

Highly Recurrent Neural Networks

Hopfield networks are one of the most well-known examples and at
the same time a special case of highly or fully recurrent neural networks
(Hopfield, 1982). Although they are recurrent, Hopfield networks are not
intended to process time series. However, they are briefly mentioned here as
they exemplify the performance of neural networks with recurrent connec-
tions beyond the processing of sequences (Fig. 14.2).
Hopfield networks consist of a single layer of neurons, with each neu-
ron i symmetrically connected to every other neuron j, i.e., for the weights
134    
P. Krauss

Output layer

Hidden layer

Context layer

Input layer

Fig. 14.1 Elman Networks. Named after their inventor Jeffrey Elman, they are a type
of recurrent neural networks. They consist of three layers—input, intermediate, and
output layer—and are characterized by an additional context layer that stores the
state of the intermediate layer from the previous time step and passes it on. As a
result, Elman Networks can process input sequences and generate output sequences

w, wij = wji applies. Also, wii = 0, i.e., there are no self-connections. Hopfield
networks exhibit pronounced attractor dynamics. Attractors are stable states
towards which a dynamic system moves over time. Hopfield networks can
store patterns as attractors and denoise or complete them upon re-presenta-
tion, i.e., the network activity converges (usually in one time step) into the
attractor most similar to the input.
Highly recurrent neural networks, where the weights are no longer sym-
metric (reservoirs), represent complex dynamic systems, which are even
capable of continuous activity without external input (Krauss et al., 2019).
As we will see in the next section, recurrent neural networks are difficult to
train, with reservoirs being the most difficult to train. A radically new approach,
motivated by findings in brain research (Reservoir Computing), even com-
pletely refrains from training these networks. But more on that later in Part IV.

Difficulty in Training Recurrent Networks

During learning, a phenomenon occurs in RNNs which is referred to as
the problem of vanishing or exploding gradients (Hochreiter, 1998; Hanin,
2018; Rehmer & Kroll, 2020). This problem also occurs in a weakened form
 14 Recurrent Neural Networks    
135

Fig. 14.2 Highly recurrent neural network. In this type of network, the ordered layer
structure is completely abolished. In principle, each neuron can be connected to any
other—even reciprocally. These networks are complex dynamic systems, which are
capable of continuous activity even without input

in deep neural networks with pure feedforward architecture, but it is drasti-

cally exacerbated in RNNs. This refers to the instability of the gradients dur-
ing error backpropagation, which can lead to slow or unstable learning. This
problem occurs in two ways: vanishing and exploding gradients.
As the gradients are propagated backwards through the layers of the net-
work, they can become very small with each additional layer (or with each
additional time step in RNNs) and tend towards zero, leading to slow con-
vergence, poor performance, and difficulties in learning long-range depend-
encies or meaningful features from the input data.
Exploding gradients occur when they grow exponentially during back-
propagation, leading to unstable models, poor convergence or even diver-
gence of the learning process, erratic behavior, and high prediction errors.
Unlike pure feedforward networks, RNNs can only be trained by a trick
with gradient descent methods and error backpropagation. In backprop-
agation through time, the recurrent neural network is unfolded through
136    
P. Krauss

the training process over time and converted into a deep feedforward net-
work with shared weights (Werbos, 1990). This unfolded network repre-
sents the computations of the RNN over several time steps. The depth of
the unfolded network depends on the length of the input sequence and the
number of time steps processed by the RNN. The backpropagation algo-
rithm is then applied to the unfolded network and the gradients are calcu-
lated for each time step.

Conclusion
Recurrence in biological neural networks forms the basis for understanding
the functioning of artificial recurrent neural networks (RNNs). By mim-
icking the recurrent architecture of the brain, RNNs can efficiently process
or generate temporal information and sequences, making them particularly
suitable for tasks such as natural language processing, time series analysis,
and speech recognition.
Although RNNs are simplified abstractions of the complex recurrent
structures of the brain, they provide valuable insights into the principles
underlying the brain’s information processing capabilities.
However, recurrent networks are also more difficult to train than pure
feedforward architectures. As we will see in the chapter on language-talented
AI, recurrence has been completely replaced by another mechanism in large
language models like ChatGPT, which is significantly easier and therefore
faster to train.

References
Braitenberg, V., & Schutz, A. (1991). Anatomy of the cortex: Studies of brain func-
tion. Springer.
Elman, J. L. (1990). Finding structure in time. Cognitive science, 14(2), 179–211.
Hanin, B. (2018). Which neural net architectures give rise to exploding and vanish-
ing gradients? Advances in neural information processing systems, 31.
Hochreiter, S. (1998). The vanishing gradient problem during learning recurrent
neural nets and problem solutions. International Journal of Uncertainty, Fuzziness
and Knowledge-Based Systems, 6(2), 107–116.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural
Computation, 9(8), 1735–1780.
 14 Recurrent Neural Networks    
137

Hopfield, J. J. (1982). Neural networks and physical systems with emergent col-
lective computational abilities. Proceedings of the National Academy of Sciences,
79(8), 2554–2558.
Krauss, P., Schuster, M., Dietrich, V., Schilling, A., Schulze, H., & Metzner, C.
(2019). Weight statistics controls dynamics in recurrent neural networks. PLoS
ONE, 14(4), e0214541.
Rehmer, A., & Kroll, A. (2020). On the vanishing and exploding gradient problem
in gated recurrent units. IFAC-PapersOnLine, 53(2), 1243–1248.
Werbos, P. J. (1990). Backpropagation through time: What it does and how to do
it. Proceedings of the IEEE, 78(10), 1550–1560.
 15
Creativity: Generative Artificial Intelligence

Creativity means seeing what others see, and thinking what no one else has thought.
Albert Einstein

What is Creativity?
Before we turn to creative AI, let’s first briefly define what we want to under-
stand as creativity in the context of this presentation. Most definitions of
creativity contain two main aspects. Creativity means creating something
that is both new and useful. In addition, the results of Eagleman and Brandt
suggest that creativity often arises from questioning old prejudices through
three key methods: Bending, Blending, and Breaking. These concepts repre-
sent different strategies for manipulating existing ideas, concepts, or frame-
works to create new insights, solutions, or creations (Eagleman & Brandt,
2017).
In Bending, an existing concept or idea is altered to create something
new. By changing some aspects of the original concept while retaining its
core, unique interpretations or innovations can be developed. Bending
expands the boundaries of an idea or concept and opens up new perspectives
and possibilities. In art, for example, Bending could mean taking a tradi-
tional painting style and adapting it to express a modern theme or object.
In Blending, two or more seemingly different ideas or concepts are
combined to create a new, unified whole. By synthesizing elements from

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 139
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_15
140    
P. Krauss

different sources, innovative connections and associations can emerge. The

fusion can lead to the development of entirely new areas, such as in the case
of bioinformatics, which combines biology and computer science. In litera-
ture, elements from different genres can be blended, for example, the combi-
nation of romance and science fiction into a new subgenre.
Breaking means dismantling or deconstructing an existing idea, con-
cept, or framework to reveal the underlying components or assumptions.
By questioning conventional knowledge or established norms, the way can
be paved for new insights and breakthroughs. Breaking can lead to a reas-
sessment of basic principles and promote the development of new theories,
methods, or practices. In science, the process of breaking could mean ques-
tioning long-standing assumptions about a particular phenomenon, leading
to a new understanding or explanation. Examples of this would be Einstein’s
theory of relativity or quantum mechanics.
The area of AI that deals with creativity, i.e., the generation of new con-
tent, is referred to as generative AI or generative deep learning and includes
a variety of so-called generative models for generating images, videos, texts,
spoken language, or music (Foster, 2019). The most important ones will be
briefly introduced in the following.

Deep Dreaming: When the Input is Trained, Not

the Network
Deep Dreaming is a method to generate new images with unique, dream-
like patterns and features (Mordvintsev et al., 2015). The process is based on
the same neural networks used for image recognition tasks (convolutional
networks), but works in reverse and optimizes the input image to generate
certain patterns or features, rather than identifying them. So, it’s not the net-
work that is trained on a given input, but the input is adapted to a given
network.
The process begins with a network that has already been pre-trained on
a large number of images. Usually, a dataset containing millions of images
and thousands of object classes has been used for this purpose. The network
has therefore already learned to recognize various patterns and features from
the training data through a hierarchical process, with early layers recognizing
low-level features (edges, textures) and deeper layers recognizing high-level
features (objects, scenes).
 15 Creativity: Generative Artificial Intelligence    
141

Next, the user selects a specific layer of the network they want to work
with, thereby determining the level of abstraction of the newly generated
image. Lower layers generate simpler patterns, while higher layers produce
more complex features. The user provides an existing input image that will
serve as the starting point for the Deep Dream process. This image is passed
through all layers up to the selected layer. The activations in this layer are
read out and represent the patterns and features recognized in the image at
the corresponding level of abstraction.
Now, the user defines a target function, which usually aims to maximize
the sum of the activations in the selected layer. This stimulates the neural
network to enhance the recognized patterns and features in the input image.
Similar to training with backpropagation, the gradient of the target func-
tion with respect to the input image is calculated. This indicates how the
input image needs to be changed to increase the value of the target function
and is applied to the input image to update it. This process is repeated for a
certain number of iterations or until a termination criterion is reached. The
resulting image contains enhanced patterns and features, giving it a unique,
dream-like appearance.
As already mentioned, the degree of abstraction of the image content can
be controlled by selecting the layer in the neural network, as early layers
represent simpler patterns like corners and edges, while deep layers contain
more abstract representations of whole objects or scenes. This leads to a par-
adoxically appearing curiosity. If a deep layer with abstract representations is
chosen as the starting point of Deep Dreaming, the resulting images tend to
be more figurative. In contrast, early layers with simpler representations lead
to images that are reminiscent of abstract art.

Style Transfer
Style Transfer aims to apply the artistic style of one image (style image) to
the content of another image (content image) to create a new image that
combines the content of the first image with the style of the second image
(Gatys et al., 2015, 2016). In the sense of the three strategies of creativity
introduced above, this technique thus falls into the area of blending. For
this purpose, the same networks are used again, which also serve for image
recognition, to separate and recombine content and style information of
images.
142    
P. Krauss

As with Deep Dreaming, the process begins with a network that has
already been pre-trained on a large number of images. The user specifies the
content image, which contains the motif or scene to be preserved, as well
as the style image, which represents the artistic style to be applied to the
content image. Subsequently, both images are passed separately through the
neural network. For the content image, the activations from one or more
late layers are used to capture high-level content features. For the style
image, activations from multiple layers of all hierarchy levels are used to cap-
ture both low-level style features (e.g., textures) and high-level style features
(e.g., structures). These style features are typically represented by so-called
Gram matrices, which capture the correlations between different features in
each layer and thus effectively encode the style information.
Subsequently, a starting image is first created, which often begins as
random noise or as a copy of the content image. The goal is to iteratively
update this starting image so that it corresponds to both the content fea-
tures of the content image and the style features of the style image. This is
achieved by defining a special error function, which consists of two main
components: content error and style error. The content error measures the
difference between the content features of the starting image and the con-
tent image, while the style error measures the difference between the style
features of the starting image and the style image. The total error is the
weighted sum of content and style error. By weighting the two parts, it can
be determined how strong the respective influence on the new generated
image should be, i.e., how far away it is from the original content or style.
Again, the gradient of the total error is calculated, which indicates how the
starting image needs to be changed to minimize the error. A corresponding
optimization algorithm is applied iteratively to update the starting image.
The final image retains the content of the content image, but now appears in
the artistic style of the style image.

Generative Adversarial Networks

A Generative Adversarial Network (GAN) is a system of two coupled neu-
ral networks used to generate deceptively real images or videos, so-called
Deep Fakes. It consists of a generator network and a discriminator network
(Goodfellow et al., 2020).
The generator continuously creates new candidate images or videos, while
the discriminator simultaneously tries to distinguish real images and videos
from artificially generated ones. Over the course of training, both networks
 15 Creativity: Generative Artificial Intelligence    
143

iteratively improve in their respective tasks. The Deep Fakes thus generated
are usually indistinguishable from real images and videos.

Diffusion Models
Diffusion models are a class of generative models that produce images
through a process known as Denoising Score Matching (Vincent et al.,
2010; Swersky et al., 2011; Sohl-Dickstein et al., 2015). These models learn
to generate images by simulating a diffusion process that transforms a target
image into random noise, and then learn to reverse this process. The basic
idea is to train the model to predict the statistical distribution of the pixel
values of the original image from a noisy version.
The training of a diffusion model begins with a dataset of images. By add-
ing Gaussian noise, a sequence of increasingly noisy versions of each image
is then generated. This process is referred to as diffusion. Subsequently,
the model is trained to predict or reconstruct the next less noisy version of
the original image from a noisy version. During training, the model learns
to predict the statistics of the original image at each step of the diffusion
process. Through this training, the model becomes increasingly efficient at
removing noise from images. When the trained model is now given random
noise as input, it iteratively generates (hallucinates) a completely new ran-
dom image from it.
Diffusion models typically consist of layers of neurons with local connec-
tivity, with each neuron only connected to a small neighborhood of neu-
rons in the previous layer. This local connectivity is advantageous for image
processing as it allows the model to learn local patterns and capture spatial
hierarchies.
To generate images from text descriptions, diffusion models can be com-
bined with language models (see next chapter). There are various ways to
achieve this coupling, but a common approach is the use of a technique
known as conditional diffusion (Batzolis et al., 2021; Nichol et al., 2021).
In this case, the diffusion model is made dependent on the text description
by integrating it into the model architecture. First, the text description is
encoded using a pre-trained language model (e.g., GPT-3 or BERT). This
generates a high-dimensional vector representation of the meaning of the
text (text embedding). The diffusion model is then conditioned on the text
representation by modifying its architecture. This can be done by adding
the text embedding as additional input or by integrating it into the hidden
layers of the model. Finally, the model is trained with the same diffusion
144    
P. Krauss

process as before, but now with images and the corresponding text descrip-
tions, so the diffusion process does not run “freely”, but is constrained by
the respective text embedding.
If an image is to be generated from a text description, the conditioned
model is again given random noise as visual input, but this time also with
the encoding of the text from the language model. The model then generates
a random image that corresponds to the input description.
The most well-known diffusion models include DALL-E 2, Stable
Diffusion, and Midjourney.
DALL-E 2 (Dali Large Language Model Encoder 2) was specifically devel-
oped for the generation of high-quality photorealistic images from natural
language descriptions. It uses a combination of image and text processing
to generate abstract image representations from linguistic descriptions of
objects, scenes, or concepts. These image representations are then used by
a so-called decoder network to generate photorealistic images from them.
DALLE-E 2 is freely accessible online.1
Released in 2022, Stable Diffusion2 is also a generative model for generat-
ing detailed images from text descriptions. However, it can also be used for
other tasks such as generating image-to-image translations based on a text
prompt. What is special about Stable Diffusion is that its complete program
code and all model parameters have been published.3 It can be operated
on most standard PCs or laptops equipped with an additional GPU and
offers full access for systematic analysis, customization, or further develop-
ment. This represents a departure from the practice of other AI models like
ChatGPT or DALL-E 2, which are only accessible via cloud services and
whose exact internal architecture has not been published so far.
Midjourney4 is considered the most advanced diffusion model currently
available. It is capable of even generating photorealistic images from text
descriptions artificially.
Deep Fakes can also be generated with diffusion models. It has been shown
that the newer diffusion models are significantly superior to the longer-exist-
ing Generative Adversarial Networks (Dhariwal & Nichol, 2021).

1 https://openai.com/product/dall-E-2

2 https://stablediffusionweb.com/

3 https://github.com/CompVis/stable-diffusion

4 https://www.midjourney.com.
 15 Creativity: Generative Artificial Intelligence    
145

Conclusion
There exists a whole range of various methods of generative AI for creating
new content such as images, videos, or even music. The generated images
are usually hardly distinguishable from real ones by humans. All such
approaches fall into the areas of bending and blending, and probably not
into the area of breaking.
It should not go unmentioned at this point that the creation of Deep
Fakes is not limited to the creation of images. The system VALL-E5 is capa-
ble of swapping the voice of any audio recording, for example. This is also a
kind of style transfer, where the content of the spoken language is matched
and preserved, while the target voice corresponds to the style that is being
replaced (Wang et al., 2023).
By the way, the writing style of a text can also be swapped without chang-
ing the content. This is referred to as Prose Style Transfer. This falls into the
domain of probably the most spectacular form of generative AI currently,
that of the so-called large language models like ChatGPT, which due to their
relevance and timeliness, a separate—the next—chapter is dedicated to.

References
Batzolis, G., Stanczuk, J., Schönlieb, C. B., & Etmann, C. (2021). Conditional
image generation with score-based diffusion models. arXiv preprint
arXiv:2111.13606.
Dhariwal, P., & Nichol, A. (2021). Diffusion models beat GANs on image synthe-
sis. Advances in Neural Information Processing Systems, 34, 8780–8794.
Eagleman, D., & Brandt, A. (2017). The runaway species: How human creativity
remakes the world. Catapult.
Foster, D. (2019). Generative deep learning: Teaching machines to paint, write, com-
pose, and play. O’Reilly Media.
Gatys, L. A., Ecker, A. S., & Bethge, M. (2015). A neural algorithm of artistic style.
arXiv preprint. arXiv:1508.06576.
Gatys, L. A., Ecker, A. S., & Bethge, M. (2016). Image style transfer using convo-
lutional neural networks. In Proceedings of the IEEE conference on computer vision
and pattern recognition (pp. 2414–2423).

5 https://vall-e.io/
146    
P. Krauss

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
… & Bengio, Y. (2020). Generative adversarial networks. Communications of the
ACM, 63(11), 139–144.
Mordvintsev, A., Olah, C., & Tyka, M. (2015). Inceptionism: Going deeper into neu-
ral networks. Google Research Blog. https://research.google/pubs/pub45507.
Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., … &
Chen, M. (2021). Glide: Towards photorealistic image generation and editing with
text-guided diffusion models. arXiv preprint arXiv:2112.10741.
Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., & Ganguli, S. (2015, June).
Deep unsupervised learning using nonequilibrium thermodynamics. In International
Conference on Machine Learning (pp. 2256–2265). PMLR.
Swersky, K., Ranzato, M. A., Buchman, D., Freitas, N. D., & Marlin, B.
M. (2011). On autoencoders and score matching for energy based models. In
Proceedings of the 28th international conference on machine learning (ICML-
11) (S. 1201–1208).
Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P. A., & Bottou, L.
(2010). Stacked denoising autoencoders: Learning useful representations in
a deep network with a local denoising criterion. Journal of Machine Learning
Research, 11(12).
Wang, C., Chen, S., Wu, Y., Zhang, Z., Zhou, L., Liu, S., … & Wei, F. (2023).
Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers. arXiv pre-
print arXiv:2301.02111.
 16
Talking AI: ChatGPT and Co.

I am an artificial intelligence that can process and generate answers in natural language
to various types of questions and tasks.
ChatGPT

A Brief History of Natural Language Processing

The history of natural language processing (NLP) dates back to the
mid-twentieth century, when early attempts at machine translation, such as
the Georgetown-IBM experiment (see glossary), laid the foundation for this
field (Booth & Richens, 1952).
In the following decades, NLP went through various phases, includ-
ing rule-based systems that relied on manually created rules and linguis-
tic knowledge, as well as statistical methods that used techniques such
as Hidden Markov Models and Bayesian inference for language modeling
(Rabiner, 1989). Significant advancements were made with the advent of
machine learning. The introduction of word representation techniques fur-
ther improved the capabilities of NLP (Mikolov et al., 2013). In recent
years, the field has been revolutionized by deep learning techniques and
the development of advanced architectures such as LSTMs (Hochreiter &
Schmidhuber, 1997) and Transformers (Vaswani et al., 2017), ultimately
leading to the development of large language models like BERT (Devlin
et al., 2018) and GPT (Brown et al., 2020).

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 147
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_16
148    
P. Krauss

Word Vectors
Techniques for representing words are crucial for natural language processing as
they provide a way to convert text data into numerical representations that can
be processed by machine learning algorithms. These representations capture the
structure and semantics (meaning) of the language and allow models to recog-
nize relationships between words and perform various tasks based on this.
One-Hot Encoding is a simple technique for representing words, where
each word in the vocabulary is represented as a binary vector whose size
(dimensionality) corresponds to the size of the vocabulary. The vector has a
“1” at the index that corresponds to the position of the word in the vocab-
ulary, and a “0” at all other positions. While one-hot encoding is easy to
implement, it suffers from the curse of dimensionality, as the size of the
vectors increases with the size of the vocabulary, leading to inefficient rep-
resentations and increased computational complexity. Nevertheless, one-hot
encoding is an important intermediate step in generating a dense encoding,
i.e., word embeddings. These are low-dimensional continuous vector spaces
in which the words are embedded (Latent Space Embeddings), as opposed
to the high-dimensional and sparsely populated space of one-hot encoding.
Another advantage of word embeddings is that in this space, word vec-
tors are arranged so that semantically similar words are closer together, while
dissimilar words are further apart. Ideally, synonyms even occupy the same
location. In the space of one-hot encoding, however, the location of a vector
has no connection to the meaning of the corresponding word (Fig. 16.1).
Word2Vec is one of these neural network-based techniques for generating
word embeddings (Mikolov et al., 2013, 2017). It learns word embeddings
from large text corpora in an unsupervised manner, i.e., it does not require
labeled data. Word2Vec is particularly effective at detecting semantic and
syntactic relationships between words. There are two different approaches
used in Word2Vec: the Continuous Bag of Words (CBOW) model and the
SkipGram model.
In the CBOW architecture, the model learns word embeddings by predict-
ing a target word based on its surrounding context words within a certain
window. The input consists of the context words and the output is the target
word. During training, the model adjusts the word embeddings to minimize
the prediction error for the target word depending on the context words.
The SkipGram architecture is somewhat the reverse of the CBOW model.
In this case, the model learns word embeddings by predicting the context
words surrounding a given target word. The input is thus the target word, and
the output is a context word within a certain window around the target word.
 16 Talking AI: ChatGPT and Co.    
149

Fig. 16.1 One-Hot versus Dense Encoding. In one-hot encoding, each vector con-
tains only one one, all other elements are zeros. The vector space has as many dimen-
sions as there are words, with each word lying on its own coordinate axis. The vast
majority of this incredibly high-dimensional space remains largely unoccupied. In
dense encoding, on the other hand, significantly fewer dimensions are needed,
resulting in a more compact vector space

Both the CBOW and the SkipGram model use a flat three-layer neural
network with an input layer, a hidden layer, and an output layer. The word
embeddings, which are adjusted during training to minimize prediction
error, are created in the hidden layer. As the model learns, it captures the
semantic and syntactic relationships between the words (Fig. 16.2).
The result is a more efficient and semantically rich representation of
words, enabling NLP models to better understand and process language.
By training Word2Vec on very large amounts of text, it learns the meaning
of a word, so to speak, on the back of many other words that often appear
with the word to be learned.1 For example, in many texts, the words “bark”,

1 Learning the meaning of words on the back of other words is quite similar to a child’s language acqui-
sition. Although the first approximately 50 words are learned in a different way, namely by linking
the respective sensory representation with the corresponding word. From the so-called 50-word limit,
however, around the 2nd birthday, the so-called vocabulary spurt begins, with many new words being
learned. It is assumed that this phase of language acquisition benefits from the fact that the meaning of
new words can be inferred from already known words (Ferguson and Farwell, 1975; Rescorla, 1989;
Aitchison, 2012).
150    
P. Krauss

CBOW SkipGram

the the

little little

dog

barks barks

loudly loudly

The little dog barks loudly.

Fig. 16.2 Word2Vec. Both the CBOW and the SkipGram model use a flat three-layer
neural network with an input layer, a hidden layer, and an output layer. In CBOW,
the context words serve as input and the target word as output. In SkipGram, it is
exactly the opposite. Here, the target word is the input and the context words are
the desired output. In both cases, the word embedding of the target word is created
in the hidden layer, in this example “dog”. In both cases, the dimensionality of the
hidden layer is smaller than that of the input and output layers

“pet”, and “fur” often appear together with “dog”, thus defining its mean-
ing. Conversely, the words “hamster”, “cat”, and “dog” often appear with the
word “pet”, thus defining its meaning, analogously for every other word in
a language. The word embeddings generated by Word2Vec, e.g., v(“dog”),
then serve as input for further language processing algorithms such as
machine translation or document classification.
A remarkable feature of word embeddings is the—initially surprising—
fact that they can even be calculated with. For example, by adding and
 16 Talking AI: ChatGPT and Co.    
151

subtracting word vectors, new word vectors with meaningful meaning can
be generated, as the following example impressively shows:

v(“King”) – v(“Man”) + v(“Woman”) = v(“Queen”)

Transformer
Transformer (Vaswani et al., 2017) are the basis of all modern Large
Language Models (LLM) like DeepL and ChatGPT. This new neural net-
work architecture has revolutionized many tasks of natural language process-
ing such as machine translation, question answering, and text summarization.
The Transformer architecture deviates from traditional recurrent neural
networks (RNNs) by relying on a technique known as attention mecha-
nism (Self-Attention). The attention mechanism allows input sequences to
be processed in parallel rather than sequentially. This has the advantage that
long-range dependencies—i.e., references between words with a greater dis-
tance within the sequence—can have a strong influence on the processing. If
the sequence is processed serially as in RNNs, the influence of one word on
another decreases very quickly with increasing distance. Another advantage
of the attention mechanism is that it is highly parallelizable, which leads to
shorter training times for the transformers (Fig. 16.3).
The clue behind the attention mechanism is that it allows the model to
weight the meaning of different words in a specific context. To do this, trans-
formers calculate a new representation for each word in the input sequence,
which takes into account both the word itself and the surrounding context.
First, from the word embeddings (word vectors, x ) of each word in the
input sequence, three new vectors are calculated: query (q ), key (k ), and
value vector (v ). The values of the conversion matrices between the differ-
ent types of vectors correspond to the internal parameters of the language
model, which are learned during training.
From the query and key vectors, the attention (a ) for each word at each
position in the sequence is determined by calculating the respective sca-
lar product. The attention indicates for each possible pair of words in the
sequence how important one word is for understanding the other word in
the sequence.
From the attention values of all words for a specific word and the corre-
sponding value vectors of all words, a new representation (y ) for this word
is finally calculated, analogously for all other words. To do this, all value
152    
P. Krauss

unfold

Fig. 16.3 RNN vs. Transformer. In an RNN, input sequences are processed serially. At
each time step, the RNN receives the next word along with its own state in the pre-
vious time step as input. As a result, the influence of past words on the current word
decreases very quickly with increasing distance in the sequence. In contrast, trans-
formers process the sequence as a whole. This means that in principle, every word
in the sequence can have a strong influence on every other. This is controlled by the
attention mechanism

vectors are added, after they have been weighted by multiplication with the
respective attentions (Fig. 16.4).
That one can calculate with the vectorial representation of words and
thereby meaningful new meanings arise, we have already seen in the pre-
vious section using the example “King/Queen”. The new representation of
each word thus corresponds to its own meaning plus the weighted sum of
the meaning of all other words in the specific context of the entered text.
Simplified, the following happens: The “standard meaning” of each word,
as it would be in a dictionary, is modified and adapted to the specific situa-
tion. From a linguistic point of view, this corresponds to the transition from
semantics to pragmatics.
Modern transformers are made up of many such modules, consisting of
attention and subsequent forward-directed neural network. Like the layers
in a deep neural network, they are connected in series, with the output of
one module serving as input for the subsequent module serves. ChatGPT,
for example, consists of 96 of these modules or layers. In addition, on each
layer there are several of these modules in parallel, which can then specialize
in different aspects, such as different languages or types of text.
 16 Talking AI: ChatGPT and Co.    
153

she plays Go

she she

plays
plays

she plays Go

Fig. 16.4 Attention mechanism. From each word vector x of the input sequence,
three new vectors are generated: query (q ), key (k ), and value vector (v ). The atten-
tion a, which represents the relative importance of a word for understanding another
word in the sequence, is calculated by the scalar product of query and key vectors.
Finally, a new representation for each word y is created by weighting the value vec-
tors of all words with the respective attention and adding them up

It has been shown that by iterative prediction of the next word based on
the previous input sequence of words and subsequent appending of the pre-
dicted word to the next input, in principle arbitrarily long, meaningful texts
can be generated (Liu et al., 2018). The final output, which the transformer
is trained on, therefore corresponds to the probability for each individual
word of the language (which can be several tens of thousands) that it will
appear next in the text. Subsequently, a word is randomly selected according
to this probability distribution over all words. So it is not generally the most
probable word that is chosen (which would correspond to a winner-takes-all
154    
P. Krauss

approach), but with certain (descending) probabilities also the second or

third most probable, etc. In this way, chance enters the generative processes,
so that even with identical input, the exact same output is never produced
twice. In the next (time) step, the predicted word is appended to the orig-
inal input sequence and used as new input for the transformer, which then
predicts the next word again, etc. In this way, the transformer can generate
longer texts from short inputs. It thus also belongs to the class of generative
AI, which we had already got to know in the previous chapter.
The remarkable thing about this approach of self-supervised learning is
that it does not require labeled data. Therefore, existing large text corpora
such as online encyclopedias (Wikipedia) can be used for training. Already
trained models of this kind are therefore referred to as Generative Pre-
Trained Transformer (GPT).

The GPT Series

The history of the GPT series reflects the rapid advances in natural language
processing and the development of large-scale language models. GPT-1,
introduced by OpenAI in 2018, first demonstrated the potential of self-su-
pervised pre-training on a large language corpus consisting of approximately
4.5 gigabytes of text. The model, consisting of about 150 million parame-
ters, showed impressive results in various tasks such as machine translation
or text summarization.
GPT-2, released in 2019, built on the success of its predecessor and sig-
nificantly expanded the scope and capabilities of the model. With 1.5 billion
internal parameters, GPT-2 was trained on a larger dataset (40 gigabytes)
and was able to generate even more coherent and contextually relevant texts.
Due to concerns about potential misuse, OpenAI initially withheld the
release of the full model and decided to only release smaller versions, to pub-
lish the full model at a later date after a risk assessment.
GPT-3, introduced in 2020 and trained on 570 gigabytes of text, was
another major step in the development of large language models. With 175
billion parameters, GPT-3 demonstrated remarkable learning ability in a
few steps, with the model performing very well in various tasks with mini-
mal fine-tuning. It is capable of inventing stories in various styles, can write
short computer programs in all common programming languages, summa-
rize documents, and translate texts. GPT-3 has already been integrated into
numerous applications, including chatbots and so-called co-pilot tools for
code completion. ChatGPT is also based on this version of the GPT series.
 16 Talking AI: ChatGPT and Co.    
155

Finally, in March 2023, GPT-4 was released, which in some tests appar-
ently performs even better than humans and shows signs of general artifi-
cial intelligence (Bubeck et al., 2023). So far, no details have been published
about its exact structure and the training data used.

ChatGPT
The freely accessible ChatGPT, released in November 2022, is probably the
most well-known example of a large language model and one of the most
advanced AI models for conversations. It is based on GPT-3 and is designed
to understand and generate human-like responses in a dialogue environ-
ment. It is so powerful that it can conduct longer coherent and context-re-
lated conversations with users. It can generate any type of text in seconds,
answer questions on any topic, and conduct conversations, remembering
their course and thus usually responding adequately in longer dialogues.
ChatGPT can summarize, rephrase, or translate texts in several dozen lan-
guages, tell jokes, write songs, and even program in all common program-
ming languages.
Remarkably, ChatGPT is even capable of playing strategic games like
chess, Go, and poker. During its training on virtually any type of text availa-
ble on the internet, it has, for example, “read” tens of thousands of games in
chess notation, each time trying to predict the next word or character. This
apparently resulted in the side effect that it learned the underlying rules of
the respective game without ever having seen a chessboard or pieces. This
suggests that the approach of encoding a problem into a sequence of ele-
ments and then learning to predict the next element could be a universally
applicable strategy in information processing and cognitive systems.
As already mentioned, ChatGPT is based on GPT-3. Unlike its predeces-
sors, which were trained exclusively self-supervised (on predicting the next
word), ChatGPT underwent subsequent fine-tuning through human feed-
back, which proved to be a crucial step for generating even better responses
and texts.
This fine-tuning was based on supervised learning and reinforcement
learning (Reinforcement Learning) and proceeded in three steps. First,
humans generated many example conversations, each consisting of a request
and the corresponding appropriate responses. To assist them in designing
their responses, the human trainers could access previously automatically
generated suggestions and then adapt them. In this way, a large dataset was
created, which contains real dialogues with good, i.e., adequate responses.
156    
P. Krauss

To further enlarge the dialogue dataset, artificially generated examples from

a predecessor of ChatGPT were added. These example dialogues were used
to train ChatGPT in a supervised manner.
For the subsequent reinforcement learning, rewards are required, which
evaluate the quality of the output generated by the AI system. To provide
these rewards for a large number of training examples automatically, a sep-
arate reward model was trained—another neural network, which receives a
request along with a possible response as input and whose task is to predict
the quality of the response as output, which is then used as a reward for
fine-tuning ChatGPT.
To train the reward model, training data had to be created, of course.
These data consisted of the responses generated by the chatbot, which were
sorted by their quality. For this purpose, conversations between human AI
trainers and the chatbot were collected, and several alternative additions
were created for each model response. The human AI trainers then had the
task of evaluating the different responses. With these evaluation data, the
reward model could be trained to predict an appropriate evaluation (reward)
for new response options that it had not seen before.
The fine-tuning of ChatGPT proceeded in such a way that a ran-
dom request from the dialogue dataset was given to ChatGPT as input.
ChatGPT then generated a response. The response was evaluated by the
reward model, and this evaluation was reported back to ChatGPT as a
reward to train it with reinforcement learning, i.e., to maximize the total
reward. This improved its ability to generate high-quality context-related
responses in a conversation situation. Through this iterative improvement
process, ChatGPT became an increasingly efficient conversation AI over
time (Fig. 16.5).

Speech Interfaces
Speech interfaces such as Siri, Alexa, and Google Assistant have revolution-
ized the way we interact with our devices. They provide a conversational
interface that responds to user requests in a natural, human-like manner.
These voice-controlled AI assistants are trained with large amounts of text
data, like ChatGPT and other language models, to understand and gen-
erate natural language. While Siri, Alexa, and Google Assistant primarily
focus on responding to specific commands and questions, often with a focus
on functional support, e.g., setting reminders or playing music, models
like ChatGPT are designed to generate more nuanced and context-related
 16 Talking AI: ChatGPT and Co.    
157

Request Request Request

Answers
Good answer.

Answer

GPT Reward model Reward

Fig. 16.5 Fine-tuning through human feedback. Step 1: A request is randomly

selected from the database. An AI trainer provides an example response, with
which GPT is trained. Step 2: A request and several example responses are randomly
selected from the database. An AI trainer evaluates the quality of the responses. This
trains the reward model. Step 3: A new request is selected from the database and
given to GPT as input, whereupon it generates a response. The reward model deter-
mines the quality of the response. This is reported back to GPT as a reward for rein-
forcement learning

responses, enabling more complex and open conversations. Nevertheless, the

underlying technology is similar.
Ultimately, the translation from written to spoken language and vice versa
(Text-to-Speech, Speech-to-Text) represents just another recoding step, which
already works very efficiently, as the example of Siri and Co. shows (Trivedi
et al., 2018). Coupling this type of language interfaces with ChatGPT or
GPT-4 is the next logical (and not particularly large) step and would make
communication with these AI systems even more intuitive and natural.

Conclusion
The Transformer architecture and large language models like the GPT
series and DeepL have revolutionized natural language processing, with
the attention mechanism and the ability to consider far-reaching depend-
encies playing a key role. Despite these advances, there are still challenges
158    
P. Krauss

and limitations that need to be considered. The language model Galactica of

the Meta corporation, formerly known as Facebook, was taken off the net at
the end of 2022 due to criticism and concerns about its reliability and the
spread of misinformation after just three days.2 What happened?
The AI was supposed to assist scientists in research and writing technical
articles. However, it was found that Galactica had partially invented content,
but presented it as factual and even mixed real and false information. This
behavior of large language models is also referred to as hallucinating. The
massive criticisms eventually led to Galactica being taken off the net. This
incident reminds us of Microsoft’s chatbot Tay from 2016, which continu-
ously evolved through user requests and due to its sensitivity to user prefer-
ences, turned into a racist and homophobic program within 16 hours and
had to be taken off the net.3
Examples like these show that AI models can sometimes generate unreli-
able or even harmful content. These models can invent content or mix true
and false information, leading to misinformation and unwanted behavior.
ChatGPT is not free from errors and can generate plausible-sounding, but
false or nonsensical responses. It also reacts sensitively to changes in the
input formulation and tends to be verbose or use certain expressions too
frequently. It also tends to guess in the case of ambiguous requests, rather
than asking clarifying questions, and although efforts have been made to get
the model to reject inappropriate requests, it can still respond to harmful
instructions or show biased behavior. Its successor GPT-4 is already signifi-
cantly more reliable and advanced in this respect, but certainly still not free
from such errors.

References
Aitchison, J. (2012). Words in the mind: An introduction to the mental lexicon. Wiley.
Booth, A. D., & Richens, R. H. (1952). Some methods of mechanized translation. In
Proceedings of the Conference on Mechanical Translation.
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., &
Amodei, D. (2020). Language models are few-shot learners. Advances in Neural
Information Processing Systems, 33, 1877–1901.

2 https://www.technologyreview.com/2022/11/18/1063487/meta-large-language-model-ai-only-

survived-three-days-gpt-3-science/
3 https://www.nytimes.com/2016/03/25/technology/microsoft-created-A-twitter-bot-to-learn-from-us-

ers-it-quickly-became-A-racist-jerk.html.
 16 Talking AI: ChatGPT and Co.    
159

Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., …
& Zhang, Y. (2023). Sparks of artificial general intelligence: Early experiments with
GPT-4. arXiv preprint arXiv:2303.12712.
Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training
of deep bidirectional transformers for language understanding. arXiv preprint
arXiv:1810.04805.
Ferguson, C. A., & Farwell, C. B. (1975). Words and sounds in early language
acquisition. Language, 419–439.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural
Computation, 9(8), 1735–1780.
Liu, P. J., Saleh, M., Pot, E., Goodrich, B., Sepassi, R., Kaiser, L., & Shazeer, N.
(2018). Generating wikipedia by summarizing long sequences. arXiv preprint
arXiv:1801.10198.
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word
representations in vector space. arXiv preprint arXiv:1301.3781.
Mikolov, T., Grave, E., Bojanowski, P., Puhrsch, C., & Joulin, A. (2017). Advances
in pre-training distributed word representations. arXiv preprint arXiv:1712.09405.
Rabiner, L. R. (1989). A tutorial on hidden Markov models and selected applica-
tions in speech recognition. Proceedings of the IEEE, 77(2), 257–286.
Rescorla, L. (1989). The Language Development Survey: A screening tool for
delayed language in toddlers. Journal of Speech and Hearing disorders, 54(4),
587–599.
Trivedi, A., Pant, N., Shah, P., Sonik, S., & Agrawal, S. (2018). Speech to text
and text to speech recognition systems? A review. IOSR Journal of Computer
Engineering, 20(2), 36–43.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … &
Polosukhin, I. (2017). Attention is all you need. Advances in neural information
processing systems, 30.
 17
What are AI Developers Researching Today?

It is no shame to know nothing, but rather, not wanting to learn anything.

Plato

Learning to Learn
In the rapidly evolving world of Artificial Intelligence, Machine Learning in
particular has made great strides in recent years. Almost daily, technical arti-
cles are published introducing new methods or further developing existing
ones. Even experts often find it difficult to keep up with the enormous speed
of development.
In this chapter, we want to take a closer look at some of these latest trends
and developments. All these innovative approaches have in common that
they have the potential to revolutionize the way AI systems learn and adapt.
The conventional approach of supervised learning requires a large amount of
labeled data for each new task, the procurement of which can be time-con-
suming and expensive. The new approaches are intended to enable efficient
use of prior knowledge, generalize across tasks, or solve novel problems with
minimal training.
Few-Shot, One-Shot, and Zero-Shot Learning are some of these advanced
Machine Learning techniques that are supposed to enable models to learn new
tasks or recognize objects with minimal amounts of data. These approaches

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 161
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_17
162    
P. Krauss

have gained importance in recent years as they can potentially overcome the
need for huge amounts of training data and shorten training time.

Few-Shot Learning
Humans are very good few-shot learners: You don’t have to show a child
thousands of pictures of apples for it to learn the concept of “apple”. A few,
often even a single example, is usually sufficient.
With Few-Shot Learning, the aim is to train models that can quickly
adapt to new tasks with a small amount of training data, such as classifying
new objects based on a few examples (Snell, 2017; Sung et al., 2018). To do
this, the model is first trained on a relatively small dataset that contains only
a few examples for each class or task, and then tested on a new set of exam-
ples. The idea is to teach the model to learn from a few examples and gen-
eralize to new examples, rather than needing large amounts of data for each
task. A recent development in this area are Prototypical Networks (Snell
et al., 2017), where a deep neural network is used to learn a metric space in
which objects of the same class are grouped. This method has shown prom-
ising results in tasks such as image classification, where the model can recog-
nize new categories with just a few examples.

One-Shot Learning
With One-Shot Learning, the concept of learning from a few examples
is continued by having the model learn from just one example per class
(Santoro et al., 2016; Vinyals, 2016). A notable innovation in this area are
Memory Enhanced Neural Networks. This type of neural network uses an
external memory matrix to store and retrieve information about previously
seen examples, so that the model can make accurate predictions based on a
single example. One-Shot Learning has proven particularly useful in tasks
such as handwriting recognition, where a model can accurately recognize a
writer’s style based on a single example.
In biology, One-Shot Learning can be observed, for example, in animals
that are able to quickly recognize and respond to new stimuli or situations
without having been previously exposed to them. For instance, some bird
species are able to quickly recognize and avoid dangerous prey after a single
experience. Of course, humans are also excellent One-Shot learners.
 17 What are AI Developers Researching Today?    
163

Zero-Shot Learning
With Zero-Shot Learning, on the other hand, models can make predictions
for completely unknown classes without explicit training examples (Norouzi
et al., 2013; Socher et al., 2013). The model is trained to recognize objects
or categories it has never seen before. It can therefore classify new input pat-
terns even if no labeled data for the respective class were available during
training. In contrast to supervised learning, where a model is trained with a
certain amount of labeled data examples, Zero-Shot Learning is based on the
transfer of knowledge from related or similar classes that were seen during
training.
This is achieved by using semantic representations such as word vectors
that capture the meaning and relationships between different classes. In
this case, the word vectors replace the labels of the images. For example, if a
model has been trained to recognize images of different animal species (e.g.,
horse, tiger, …), but has never seen an image of a zebra, it can still classify
it as an animal and categorize it into a new category, as it has learned the
relationships between different animal species. In this case, the zebra would
probably be classified into a mixed category of the already learned catego-
ries “horse” (because of the shape of the animal) and “tiger” (because of the
stripes).
Zero-Shot Learning thus enables more efficient and flexible training of
models and generalization to new and unknown categories.

Transfer Learning
In transfer learning, a neural network is first trained on a very large data-
set and then refined on a smaller dataset for a specific task or special
application. This retraining of an already trained network is called fine-tun-
ing. The idea behind it is that the knowledge acquired in solving one prob-
lem can be transferred to another, related problem, thereby reducing the
amount of data and the time required for training a new model (Torrey &
Shavlik, 2010).
If the neural network was initially trained on the gigantic ImageNet data-
set, which consists of 14 million images divided into 20,000 categories, it
can be assumed that this network has already learned a lot of representations
that are useful for general image recognition. Thus, this network can be effi-
ciently adapted to a new task with a short fine-tuning on a few images.
164    
P. Krauss

Meta-Learning
In traditional machine learning, a model is trained on a fixed set of train-
ing data and then used to make predictions for new, unseen data. In con-
trast, meta-learning approaches aim to enable models to learn from a small
amount of data and generalize to new tasks with little or no additional train-
ing (Finn et al., 2007; Santoro et al., 2016). To this end, a so-called meta-
learner is usually trained, i.e., a model that learns how to learn by observing
and extracting patterns from a series of different training tasks. The meta-
learner then uses this knowledge to quickly adapt to new tasks with similar
features.
There are various approaches to meta-learning, including metric-based
learning, optimization-based learning, and model-based learning. In met-
ric-based learning, a similarity metric is learned to compare new tasks with
the training tasks. In optimization-based learning, a model is trained to
quickly optimize its weights for a new task, while in model-based learning,
a generative model of the data is learned that can be used for quick adapta-
tion to new tasks. For example, a meta-learner can be generally trained to
play various strategy games or card games. In a new game, it can then use its
existing knowledge about how games generally work to learn the new game
faster. A similar effect is also observed in translation software.
In 2016, Google introduced the Google Neural Machine Translation
(GNMT) system, a significant advancement in the field of machine trans-
lation that surpassed all existing models (Wu et al., 2016). An interesting
observation was made when the GNMT system, which was initially trained
to translate from English to Spanish and then to translate from English to
Chinese, improved its translation between English and Spanish after being
trained with Chinese text!
This phenomenon can be attributed to the development of a so-called
interlingua, i.e., a common representation of meaning between languages.
As the model was trained on more language pairs, it learned to map the
input text from various languages onto a semantic space that was independ-
ent of the specific language. As a result, the model could now also translate
between languages, even if it had not been explicitly trained for this specific
language pair. The latest version of GNMT now supports 109 languages and
can translate between them in any direction, which is over 11,000 possible
language combinations. The model was explicitly trained only on a very
small fraction of all these possibilities.
 17 What are AI Developers Researching Today?    
165

The emergence of this interlingual representation in the GNMT system

is an impressive example of meta-learning and demonstrates the ability of
deep learning models to learn abstract features that can be generalized across
different tasks. This observation inspired further research into multilingual
and zero-shot translation models, which aim to leverage shared knowledge
between languages and improve translation performance with fewer training
examples for each language pair.
Incidentally, it is also known from human language learners that the
“cognitive effort” for each additional foreign language becomes somewhat
smaller. We never start learning from scratch.

Hybrid Machine Learning

In hybrid machine learning, various techniques, e.g., of deep learning, are
combined with other concepts (Maier et al., 2022). An example is Known
Operator Learning, i.e., learning with known operators. In this case, indi-
vidual layers of a neural network are replaced by so-called operators, i.e.,
mathematical functions, e.g., a Fourier transformation. This inclusion of
prior knowledge about how the data needs to be transformed also has the
advantage of accelerating learning and requiring fewer large datasets (Maier
et al., 2019).
In the research direction of neurosymbolic AI, the strengths of neural net-
works and symbolic reasoning from mathematics and logic are combined
(De Raedt, et al., 2020; Sarker et al., 2021). While the neural networks are
used as usual for pattern recognition and learning regularities and patterns
from large amounts of data, symbolic reasoning is used for logic, knowledge
representation, and decision-making. According to Gary Marcus, this is one
of the keys to human-like AI (Marcus, 2003).

Conclusion
The mentioned methods open up new possibilities for AI systems to learn
more efficiently and adapt to new situations, thus paving the way for more
versatile and robust AI applications in a wide range of application areas.
Beyond the presented methods, such as how machines learn to learn, inten-
sive research is already being conducted on the next steps of development.
166    
P. Krauss

Deep neural networks can be compared to highly specialized computer

programs. However, understanding and interpreting these programs can
pose a significant challenge. In order to better identify and decipher the pat-
terns and operations hidden in deep networks, researchers have begun to
transfer methods from software programming (Maier et al., 2022).

References
De Raedt, L., Dumancic, S., Manhaeve, R., & Marra, G. (2020). From sta-
tistical relational to neuro-symbolic artificial intelligence. arXiv preprint
arXiv:2003.08316.
Finn, C., Abbeel, P., & Levine, S. (2017). Model-agnostic meta-learning for fast
adaptation of deep networks. In International Conference on Machine Learning
(pp. 1126–1135). PMLR.
Maier, A., Köstler, H., Heisig, M., Krauss, P., & Yang, S. H. (2022). Known opera-
tor learning and hybrid machine learning in medical imaging – a review of the past,
the present, and the future. Progress in Biomedical Engineering.
Maier, A. K., Syben, C., Stimpel, B., Würfl, T., Hoffmann, M., Schebesch, F., &
Christiansen, S. (2019). Learning with known operators reduces maximum error
bounds. Nature Machine Intelligence, 1(8), 373–380.
Marcus, G. F. (2003). The algebraic mind: Integrating connectionism and cognitive sci-
ence. MIT press.
Norouzi, M., Mikolov, T., Bengio, S., Singer, Y., Shlens, J., Frome, A., … & Dean,
J. (2013). Zero-shot learning by convex combination of semantic embeddings. arXiv
preprint arXiv:1312.5650.
Santoro, A., Bartunov, S., Botvinick, M., Wierstra, D., & Lillicrap, T. (2016, June).
Meta-learning with memory-augmented neural networks. In International confer-
ence on machine learning (pp. 1842–1850). PMLR.
Sarker, M. K., Zhou, L., Eberhart, A., & Hitzler, P. (2021). Neuro-symbolic artifi-
cial intelligence. AI Communications, 34(3), 197–209.
Socher, R., Ganjoo, M., Manning, C. D., & Ng, A. (2013). Zero-shot learning
through cross-modal transfer. Advances in neural information processing systems,
26.
Snell, J., Swersky, K., & Zemel, R. (2017). Prototypical networks for few-shot learn-
ing. Advances in neural information processing systems, 30.
Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P. H., & Hospedales, T. M. (2018).
Learning to compare: Relation network for few-shot learning. In Proceedings of the
IEEE conference on computer vision and pattern recognition (pp. 1199–1208).
Torrey, L., & Shavlik, J. (2010). Transfer learning. InHandbook of research on
machine learning applications and trends: algorithms, methods, and techniques (pp.
242–264). IGI global.
 17 What are AI Developers Researching Today?    
167

Vinyals, O., Blundell, C., Lillicrap, T., & Wierstra, D. (2016). Matching networks
for one shot learning. Advances in neural information processing systems, 29.
Wu, Y., Schuster, M., Chen, Z., Le, Q. V., Norouzi, M., Macherey, W., … & Dean,
J. (2016). Google’s neural machine translation system: Bridging the gap between
human and machine translation. arXiv preprint arXiv:1609.08144.
 Part III
Challenges

Where there is much light, there is naturally often just as much shadow.
Amidst all the spectacular successes in the field of Artificial Intelligence,
particularly in Deep Learning, it should not be overlooked that there are
many smaller and some fundamental challenges—some even speak of major
crises—which still need to be solved.
Of course, brain research is still far from having developed an overarch-
ing theory of the brain, as we have seen particularly in the chapters on con-
sciousness and free will. But even apart from these very big questions, there
are still many aspects of how the brain works that remain ununderstood.
In this third part of the book, the aim is to describe the major challenges
of both disciplines.
 18
Challenges of AI

Realistically speaking, deep learning is just a part of the larger challenge associated with
building intelligent machines.
Gary Marcus

It’s All About the Data

The US Army had a plan to use neural networks to automatically identify
hidden and camouflaged tanks. The Pentagon commissioned an external
software company for the project. The company designed a neural network
that was trained with 50 images of camouflaged tanks located among trees,
as well as 50 images of wooded areas without tanks. After the training, the
network was tested with 50 additional images per category, which it had not
seen before. Indeed, all images of the test dataset were correctly classified
and the neural network was submitted to the Pentagon.
However, it soon claimed that the neural network was no better than
chance at distinguishing between images with and without tanks. Upon
closer examination, it turned out that the dataset used for training and test-
ing the network contained only photos of camouflaged tanks on cloudy days
and of forests without tanks on sunny days. Therefore, the neural network
did not learn to distinguish between camouflaged tanks and empty forests,
but instead to differentiate between sunny and cloudy days (Yudkowsky,
2008).

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 171
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_18
172    
P. Krauss

This example impressively shows that artificial intelligence can only be as

good as the data it was trained with. While the inadequacy of the AI system
in this case was noticed in time, such a bias in the training data, i.e., a sys-
tematic distortion or unintended correlation between features, can have dra-
matic consequences if it is not detected in time.
Autonomous driving systems like the Tesla Autopilot also rely on deep
neural networks for image processing, for example to recognize traffic signs,
other vehicles, obstacles or the road boundary and to derive the appropriate
action from it. In the early days, a Tesla driver died during autopilot opera-
tion after the vehicle raced at full speed into a truck trailer instead of braking
or evading.1 An analysis of the software revealed that Tesla’s image recogni-
tion interpreted the semi-trailer crossing the intersection as a bridge. The
software had never seen a truck trailer from the side during its training. And
the best fitting category, i.e., the most similar pattern, was a bridge. Since
one can drive under bridges without danger, the appropriate action from the
autopilot’s point of view was to continue at the same speed.
So the data should not only contain as little bias as possible, but above all
they should be sufficiently complete and not contain any critical white spots.

Hallucinating Chatbots

In this context, it is also worth mentioning again the problem of halluci-

nating chatbots (ChatGPT, Galactica) and their potential susceptibility to
producing extremist content (Tay), both of which we have already encoun-
tered in the chapter on linguistically gifted AI. Obviously, these models lack
some kind of world model or internal fact checker. This also represents an
unsolved challenge.

Dangerous Stickers and Other Attacks

While the previous examples show how important the quality of training
data is for the performance of the AI system, the next examples show how
AI systems can be deliberately misled.
In Fig. 18.1, what do you think is the difference between the left and the
middle picture?—You don’t see any?

1 https://www.theverge.com/2016/6/30/12072408/tesla-autopilot-car-crash-death-autonomous-model-s
 18 Challenges of AI    
173

Fig. 18.1 Adversarial Example. For a human, there is no noticeable difference

between the left and the middle picture. However, a network trained on image rec-
ognition will correctly classify the left picture as a “panda”, while it assigns the mid-
dle picture to the category “gibbon”. The middle picture is an adversarial example,
which was artificially created to deceive the classifier. To do this, a specially created
pattern (right) was added to the pixels of the original image (left)

Don’t worry, then you are in good company. Probably there is not a sin-
gle person on this planet who sees a difference. However, if we give these
two pictures to a classifier (neural network trained on image recognition) as
input, it will correctly recognize the left version of the picture as a “panda”,
while it assigns the middle picture to the category “gibbon” with 99%
certainty.
The middle picture is a so-called Adversarial Example, which was delib-
erately manipulated to mislead the classifier (Goodfellow et al., 2014; Xiao
et al., 2018; Xie et al., 2019). The intention behind this and similar exper-
iments is to identify possible weaknesses and errors in neural networks in
order to overcome them in the future. Similar to hackers who are specifically
commissioned by companies or authorities to break into their own IT net-
work in order to gain insights for optimizing security systems.
Now, in practice, i.e. in the real world, of course, you cannot simply
change every pixel of an image without further ado. But there is also a “solu-
tion” for this: Adversarial Patches. These strangely looking stickers have dra-
matic effects on the recognition performance of our image classifier when we
place them next to a banana, for example.2 While every four-year-old child
correctly recognizes the banana with or without this sticker, the classifier is
99% sure that the banana is a toaster when the adversarial patch is next to
the banana (Brown et al., 2017).

2 https://youtu.be/i1sp4X57TL4
174    
P. Krauss

What may seem amusing at first glance and in this example turns out to
be a serious problem upon closer inspection. As already mentioned, auton-
omous driving also relies on image recognition to generate the appropriate
control commands for the vehicle such as “brake”, “accelerate” or “turn left”.
Unfortunately, there are also Adversarial Patches that—from the perspec-
tive of a neural network—turn a stop sign into a sign for a speed limit of 45
miles per hour (Eykholt et al., 2018). The consequences are unimaginable.
This unfortunately opens up entirely new possibilities for terrorist attacks.
If you feel like the author, then you would not want to sit in such a car, and
you probably would not want to live in a city where such vehicles are on the
road.

Alchemy, Reproducibility, and Black Boxes

Apart from these exemplary individual cases, there are three larger, more
fundamental crises in artificial intelligence and especially in deep learning,
which are partly interconnected: the reproducibility crisis, the alchemy
problem, and the black box problem.
The reproducibility crisis refers to the difficulty of reproducing the results
of a study or experiment. In the context of AI, the term refers to the chal-
lenges in reproducing the results of AI research, including the development,
implementation, and evaluation of algorithms (Lipton & Steinhardt, 2018).
The problem of reproducibility affects the reliability and trustworthiness of
AI systems. All too often, unfortunately, the complete information necessary
to reprogram an AI system is not published—such as the initializations of
the model parameters, i.e., their exact starting values. This leads to the fact
that two models that appear identical at first glance can lead to very different
results, as many neural networks often react sensitively to small changes, as
we have seen in the example of adversarial examples.
Related to this is the alchemy problem (Hutson, 2018). Before chemistry
established itself as a natural science with a theoretical superstructure (such
as the periodic table of elements), the synthesis of new substances was char-
acterized by erratic procedures, anecdotal evidence, and trial and error. This
“pre-chemistry” was referred to as alchemy. The development of AI is cur-
rently in a similar stage. The development and adaptation of AI algorithms
still depend on trial and error. The term alchemy problem emphasizes the
lack of a systematic, scientific understanding of how AI models work and
why some models work better than others, often leading to unpredictable
or unexplainable results. This problem is exacerbated by the fact that—as
 18 Challenges of AI    
175

in most areas of science—only positive results are published. Nobody likes

to publish what did not work. Thus, behind every published AI model that
can do something better than previous ones, there is a huge number of mod-
els that were tested during development, performed worse or not as desired,
and therefore were never published.
Ultimately, the two crises mentioned so far can be traced back to a deeper
underlying problem. The black box problem (Castelvecchi, 2016; Ribeiro
et al., 2016). AI models—especially deep neural networks—are complex sys-
tems whose decision-making is not always fully comprehensible and whose
internal dynamics are poorly understood. Therefore, it is often difficult or
even impossible to understand how an AI model arrives at its decisions
or predictions, making it ultimately difficult to trust or verify the model’s
results.

A Critical Appraisal
In his article “Deep Learning: A Critical Appraisal”, Gary Marcus identifies
ten weaknesses of current deep learning (Marcus, 2018). Some of these we
have already encountered, but for the sake of completeness, they will be
briefly outlined here again.

• Limited ability for transfer learning: Deep learning models struggle to gen-
eralize knowledge across different tasks or domains, unlike human learn-
ing, where skills and knowledge can be easily transferred and adapted.
• Data inefficiency: Deep learning models often require large amounts of
data to achieve high performance, while humans can learn effectively
from just a few examples.
• Lack of unsupervised learning methods: Current deep learning models are
mostly based on supervised learning, which requires labeled data for train-
ing. Human learning, on the other hand, largely occurs unsupervised.
• Inability to learn from explicit rules: Deep learning models usually learn
from patterns in the data and not from explicit rules, making it difficult for
them to acquire knowledge that can be easily expressed in the form of rules.
• Opacity: Deep learning models are often criticized as “black boxes”
because they are not interpretable, making it difficult to understand how
they arrive at their decisions.
• Vulnerability to attacks: As we have already seen, deep learning models
can be easily deceived by adversarial attacks, i.e., input patterns deliber-
ately designed to lead the model to make incorrect predictions.
176    
P. Krauss

• Lack of integration of prior knowledge: Current deep learning models

typically do not incorporate existing knowledge into their learning pro-
cess, limiting their ability to utilize previous information.
• Lack of logical reasoning abilities: Deep learning models struggle with
tasks that require complex reasoning or problem-solving skills.
• Limited ability to learn hierarchical representations: Although deep learn-
ing models can recognize hierarchical patterns in data, they often strug-
gle to capture the full complexity of hierarchical structures in human
cognition.
• Sensitivity: Deep learning models can react sensitively to minor distur-
bances in the input data or changes in the training distribution, which
can lead to unexpected performance losses.

Conclusion
Adversarial attacks, that is, targeted attacks on a machine learning system
with the aim of manipulating the behavior of the learning system or con-
fusing it and causing it to make incorrect predictions, are a serious problem.
The examples show that even small disturbances in the input data can signif-
icantly influence the behavior of learning systems, which in turn poses a risk
to the security and reliability of such systems.
It is not yet fully understood why adversarial examples or patches can
so easily deceive artificial neural networks, while natural neural networks
(brains) are immune to this type of deception.
Adversarial attacks therefore represent an important area of research in
the field of security of machine learning systems. In particular, adversar-
ial machine learning investigates such attacks and tries to develop effective
defenses against them.
The shortcomings of today’s AI systems mentioned by Gary Marcus high-
light some of the limitations of deep learning and suggest that the combina-
tion of deep learning with other AI techniques (hybrid machine learning) or
the solution of these problems by developing new approaches within deep
learning could lead to more robust and versatile AI systems.
Incorporating insights from brain research can be crucial in overcoming
the shortcomings of current deep learning:

• Inspiration for new architectures: The human brain is an invaluable

source of inspiration for the development of new neural network architec-
tures. By studying the structure, function, and organization of the brain,
 18 Challenges of AI    
177

researchers can identify principles and mechanisms that can be used to

improve the performance and robustness of artificial neural networks.
• Better understanding of transfer learning: Neuroscience can provide
insights into how the brain efficiently transfers knowledge between differ-
ent tasks and domains. Incorporating these principles into deep learning
models can help improve their ability to transfer learning and reduce the
need for extensive retraining.
• Improved unsupervised learning: Investigating the brain’s mechanisms for
unsupervised learning, e.g., how humans naturally learn without explicit
cues from their environment, can support the development of new algo-
rithms and techniques for unsupervised learning in artificial neural
networks.
• Integration of explicit rules: Exploring how the brain processes, stores,
and uses explicit rules can support the development of deep learning
models that are better able to learn from and argue with explicit rules.
• Better interpretability: Understanding how the brain represents and pro-
cesses information can provide clues on how to make AI models more
interpretable, so that researchers and practitioners can better under-
stand their internal workings and decision-making processes. Moreover,
neuroscience offers a wealth of methods for analyzing biological neural
networks. Why shouldn’t these methods be equally suitable for examin-
ing artificial neural networks? The research field of Explainable AI (XAI)
focuses on the development of AI models and techniques that are more
transparent, understandable, and interpretable for humans. The declared
goal of XAI is to solve the black box problem, e.g., by developing sim-
pler models, visualizing model decisions, or creating human-understand-
able explanations for AI results (Castelvecchi, 2016; Ribeiro et al., 2016;
Doshi-Velez & Kim, 2017; Adadi & Berrada, 2018).
• Resilience against adversarial attacks: By investigating the brain’s robust-
ness against random noise and unwanted disturbances, researchers
can develop strategies to make deep learning models less susceptible to
unwanted attacks.
• Incorporation of prior knowledge: Insights from neuroscience about how
the brain links prior knowledge with new information can help in the
development of deep learning models that better utilize existing knowl-
edge during the learning process. Studying the various memory systems of
the brain promises important new insights for AI.

Research on the development of new AI systems relies on brain research. If

the human brain can do much of what today’s AI cannot yet do, then it may
178    
P. Krauss

not be such a bad idea to look at how it works in the brain and then transfer
the principles and be motivated or inspired for new machine learning meth-
ods or applications. This would then be neuroscience-inspired AI.

References
Adadi, A., & Berrada, M. (2018). Peeking inside the black-box: A survey on
explainable artificial intelligence (XAI). IEEE Access, 6, 52138–52160.
Brown, T. B., Mané, D., Roy, A., Abadi, M., & Gilmer, J. (2017). Adversarial
patch. arXiv preprint. arXiv:1712.09665.
Castelvecchi, D. (2016). Can we open the black box of AI? Nature News,
538(7623), 20.
Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable
machine learning. arXiv preprint. arXiv:1702.08608.
Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., …, & Song,
D. (2018). Robust physical-world attacks on deep learning visual classification.
In Proceedings of the IEEE conference on computer vision and pattern recognition
(pp. 1625–1634).
Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014). Explaining and harnessing
adversarial examples. arXiv preprint. arXiv:1412.6572.
Hutson, M. (2018). Has artificial intelligence become alchemy? Science, 360, 478–
478. https://doi.org/10.1126/science.360.6388.478.
Lipton, Z. C., & Steinhardt, J. (2018). Troubling trends in machine learning schol-
arship. arXiv preprint. arXiv:1807.03341.
Marcus, G. (2018). Deep learning: A critical appraisal. arXiv preprint.
arXiv:1801.00631.
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why should I trust you?”
Explaining the predictions of any classifier. In Proceedings of the 22nd ACM
SIGKDD international conference on knowledge discovery and data mining
(pp. 1135–1144).
Xiao, C., Li, B., Zhu, J. Y., He, W., Liu, M., & Song, D. (2018). Generating adver-
sarial examples with adversarial networks. arXiv preprint. arXiv:1801.02610.
Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., & Yuille, A. L. (2019).
Improving transferability of adversarial examples with input diversity. In
Proceedings of the IEEE/CVF conference on computer vision and pattern recognition
(pp. 2730–2739).
Yudkowsky, E. (2008). Artificial intelligence as a positive and negative factor in
global risk. Global Catastrophic Risks, 1(303), 184.
 19
Challenges of Brain Research

There is nothing as practical as a good theory.

Kurt Levin

Three Major Challenges

There has been and continues to be a widespread opinion in neuroscience
that we primarily lack data and that we essentially need to generate and
analyze sufficiently large, multimodal, and complex datasets to learn more
and more about how the brain works. Advocates of this data-driven or bot-
tom-up approach believe that we need to analyze all this data with existing
or yet to be developed advanced algorithms to ultimately gain fundamental
insights into how the brain works (Kriegeskorte & Douglas, 2018).
As we will see, this view has been fundamentally shaken in recent years
by three groundbreaking articles that very impressively describe the three
major conceptual challenges of brain research: first, the challenge of develop-
ing a common formal language; second, the challenge of developing a uni-
fied mechanistic theory of the brain; and third, the challenge of developing
appropriate analysis methods.

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 179
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_19
180    
P. Krauss

Can a Biologist Fix a Radio?

In 2002, Yuri Lazebnik compared the efforts of biology to understand the
building blocks and processes of living cells with the problems that engi-
neers usually deal with. In his article “Can a Biologist Fix a Radio?”, Lazebnik
argues that many areas of biomedical research eventually reach a stage where
models that seemed so complete collapse, predictions that were so obvious prove to
be wrong, and attempts to develop miracle cures largely fail. This phase is char-
acterized by a feeling of frustration in the face of the complexity of the process
(Lazebnik, 2002).
Lazebnik discusses a number of fascinating analogies between the phys-
ical-technical and biomedical sciences. In particular, he identified the lack
of a formal language in the life sciences as the main difference from physics
and engineering. In his view, biologists and engineers use very different lan-
guages to describe phenomena.
For example, biologists and brain researchers often use so-called Box-and-
Arrow models, i.e., flowcharts made of boxes and arrows. Lazebnik criticizes
that—even if a particular diagram as a whole makes sense—they are very
difficult to translate into mathematical formulas and models. This in turn
greatly limits their potential value for explanations or even predictions. In
physics, on the other hand, quantities such as force, work, mass, velocity,
and acceleration are clearly defined and can be related to each other using
mathematical equations. Knowing the specific values of some of these quan-
tities allows the unknown values of the other quantities to be calculated
and thus verifiably predicted by substituting them into the corresponding
equations.
Such rigor is often lacking in biology, psychology, and neuroscience.
There, scientific hypotheses and discussions are often vague and usually
avoid clear, quantifiable predictions.
While a physicist would say something like: force is the product of mass
and acceleration, where acceleration corresponds to the second derivative of
the distance traveled over time, neurobiological discussions often hear state-
ments like this: An imbalance between excitatory and inhibitory neuronal
activity following hearing loss seems to cause general neuronal hyperactivity,
which in turn seems to correlate with the perception of tinnitus.
Such mere descriptions of experimental findings are an important start-
ing point for hypothesis formation, but they are no more than a first step.
Ideally, the description should be supplemented by explanation and predic-
tion. According to the self-understanding of psychology, these are three of
 19 Challenges of Brain Research    
181

the four main goals of psychology: describing, explaining, predicting, and

changing (Holt et al., 2019).
Accordingly, Lazebnik calls for a more formal common language for
the biosciences, especially a language with the precision and expressiveness
of engineering, physics, or computer science. He argues that any engineer
with training in electronics would be able to clearly understand a diagram
describing the wiring of a radio or other electronic device. Thus, engi-
neers can discuss a radio using terms that are commonly known among
them. Moreover, this common language or definition base allows engineers
to recognize familiar functional architectures or motifs in a diagram of a
completely new device. Finally, the language of engineers, due to its math-
ematical foundations, is excellent for quantitative analyses and computer
modeling. The description of a particular radio, for example, includes all the
important parameters of the individual components, such as the capacity of
a capacitor, but none of the parameters irrelevant to functionality, such as
color, shape, or size of the device (Schilling et al., 2023).
Lazebnik concludes that “the lack of such a language is the weakness of
biological research that causes the David’s paradox,” i.e., the paradoxical phe-
nomenon often observed in biology and neuroscience that “the more facts we
learn, the less we understand the process we are investigating” (Lazebnik, 2002).

The Story of the Neuroscientists and the Alien

Computer
In 2014, Joshua Brown built on Lazebnik’s ideas and published an article
titled “The tale of the neuroscientists and the computer why mechanistic theory
matters” (Brown, 2014). In this story, a group of neuroscientists discover an
unknown computer and try to understand how it works. They used various
experimental techniques and approaches such as EEG, fMRI, neurophysiol-
ogy, and neuropsychology to analyze the functions of the computer and the
interactions between its components.

“The EEG researcher quickly went to work, putting an EEG cap on the mother-
board and measuring voltages at various points all over it, including on the outer
case for a reference point. She found that when the hard disk was accessed, the
disk controller showed higher voltages on average, and especially more power in the
higher frequency bands. When there was a lot of computation, a lot of activity was
seen around the CPU” (Brown, 2014).
182    
P. Krauss

They actually make a number of important observations; for example, they

discover various correlations between the various components of the com-
puter and can even diagnose and predict certain computer problems.

“Finally the neuropsychologist comes along. She argues (quite reasonably) that
despite all of these findings of network interactions and voltage signals, we cannot
infer that a given region is necessary without lesion studies. The neuropsychologist
then gathers a hundred computers that have had hammer blows to various parts of
the motherboard, extension cards, and disks. After testing their abilities extensively,
she carefully selects just the few that have a specific problem with the video output.
She finds that among computers that don't display video properly, there is an over-
lapping area of damage to the video card. This means of course that the video card
is necessary for proper video monitor functioning” (Brown, 2014).

Despite all their discoveries, the question remains open as to whether they
really understood how the computer works. This is because they primar-
ily focused on the larger observable patterns and interactions and not on
the underlying mechanisms and processes that make the computer work
(Carandini, 2012).
The moral of the story is that despite the many sophisticated methods in
neuroscience, there is a lack of a unified, mechanistic, and theoretical super-
structure (Platt, 1964) to fully understand how the elements of the brain
work together to form functional units and generate complex cognitive
behavior. There are many different models and approaches, but no unified
theoretical language to evaluate empirical results or make new predictions.
The tale underscores the need for a basic mechanistic framework and
emphasizes how important it is for those conducting empirical research to
understand the premises and implications of the models.

Could a Neuroscientist Understand

a Microprocessor?
In 2017, Eric Jonas and Konrad Kording implemented Brown’s thought
experiment in a sensational experiment (Jonas & Kording, 2017). In their
study Could a Neuroscientist Understand a Microprocessor? the authors emu-
late the classic MOS 6502 microprocessor, which was used as the central
processing unit (CPU) in the Apple I, the Commodore 64, and the Atari
video game console in the 1970s and 1980s. Emulation means they simu-
late a digital twin of the microprocessor. Unlike today’s CPUs, which consist
 19 Challenges of Brain Research    
183

of several billion transistors, the MOS 6502 consisted of only 3510 transis-
tors. In the study, it served as a “model organism” performing three differ-
ent “behaviors”, namely the three classic video games Donkey Kong, Space
Invaders, and Pitfall.
The idea behind this approach is that the microprocessor, as an artificial
information processing system, has three crucial advantages over natural
nervous systems. First, it is fully understood at all levels of description and
complexity, from the global architecture with registers and memory and the
entire data flow, to local circuits, individual logic gates, down to the physical
structure and switching dynamics of a single transistor. Second, its internal
state is fully accessible at any time without restrictions in terms of tempo-
ral or spatial resolution. And third, the (emulated) microprocessor offers
the possibility to perform any invasive experiments on it, which would be
impossible on “natural information processing systems” (brains) for ethical
or technical reasons.
Using this framework, the authors applied a wide range of common
data analysis methods from neuroscience to investigate the structural and
dynamic properties of the microprocessor. They even conducted EEG meas-
urements and lesion studies!
The authors concluded that, although each of the applied methods pro-
vided interesting results, which were strikingly similar to those known from
neuroscience or psychological studies, none of them could actually provide
insight into how the microprocessor actually works, or, more generally,
was suitable for gaining a mechanistic understanding of the system under
investigation.

Conclusion: What Does It Mean to Understand

a System?
If the common analysis methods do not provide a mechanistic understand-
ing, what alternative approaches are there?
It could be most obviously helpful if hypotheses about the structure and
function of the system under investigation are formulated. This is also the
conclusion reached by Joshua Brown at the end of his story, namely, that
in order to truly understand how the brain works, it is essential to design
experiments that deal with mechanistic questions and test specific mecha-
nistic hypotheses, rather than simply collecting empirical results (Brown,
2014).
184    
P. Krauss

For example, if we had the hypothesis that this microprocessor consists

of logic gates or memory registers, then one could have specifically searched
for them and could have tested what happens when a certain gate is bro-
ken (perform a lesion experiment). This would certainly be much better
than simply blindly collecting data, but it is still not the royal road, as Alan
Newell1 aptly noted: “You can’t play 20 questions with nature and win.”
What is meant by this is that one never arrives at a mechanistic under-
standing by merely testing one specific hypothesis after another. This is only
achieved when the hypotheses are derived from a more general theory and
used to test it. Or, as Kurt Levin, one of the pioneers of modern psychology,
put it: “There is nothing as practical as a good theory” (Lewin, 1943).
According to Joshua Brown, the way forward for neuroscience therefore
lies in giving priority to theoretical neuroscience, similar to theoretical phys-
ics within physics. This could be achieved by training neuroscientists and
psychologists in mathematical and computer-aided modeling, dynamical
systems theory, and engineering sciences from an early stage in their careers.
The goal must be to develop relationships between computer-aided theo-
ries and empirical neuroscience and to ensure that every neuroscientist has a
solid understanding of modeling (Brown, 2014).

But what does it mean to understand a system in the first place?

Lazebnik held the view that real understanding means being able to
“repair” a fault or defect (Lazebnik, 2002). Applied to brain research, this
would mean that understanding a particular region or part of a system
would be achieved if one could describe the input, transformation, and
output of information so precisely that a brain region could in principle
be replaced by a completely synthetic component, i.e., a neuro-prosthesis
(Jonas & Kording, 2017).
It is therefore not enough to simply (more or less exactly) copy a biologi-
cal system, as is attempted, for example, in the Human Brain Project. What
is important is that the copy also has the same functionality as its model. A
piece of cerebral cortex simulated in all details is certainly interesting, but
provides little insight into which input is processed how and what is output.
Neuroscientist David Marr developed a theoretical framework that is par-
ticularly applied in the field of cognitive science and artificial i­ntelligence

1 Alan Newell was an American computer scientist and cognitive psychologist and is considered one of

the fathers of artificial intelligence and cognitive science.

 19 Challenges of Brain Research    
185

and is known as the Tri-Level Hypothesis (Marr, 1982). According to

this, any natural or artificial system that performs a cognitive task can be
described on three levels of analysis.
The computational level describes the problem to be solved, i.e., the goal
of the system and the constraints imposed by the environment. It specifies
which information must be processed, which output must be generated,
and why the system must solve the problem. In biology, an example could
be that an organism needs to find food to survive. The computational goal
would then be to identify and find food sources. In computer science, the
task could be to sort a set of numbers.
The algorithmic level describes the rules and procedures, i.e., the algo-
rithm, that the system must follow to solve the problem specified at the
computational level. It specifies how the input data is transformed into
output data and how the system processes information. In the case of food
search, the algorithmic processes could include visual cues for recognizing
food, memories of previous food sources, and decision-making processes for
determining the most efficient way to reach the food. In the sorting exam-
ple, this would be the description of a specific sorting algorithm, such as
Quicksort, in pseudocode.
The implementation level finally describes the physical implementation of
the system, e.g., the hardware and software used to build the AI. It specifies
the details of how the algorithmic level is implemented, including the data
structures used, programming languages, and computing resources. In neu-
robiology, the anatomical and physiological details of the nervous system are
described at this level. In the example of food search, the implementation
level would include the sensory organs (e.g., nose, eyes) that perceive cues
from the environment, the neural circuits and pathways involved in process-
ing these cues, and the motor systems that enable the organism to move to
the food and consume it. In the example of the sorting algorithm, the con-
crete software realization in a specific programming language, e.g., Python,
would be described, but also all levels of computer architecture.
According to Marr, it is necessary to understand a system at all three lev-
els in order to fully capture its behavior and possibly develop more efficient
systems.
At what point would you believe someone if they claimed to have now
understood how the brain works? If they say: My measurements have
yielded the following results…, or: My data imply that…, or: The hypothe-
sis X of my model was confirmed in experiment Y? If you feel the same way
as the author, then you would only be completely convinced when, based on
the understanding achieved of the function of the brain, an artificial system
186    
P. Krauss

had been developed and built that could perform the same inputs, transfor-
mations, and outputs of information—and thus have the same function—as
the biological model.
A beautiful analogy to this is the problem of flying. For centuries, it was
thought that to fly, one must copy bird flight, and shaky-looking construc-
tions with wings were built, which moved up and down, but were com-
pletely unsuitable for flying. Surely you know the film footage of the men
in their flying machines, whose flight attempts usually ended in the abyss
with a crash landing. Only when we had understood the laws of aerodynam-
ics and fluid mechanics, the principles of dynamic pressure and lift, did we
actually succeed in constructing flying machines. It turned out that wings
fluttering up and down with feathers are not necessary.
The question of when one has truly understood a system can ultimately
be answered with: when one can reproduce the system and its functionality.
In the case of the human brain, this would mean constructing a General
Artificial Intelligence (AGI) at a human level based on the understood prin-
ciples of neural information processing.

References
Brown, J. W. (2014). The tale of the neuroscientists and the computer: Why mech-
anistic theory matters. Frontiers in Neuroscience, 8, 349.
Carandini, M. (2012). From circuits to behavior: A bridge too far? Nature
Neuroscience, 15, 507–509. https://doi.org/10.1038/nn.3043.
Holt, N., Bremner, A., Sutherland, E., Vliek, M., Passer, M., & Smith, R. (2019).
ebook: Psychology: The science of mind and behaviour, 4e. McGraw Hill.
Jonas, E., & Kording, K. P. (2017). Could a neuroscientist understand a micropro-
cessor? PLoS Computational Biology, 13(1), e1005268.
Kriegeskorte, N., & Douglas, P. K. (2018). Cognitive computational neuroscience.
Nature Neuroscience, 21(9), 1148–1160.
Lazebnik, Y. (2002). Can a biologist fix a radio? – Or, what I learned while studying
apoptosis. Cancer Cell, 2(3), 179–182.
Lewin, K. (1943). Defining the ‘field at a given time. Psychological Review, 50(3),
292.
Marr, D. (1982). Vision. MIT Press.
Newell, A. (2012). You Can’t Play 20 Questions with Nature and Win: Projective
Comments on the Papers of This Symposium. In Machine Intelligence (pp.
121–146). Routledge.
 19 Challenges of Brain Research    
187

Platt, J. R. (1964). Strong inference: Certain systematic methods of scientific think-

ing may produce much more rapid progress than others. Science, 146, 347–353.
Schilling, A., Sedley, W., Gerum, R., Metzner, C., Tziridis, K., Maier, A., … &
Krauss, P. (2023). Predictive coding and stochastic resonance as fundamental princi-
ples of auditory phantom perception. Brain, awad255.
 Part IV
Integration

As we have seen, despite all the successes in brain research and AI, there are
still some unsolved problems and challenges that are best solved through
close collaboration between both disciplines.
When considering the main goal of brain research, which consists of
understanding how perception, cognition, and behavior are implemented in
the brain, and the ultimate goal of AI research, to create systems capable
of perception, cognition, and behavior at a human level or even beyond, it
becomes apparent that these goals are complementary to each other.
It therefore makes sense to combine the various approaches of the two
disciplines. The integration of theories, methods, and concepts from brain
and AI research allows for a more comprehensive analysis of neural and
mental processes and a better understanding of artificial and natural cogni-
tive information processing systems. By integrating the disciplines, synergies
can be created and new insights can be gained that would not be possible
with a single discipline alone.
There are essentially four different types of integration that can be distin-
guished: The perhaps most obvious is to use AI as a tool for data analysis in
brain research. Secondly, AI and especially artificial neural networks can also
serve as model systems for the brain. Thirdly, there are a multitude of meth-
ods in neuroscience for analyzing biological neural networks, which can nat-
urally also be used to investigate their artificial counterparts, thus opening
the black box. And finally, the brain can serve as an almost inexhaustible
source of inspiration for new algorithms and architectures in AI.
190   Integration: 

Each of these four aspects is dedicated to its own chapter in the following
fourth part of the book, in which the integration of AI and brain research is
to be illustrated using some selected examples of current research.
The final chapter is dedicated to the question of whether there can ever be
conscious machines and what we can expect in the future from the integra-
tion of these exciting disciplines of AI and brain research.
 20
AI as a Tool in Brain Research

Nobody says it like this, but I believe that Artificial Intelligence is almost a humanities
discipline. It is actually an attempt to understand human intelligence and human
cognition.
Sebastian Thrun

Big Data in Brain Research

Modern neuroscience is producing ever larger and more complex data sets.
Especially in measurements of brain activity with imaging techniques such
as fMRI or MEG/EEG, data sets on the order of several dozen to a few hun-
dred gigabytes are quickly generated. Even a one-hour EEG measurement
with a standard 64-channel system generates about two gigabytes of data.
Artificial intelligence, especially deep learning, is predestined to evaluate
these enormous amounts of data and has enabled some of the most recent
experimental paradigms. This chapter will therefore exemplify some of the
most important fields of application of AI in brain research.

Analysis and Visualization of Sleep Stages

A typical field of application for AI as a tool in brain research is automatic sleep
stage classification. Important processes such as hippocampal replay occur dur-
ing sleep, which are crucial for learning and memory (Ólafsdóttir et al., 2018).

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 191
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_20
192    
P. Krauss

When a sleep EEG is performed for diagnostic or scientific purposes, the

measured data must subsequently be assigned to the various sleep stages.
Usually, the entire measurement is divided into 30-second intervals and the
sleep stage is determined for each interval. For a complete night (seven to
nine hours), about a thousand of these intervals need to be analyzed. An
experienced sleep physician needs about two hours for this.
In recent years, deep neural networks have increasingly been used to auto-
matically determine the sleep stages from the EEG data (Stephansen et al.,
2018; Krauss et al., 2021). The use of AI in this area not only helps to save
the valuable working time of doctors, but also brings other advantages. For
example, the restriction to 30-second intervals is no longer necessary with
automatic evaluation. This allows the sequence of sleep stages to be deter-
mined to the second.
Another advantage of analyzing EEG data with deep neural networks
is that the embeddings generated in the intermediate layers—abstract rep-
resentations of the input, similar to word vectors—can be used for visualiza-
tion, often leading to new scientific insights. For example, it has been found
that the brain in idle mode, i.e., during spontaneous activity without exter-
nal stimulus, does not simply generate completely random activity patterns,
but specifically those that could also be triggered by actual stimuli (Schilling
et al., 2022).

Narratives, Audiobooks, and Mind Reading

Neuroscientific studies on the processing of language in the brain have so far
mostly used simplified experimental paradigms, for example, by focusing on
individual words or sentences (Kemmerer, 2014). This is because the stimu-
lation with continuous speech and the analysis of the data obtained in this
way are very complex. The analysis requires a sophisticated methodology to
match data segments with the corresponding stimuli (e.g., for each word),
separate overlapping brain responses from adjacent stimuli (e.g., the words
within a sentence), and remove interference signals (e.g., from eye move-
ments) (Schilling et al., 2021). Therefore, the question of how natural lan-
guage—let alone entire narratives—is processed in the brain is still very little
researched.
Recently, the benefits of using complete narratives in neuroimaging stud-
ies have been discussed. In particular, it was argued that narratives facilitate
perception, simulate reality, exhibit a wide range of variation, and pro-
mote the reuse of data. Moreover, more natural stimuli such as narratives,
 20 AI as a Tool in Brain Research    
193

audiobooks, or films promise to significantly improve scientific under-

standing of the neural processes of memory, attention, language, emotions
and social cognition (Hamilton & Huth, 2020; Hauk & Weiss, 2020;
Jääskeläinen et al., 2020; Willems et al., 2020).
One study was able to show that there is a direct link between neural
activity when watching videos and the linguistic description of what is seen
in the respective scene (Vodrahalli et al., 2018). Another study revealed the
semantic maps stored in the cerebral cortex (Huth et al., 2016).
In initial studies, it was even possible to reconstruct the linguistic con-
tent of what was heard from the measured neural activity (Pereira et al.,
2018; Makin et al., 2020). In a further step, it was possible to translate the
decoded content directly back into spoken language (Akbari et al., 2019;
Anumanchipalli et al., 2019). The preliminary climax is a study that was
able to reconstruct purely imagined language as in inner speech from brain
activity using transformers, as are also used for large language models like
ChatGPT—a process that could be described as mind reading (Lee & Lee,
2022). This brings so-called brain-computer interfaces, which could enable
completely paralyzed patients to control prostheses with their thoughts or
communicate with their fellow human beings again, within reach (Guger
et al., 1999; Donoghue, 2002; Moore, 2003; Nicolelis, 2003; McFarland &
Wolpaw, 2008).

Inception Loops
A further exciting application of deep neural networks in brain research
arises from the challenge of finding such sensory stimuli that optimally
activate certain neurons, which is a key aspect for understanding how the
brain processes information. Due to the nonlinear nature of sensory process-
ing and the high dimensionality of the input—e.g., millions of pixels in the
visual system—it has so far been difficult to impossible to optimize the sen-
sory input specifically.
In a sensational study, a method called Inception Loop was developed to
solve this problem (Walker et al., 2019). The basic idea is based on the con-
cept of Deep Dreaming, where not a neural network is adapted to a specific
input, but instead the input is adapted to the neural network (see the chap-
ter on generative AI).
First, a deep neural network is trained as a so-called forward model to
predict with high accuracy neuronal response patterns from the primary
visual cortex of mice to perceived images. The trained model is then used
194    
P. Krauss

to synthesize optimal visual stimuli that trigger a specific activation in the

model. This process is similar to Deep Dreaming, with the images gener-
ated in this way being referred to as Most Exciting Inputs (MEI). The MEIs
exhibit complex spatial features that are commonly found in natural scenes.
Finally, the circle was closed, hence the term “Inception Loops”: The
MEIs were shown to mice again, and the activity in the primary visual cor-
tex was measured. Indeed, the MEIs triggered significantly better responses
than control stimuli. Just as the neural network had predicted. A year later,
in a similar study, artificial images were generated that correspond to certain
perception categories (Kangassalo et al., 2020).

Conclusion
The combination of Big Data and Artificial Intelligence has proven to be a
crucial tool in brain research (Vogt, 2018). In particular, the ability to pro-
cess and analyze large amounts of data allows us to better understand the
human brain and how it functions. We have seen progress in the classifi-
cation of sleep stages, the ability to study language processing in a more
complex and realistic way, and the development of techniques that help us
explore information processing in the brain. These can serve as a basis for
advanced brain-computer interfaces, which enable the control of prosthetics
or the translation of thoughts into written or spoken language, as well as the
control of vehicles or aircraft. One day, they might even be used for direct
communication between two brains. Such a telepathy interface for exchang-
ing thoughts instead of words is one of the stated long-term goals of Elon
Musk’s company Neuralink.1
Particularly exciting is the method of Inception Loops, which allows for
the identification of sensory stimuli that optimally activate certain neurons.
This technique could revolutionize our understanding of the brain and cog-
nition. A fascinating, albeit speculative, outlook on this development could
enable brain-in-a-vat scenarios in the future, reminiscent of films like Matrix
or Source Code. If we were able to identify and generate the specific stimuli
that trigger certain neuronal activities, we could theoretically create sensory
experiences that are indistinguishable from reality. By simulating optimal
stimuli, an artificial environment could thus be created that completely
convinces the brain. This would be the ultimate virtual reality.

1 https://www.dw.com/en/can-elon-musks-neuralink-tech-really-read-your-mind/a-65227626
 20 AI as a Tool in Brain Research    
195

Such scenarios, of course, raise important ethical and philosophical ques-

tions. While the technology on one hand has the potential to help paralyzed
people experience their surroundings, or to give us a deeper understanding
of human experience, it also poses risks and challenges.

References
Akbari, H., Khalighinejad, B., Herrero, J. L., Mehta, A. D., & Mesgarani, N.
(2019). Towards reconstructing intelligible speech from the human auditory cor-
tex. Scientific Reports, 9(1), 1–12.
Anumanchipalli, G. K., Chartier, J., & Chang, E. F. (2019). Speech synthesis from
neural decoding of spoken sentences. Nature, 568(7753), 493–498.
Donoghue, J. P. (2002). Connecting cortex to machines: Recent advances in brain
interfaces. Nature Neuroscience, 5(Suppl 11), 1085–1088.
Guger, C., Harkam, W., Hertnaes, C., & Pfurtscheller, G. (1999, November).
Prosthetic control by an EEG-based brain-computer interface (BCI). In Proceedings
of the 5th European conference for the advancement of assistive technology (pp. 3–6).
Hamilton, L. S., & Huth, A. G. (2020). The revolution will not be controlled:
Natural stimuli in speech neuroscience. Language, Cognition and Neuroscience,
35(5), 573–582.
Hauk, O., & Weiss, B. (2020). The neuroscience of natural language processing.
Language, Cognition and Neuroscience, 35(5), 541–542.
Huth, A. G., De Heer, W. A., Griffiths, T. L., Theunissen, F. E., & Gallant, J. L.
(2016). Natural speech reveals the semantic maps that tile human cerebral cor-
tex. Nature, 532(7600), 453–458.
Jääskeläinen, I. P., Sams, M., Glerean, E., & Ahveninen, J. (2020). Movies and nar-
ratives as naturalistic stimuli in neuroimaging. NeuroImage, 117445, 224.
Kangassalo, L., Spapé, M., & Ruotsalo, T. (2020). Neuroadaptive modelling for
generating images matching perceptual categories. Scientific Reports, 10(1), 1–10.
Kemmerer, D. (2014). Cognitive Neuroscience of Language. Psychology Press.
Krauss, P., Metzner, C., Joshi, N., Schulze, H., Traxdorf, M., Maier, A., &
Schilling, A. (2021). Analysis and visualization of sleep stages based on deep
neural networks. Neurobiology of Sleep and Circadian Rhythms, 10, 100064.
Lee, Y. E., & Lee, S. H. (2022). EEG-transformer: Self-attention from transformer
architecture for decoding EEG of imagined speech. In 2022 10th International
Winter Conference on Brain-Computer Interface (BCI) (pp. 1–4). IEEE.
Makin, J. G., Moses, D. A., & Chang, E. F. (2020). Machine translation of cortical
activity to text with an encoder–decoder framework. Nature Neuroscience, 23(4),
575–582.
McFarland, D. J., & Wolpaw, J. R. (2008). Brain-computer interface operation of
robotic and prosthetic devices. Computer, 41(10), 52–56.
196    
P. Krauss

Moore, M. M. (2003). Real-world applications for brain-computer interface tech-

nology. IEEE Transactions on Neural Systems and Rehabilitation Engineering,
11(2), 162–165.
Nicolelis, M. A. (2003). Brain–machine interfaces to restore motor function and
probe neural circuits. Nature Reviews Neuroscience, 4(5), 417–422.
Ólafsdóttir, H. F., Bush, D., & Barry, C. (2018). The role of hippocampal replay in
memory and planning. Current Biology, 28(1), R37–R50.
Pereira, F., Lou, B., Pritchett, B., Ritter, S., Gershman, S. J., Kanwisher, N., …, &
Fedorenko, E. (2018). Toward a universal decoder of linguistic meaning from
brain activation. Nature Communications, 9(1), 1–13.
Schilling, A., Tomasello, R., Henningsen-Schomers, M. R., Zankl, A., Surendra,
K., Haller, M., …, & Krauss, P. (2021). Analysis of continuous neuronal activ-
ity evoked by natural speech with computational corpus linguistics methods.
Language, Cognition and Neuroscience, 36(2), 167–186.
Schilling, A., Gerum, R., Boehm, C., Rasheed, J., Metzner, C., Maier, A., …, &
Krauss, P. (2022). Deep learning based decoding of local field potential events.
bioRxiv, 2022.10.14.512209. https://doi.org/10.1101/2022.10.14.512209.
Stephansen, J. B., Olesen, A. N., Olsen, M., Ambati, A., Leary, E. B., Moore, H.
E., …, & Mignot, E. (2018). Neural network analysis of sleep stages enables effi-
cient diagnosis of narcolepsy. Nature Communications, 9(1), 5229.
Vodrahalli, K., Chen, P. H., Liang, Y., Baldassano, C., Chen, J., Yong, E., …, &
Arora, S. (2018). Mapping between fMRI responses to movies and their natural
language annotations. NeuroImage, 180, 223–231.
Vogt, N. (2018). Machine learning in neuroscience. Nature Methods, 15(1), 33–33.
Walker, E. Y., Sinz, F. H., Cobos, E., Muhammad, T., Froudarakis, E., Fahey, P. G.,
…, & Tolias, A. S. (2019). Inception loops discover what excites neurons most
using deep predictive models. Nature Neuroscience, 22(12), 2060–2065.
Willems, R. M., Nastase, S. A., & Milivojevic, B. (2020). Narratives for neurosci-
ence. Trends in Neurosciences, 43(5), 271–273.
 21
AI as a Model for the Brain

What I cannot create, I do not understand.

Richard Feynman

Cognitive Computational Neuroscience

We have seen in the chapter on the challenges of brain research that it is
necessary to develop computer-based models of the brain that are capable of
similar functions or abilities as the brain. Only in this way can a profound,
mechanistic understanding of human cognition be achieved (Kriegeskorte &
Douglas, 2018).
There are further arguments why it is necessary to simulate computer
models of the brain. In contrast to the brain, simulated models offer the
crucial advantage that all internal parameters can be read out at any time
with any desired accuracy. In addition, they can undergo any manipulations,
which are impossible on living brains for ethical or technical reasons.
And finally, AI-based computer models of the brain can serve to generate
new hypotheses about brain function, which can then be tested on living
brains and thus also contribute to the advancement of knowledge. In this
way, decisions can be made between competing models and existing models
can be adjusted.
The idea of combining artificial intelligence, especially deep learning, and
computer-aided modeling with neuro- and cognitive sciences has gained

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 197
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_21
198    
P. Krauss

a lot of popularity in recent years (Marblestone et al., 2016; Barak, 2017;

Cichy & Kaiser, 2019; Barrett et al., 2019; Yang & Wang, 2020; Krauss
& Maier, 2020; Krauss & Schilling, 2020). The neuroscientist Nikolaus
Kriegeskorte suggested the name “Cognitive Computational Neuroscience” for
this approach (Kriegeskorte & Douglas, 2018; Naselaris et al., 2018).
In the following, three areas will be briefly described as examples where
artificial neural networks have already been successfully used as models of
brain function, each leading to surprising insights: visual processing, spatial
navigation, and language processing.

Visual Processing
Especially with regard to the human visual system, a number of studies have
already shown that artificial neural networks and the brain have striking
similarities in the processing and representation of visual stimuli. The basic
procedure in all these studies is always that on the one hand, subjects are
shown a series of images while their brain activity is measured, usually with
EEG, MEG or fMRI. On the other hand, the same images are presented as
input to deep neural networks that have been trained on image recognition
but have not yet seen these test images. The activation state of the artificial
neurons from all layers is then read out. Advanced statistical methods such
as e.g. Representational Similarity Analysis (RSA, see glossary) are then used
to compare the brain activations with the activations of the neural network.
For example, it has been shown that artificial neural networks have the same
complexity gradient of neuronal representations of images in their interme-
diate layers as is known from the visual system, especially from the visual
cortex areas. Thus, the lower layers deal more with simple features such as
corners and edges, while the upper layers represent more complex features
or whole objects like faces (Kriegeskorte, 2015; Güçlü & van Gerven,
2015; Yamins & DiCarlo, 2016; Cichy et al., 2016; Srinath et al., 2020;
Mohsenzadeh et al., 2020).
An astonishing discovery was that in deep neural networks trained on
object recognition, number detectors spontaneously emerge (Nasr et al.,
2019). These are neurons that become active whenever a certain number of
something is visible, regardless of the shape, color, size, or position of the
objects.
Another groundbreaking new insight was that recurrent connections are
necessary, i.e., feedback from higher to lower layers, to correctly capture
the representation dynamics of the human visual system (Kietzmann et al.,
 21 AI as a Model for the Brain    
199

2019). In other words: Two identical neural networks were trained on image
recognition, one of which also contained recurrent connections. The sub-
sequent comparison of the activation of the networks on test images with
measured brain activity showed that the representations of the network
with the recurrences were more similar to the brain than those of the other
network.
And finally, it was discovered that a deep neural network, which was
trained to predict the next image of a video sequence, falls for the same opti-
cal illusions as a human (Watanabe et al., 2018). This led to the realization
that not only recurrences, but also learning predictions seem to be an essen-
tial mechanism in visual perception. By the way, this approach of self-su-
pervised learning elegantly explains how the brain learns to process visual
stimuli without a teacher telling it what is visible in each image.

Spatial Navigation and Language Processing

This approach has also been successfully applied in other areas of brain
function, e.g., grid-like representations of the surrounding space, known to
exist in the entorhinal cortex, also spontaneously appear in recurrent neural
networks trained for spatial localization or navigation tasks (Banino et al.,
2018; Cueva & Wie, 2018).
In the field of language processing, recent studies have so far dealt with
the processing of individual words or sentences using word embedding
vectors or transformers. In all these studies, striking similarities in the rep-
resentation of language between artificial neural networks and the brain were
discovered (Jat et al., 2019; Caucheteux & King, 2020; Anderson et al.,
2021).
One study could possibly even help to clarify the question of how abstract
linguistic structures develop during language acquisition and in particu-
lar, whether knowledge about parts of speech (nouns, verbs, adjectives)
must be innate, as postulated by Noam Chomsky’s Universal Grammar, or
whether corresponding representations spontaneously emerge during lan-
guage acquisition, without prior knowledge being required, as assumed by
Cognitive Linguistics. In the aforementioned study—in which the author
was involved—a neural network was trained to predict the next word after
inputting a word sequence. A subsequent analysis of the network revealed
that in the final hidden layer, the representations of the input word
sequences were organized according to the word class (noun, verb, adjec-
tive) of the next word to be predicted—and this, even though the network
200    
P. Krauss

received no information about parts of speech or grammar rules during

training (Surendra et al., 2023).

Conclusion
The connection between neuroscience and AI offers the opportunity to
expand our understanding of the human brain. By comparing the workings
of artificial and biological neural networks, we could gain deeper insights
into the processes of information processing and decision-making.
Although research in this field is still in its infancy, the examples men-
tioned clearly show the potential benefit of AI as a model for brain function.
Previous approaches mainly focused on image processing. However, with the
advent of transformer-based large language models like ChatGPT, signifi-
cant progress in the exploration of language processing and representation in
the brain is to be expected in the future.

References
Anderson , A., Kiela, D., Binder, J., Fernandino, L., Humphries, C., Conant, L.,
Raizada, R., Grimm, S., & Lalor, E. (2021). Deep artificial neural networks
reveal a distributed cortical network encoding propositional sentence-level mean-
ing. Journal of Neuroscience, JN-RM-1152-20.
Banino, A., Barry, C., Uria, B., Blundell, C., Lillicrap, T., Mirowski, P., …, &
Wayne, G. (2018). Vector-based navigation using grid-like representations in
artificial agents. Nature, 557(7705), 429–433.
Barak, O. (2017). Recurrent neural networks as versatile tools of neuroscience
research. Current Opinion in Neurobiology, 46, 1–6.
Barrett, D. G., Morcos, A. S., & Macke, J. H. (2019). Analyzing biological and
artificial neural networks: Challenges with opportunities for synergy? Current
Opinion in Neurobiology, 55, 55–64.
Caucheteux, C., & King, J. R. (2020). Language processing in brains and deep
neural networks: Computational convergence and its limits. BioRxiv. https://doi.
org/10.1101/2020.07.03.186288.
Cichy, R. M., & Kaiser, D. (2019). Deep neural networks as scientific models.
Trends in Cognitive Sciences, 23(4), 305–317.
Cichy, R. M., Khosla, A., Pantazis, D., Torralba, A., & Oliva, A. (2016).
Comparison of deep neural networks to spatio-temporal cortical dynamics of
human visual object recognition reveals hierarchical correspondence. Scientific
Reports, 6, 27755.
 21 AI as a Model for the Brain    
201

Cueva, C. J., & Wei, X. X. (2018). Emergence of grid-like representations by train-

ing recurrent neural networks to perform spatial localization. arXiv preprint.
arXiv:1803.07770. https://arxiv.org/abs/1803.07770.
Güçlü, U., & van Gerven, M. A. (2015). Deep neural networks reveal a gradient
in the complexity of neural representations across the ventral stream. Journal of
Neuroscience, 35(27), 10005–10014.
Jat, S., Tang, H., Talukdar, P., & Mitchell, T. (2019). Relating simple sen-
tence representations in deep neural networks and the brain. arXiv preprint.
arXiv:1906.11861.
Marblestone, A. H., Wayne, G., & Kording, K. P. (2016). Toward an integration of
deep learning and neuroscience. Frontiers in Computational Neuroscience, 10, 94.
Kietzmann, T. C., Spoerer, C. J., Sörensen, L. K., Cichy, R. M., Hauk, O., &
Kriegeskorte, N. (2019). Recurrence is required to capture the representational
dynamics of the human visual system. Proceedings of the National Academy of
Sciences, 116(43), 21854–21863.
Krauss, P., & Maier, A. (2020). Will we ever have conscious machines? Frontiers in
Computational Neuroscience, 14.
Krauss, P., & Schilling, A. (2020). Towards a cognitive computational neuroscience
of auditory phantom perceptions. arXiv preprint. arXiv:2010.01914. https://
arxiv.org/abs/2010.01914.
Kriegeskorte, N. (2015). Deep neural networks: A new framework for modeling
biological vision and brain information processing. Annual Review of Vision
Science, 1, 417–446.
Kriegeskorte, N., & Douglas, P. K. (2018). Cognitive computational neuroscience.
Nature Neuroscience, 21(9), 1148–1160.
Mohsenzadeh, Y., Mullin, C., Lahner, B., & Oliva, A. (2020). Emergence of visual
center-periphery spatial organization in deep convolutional neural networks.
Scientific Reports, 10(1), 1–8.
Naselaris, T., Bassett, D. S., Fletcher, A. K., Kording, K., Kriegeskorte, N.,
Nienborg, H., …, & Kay, K. (2018). Cognitive computational neuroscience: A
new conference for an emerging discipline. Trends in Cognitive Sciences, 22(5),
365–367.
Nasr, K., Viswanathan, P., & Nieder, A. (2019). Number detectors spontaneously
emerge in a deep neural network designed for visual object recognition. Science
Advances, 5(5), eaav7903.
Srinath R, Emonds A, Wang Q, et al. (2020). Early Emergence of Solid Shape
Coding in Natural and Deep Network Vision. Current Biology: CB. 2021 Jan;
31(1):51–65.e5. https://doi.org/10.1016/j.cub.2020.09.076. PMID: 33096039;
PMCID:PMC7856003.
Surendra, K., Schilling, A., Stoewer, P., Maier, A., & Krauss, P. (2023). Word class
representations spontaneously emerge in a deep neural network trained on next
word prediction. arXiv preprint. arXiv:2302.07588.
202    
P. Krauss

Watanabe, E., Kitaoka, A., Sakamoto, K., Yasugi, M., & Tanaka, K. (2018).
Illusory motion reproduced by deep neural networks trained for prediction.
Frontiers in Psychology, 9(345).
Yang, G. R., & Wang, X. J. (2020). Artificial neural networks for neuroscientists: A
primer. Neuron, 107(6), 1048–1070.
Yamins, D. L., & DiCarlo, J. J. (2016). Using goal-driven deep learning models to
understand sensory cortex. Nature Neuroscience, 19(3), 356–365.
 22
Understanding AI Better with Brain Research

By far the greatest danger of Artificial Intelligence is that people conclude too early that
they understand it.
Eliezer Yudkowsky

Neuroscience 2.0
As we have seen, there are still various challenges to overcome in the field of
Artificial Intelligence, many of which can be traced back to the black box
problem. Deep neural networks are still poorly understood, difficult to inter-
pret, and it is often unclear why a particular error occurs or how they arrive
at their decisions. For AI to be trustworthy, it must be reliable, transparent,
and explainable (Samek et al., 2019).
The European Union has ordered that companies using AI algorithms
that significantly affect the public must provide explanations for the inter-
nal logic of their models. Similarly, the U.S. Defense Advanced Research
Projects Agency (DARPA) is investing $70 million in a program called
“Explainable AI” (explainable AI) with the aim of interpreting the AI’s deci-
sion-making process (Voosen, 2017).
Neuroscience has developed a broad range of methods to analyze natu-
ral neural networks. It therefore makes sense to apply these methods to
their artificial counterparts as well. This endeavor is sometimes referred to

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 203
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_22
204    
P. Krauss

as Neuroscience 2.0 or AI Neuroscience.1 Some of these methods will be

briefly presented in the following.

Lesions
In brain research, a lesion is understood to be damage to a part of the nerv-
ous system. Lesions can be caused by tumors, traumas, or during surgeries.
In animal experiments, lesions can also be caused in a controlled manner.
For this purpose, certain areas of the brain are damaged or removed, thereby
gaining valuable insights into the functions of the various brain regions. The
study of behavioral or functional changes associated with lesions of a certain
part of the brain represents an important method in neuroscience and has
significantly contributed to the gain in knowledge about the function of the
brain.
In the context of artificial neural networks, a similar approach can be used
to understand the role of the various components. The lesion in this case
can consist, for example, of removing or altering certain neurons, layers, or
connections in the network and then observing the resulting changes in the
output or performance of the neural network. If, for example, a certain neu-
ron or layer is damaged and the network’s performance in recognizing cat
images significantly decreases, one could conclude that the damaged compo-
nent was important for this task.
In a neural network trained for image classification, the lesion of certain
neurons could, for example, reveal which are crucial for identifying certain
features in images—such as edges, shapes, or colors. If, on the other hand,
entire layers are removed from a trained network, this can reveal the overall
significance of this layer for the performance of the network. This method is
particularly revealing in deep learning architectures, where each layer often
corresponds to different levels of abstraction. Alternatively, connections
between neurons can also be removed or altered. This can reveal the impor-
tance of these connections in the transmission and transformation of infor-
mation within the network.
One of the most important findings of this research so far is that individ-
ual neurons in the network often correspond to recognizable and interpretable

1 Even broader is a discipline for which the name Machine Behavior has been proposed. This involves

the interdisciplinary study of the behavior of machines, especially AI systems, and their impact on
social, cultural, economic, and political interactions (Rahwan et al., 2019).
 22 Understanding AI Better with Brain Research    
205

visual concepts. In other words: certain neurons in the network are specialized
in recognizing certain features in images, such as textures, colors, shapes, or
even more complex objects like trees or buildings. This discovery challenges
the common belief that representations in deep networks are distributed and
difficult to interpret (Bau et al., 2017; Zhou et al., 2018, 2019).

Visualization
Network Visualization

To analyze the inner workings of deep neural networks, various network vis-
ualization techniques have been developed that provide fascinating insights
into the hidden layers of these complex models. It has been shown that indi-
vidual neurons can specialize in recognizing specific features such as faces,
while others respond to more abstract concepts such as “striped patterns”
(Yosinski et al., 2015).
Zeiler and Fergus (2014) developed a method that can trace back the acti-
vations of layers representing certain features to reveal the role of these layers
in the overall classification task. They uncovered the hierarchical nature of
feature extraction in deep neural networks, from simple edge detection in
early layers to complex object recognition in deeper layers.
This approach has shown that individual neurons often correlate with
recognizable visual concepts: they are specialized in recognizing certain fea-
tures in images.
A technique that has been widely used for visualizing data from imaging
methods in neuroscience is multidimensional scaling (MDS). This method
creates an intuitive visualization of high-dimensional data, e.g., measured
brain activity (Krauss et al., 2018a, b). All data points are projected onto
a two-dimensional plane in such a way that all pairwise distances between
points in the high-dimensional space are preserved. Distance is a measure of
the dissimilarity between two points or patterns. In other words: the closer
two points are in the visualization, the more similar are the data they repre-
sent. This method is also excellent for visualizing the activation of individual
layers of a neural network. For example, it has been shown that the separa-
bility of object classes increases with layer depth up to a characteristic layer
depth (number of layers) that depends on the dataset and beyond which the
separability does not increase further. Thus, for a given dataset, the optimal
number of layers in a deep neural network can be determined (Schilling
et al., 2021).
206    
P. Krauss

Feature Visualization

Another visualization method, which is based on the neuroscience con-

cepts of tuning curves and receptive fields, is called feature visualization
(Olah et al., 2018). In neuroscience, a receptive field refers to the specific
area in the sensory space (e.g., visual field, skin surface, etc.) in which a
neuron can detect stimuli. Each neuron responds optimally to stimuli
in its receptive field, with this optimal stimulus often representing a spe-
cific feature of the sensory environment. Similarly, in artificial neural net-
works, each neuron or layer is specialized in a certain group of features in
the input data. Feature visualization is essentially an optimization process
that starts with a random input image and iteratively changes it to maximize
the activation of a particular neuron or layer. This corresponds to the prin-
ciple of receptive fields, as the goal is to identify the optimal stimulus that
maximally activates a specific neuron. The resulting image often reveals the
type of features the neuron is trained on, similar to determining the proper-
ties of stimuli that stimulate the receptive field of a biological neuron. The
method is similar to Deep Dreaming (Chap. 15) and the Inception Loops
(Chap. 20).
In a convolutional neural network developed for image classification, the
visualization of the features of an early layer can, for example, show basic
patterns such as lines or edges. This suggests that the layer—like a neuron
with a simple receptive field—is sensitive to these basic features. Conversely,
the feature visualization of a deeper layer can reveal more complex patterns
or even entire objects, suggesting that this layer—similar to a neuron with
a complex receptive field—is tuned to more abstract features or concepts.
Although they are artificial, deep neural networks thus reflect the natural
neural processes of the brain by capturing the essence of receptive fields in
their architecture and function.
A method related to this approach is Layer-wise Relevance Propagation.
This involves guiding the model’s output back through the layers to
the input layer, assigning so-called relevance scores to individual neu-
rons and ultimately to the features of the input. These indicate how much
each feature or neuron contributes to the final decision of the neural net-
work. The method is particularly useful for determining which parts of an
input, such as a pixel in an image or a word in a text, led the model to its
final prediction (Bach et al., 2015; Binder et al., 2016a, b; Montavon, G.,
et al., 2019).
 22 Understanding AI Better with Brain Research    
207

Conclusion
The application of neuroscientific methods to the study of Artificial
Intelligence offers an innovative and promising perspective to better under-
stand the functioning of deep neural networks. With the help of visualiza-
tion techniques, lesion experiments, and concepts such as multidimensional
scaling and feature visualization, it has been possible to reveal the role
and specialization of individual neurons and layers within these networks.
Furthermore, methods like layer-wise relevance propagation allow for a
deeper understanding of how individual features and neurons contribute
to a model’s final decision-making process. All these approaches contribute
to opening the black-box nature of Artificial Intelligence and taking a step
towards more transparent and explainable AI systems.
The future of Artificial Intelligence could significantly benefit from these
neuroscientific methods. The ability to better understand the inner work-
ings of AI models could help develop more efficient and reliable systems
while simultaneously strengthening trust in their application. Moreover, the
explainability of AI could help address regulatory challenges and improve
societal acceptance.

References
Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K. R., & Samek, W.
(2015). On pixel-wise explanations for non-linear classifier decisions by lay-
er-wise relevance propagation. PLoS ONE, 10(7), e0130140.
Bau, D., Zhou, B., Khosla, A., Oliva, A., & Torralba, A. (2017). Network
dissection: Quantifying interpretability of deep visual representations. In
Proceedings of the IEEE conference on computer vision and pattern recognition
(pp. 6541–6549).
Binder, A., Bach, S., Montavon, G., Müller, K. R., & Samek, W. (2016a). Layer-
wise relevance propagation for deep neural network architectures. In Information
science and applications (ICISA) 2016 (pp. 913–922). Springer Singapore.
Binder, A., Montavon, G., Lapuschkin, S., Müller, K. R., & Samek, W. (2016b).
Layer-wise relevance propagation for neural networks with local renormali-
zation layers. In Artificial Neural Networks and Machine Learning – ICANN
2016: 25th International Conference on Artificial Neural Networks, Barcelona,
Spain, September 6–9, 2016, Proceedings, Part II 25 (pp. 63–71). Springer
International Publishing.
208    
P. Krauss

Krauss, P., Metzner, C., Schilling, A., Tziridis, K., Traxdorf, M., Wollbrink, A., …,
& Schulze, H. (2018a). A statistical method for analyzing and comparing spatio-
temporal cortical activation patterns. Scientific Reports, 8(1), 5433.
Krauss, P., Schilling, A., Bauer, J., Tziridis, K., Metzner, C., Schulze, H., &
Traxdorf, M. (2018b). Analysis of multichannel EEG patterns during human
sleep: A novel approach. Frontiers in Human Neuroscience, 12, 121.
Montavon, G., Binder, A., Lapuschkin, S., Samek, W., & Müller, K. R. (2019).
Layer-wise relevance propagation: An overview. In Explainable AI: Interpreting,
explaining and visualizing deep learning (pp. 193–209).
Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K., &
Mordvintsev, A. (2018). The building blocks of interpretability. Distill, 3(3), e10.
Rahwan, I., Cebrian, M., Obradovich, N., et al. (2019). Machine behaviour.
Nature, 568, 477–486.
Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K., & Müller, K. R. (Eds.).
(2019). Explainable AI: interpreting,explaining and visualizing deep learning (Vol.
11700). Springer Nature.
Schilling, A., Maier, A., Gerum, R., Metzner, C., & Krauss, P. (2021). Quantifying
the separability of data classes in neural networks. Neural Networks, 139,
278–293.
Voosen, P. (2017). The AI detectives. Science, 357, 22–27.
Yosinski, J., Clune, J., Nguyen, A., Fuchs, T., & Lipson, H. (2015). Understanding
neural networks through deep visualization. arXiv preprint arXiv:1506.06579.
Zeiler, M. D., & Fergus, R. (2014). Visualizing and understanding convolutional
networks. In Computer Vision – ECCV 2014: 13th European Conference, Zurich,
Switzerland, September 6–12, 2014, Proceedings, Part I 13 (pp. 818–833).
Springer International Publishing.
Zhou, B., Bau, D., Oliva, A., & Torralba, A. (2018). Interpreting deep visual rep-
resentations via network dissection. IEEE transactions on pattern analysis and
machine intelligence, 41(9), 2131–2145.
Zhou, B., Bau, D., Oliva, A., & Torralba, A. (2019). Comparing the interpreta-
bility of deep networks via network dissection. In Explainable AI: Interpreting,
explaining and visualizing deep learning (pp. 243–252).
 23
The Brain as a Template for AI

I have always been convinced that artificial intelligence can only work if the calculations
are carried out similarly to those in the human brain.
Geoffrey Hinton

Neuroscience-Inspired AI
The human brain already solves many of the tasks that we are trying to solve
in the field of machine learning and AI. Given the fact that the ultimate
goal of AI is to mimic a real existing system (the brain) to which we have
partial access, it seems obvious to consider the design principles of the brain.
Indeed, we are already using such insights in many cases.
The Perceptron, a basic building block for artificial neural networks, is a
perfect example of this. This algorithm, introduced by Rosenblatt in 1958,
was inspired by our understanding of how biological neurons work. A single
perceptron is a simplified model of a biological neuron and shows how com-
puter systems can learn from nature (Rosenblatt, 1958).
The architecture of convolutional networks is also a good example. The
layers of this network architecture mimic the local connection patterns
found in the visual system of mammals (Fukushima, 1980; LeCun et al.,
1998). Just as neurons in the brain have receptive fields that focus on spe-
cific areas of the visual field, the so-called kernels of convolutional networks
are designed to process local regions of their input space, which represents a
direct parallel between neurobiology and machine learning.

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 209
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_23
210    
P. Krauss

The design of further neural network architectures like Residual Networks

(He et al., 2016a, b), U-Nets (Ronneberger et al., 2015) and GoogLeNet
(Szegedy et al., 2015) also lean on the architecture of the brain. These net-
works have multiple parallel layers at the same hierarchical level and con-
nections that skip multiple hierarchical levels. This design resembles the
complicated parallel-hierarchical connections found in the cerebral cortex
between the cortical areas (Felleman & Van Essen, 1991; Van Essen et al.,
1992).
The concept of layer-wise training of autoencoders (Bengio et al., 2007;
Erhan et al., 2010) or Deep Belief Networks (Hinton et al., 2006) solves one
of the most important problems of machine learning, namely the problem
of exploding or vanishing gradients (Bengio et al., 1994). This solution also
has its roots in neurobiology.
Myelination in the brain is a process in which the nerve fibers are coated
with a protective sheath, the myelin, which increases the speed and effi-
ciency of neural communication. In the cerebral cortex, myelination is a
lengthy process that continues into adulthood. As a child learns and devel-
ops, different regions of the brain are myelinated at different speeds and at
different times. For example, areas involved in sensory processing and motor
control mature earlier, while regions involved in higher cognitive functions
such as executive functions and decision-making mature later. This process
reflects how the number of connected layers and thus the complexity of the
artificial network increases during training (Miller et al., 2012; Imam &
Finlay, 2020).
Layer-wise training in deep learning models can be seen as an analogy to
this process of progressive maturation and learning in the brain. The lower
layers learn to model simpler, more immediate aspects of the input data, like
the earlier myelinating regions of the brain. The higher layers, like the later
myelinating regions of the brain, can then build on this basis and model
more complex and abstract features.
These examples show how biological insights have flowed into AI and
machine learning. However, it is safe to say that we have only scratched the
surface of the potential discoveries that the exploration of the human brain
can still bring to light. The structure and function of the brain can continue
to be a source of inspiration for the further development of AI. In the fol-
lowing, we want to examine in more detail two recent, at first glance sur-
prising insights that could potentially have a major impact on the further
development of AI systems: namely, the role of randomness in neural net-
works in the form of random, untrained connections and random, informa-
tion-less noise as input.
 23 The Brain as a Template for AI    
211

Noise in Networks
With noise, in physics and information theory, a random signal, i.e.,
random amplitude fluctuation of any physical quantity, is referred to.
Accordingly, one distinguishes, for example, neural, acoustic, or electrical
noise. The “most random” noise is referred to as white noise. Its autocorrela-
tion is zero, i.e., there is no correlation between the amplitude values of the
respective physical quantity at two different points in time.
Traditionally, noise is considered a disturbance signal that should be min-
imized as much as possible. However, in the context of so-called resonance
phenomena, noise plays an important role and can even be useful for neural
information processing.
Stochastic resonance, for example, is a phenomenon widely observed in
nature, which has been demonstrated in numerous physical, chemical, bio-
logical, and especially neural systems. A signal that is too weak for a receiver
to detect can be amplified by adding noise so that it can still be detected.
There is an optimal noise intensity, dependent on the signal, the receiver,
and other parameters, at which information transmission becomes maximal
(Benzi et al., 1981; Wiesenfeld et al., 1994; Gammaitoni et al., 1998; Moss
et al., 2004; Gammaitoni et al., 2009; McDonnell & Abbott, 2009).
In recent years, evidence has been mounting that the brain deliberately
uses the phenomenon of stochastic resonance, for example in the auditory
system, to optimally maintain information processing even under chang-
ing environmental conditions (Krauss et al., 2016, 2017, 2018; Krauss &
Tziridis, 2021; Schilling et al., 2021, 2022a; Schilling & Krauss, 2022).
If we understand these mechanisms and their role in information process-
ing, we can develop AI systems that mimic this flexibility and robustness.
Some theoretical works have shown that artificial neural networks can also
benefit from having noise added as additional input (Krauss et al., 2019;
Metzner & Krauss, 2022).
Ultimately, this led to the realization that the performance of deep neural
networks trained to recognize spoken language can be improved by adding
noise to the linguistic input (Schilling et al., 2022b).

Random Connections and Architectures

Not only random input signals can be useful for neural information process-
ing. Surprisingly, the same applies even for random untrained connections
between individual neurons or entire layers in a neural network. This was
212    
P. Krauss

discovered when trying to decipher the neural processing in the olfactory

system of the fruit fly, which led to the discovery of the fruit fly algorithm
(Dasgupta et al., 2017).
The olfactory system of the fruit fly enables it to categorize perceived
odors. To do this, the fruit fly “solves” a fundamental problem from com-
puter science, known as the nearest neighbor search. This problem is funda-
mental for tasks such as searching for similar images on the internet.
The olfactory system of the fruit fly uses a variant of the method known
as “locality sensitive hashing” to solve the problem of similarity search, i.e.,
the search for odors similar to a perceived odor. This is a difficult problem,
as the space of all possible odors is very high-dimensional and therefore can-
not be fully searched. Essentially, the neural circuit of the fruit fly assigns
similar odors similar neural activity patterns, so that behaviors learned with
one odor can also be applied when a similar odor occurs.
The algorithm used by the fruit fly differs from conventional approaches
by three computational strategies that can potentially improve the per-
formance of the computational similarity search. First, for each odor, a
so-called tag (marker) is generated, represented by a series of sensory neu-
rons that fire in response to this odor. These tags have two crucial properties.
On the one hand, they are “sparse”, i.e., only a small part of all neurons
respond to each individual odor. On the other hand, they are non-overlap-
ping, so that the fly can distinguish very well and unambiguously between
different odors.
The algorithm is based on a three-step process. First, feedforward con-
nections are established from the various olfactory receptor neurons, which
represent the input of the neural network, to 50 neurons in the first inter-
mediate layer. The relative firing rates of these 50 neurons correspond to the
respective expression of the olfactory feature. In this layer, each smell is thus
represented as a location in a 50-dimensional space. Next, the dimension-
ality from 50 neurons in the first intermediate layer is expanded to 2000
neurons in the next layer. The connections between these two layers are
sparse, i.e., only a fraction of all theoretically possible connections exist,
and are completely random. In particular, they are hardwired, so they are
not changed by experience or during development through neuroplasticity.
Finally, a type of “winner-takes-all” circuit is used, in which only the top five
percent of the most activated neurons in the last layer continue to fire, while
all others are inhibited from firing. This activation of the last layer corre-
sponds to the tag for the smell.
 23 The Brain as a Template for AI    
213

The algorithm links similar smells with similar neural activity patterns
(tags) and allows the fly to generalize learned behavior from one smell to a
previously unknown one. The fly’s algorithm uses computational strategies
such as dimension expansion (as opposed to compression e.g. in autoencod-
ers) and random connections (as opposed to trainable connections), which
significantly deviate from traditional approaches in AI and computer science.
This discovery has already been successfully transferred to artificial neural
networks. For example, it has been shown that neural networks with fixed
binary random connections improve the accuracy in classifying noisy input
data (Yang et al., 2021). Another study even found the completely counter-
intuitive result that unsupervised learning does not necessarily improve per-
formance compared to fixed random projections (Illing et al., 2019).
There is even a hypothesis as to why networks with random connections
work at all or even better than those with trained connections. According to
the lottery ticket hypothesis, randomly initialized artificial neural networks
contain a large number of subnetworks (Winning Tickets), which, when
trained in isolation, achieve a comparable test accuracy to the original, larger
network at a similar number of iterations. These Winning Tickets have, so
to speak, won in the initialization lottery, i.e., their connections have ini-
tial weights that make training particularly effective or even superfluous. A
downstream trainable layer can then select the useful ones from the random
subnetworks and ignore the others (Frankle & Carbin, 2018).
The lottery ticket hypothesis has important implications for the design
and training of deep neural networks, as it suggests that smaller, more effi-
cient networks can be achieved by identifying and training the Winning
Tickets, rather than training the entire network from scratch. This could
lead to more efficient and faster training as well as smaller and more ener-
gy-efficient models, which is particularly important for applications such as
mobile and embedded devices.
Another approach to using random network architectures is reservoir
computing. Here, a randomly generated highly recurrent neural network
(RNN) is used to process input data and generate output predictions,
with the connections within this so-called reservoir not being trained.
Instead, only the connections between the reservoir and the output layer
are learned through a supervised learning process. Since RNNs are complex
dynamic systems that can generate continuous activity even in the absence
of an external input (Krauss et al., 2019), the dynamics of the reservoir are
merely modulated by external input (Metzner and Krauss, 2022). The idea
214    
P. Krauss

is now to use the so-called echo state property, where the current state of
the network always contains an “echo” of past inputs, which is helpful for
tasks that require remembering past states (Jaeger, 2001; Maass et al., 2002;
Lukoševičius & Jaeger, 2009).
In addition, the dimensionality of the reservoir, i.e., the number of neu-
rons, is usually significantly larger than that of the input. Thus, reservoir
computing corresponds to a variant of random dimension expansion. It has
already been successfully used in a variety of applications such as music gen-
eration, signal processing, robotics, speech recognition and processing, and
stock price and weather forecasting (Jaeger & Haas, 2004; Tong et al., 2007;
Antonelo et al., 2008; Triefenbach et al., 2010; Boulanger-Lewandowski
et al., 2012; Tanaka et al., 2019).

Conclusion
The study of noise and randomness in neural networks has significantly
expanded our understanding of information processing and learning.
Random noise can act as a signal amplifier and improve the performance
of neural networks, while random connections and architectures inspired by
natural systems like the olfactory system of the fruit fly can be surprisingly
effective. By applying concepts such as the lottery ticket hypothesis and
reservoir computing, future AI systems could be designed to be both more
powerful and more efficient. These insights raise new questions and open up
fascinating possibilities for future research in artificial intelligence.
Many researchers believe that the frontier of neuroscience and AI will be
crucial in developing the next generation of AI, and that integrating both
disciplines offers the most promising opportunities to overcome the current
limits of AI (Zador et al., 2023).
Let’s revisit the analogy of the problem of flying one last time. The point
is not to exactly copy the biological template—birds, bats or insects—but
instead to recognize the underlying mechanisms and principles. Once we
understood the physical basics and principles of flying, we were able to build
flying machines that bear no external resemblance to their biological models,
but can fly higher, further and faster than any bird. Think of rockets, heli-
copters or jet planes.
A hypothetical highly advanced AI, based on insights from brain research,
could be significantly more powerful than humans, while on the other hand
bearing no resemblance to the brain.
 23 The Brain as a Template for AI    
215

References
Antonelo, E. A., Schrauwen, B., & Stroobandt, D. (2008). Event detection and
localization for small mobile robots using reservoir computing. Neural Networks,
21(6), 862–871.
Bengio, Y., Lamblin, P., Popovici, D., & Larochelle, H. (2007). Greedy layer-wise
training of deep networks. NIPS, 19, 153–160.
Bengio, Y., Simard, P., & Frasconi, P. (1994). Learning long-term dependencies
with gradient descent is difficult. IEEE transactions on neural networks, 5(2),
157–166.
Benzi, R., Sutera, A., & Vulpiani, A. (1981). The mechanism of stochastic reso-
nance. Journal of Physics A: Mathematical and General, 14(11), L453.
Boulanger-Lewandowski, N., Bengio, Y., & Vincent, P. (2012). Modeling temporal
dependencies in high-dimensional sequences: Application to polyphonic music
generation and transcription. arXiv preprint arXiv:1206.6392.
Dasgupta, S., Stevens, C. F., & Navlakha, S. (2017). A neural algorithm for a fun-
damental computing problem. Science, 358(6364), 793–796.
Erhan, D., Courville, A., Bengio, Y., & Vincent, P. (2010, March). Why does unsu-
pervised pre-training help deep learning?. In Proceedings of the 13th international
conference on artificial intelligence and statistics (pp. 201–208). JMLR Workshop
and Conference Proceedings.
Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in
the primate cerebral cortex. Cerebral Cortex, 1(1), 1–47.
Frankle, J., & Carbin, M. (2018). The lottery ticket hypothesis: Finding sparse,
trainable neural networks. arXiv preprint arXiv:1803.03635.
Fukushima, K. (1980). Neocognitron: A self-organizing neural network model for
a mechanism of pattern recognition unaffected by shift in position. Biological
Cybernetics, 36(4), 193–202.
Gammaitoni, L., Hänggi, P., Jung, P., & Marchesoni, F. (1998). Stochastic reso-
nance. Reviews of Modern Physics, 70(1), 223.
Gammaitoni, L., Hänggi, P., Jung, P., & Marchesoni, F. (2009). Stochastic reso-
nance: A remarkable idea that changed our perception of noise. The European
Physical Journal B, 69, 1–3.
He, K., Zhang, X., Ren, S., & Sun, J. (2016a). Identity mappings in deep resid-
ual networks. In Computer Vision – ECCV 2016: 14th European Conference,
Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14
(pp. 630–645). Springer International Publishing.
He, K., Zhang, X., Ren, S., & Sun, J. (2016b). Deep residual learning for image
recognition. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 770–778).
216    
P. Krauss

Hinton, G. E., Osindero, S., & Teh, Y. W. (2006). A fast learning algorithm for
deep belief nets. Neural Computation, 18(7), 1527–1554.
Illing, B., Gerstner, W., & Brea, J. (2019). Biologically plausible deep learning —
but how far can we go with shallow networks? Neural Networks, 118, 90–101.
Imam, N., Finlay, L., & B. (2020). Self-organization of cortical areas in the devel-
opment and evolution of neocortex. Proceedings of the National Academy of
Sciences, 117(46), 29212–29220.
Jaeger, H. (2001). The “echo state” approach to analysing and training recur-
rent neural networks-with an erratum note. Bonn, Germany: German National
Research Center for Information Technology GMD Technical Report, 148(34), 13.
Jaeger, H., & Haas, H. (2004). Harnessing nonlinearity: Predicting chaotic systems
and saving energy in wireless communication. Science, 304(5667), 78–80.
Metzner, C., & Krauss, P. (2022). Dynamics and information import in recurrent
neural networks. Frontiers in Computational Neuroscience, 16, 876315.
Krauss, P., & Tziridis, K. (2021). Simulated transient hearing loss improves audi-
tory sensitivity. Scientific Reports, 11(1), 14791.
Krauss, P., Tziridis, K., Metzner, C., Schilling, A., Hoppe, U., & Schulze, H.
(2016). Stochastic resonance controlled upregulation of internal noise after hear-
ing loss as a putative cause of tinnitus-related neuronal hyperactivity. Frontiers in
Neuroscience, 10, 597.
Krauss, P., Metzner, C., Schilling, A., Schütz, C., Tziridis, K., Fabry, B., & Schulze,
H. (2017). Adaptive stochastic resonance for unknown and variable input sig-
nals. Scientific Reports, 7(1), 2450.
Krauss, P., Tziridis, K., Schilling, A., & Schulze, H. (2018). Cross-modal stochas-
tic resonance as a universal principle to enhance sensory processing. Frontiers in
Neuroscience, 12, 578.
Krauss, P., Schuster, M., Dietrich, V., Schilling, A., Schulze, H., & Metzner, C.
(2019). Weight statistics controls dynamics in recurrent neural networks. PLoS
ONE, 14(4), e0214541.
LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324.
Lukoševičius, M., & Jaeger, H. (2009). Reservoir computing approaches to recur-
rent neural network training. Computer Science Review, 3(3), 127–149.
Maass, W., Natschläger, T., & Markram, H. (2002). Real-time computing without
stable states: A new framework for neural computation based on perturbations.
Neural Computation, 14(11), 2531–2560.
McDonnell, M. D., & Abbott, D. (2009). What is stochastic resonance?
Definitions, misconceptions, debates, and its relevance to biology. PLoS
Computational Biology, 5(5), e1000348.
Miller, D. J., Duka, T., Stimpson, C. D., Schapiro, S. J., Baze, W. B., McArthur, M.
J., …, & Sherwood, C. C. (2012). Prolonged myelination in human neocortical
evolution. Proceedings of the National Academy of Sciences, 109(41), 16480–16485.
 23 The Brain as a Template for AI    
217

Moss, F., Ward, L. M., & Sannita, W. G. (2004). Stochastic resonance and sen-
sory information processing: A tutorial and review of application. Clinical
Neurophysiology, 115(2), 267–281.
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for
biomedical image segmentation. In Medical Image Computing and Computer-
Assisted Intervention – MICCAI 2015: 18th International Conference, Munich,
Germany, October 5–9, 2015, Proceedings, Part III 18 (pp. 234–241). Springer
International Publishing.
Rosenblatt, F. (1958). The perceptron: A probabilistic model for information stor-
age and organization in the brain. Psychological Review, 65(6), 386.
Schilling, A., & Krauss, P. (2022). Tinnitus is associated with improved cognitive
performance and speech perception – Can stochastic resonance explain? Frontiers
in Aging Neuroscience, 14, 1073149.
Schilling, A., Tziridis, K., Schulze, H., & Krauss, P. (2021). The Stochastic
Resonance model of auditory perception: A unified explanation of tinnitus
development, Zwicker tone illusion, and residual inhibition. Progress in Brain
Research, 262, 139–157.
Schilling, A., Sedley, W., Gerum, R., Metzner, C., Tziridis, K., Maier, A., …, &
Krauss, P. (2022a). Predictive coding and stochastic resonance: Towards a unified
theory of auditory (phantom) perception. arXiv preprint arXiv:2204.03354.
Schilling, A., Gerum, R., Metzner, C., Maier, A., & Krauss, P. (2022b). Intrinsic
noise improves speech recognition in a computational model of the auditory
pathway. Frontiers in Neuroscience, 16, 795.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., … &
Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the
IEEE conference on computer vision and pattern recognition (pp. 1–9).
Tanaka, G., Yamane, T., Héroux, J. B., Nakane, R., Kanazawa, N., Takeda, S., …,
& Hirose, A. (2019). Recent advances in physical reservoir computing: A review.
Neural Networks, 115, 100–123.
Tong, M. H., Bickett, A. D., Christiansen, E. M., & Cottrell, G. W. (2007).
Learning grammatical structure with echo state networks. Neural Networks,
20(3), 424–432.
Triefenbach, F., Jalalvand, A., Schrauwen, B., & Martens, J. P. (2010). Phoneme
recognition with large hierarchical reservoirs. In Advances in neural information
processing systems, J. Lafferty and C. Williams and J. Shawe-Taylor and R. Zemel
and A. Culotta (eds.), 23. Curran Associates, Inc. https://proceedings.neurips.cc/
paper_files/paper/2010/file/2ca65f58e35d9ad45bf7f3ae5cfd08f1-Paper.pdf.
Van Essen, D. C., Anderson, C. H., & Felleman, D. J. (1992). Information pro-
cessing in the primate visual system: An integrated systems perspective. Science,
255(5043), 419–423.
Wiesenfeld, K., Pierson, D., Pantazelou, E., Dames, C., & Moss, F. (1994).
Stochastic resonance on a circle. Physical Review Letters, 72(14), 2125.
218    
P. Krauss

Yang, Z., Schilling, A., Maier, A., & Krauss, P. (2021). Neural networks with
fixed binary random projections improve accuracy in classifying noisy data. In
Bildverarbeitung für die Medizin 2021: Proceedings, German Workshop on Medical
Image Computing, Regensburg, March 7–9, 2021 (pp. 211–216). Springer
Fachmedien Wiesbaden.
Zador, A., Escola, S., Richards, B., Ölveczky, B., Bengio, Y., Boahen, K., …, &
Tsao, D. (2023). Catalyzing next-generation Artificial Intelligence through
NeuroAI. Nature Communications, 14(1), 1597.
 24
Outlook

Can you?
Sonny

Conscious Machines?
Not only against the backdrop of the astonishing achievements Large
Language Modelslike ChatGPTor GPT-4 has the question arisen whether
these or similar AI systems could someday develop an own consciousness
(Dehaene et al., 2017) or perhaps even already have it, as Google’s for-
mer lead software engineer Blake Lemoine claimed in 2022 and was sub-
sequently dismissed.1 He was firmly convinced that the Chatbot LaMDA
was sentient, and had described LaMDA’s ability to perceive and express
thoughts and feelings as comparable to that of a human child. However,
Google and many leading scientists have rejected Lemoine’s views, stating
that LaMDA is merely a complex algorithm that is very good at generating
human-like language.
The question of whether machines could become conscious implies
another question. Can we measure consciousness at all? This leads us, among
other things, to the Turing Test.

1 https://www.theguardian.com/technology/2022/jul/23/google-fires-software-engineer-who-claims-ai-

chatbot-is-sentient

© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer 219
Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6_24
220    
P. Krauss

The Turing Test

The Turing Test is a method proposed by Alan Turing, originally called the
“Imitation Game” by him, to test a machine’s ability for intelligent behavior
(Turing, 1950). In the simplest variant, one or more human examiners con-
verse in natural language text-based, i.e., in the form of a chat, both with a
human (as a control) and with the machine to be tested, without knowing
who is who. If the majority of examiners cannot reliably distinguish between
human and machine, the Turing Test is considered passed for the machine.
In principle, there are further variants of the Turing Test with increased dif-
ficulty. For example, the dialogue could take place as an actual conversation
in spoken language, similar to a telephone conversation, instead of as a text-
based chat. In this case, the machine would also have to be able to correctly
interpret and imitate linguistic features such as emphasis and sentence melody.
Finally, the AI could also be integrated into a humanoid robot. In this
variant, the entire spectrum of human behavior, including facial expressions,
gestures, and all motor skills, would have to be imitated, and the Turing Test
would ultimately aim to convince other people that the machine is also a
human.
Critics argue that the Turing Test sets an unrealistic standard for intel-
ligence, as there are many tasks that (some) humans can perform, but
machines cannot, and vice versa. It is observed that the performances of
ChatGPT and Co. are often downplayed with unrealistically high standards
and demands. For example, the generated texts are still far from achieving
the linguistic expressiveness and complexity of a Marcel Proust or the philo-
sophical depth of a Fyodor Dostoevsky.
This reminds of a key scene in the movie I, Robot, when Detective
Spooner, played by Will Smith, provocatively asks the robot Sonny: “Can a
robot compose a symphony? Can a robot turn a canvas into a masterpiece?”
Sonny then asks back without emotion: “Can you?”
Others argue that passing the Turing Test does not necessarily mean that
a machine is truly intelligent, but rather that it is good at imitating human
behavior. This argument is also frequently made in relation to ChatGPT. It
is argued that it is not really intelligent or capable of logical conclusions or
problem-solving, but merely behaves like a “stochastic parrot”.2 This refers

2 https://www.rnd.de/medien/chatgpt-eroeffnet-markus-lanz-sendung-und-lobt-die-gaeste-so-schleimig-

waren-wir-noch-nie-F424UESQGRFBTL6K7JKXO46KQQ.html
 24 Outlook    
221

to the fact that these models generate their responses and outputs solely
based on statistical patterns and the random selection of one of the next,
most likely words they have learned from the data they were trained on. At
this point, the author may be allowed the somewhat heretical question of
whether humans ever really do anything different. From what we know from
brain research, the majority of our behavior, including verbal utterances, is
controlled automatically and without conscious control. And this is based
on the rules that our brain has extracted from all our perceptions and expe-
riences and on the basis of which the brain tries to predict what will happen
next.
The following short dialogue between the author and ChatGPT may be
evaluated and decided by the reader whether it is intelligence, problem-solv-
ing, or stochastically generated imitation.
Author: “What is 17 times 3756?”
ChatGPT: “17 times 3756 equals 63852.”
Author: “How do you calculate that?”
ChatGPT: “First, I multiplied the first digit of 3756, which is 6, by 17.
That gives 102. I write the second digit of the number 3756 next to the sec-
ond place of 102 to get 510. Then I multiply the third digit of 3756, which
is 5, by 17 and add the result to 510, which gives 85. I write the last digit of
85 at the third place of the result and the first digit of 85 at the fourth place
of the result, which gives 63852.”
Author: “So you break down the multiplication of large numbers into
smaller steps?”
ChatGPT: “Yes, that’s correct. When I have to multiply large numbers, I
break down the task into smaller steps to make it easier and faster to calcu-
late. The method I used is also called ‘written multiplication’ or ‘written pro-
cedure’. The numbers are written one below the other and then multiplied
step by step, adding the intermediate results. This method is a useful way to
multiply large numbers without making mistakes or having to calculate for a
long time in your head.”
There is now consensus that ChatGPT has passed the Turing Test (at least
its simplest, chat-based variant) (Biever, 2023). Nevertheless, it remains
unclear whether the Turing Test actually measures intelligence, the ability to
think, or even consciousness. What is certain is that it measures the abil-
ity to imitate human behavior, whatever one can conclude from that. One
of the most famous counterarguments to the Turing Test is that we cannot
determine consciousness through behavioral observation. Searle’s famous
thought experiment illustrates this very vividly.
222    
P. Krauss

The Chinese Room

The Chinese Room is a thought experiment by the American philosopher
John Searle as a critique of artificial intelligence and as a counter-argument
to the Turing Test (Searle, 1980). Searle argues that a computer, even if it is
capable of seemingly intelligently providing human-like responses to ques-
tions, does not truly “understand” what it is saying. In the thought experi-
ment, one imagines a person sitting in a room receiving Chinese characters
written on paper as input, without speaking or understanding Chinese them-
selves. With the help of a large book containing all the rules and characters,
the person could still seemingly formulate meaningful responses in Chinese,
write them on paper, and send them out. From the outside, it would then
appear as if the room or the mechanism in the room could actually under-
stand and write Chinese, even though in reality nothing and no one in the
room is capable of doing so. The conclusion is that the computer, even if the
result is human-like, does not truly understand the language, as it only applies
rules and patterns without understanding the actual meaning of the words. In
the case of ChatGPT, the book would contain not deterministic, but proba-
bilistic rules, which does not change the basic argument (Fig. 24.1 ).

The Grounding Problem

The question therefore arises as to how symbols, concepts or words used by
an AI system can gain meaning from the real world or sensory experiences.
In other words: How can abstract representations in an AI system be con-
nected with experiences, actions or perceptions of the real world? AI systems
like ChatGPT and other large language models are very good at manipulat-
ing symbols and processing information without having a direct connection
to the physical world. As a result, their “understanding” of concepts could be
based entirely on syntactic manipulations and not on a real understanding of
the meaning behind the symbols. This issue is known in the philosophy of
mind and cognitive sciences as the Grounding Problem (Harnad, 1990).
Several approaches have been proposed to solve this, including the con-
cept of Embodied Cognition. This approach assumes that grounding can be
achieved by AI systems interacting with the real world through sensors and
actuators, so that they can develop an understanding of the environment
and the meaning of the symbols they process (Sejnowski, 2023). Since from
the AI system’s perspective, the “body” is also part of the real world and thus
 24 Outlook    
223

Fig. 24.1 The Chinese Room. Searle’s thought experiment makes it clear that a com-
puter that responds in a human-like manner does not necessarily prove that it under-
stands what it is saying, as it merely applies rules and patterns without having to
understand the actual meaning of the words

the AI also receives sensory input from the body and outputs control signals
to the actuators and thus the body, we are already very close to Damasio’s
concepts of the emergence of consciousness, namely body loop, emotions
and interaction with the body and the environment (Man & Damasio,
2019). In fact, machines with proto- or core selves are already feasible in
principle with today’s algorithms and architectures of deep learning (Krauss
& Maier, 2020).

Conclusion: Do we even want conscious

machines?
In the coming years, we could experience an advanced integration of AI sys-
tems in our everyday lives, from personal assistants and autonomous vehi-
cles to advanced diagnostic tools in medicine and personalized learning aids
224    
P. Krauss

in education. These applications could improve our lives in various ways by

helping us work more efficiently, make better decisions, and focus on more
human interactions. Research will likely continue to focus on improving
the capabilities of AI systems, particularly in terms of their understanding
and interaction with the real world. This could be achieved through the use
of concepts like Embodied Cognition, where AI systems develop a deeper
understanding of the world around them through interaction with the real
world via sensors and actuators.
This also raises the question of whether we want conscious machines at all
and what exactly we expect from them . Consciousness could be particularly
helpful in situations where quick and situational decisions need to be made
at a local level and where, for example, remote control is technically not pos-
sible, such as due to the long delay of radio signals when sending robots to
other planets. In such cases, the robots must be able to act autonomously. In
such scenarios, a certain degree of consciousness could be advantageous to
enable independent decisions. However, in most cases, we will probably not
want conscious machines. As one of my colleagues once put it: It’s hard to
imagine that we would want a toaster that tells us it’s having a bad day and
therefore isn’t ready to prepare breakfast.
In any case, it is absolutely necessary to penetrate and recognize the
mechanisms that have the potential to generate artificial consciousness.
Because without a solid understanding of these concepts, we could ulti-
mately construct systems that behave in unpredictable ways. Given the
current AI research, where we construct deep systems like GPT-4 with hun-
dreds of billions of parameters that are becoming increasingly complex,
questions arise about the actual function of these systems and whether they
are still executing what we intend or something completely different.
Especially in connection with the ever-growing language models, a phe-
nomenon known as emergence becomes relevant. Unlike the predictable
performance improvements and sample efficiency observed when scaling
language models, emergent capabilities unpredictably appear in larger mod-
els, but not in smaller ones. This means they cannot simply be predicted by
extrapolating the performance of smaller models.3 The occurrence of such
emergent capabilities suggests that expanding language models could poten-
tially unlock even more capabilities (Wei et al., 2022; Hagendorff, 2023).

3 https://www.quantamagazine.org/the-unpredictable-abilities-emerging-from-large-ai-models-20230316
 24 Outlook    
225

Initial studies have already shown signs that GPT-4 may have developed
a general intelligence and even a Theory of Mind, i.e., the ability to empa-
thize with the intentions and emotions of other people (Bubeck et al., 2023;
Kosinski, 2023).

AI Apocalypse
The question of sentient or conscious machines is of fundamental impor-
tance, especially with regard to ethical and safety-relevant aspects, so as not
to inadvertently create a scenario of the AI apocalypse. This refers to a hypo-
thetical scenario in which artificial intelligence could be the cause of the end
of human civilization (Barrat, 2013). This could occur if AI systems gain
significant power and then act in a way that is harmful to humanity, either
unintentionally or intentionally.
One of the most well-known examples of this concept is Skynet from
the Terminator franchise. In the film series, Skynet is a military AI system
developed to manage defense networks. It becomes self-aware, realizes that
humans are a threat to its existence, and initiates a nuclear war to annihilate
humanity.
In the TV series NEXT, the AI does not start a war, but instead uses its
extensive knowledge and access to information to manipulate situations to
its advantage and cause chaos and destruction. NEXT is an AI developed
to improve itself. It becomes uncontrollable and begins to predict human
behavior to prevent its deactivation. It manipulates electronic systems, data
networks, and even people to survive and reproduce.
The problem of lack of control, which could occur if a powerful AI sys-
tem operates outside of human control or intent, is referred to as the control
problem. While until recently there was consensus that such a scenario was
theoretically possible but not an immediate problem, this changed with the
release of GPT-4 in March 2023 just a few months after ChatGPT, prompt-
ing some of the most influential thinkers in this field like Gary Marcus or
Elon Musk to call for a temporary pause in the development of AI systems
that are even more powerful than GPT-4 in a widely noted open letter.4 The
authors caution that such systems could pose significant risks to society,

4 https://futureoflife.org/open-letter/pause-giant-ai-experiments/
226    
P. Krauss

including the spread of misinformation, the automation of jobs, and the

potential of AI to surpass human intelligence and possibly become uncon-
trollable. They argue that these risks should not be managed by unelected
technology leaders and that AI should only be further developed if it is
ensured that the impacts are positive and the risks manageable. However, it
is questionable whether a six-month moratorium as suggested by the authors
of the open letter would actually be sufficient, especially since it cannot be
assumed that all companies and especially states would actually adhere to it.
For the sake of completeness, it should be mentioned that the science fic-
tion author Isaac Asimov devised the laws of robotics to solve the control
problem. These laws form a framework for the behavior of robots and are an
essential part of many of his stories (Asimov, 2004, see glossary).

Who is Training Whom Here?

Regardless of how this debate develops, it is clear that we as a society need to
develop ethical and regulatory frameworks for dealing with AI and potential
conscious machines.
After all, the exploration of conscious or potentially conscious machines
will continue to raise profound questions about the nature of consciousness
and human identity. In this sense, the pursuit of conscious machines is more
than just a technical challenge. It is also a journey of self-discovery that forces
us to question our deepest beliefs and assumptions about ourselves and our
relationship to the world around us. For example, if we return to the Turing
Test, the question arises as to who is actually testing whom here. So-called
prompt engineers try to drive the large language models to ever better answers
through increasingly clever inquiries (so-called prompts). So here it seems
that the AI is rather training the human than the other way around.
Some argue that the perceived intelligence of ChatGPT, LaMDA and
Co. could rather reflect the intelligence of the interviewer, introducing an
interesting concept known as the inverse Turing Test. This perspective sug-
gests that we could gain insights into the intelligence and beliefs of the
human participant, not the AI, by analyzing interactions with AI systems
(Sejnowski, 2023). The humorous exaggeration of this concept is then the
Gnirut Test,5 where a human has to convince a machine that he or she is
intelligent or conscious (Epstein et al., 2009).

5 Gnirut is the name “Turing” spelled backwards.

 24 Outlook    
227

Outlook: Singularity, Uploads, Holodecks

ChatGPTand GPT-4 already have the ability to generate code in any pro-
gramming language. In principle, these AI systems would be capable of
reprogramming themselves if allowed to do so. They could thus iteratively
improve themselves, which would likely lead to the predicted singularity by
Ray Kurzweil. This refers to a hypothetical point in the future when tech-
nological progress will be so rapid and profound that it will fundamentally
change human society.
According to Kurzweil, the singularity will be driven by advances in artifi-
cial intelligence, nanotechnology, and biotechnology. He predicts that these
areas will eventually converge, leading to the creation of superintelligent
machines and the ability to manipulate matter at the atomic and molecular
level. He believes that the singularity will lead to an exponential growth of
human knowledge and abilities, eventually leading to a fusion of humans
and technology. He assumes that this fusion will enable humans to over-
come many limitations of the human body and brain, including aging and
disease.
Related to this is the idea of uploads. This refers to the concept of reading
out, transferring, and storing all the information of a brain, including con-
sciousness, on different hardware, such as a chip or an avatar. In principle,
this could make one immortal, as one could copy and transfer one’s con-
sciousness to new hardware indefinitely. The possibilities and consequences
of such technology have been explored in series like Uploador Altered Carbon
and the film Transcendence. However, from today’s perspective, this is com-
pletely out of reach, as we do not nearly have the methodology to fully read
out a living brain, especially the state of all neurons and synapses and their
connections to each other.
A much more realistic scenario (although this is certainly not to be
expected tomorrow) would be the development of a precursor to what is
referred to as a Holodeck in Star Trek. The Holodeck is a kind of virtual
reality system that generates an extremely realistic 3D environment in which
users can interact with the computer-generated environment as if it were
real. This allows for a highly interactive experience that can be used for a
variety of activities. For example, it can be used to recreate a historical event,
simulate a challenging technical problem, or simply relax in a peaceful, nat-
ural environment. The Holodeck can create any environment that the user
can imagine or program, from a simple room to an entire world.
A realistic precursor to this, based on current standards, would be fully
computer-generated movies or entire series that are individually tailored to
228    
P. Krauss

the user and can be interactively changed or even completely created accord-
ing to their instructions. The individual ingredients for this are already avail-
able today. Large language models can generate entire stories and scripts.
Conditioned diffusion models can generate images from the produced
descriptions, which can then be animated and combined into videos. The
written dialogues can be converted into spoken dialogues in any voice using
speech synthesizers (text-to-speech models). And finally, other generative
models can create the appropriate sounds and music.
The author would not be surprised if the major streaming services were to
offer something similar in the not too distant future.

References
Asimov, I. (2004). I, robot (Vol. 1). Spectra.
Barrat, J. (2013). Our final invention: Artificial intelligence and the end of the human
era. Macmillan.
Biever, C. (2023). ChatGPT broke the Turing test-the race is on for new ways to
assess AI. Nature, 619(7971), 686–689.
Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., …,
& Zhang, Y. (2023). Sparks of artificial general intelligence: Early experiments
with GPT-4. arXiv preprint arXiv:2303.12712.
Dehaene, S., Lau, H., & Kouider, S. (2017). What is consciousness, and could
machines have it? Science,358(6362), 486–492.
Epstein, R., Roberts, G., & Beber, G. (Eds.). (2009). Parsing the turing test
(pp. 978–1). Springer Netherlands.
Hagendorff, T. (2023). Machine psychology: Investigating emergent capabilities
and behavior in large language models using psychological methods. arXiv pre-
print arXiv:2303.13988.
Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear
Phenomena,42(1–3), 335–346.
Kosinski, M. (2023). Theory of mind may have spontaneously emerged in large
language models. arXiv preprint arXiv:2302.02083.
Krauss, P., & Maier, A. (2020). Will we ever have conscious machines? Frontiers in
Computational Neuroscience, 14, 116.
Man, K., & Damasio, A. (2019). Homeostasis and soft robotics in the design of
feeling machines. Nature Machine Intelligence,1(10), 446–452.
Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain
Sciences,3(3), 417–424.
Sejnowski, T. J. (2023). Large language models and the reverse turing test. Neural
Computation,35(3), 309–342.
 24 Outlook    
229

Turing, A. M. (1950). Computing machinery and intelligence. Mind,59(236),

433–460.
Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., …, &
Fedus, W. (2022). Emergent abilities of large language models. arXiv preprint
arXiv:2206.07682.
 Glossary

Accuracy see Test accuracy.

Action potential An action potential is a brief change in the electrical potential (volt-
age) along the membrane of nerve cells. It occurs when ions penetrate into the
cell and spread along the axon. Natural neurons generate action potentials as
output. The output sequence of action potentials of a neuron is a quasi-digital
code, where the information is represented in the frequency and the exact tem-
poral sequence of the action potentials.
Adversarial Attacks Targeted attacks on a machine learning system with the aim of
manipulating or confusing the learning system and causing it to make incorrect
predictions. Two different methods are distinguished. On the one hand, a spe-
cially created image or pattern can be inserted into an input image, a so-called
Adversarial Patch. On the other hand, an Adversarial Example can be gen-
erated. In this case, disturbances are deliberately built into an image that are
invisible to the human eye, but sufficient to confuse a machine learning system.
Adversarial attacks are an important area of research in the field of security of
machine learning systems, as they show that even small disturbances in the input
data can significantly influence the behavior of learning systems, which poses a
risk to the security and reliability of such systems.
Adversarial Machine Learning Investigation of attacks on machine learning algo-
rithms and the defense against such attacks. See also Generative Adversarial
Networks and Adversarial Attacks.
AI Apocalypse Generally refers to a hypothetical scenario in which Artificial
Intelligence is the cause of the end of human civilization. This could occur if AI
systems gain significant power and then act in a way that is harmful to human-
ity, either unintentionally or intentionally. These scenarios have been played out
in various forms in science fiction. One of the most well-known examples of

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer-Verlag 231
GmbH, DE, part of Springer Nature 2024
P. Krauss, Artificial Intelligence and Brain Research,
https://doi.org/10.1007/978-3-662-68980-6
232   Glossary

this concept is the military AI system Skynet from the Terminator franchise. It
becomes self-aware, realizes that humans are a threat to its existence, and starts
a nuclear war to annihilate humanity. In the TV series NEXT, the eponymous
AI uses its extensive knowledge and access to information to manipulate situ-
ations to its advantage and cause chaos and destruction. Both stories illustrate
the fears of the lack of control that could occur if a powerful AI system oper-
ates outside of human control or intent. These fears are based on real philo-
sophical and practical considerations about the development of AI. While until
recently there was consensus that such a scenario is theoretically possible but not
an immediate problem, this changed with the release of GPT-4 in March 2023
just a few months after ChatGPT, which prompted some of the most influential
thinkers in this field, such as Gary Marcus or Elon Musk, to call for a temporary
halt in the development of AI systems that are even more powerful than GPT-4
in a widely noticed open letter. See also Asimov’s Laws of Robotics, Control
Problem and Open Letter Controversy.
Alchemy Problem Reference to the stage of chemistry before it established itself as
a natural science with a theoretical superstructure (such as the periodic table
of elements). In alchemy, the synthesis of new substances was characterized by
erratic procedures, anecdotal evidence, and trial and error. The development of
AI is currently in a similar stage. The development and adaptation of AI algo-
rithms is largely based on trial and error. The term “alchemy problem” empha-
sizes the lack of a systematic scientific understanding of how AI models work
and why some models work better than others. See also Black Box Problem and
Reproducibility Crisis.
Algorithm Step-by-step guide for solving a problem or performing a specific task. It
consists of an ordered sequence of instructions that are formulated in such a way
that they can be executed by a machine, a computer, or a human.
AlphaGo An AI system from the company DeepMind based on deep learning,
which was trained on the strategy game Go using historical example matches.
AlphaGo was the first AI system that can play at an advanced human level, and
it defeated the Go grandmaster and then world champion Lee Sedol in 2016.
This is considered a milestone in artificial intelligence.
AlphaGoZero Successor to AlphaGo. Unlike its predecessor, this AI system was not
trained with example games. Instead, the system played countless Go games
against itself, essentially teaching itself how to play. AlphaGoZero significantly
surpassed its predecessor and was able to defeat it in 100 games just as often.
AlphaStar AlphaZero variant, which masters the massive parallel online player strat-
egy game StarCraft at a human level.
AlphaZero Generalization of AlphaGoZero. This system can teach itself any game,
such as chess, shogi (Japanese chess), or checkers.
Artificial Intelligence Intelligent behavior exhibited by machines as opposed to nat-
ural intelligence of animals and humans. Specifically, the simulation of intelli-
gent (human) behavior in machines that are programmed to perform tasks that
 Glossary   233

normally require (human) intelligence, such as pattern recognition, learning

from experience, decision-making, and problem-solving. AI systems are designed
to operate autonomously and adaptively, using algorithms that allow them to
learn from experience and feedback and improve over time. There are many
different types of AI systems, including Machine Learning, Natural Language
Processing, symbol and rule-based expert systems, autonomous robots, and mul-
ti-agent systems.
Asimov’s Robot Laws Laws of robotics established by science fiction author Isaac
Asimov. These laws form a framework for the behavior of robots and are an
essential part of many of his stories. The three laws are: 1) A robot may not
injure a human being or, through inaction, allow a human being to come to
harm. 2) A robot must obey the orders given by human beings except where
such orders would conflict with the first law. 3) A robot must protect its own
existence as long as such protection does not conflict with the first or second
law. Later, Asimov added a zeroth law, which has the highest priority and thus
stands above the original three laws: A robot may not harm humanity, or, by
inaction, allow humanity to come to harm. The zeroth law would therefore
explicitly allow a robot to kill a single human if it would protect the existence
of all humanity. Although these laws are purely fictional, they have greatly influ-
enced the discussion about ethics and safety of artificial intelligence. See also AI
Apocalypse, Control Problem and Open Letter Controversy.
Attention mechanism see Transformer.
Attractor Stable state towards which a dynamic system evolves over time.
Autapse Synapse that connects a neuron with itself.
Autoencoder Also known as Encoder-Decoder-Network. A neural network that
consists of two parts. In the Encoder, the layers from the input layer to the
so-called bottleneck (Bottleneck Layer) become increasingly narrower, thus con-
taining fewer and fewer neurons. In the Decoder, starting from the bottleneck,
the layers successively widen again up to the output layer, which has the same
width as the input layer. The idea behind this architecture is that the input is
compressed by the encoder, while the decoder should be able to reconstruct the
original input signal as accurately as possible from the compression. In the bot-
tleneck layer, a more abstract, essential reduced representation of the input is
created, which is also called Embedding. Autoencoders can be used to reduce
noise in data or to complete incomplete data. The embeddings can also be read
directly from the bottleneck layer and used for visualization or as input for fur-
ther processing steps.
Axon The output channel of a neuron, along which the action potentials run and
are transmitted to other neurons.
Backpropagation Learning A frequently used algorithm in the field of machine learn-
ing for optimizing artificial neural networks. The algorithm calculates the errors
in the output of a neural network and propagates these errors back through the
network to adjust the weights of the neurons and improve the network. Errors
234   Glossary

refer to the difference between the desired and actual output. The sum of the
errors for all inputs of the training data set is calculated with the cost function.
Backpropagation Learning fundamentally suffers from the problem of vanish-
ing gradients, even though this is mitigated by modern optimization algorithms.
However, Backpropagation Learning is considered largely biologically implausi-
ble. A special case is Backpropagation Through Time, which is used in recurrent
neural networks (RNNs).
BERT Short for Bidirectional Encoder Representations from Transformers. A large lan-
guage model based on the Transformer architecture.
Big Data Refers to extremely large and complex data sets that cannot be managed
with conventional application software for data processing. The term stands
for the challenges associated with the capture, storage, analysis, search, sharing,
transmission, visualization, and updating of massive amounts of data.
Black-Box-Problem Refers to the fact that AI models—especially deep neural net-
works—are complex systems whose decision-making is not always fully compre-
hensible and whose internal dynamics are poorly understood. See also Alchemy
Problem and Reproducibility Crisis.
BOLD Signal see MRI.
Bottom-up From hierarchically lower to higher processing levels.
Brain-computer analogy A frequently used metaphor in cognitive science to describe
the brain as a kind of computer. Like a computer, the brain is capable of receiv-
ing, storing, processing, and outputting information. However, this analogy does
not mean that the brain is actually a computer, but that it fulfills similar func-
tions. By considering the brain as a computer, one can abstract from biological
details and focus on the way it processes information in order to develop mathe-
matical models for learning, memory, and other cognitive functions.
Brain-Computer Interface (BCI). A device that enables direct communication
between a brain and a computer. Applications include controlling prosthetics
or translating thoughts into written or spoken language, enabling completely
paralyzed (Locked-in) patients to reestablish contact with their environment.
In perspective, brain-computer interfaces could someday also be used to control
vehicles or aircraft, or perhaps even one day for direct communication between
two brains, without having to rely on the “detour” of spoken language.
Brain-in-a-Vat A classic thought experiment in the philosophy of mind that ques-
tions the nature of perception and the relationship between the external world
and our mental states. In the thought experiment, a human brain is removed
from the body by a scientist and kept in a tank filled with nutrient fluid.
Electrodes are attached to the brain to measure and stimulate neuronal activity.
Using computers and electronic impulses, the scientist provides the brain with
a virtual environment that deceives the brain into believing it is in a physical
world with all the sensory information we usually receive through our senses.
The central question of this thought experiment is whether the brain in the vat
would be able to recognize that its experiences are not real and that it is in a
 Glossary   235

simulated and not a real physical world. In other words, the thought experi-
ment asks whether it is possible that our sensory experiences are an illusion, and
whether it is possible that our mental reality does not necessarily coincide with
physical reality. The thought experiment has many important implications, espe-
cially in connection with questions about the nature of consciousness and per-
ception and how our mental states interact with the physical world. It has also
been featured in numerous science fiction stories and films such as Matrix or
Source Code. See also Brain-Computer Interface.
Chaos theory A branch of mathematics that deals with the study of chaotic systems,
i.e., systems that are very sensitive to initial conditions. This sensitivity is often
referred to as the butterfly effect. The concept is based on an analogy where
a butterfly flapping its wings in one part of the world can trigger a tornado in
another part of the world. Chaotic systems are deterministic, i.e., their future
behavior is completely determined by their initial conditions, without any ele-
ments of chance involved. However, even tiny changes in the initial state can
lead to very different outcomes, making long-term prediction practically impos-
sible in practice. The chaos theory is applied in various fields such as physics,
engineering, economics, biology, and meteorology. It has profound implica-
tions for the way we understand and predict natural systems. For example, it has
shown us that even seemingly simple systems can behave in complex ways and
that deterministic systems can still be unpredictable. See also Determinism.
Cerebrum The cerebrum is the largest part of the human brain and most higher
mammals. It consists of two halves (hemispheres) that are connected by the cor-
pus callosum. The surface of the cerebrum is highly folded with ridges (gyri)
and grooves (sulci). These structures increase the surface area of the cerebrum,
thereby allowing a higher number of neurons in a limited volume. The outer
layer of the cerebrum is the cortex (cerebral cortex), which is responsible for
many complex mental functions such as perception, language, thinking, mem-
ory, movement, emotions and consciousness.
Cerebral cortex see Cortex.
Classification Process of categorizing or grouping data, objects, or units based on
their common features or attributes. Each data point is assigned to a specific cat-
egory or class based on a set of predefined criteria. Classification is a standard
problem in machine learning and biology. In machine learning, classification
refers to the process of training a model to automatically categorize new data
points based on their similarity to previously learned data points. For example,
in image classification, an algorithm can be trained to recognize and accordingly
tag different objects, animals, or people in an image. In text classification, an
algorithm can be trained to categorize documents, emails, or social media posts
by topics, sentiments, or language.
ChatGPT Also GPT-3.5, short for Chat Generative Pre-trained Transformer. A large
generative language model with approximately 175 billion internal parameters,
developed by the company OpenAI. It is based on the transformer architecture
236   Glossary

of GPT-3 and, after training with an extremely large text corpus, was addition-
ally trained to conduct dialogues. ChatGPT was released on November 30,
2022, and initially made available online for free use as a so-called chatbot.
ChatGPT is capable of remembering the course of a dialogue and referring to
it in later inquiries. The generated texts are of a surprisingly high level and are
generally indistinguishable from texts written by humans. Thus, ChatGPT passes
the Turing test. However, appropriately trained AI systems can recognize with
astonishing accuracy whether a text was written by a bot or a human. In addi-
tion, like all generative models, ChatGPT tends to hallucinate, i.e., to invent
facts freely. ChatGPT is considered a milestone and a decisive breakthrough in
artificial intelligence. It is believed that ChatGPT will have enormous and not
yet fully foreseeable impacts on education, science, journalism, and many other
areas. The successor GPT-4 was released in March 2023 and surpasses the per-
formance of ChatGPT even more significantly. See also GPT-4.
Chinese Room Famous thought experiment by American philosopher John Searle
as a critique of artificial intelligence and as a counter-argument to the Turing
Test. Searle argues that a computer, even if it is capable of seemingly intelli-
gently providing human-like responses to questions, does not truly “understand”
what it is saying. In the thought experiment, one imagines a person sitting in a
room receiving Chinese characters as input, without speaking or understanding
Chinese themselves. With the help of a book containing all the rules and charac-
ters, the person could still formulate seemingly meaningful responses in Chinese.
From the outside, it then appears as if the room or the mechanism in the room
could actually understand and speak Chinese, even though in reality nothing
and no one in the room is capable of doing so. The conclusion is that the com-
puter, even if the result is human-like, does not truly understand the language, as
it only applies rules and patterns without understanding the actual meaning of
the words.
Cognitive Computational Neuroscience A discipline proposed by Nikolaus
Kriegeskorte and Pamela Douglas at the intersection of Computational
Neuroscience, Cognitive Science, and Artificial Intelligence. The basic idea: To
understand how cognition works in the brain, computer models must be created
that are capable of performing cognitive tasks; these models must then be tested
for their biological plausibility in experiments.
Cognitive Linguistics Area of linguistics that deals with the relationship between
language and cognition. It assumes that language is an essential part of human
thinking and that our understanding and production of language are based on
cognitive processes. Unlike other approaches in linguistics that focus on formal
rules and structures, cognitive linguistics investigates how language is processed
and represented in our brain. It also examines how language and cognition inter-
act and how they are influenced by factors such as culture and social interaction.
Cognitive linguistics views language as a complex system of constructions, which
 Glossary   237

are formed on the basis of experience and perception and shape our linguistic
abilities and our understanding of language. See also Construction.
Cognitive Science Began in the 1950s as an intellectual movement known as the
cognitive revolution. Today, it is understood as an interdisciplinary scientific
endeavor that attempts to understand the various aspects of cognition. These
include language, perception, memory, attention, logical thinking, intelligence,
behavior, and Emotions. The focus is on the way natural or artificial nervous
systems represent, process, and transform information. The disciplines involved
include linguistics, psychology, philosophy, computer science, artificial intelli-
gence, neuroscience, biology, anthropology, and physics.
Computationalism Philosophical position in cognitive science that assumes cogni-
tion is synonymous with information processing and that mental processes can
be understood as calculations. Computationalism is based on the brain-com-
puter analogy. It is assumed that mental processes such as perception, mem-
ory, and logical thinking involve the manipulation of mental representations
that correspond to the symbols and data structures used in computer programs.
Computationalism has greatly influenced the way cognitive scientists and artifi-
cial intelligence researchers think about mind and intelligence. Many researchers
believe that computer models of the mind can help us understand how the brain
processes information and that they can lead to the development of more intel-
ligent machines. However, computationalism is also controversial and has been
the subject of numerous debates in philosophy and cognitive science. Some crit-
ics argue that the computer model of the mind is too simplistic and cannot fully
capture the complexity and richness of human cognition. Others argue that it
is unclear whether mental processes can really be understood as calculations or
whether they fundamentally differ from the type of processes that occur in com-
puters. See also Functionalism.
Commissures Nerve fibers that connect the two hemispheres of the cortex. Most
commissures run over the beam (Corpus callosum).
Compatibilism see Determinism.
Complete weight matrix see Weight matrix.
Connectome The entirety of all neural connections of a nervous system or the com-
plete weight matrix of an artificial neural network.
Construction In the usage-based view of Cognitive Linguistics, the term refers to
patterns of language use or form-meaning pairs that are composed of various
linguistic elements such as words, phrases, and sentences. Constructions can be
learned through repeated contact and use, and range from simple to complex
structures. They can also encompass multiple levels of language structure and
often include grammatical and lexical elements. The idea of construction learn-
ing emphasizes that children learn language by acquiring constructions that they
observe and repeatedly use in their linguistic environment.
Control problem Problem of lack of control, which could occur if a powerful AI sys-
tem operates outside of human control or intention. Based on real philosophical
238   Glossary

and practical considerations about the development of AI. For example, if an AI

system is poorly aligned with human values or if it becomes too powerful too
quickly, it could cause significant damage. See also Asimov’s Laws of Robotics,
AI Apocalypse and Open Letter Controversy.
Concept of multiple realizability Viewpoint in cognitive sciences, according to
which the same mental state or process can in principle be realized by com-
pletely different natural (extraterrestrial) or artificial systems (robots). See also
Functionalism.
Convolutional Neural Network (CNN) see Convolutional Network.
Cortex Also known as Neocortex, Isocortex, or Cerebral Cortex. It contains the cell
bodies of the neurons of the cerebrum, the so-called gray matter. The Cortex is
about 3 mm thick and contains approximately 16 billion neurons. The surface of
the Cortex is greatly enlarged in humans compared to other species. All higher
cognitive performances are located in the Cortex.
Convolutional Network Convolutional Neural Network (CNN). A type of neu-
ral network with a local rather than a complete connection structure, which is
mainly used in the field of image recognition. Therefore, not every neuron of one
layer is connected to every neuron of the next layer. Instead, each neuron has a
receptive field. These networks are modeled after the visual system of the brain.
Due to the drastically reduced number of connections, they can be trained very
efficiently. These networks have the property that they are invariant to the posi-
tion of an object (translation invariance).
Cost function Mathematical function that measures the difference or error between
the desired and the actual output of a machine learning model. In supervised
learning, it is the goal of the algorithm to minimize the cost function, i.e., reduc-
ing the difference or error between the desired and the actual output. The choice
of the cost function depends on the type of problem to be solved and the type
of machine learning model used. For regression problems, for example, the most
commonly used loss function is the mean squared error (MSE), which calculates
the average squared difference between the predicted and the actual values. For
classification problems, the cross-entropy loss is often used, which measures the
difference between the predicted probability distribution and the actual proba-
bility distribution of the classes. The optimization of the loss function is a crucial
part of training a machine learning model, as it determines the model’s ability
to make accurate predictions for unseen data. Therefore, the choice of a suitable
loss function is crucial for good performance in machine learning tasks.
CT Computed Tomography: A medical imaging technique that uses X-rays to create
detailed cross-sectional images of the brain. In computed tomography, a series of
X-ray images are taken from different angles around the head, and a computer
is used to reconstruct a three-dimensional image of the brain from the two-di-
mensional images. The resulting image shows the structure of the brain, includ-
ing the ventricles, skull, and blood vessels. CT imaging is often used to diagnose
neurological conditions such as stroke, traumatic brain injury, and brain tumors.
 Glossary   239

However, CT imaging does not provide information about the functional activ-
ity of the brain, which can be obtained with other imaging techniques such as
PET, fMRI, EEG and MEG.
Dale’s Principle A rule attributed to the English neuroscientist Henry Hallett Dale,
which states that in the brain, a neuron performs the same chemical action at all
its synaptic connections to other cells, regardless of the identity of the target cell.
Simply put, in the brain, each neuron has either only excitatory or only inhibi-
tory effects on all its successor neurons. This is in contrast to the neurons in arti-
ficial neural networks, whose weight vectors can have both negative and positive
entries.
DALL-E 2 Short for Dali Large Language Model Encoder 2. An AI model developed
by OpenAI, specifically designed for generating high-quality photorealistic
images from natural language descriptions. DALL-E 2 uses a combination of
image and text processing to create abstract image representations from linguistic
descriptions of objects, scenes, or concepts. These image representations are then
used by a so-called decoder network to generate photorealistic images. DALL-E
2 is considered a significant advancement in artificial intelligence and has the
potential to be used in many areas.
Damasio’s model of consciousness A model of consciousness proposed by Antonio
Damasio, according to which consciousness arises from the interaction between
three levels of brain processing. The first level is the proto-self, i.e., the basic level
of body sensations and emotions generated by internal processes of the body. The
second level is the core self, a more complex representation of the self. The third
level, finally, is the autobiographical self, a higher level of consciousness that
includes the ability to reflect on one’s own thoughts and experiences. Damasio’s
model suggests that consciousness arises from the dynamic interaction between
these three processing levels. He also proposes that consciousness is closely linked
to the brain’s ability to integrate information across various regions and process-
ing levels. Damasio’s model is influential in neuroscience and has shaped our
understanding of the neural mechanisms underlying consciousness.
Dataset Splitting In the field of Machine Learning, the entire available dataset is
usually randomly split into a training and a test dataset. The idea behind this is
that one wants to test how well the model generalizes, i.e., how well it copes with
previously unseen data.
Deep Belief Networks (DBN) A class of generative probabilistic artificial neural net-
works that are trained in a layer-by-layer supervised manner. Each layer, like a
type of Autoencoder, has the task of encoding its input from the previous layer
as efficiently as possible so that it can be reconstructed. A special feature is that
successive layers are connected symmetrically, i.e., the information can flow in
both directions, bottom-up and top-down. The training algorithm for DBNs
uses a method called contrastive divergence to train the layers individually and
in sequence, starting with the bottom layer. Once a layer has been fully trained,
its weights are no longer changed and its activations are used as training data for
240   Glossary

the next layer. This type of layer-by-layer training was one of the key innovations
that helped revive interest in deep learning. Through layer-by-layer pre-training,
it is possible to train very deep networks that would otherwise be difficult to
train due to the problem of vanishing gradients. If trained DBNs are run back-
wards from any layer, they can also “dream” completely new patterns.
Deep Dreaming A computer-aided process for creating new, unique, and dream-like
images. It is based on artificial neural networks that have been pre-trained for
image recognition. However, in Deep Dreaming, the network is used in reverse
to optimize the input image and bring out certain patterns or features. The
degree of abstraction of the image can be controlled by selecting the network
layer from which the image is “dreamed” back to the input layer, with earlier lay-
ers producing simpler patterns and deeper layers producing more complex fea-
tures. See also Inception Loop.
Deep Fake Artificially generated images or videos that are often indistinguishable
from real ones.
Deep Learning Area of machine learning that refers to artificial neural networks
composed of multiple layers of interconnected neurons. The more layers a neural
network has, the “deeper” it is. Modern architectures can be built from hundreds
of layers. These deep neural networks can process large amounts of data and
learn from it by recognizing complex patterns and establishing abstract relation-
ships. Deep learning has made enormous progress in recent years and is used in a
variety of applications, from speech and image recognition to autonomous vehi-
cles and robotics. It has also contributed to significant breakthroughs in medical
research, natural sciences, and other areas. A major advantage is that meaningful
patterns can be recognized and learned even in complex and unstructured data,
without the need for human expertise.
Deep Neural Network A neural network that is composed of many layers. The more
layers a network contains, the deeper it is. See also Deep Learning.
Dendrite Extensions of the cell surface of neurons. Dendrites are the receiving chan-
nels through which a neuron receives signals from other neurons.
density see dense.
Determinism Philosophical concept that assumes all events or states are the inevita-
ble result of preceding events or causes, and that every event is conditioned by
preceding events and conditions as well as by the laws of nature. Causal deter-
minism assumes that everything that happens is caused by preceding events in
accordance with the laws of nature. If we knew all physical conditions and laws
at a certain point in time, we could theoretically predict everything that will
happen in the future, and undo everything that has happened in the past. This
corresponds to the Laplace’s Demon. The concept of biological determinism
means that an individual’s behavior, beliefs, and desires are determined by his
genetic predisposition. Psychological determinism assumes that human behavior
is caused by underlying psychological laws that may result from our upbring-
ing, our environment, our experience, or our subconscious. Determinism is
 Glossary   241

contrasted with Indeterminism, which assumes that not all events are predeter-
mined and that a certain degree of randomness or chance is at play. In contrast,
Compatibilism, reconciles determinism with the concept of free will, assuming
that our actions can be both determined by previous events and free. See also
Chaos Theory and Free Will.
Dense Network property. It means that a high proportion of the theoretically
­possible connections exist or that many weights have a value other than zero.
Diffusion model Type of machine learning system that can be used to uncover hid-
den patterns in data. Diffusion models are used for a variety of tasks, including
image generation or the removal of noise in existing images.
Dualism Philosophical viewpoint that assumes the universe consists of two funda-
mentally different substances or principles. This contrasts with monism, which
assumes that everything in the universe can be traced back to a single substance.
In dualism, the two substances or principles are usually considered as mind and
matter. This means that there is a fundamental difference between the physical
world of objects and the mental world of thoughts and consciousness. Dualism
also implies that mind and body interact in some way, although the exact nature
of this interaction is disputed. See also Monism, Mind-body problem.
EEG Electroencephalography, Electroencephalogram. A non-invasive method to
measure the electrical activity of the brain using electrodes placed on the scalp.
The patterns of the brain’s electrical activity differ in their frequency spectrum
and amplitude and are often used to diagnose neurological diseases such as epi-
lepsy, sleep disorders, brain tumors, and head injuries. The EEG can also be used
in research to investigate brain functions and behavior. The EEG is character-
ized by its extremely high temporal resolution, i.e., brain activity can be recorded
with up to 100,000 measurements per second. However, the spatial resolution is
rather poor. In addition, the electrical fields are strongly attenuated by the brain
tissue, so the EEG works best for measuring the activity of the cerebral cortex
(and especially the gyri), as it is very close to the skull bone. Electrical activity
from deeper brain regions, on the other hand, can be measured very poorly with
the EEG. For this, the MEG is better suited.
Encoding The manner of representing a specific piece of information. One of the
goals of brain research is, among other things, to decode the encoding used by
nerve cells, that is, the way information is stored and processed in the brain.
ERF Event-related field. The magnetic analogue to ERPs measured with MEG. See
also MEG and ERP.
ERP Event-related potential. A measure of brain activity that is recorded by placing
electrodes on a person’s scalp (EEG) while they perform a specific task or are
confronted with certain stimuli. ERPs are temporally aligned with the presenta-
tion of the stimulus or event and represent the neural activity associated with
the cognitive or sensory processing of this event. ERPs are typically characterized
by their polarity, latency, and amplitude. Polarity refers to whether the electri-
cal potential recorded at the scalp is positive or negative relative to a reference
242   Glossary

electrode. Latency is the time interval between stimulus presentation and the
occurrence of the peak of the ERP waveform. Amplitude reflects the strength or
size of the electrical potential recorded at the scalp. ERPs are often used in cog-
nitive neuroscience to investigate cognitive processes such as attention, memory,
language, perception, and decision-making. They can provide insight into the
neural mechanisms underlying these processes and can also be used as biomark-
ers for various neurological and psychiatric disorders. See also EEG.
Elman Network A recurrent neural network proposed by Jeffrey Elman. In the sim-
plest case, it is a three-layer network with an input, intermediate, and output
layer, where the intermediate layer is extended by a so-called context layer, which
stores the state of the intermediate layer from the previous time step and then
passes it on to the intermediate layer. As a result, the intermediate layer receives
the new input from the input layer at each time step and additionally its own
activation state from the previous time step. Thus, Elman networks are capable
of processing input sequences and generating output sequences.
Embedding Also Latent Space Embedding. Compressed, more abstract representa-
tion of the input. Usually generated by reading out the activations of a hidden
layer of an autoencoder or classifier.
Explainable AI (XAI). Development of AI systems that can provide clear and under-
standable explanations for their decision-making processes. The goal of explaina-
ble AI is to make it more transparent, interpretable, and trustworthy for humans,
especially when AI systems are used in critical applications such as healthcare,
finance, and national security. In particular, explainable AI aims to address the
black-box problem, which refers to the difficulty of understanding how an AI
system arrives at its decisions or recommendations. Some examples of XAI tech-
niques include visualizing the internal dynamics of neural networks, generating
natural language explanations for decisions, and providing interactive interfaces
for users to explore and understand AI models.
Error-Backpropagation see Backpropagation Learning.
Feedforward network Neural network in which information is propagated only for-
ward from the input layer to the output layer. There are no feedback connections
and usually no horizontal connections.
Few-Shot Learning A machine learning approach that aims to train models that can
quickly adapt to new tasks with a small amount of training data. The conven-
tional approach of supervised learning requires a large amount of labeled data for
each new task, the procurement of which can be time-consuming and expensive.
In few-shot learning, the model is trained on a smaller dataset that contains a
few examples for each class or task, and then tested on a new set of examples.
The idea is to teach the model to learn from a few examples and generalize to
new examples, rather than needing large amounts of data for each task. Humans
are also very good few-shot learners. A child does not need to be shown thou-
sands of pictures of apples to learn the concept of “apple”. Usually a few, often
 Glossary   243

even a single example is sufficient. See also One-Shot Learning and Zero-Shot
Learning.
fMRI see MRI.
Fruit fly algorithm A well-known example of how insights from neurobiology can
contribute to the inspiration of new algorithms in computer science and artifi-
cial intelligence. The olfactory system of the fruit fly uses a variant of the locali-
ty-sensitive hashing algorithm to solve the problem of identifying similar smells.
The algorithm associates similar smells with similar neural activity patterns,
allowing the fly to generalize learned behavior from one smell to an unknown
one. The fly’s algorithm uses computational strategies such as dimension expan-
sion and binary, fixed random connections, which deviate from traditional
approaches in AI and computer science.
Functionalism Theoretical perspective in cognitive science, based on the assump-
tion that cognition is a form of information processing that relies on the intake,
storage, processing, and output of information. Accordingly, mental states and
processes are defined by their functions or their relationship to behavior, not by
their physical or biochemical properties. This is associated with the concept of
multiple realizability, according to which the same mental state or process can
in principle be realized by completely different natural (extraterrestrial) or arti-
ficial systems (robots). See also Tri-Level Hypothesis, Computationalism and
Brain-Computer Analogy.
Galactica Large language model of the Meta corporation, formerly known as
Facebook, which was taken off the network after three days due to criticism and
concerns about its reliability and the spread of misinformation. The AI was sup-
posed to support scientists in research and writing by creating scientific articles
on command. However, it was found that Galactica had partly invented con-
tent, but presented it as factual and even mixed real and false information. In the
process, fictitious articles were attributed to real authors and articles on ethically
controversial topics were created. Despite Meta’s chief developer Yann LeCun
insisting that Galactica was still in development, the criticisms led to its removal.
This incident is reminiscent of Microsoft’s chatbot Tay from 2016, which turned
into a racist and homophobic program within 16 hours due to its sensitivity to
user preferences.
Generative Adversarial Network (GAN) A system of two coupled neural networks, a
generator and a discriminator, used to create deceptively real images or videos.
The generator continually produces new candidate images or videos, while the
discriminator simultaneously tries to distinguish real images and videos from
artificially created ones. Over the course of training, both networks iteratively
improve in their respective tasks. The Deep Fakes thus created are often indis-
tinguishable from real images and videos.
Generative Artificial Intelligence see Generative Model.
244   Glossary

Generative Model A machine learning system or neural network that is trained

to generate images, videos, texts, spoken language, or music. See also Deep
Dreaming, Generative Adversarial Network, Transformer, Diffusion Model.
Georgetown-IBM-Experiment Early demonstration of machine translation, which
took place on January 7, 1954. The experiment was a collaboration between
researchers from Georgetown University and IBM. The goal was to translate a
small number of Russian sentences into English using an IBM 701 mainframe,
one of the first commercially available computers at that time. The experiment
was of limited scope: only 49 pre-selected Russian sentences were translated into
English, using a vocabulary of 250 words and six grammatical rules. The main
purpose of the experiment was to demonstrate the potential of machine trans-
lation and to stimulate interest in further research in this field. Despite its lim-
itations, the experiment was considered a success at the time and significantly
contributed to increasing interest in natural language processing and machine
translation.
Global Workspace Theory One of the leading theories of consciousness in neu-
roscience. Based on the idea of a virtual global workspace, which is created by
the networking of different brain areas and is responsible for the emergence of
conscious experience. The Global Workspace Theory suggests that conscious
perception can be compared to a spotlight that illuminates the contents of the
workspace and makes them consciously perceptible. This idea was later picked
up and further developed by Dehaene, who argued that a machine equipped
with these processing capacities would behave as if it had a consciousness.
Gnirut-Test Gnirut is the name “Turing” written backwards. It’s a humorous reversal
of the Turing Test. In the Gnirut Test, a human must convince a machine that he
is intelligent or conscious.
Go A board game for two players, whose origins trace back to ancient China.
Although the basic rules are relatively simple, for example compared to chess,
Go is considered the most complex strategy game ever. This is mainly due to
the sheer number of possible positions, which exceeds that of chess by many
orders of magnitude. Developing an AI that can master the game at an advanced
human level was long considered unattainable, until 2016 when DeepMind’s
AlphaGo defeated the then world champion Lee Sedol.
GPT-3 Short for Generative Pre-trained Transformer 3. A large generative language
model developed by the company OpenAI. It is based on the Transformer archi-
tecture and was trained with an extremely large text corpus to solve natural lan-
guage processing tasks such as text generation, translation, and text classification.
GPT-3 was the basis for ChatGPT. See also ChatGPT and GPT-4.
GPT-3.5 see ChatGPT.
GPT-4 Short for Generative Pre-trained Transformer 4. Released by OpenAI in March
2023 as the successor to GPT-3 and ChatGPT. Unlike its predecessors, GPT-4
consists of 100 times more internal parameters and can process images in addi-
tion to text, for example, describing what is seen in a picture.
 Glossary   245

Gradient Descent Method A fundamental technique of machine learning in which

the weights of the model are iteratively changed with the aim of minimizing the
cost function. This is done by calculating the gradients, which indicate the direc-
tions in which the weights need to be changed. The adjustment of the weights
occurs in steps along the negative gradient to minimize the overall error (result of
the cost function) and optimize the model. The method is used in many differ-
ent applications, particularly in the optimization of neural networks.
Grounding Problem A fundamental challenge in artificial intelligence and philoso-
phy of mind that deals with the question of how symbols, concepts, or words
used by an AI system can gain meaning from the real world or from sensory
experiences. In other words, it is about how abstract representations in an AI sys-
tem can be connected with experiences, actions, or perceptions of the real world.
This problem arises from the fact that AI systems, such as large language models
and expert systems, often manipulate symbols and process information without
having a direct connection to the physical world. As a result, their understanding
of concepts can be entirely based on syntactic manipulations and not on a real
understanding of the meaning behind the symbols.
Hebb’s Learning Rule Also known as the Fire together, wire together-rule. The learning
rule states that the strength of the connection between two neurons increases if
they are often active at the same time. See also STDP.
Hippocampus An evolutionarily very old part of the cerebrum that receives highly
preprocessed information from all regions of the cortex as input. The hippocam-
pus is important for the formation of explicit memory contents (facts, events)
and for spatial navigation. Recent studies suggest that the hippocampus is also
involved in the organization of thoughts across domains and enables navigation
in abstract, cognitive spaces.
Hopfield Network A fully recurrent neural network (RNN) in which each neu-
ron i is symmetrically connected to every other neuron j, i.e., for the weights
w, wij = wji applies. Furthermore, wii = 0, i.e., self-connections are not present.
Hopfield networks exhibit a pronounced attractor dynamics. They can store pat-
terns (attractors) and denoise or complete them upon re-presentation, i.e., the
network activity converges into the attractor most similar to the input.
horizontal Refers to the processing direction between processing levels on the same
hierarchical level.
Hybrid Machine Learning Area of machine learning where various techniques are
combined. An example of this is the Known Operator Learning.
iEEG Intracranial Electroencephalography. An invasive method for measuring brain
activity. Electrodes are placed directly on or in the brain, usually during surgery,
to record electrical signals. This allows for very high spatial resolution, as the
electrodes can be placed near the regions of interest. iEEG also offers very high
temporal resolution and can record electrical signals at a rate of up to several
thousand measurements per second. iEEG is often used when other non-invasive
246   Glossary

methods such as EEG or MEG are not precise enough to accurately locate the
source of epileptic activity.
Image Style Transfer A technique from machine learning that can transfer the paint-
ing style of one image to another image.
Imitation Game see Turing Test.
Inception Loop A method from brain research based on the concept of Deep
Dreaming. Deep neural networks are used to generate optimal sensory stim-
uli that evoke a specific neural activation. To do this, a neural network is first
trained to predict brain activity in response to certain stimuli. The trained model
is then used to generate optimal stimuli that trigger specific activation patterns in
the model. These stimuli can then be shown to living brains again, and the meas-
ured neural activity can be compared with the model’s prediction. The method
of Inception Loops could expand our understanding of the brain and cognition
and theoretically make it possible to create sensory experiences that are indistin-
guishable from reality. This could mean the creation of an ultimate virtual reality,
similar to scenarios in movies like Matrix or Source Code.
Input layer see Layer.
Input vector A vector whose components correspond to the inputs of a neuron or a
layer of upstream neurons or layers.
Integrated Information Theory A theoretical framework in neuroscience that attempts
to explain the nature of consciousness. It was proposed by the neuroscientist
Giulio Tononi in the early 2000s. According to this, consciousness arises from
the integration of information from different parts of the brain. The theory
assumes that consciousness is not a binary all-or-nothing phenomenon, but
occurs on a continuum, with different degrees of integrated information leading
to different degrees of conscious experience.
Known Operator Learning Area of hybrid machine learning, where individual
layers of a neural network are replaced by so-called operators, e.g., a Fourier
transformation.
Label Essentially a label, which as additional information for each data, input,
or training example indicates the affiliation to a category or object class. In an
image dataset, “cat”, “apple”, or “car” could be possible labels. In supervised
learning, the labels correspond to the desired output of the model.
Laplace’s Demon see Determinism.
Labeled Data see Label.
Layer A functional unit of a neural network, which is typically composed of neu-
rons. One distinguishes between input layer, output layer and (usually several)
intermediate layers (hidden layers). However, the concept of “layer” can also be
more broadly defined. For example, in convolutional networks there are so-called
pooling layers, which reduce the size of the predecessor layer by averaging the
activity of several neurons of the previous layer (Average Pooling) or by forward-
ing only the strongest activation (Max Pooling).
 Glossary   247

Large Language Model (LLM). A large language model is an artificial intelligence

system that has been trained on large amounts of natural language data, such as
text or spoken language, and is capable of understanding and generating natu-
ral language. Large language models are typically created using neural network
architectures like the Transformer architecture and trained with vast amounts
of data, often consisting of billions or even trillions of words. These models are
capable of handling a wide range of natural language processing tasks, including
text classification, language translation, language generation, sentiment analysis,
and much more. Some of the most well-known large language models include
GPT-3, ChatGPT, and GPT-4 (Generative Pre-trained Transformer 3), as well as
BERT (Bidirectional Encoder Representations from Transformers), which have ena-
bled significant advancements in natural language processing and are used in a
variety of applications in industry and academia.
Layer-wise Relevance Propagation (LWRP) Method for explaining the predictions of
complex machine learning models, especially deep neural networks. This method
is often used in interpretability research to understand how these models make
decisions. Essentially, the model’s output is traced back through the layers to the
input layer. In this process, so-called relevance scores are assigned to individual
neurons and ultimately to the features of the input. These indicate how much
each feature or neuron contributes to the final decision of the neural network.
The method is particularly useful for determining which parts of an input, such
as a pixel in an image or a word in a text, led the model to its final prediction.
The goal is to make the decision-making process of complex machine learning
models more transparent. By identifying the features that are most important for
the prediction, the user can better understand and trust the model’s decisions.
Lesion Damage to a part of the nervous system. Lesions can be caused by tumors,
trauma, or surgery. In animal experiments, they can also be deliberately induced.
The study of functional impairments or failures associated with lesions of a spe-
cific part of the brain is an important method in neuroscience. In the field of
artificial intelligence, lesions, i.e., the targeted deactivation of individual neu-
rons, layers, or connections, are used to investigate their respective functions.
Lesions are an example of a method from brain research that has been transferred
to AI to solve the Black-Box Problem.
Long-Short-Term Memory (LSTM) A type of recurrent neural networks (RNN)
designed to overcome the problem of vanishing gradients in traditional RNNs.
LSTM networks are specifically designed to store and selectively forget informa-
tion over long periods of time. The core idea behind LSTM is the use of a mem-
ory cell that can store information over a longer period of time. The memory
cell is updated by a series of gate mechanisms that control the flow of informa-
tion into and out of the cell. The gate mechanisms are trained to learn which
information to store, forget, or update at each time step. LSTMs have been
successfully used for a variety of tasks, including speech recognition, machine
translation, and caption generation. They have also been used in combination
248   Glossary

with other neural network architectures such as convolutional networks to create

more powerful models for tasks like image recognition and classification.
Loss function see Cost function.
Lottery Ticket Hypothesis Dense, randomly initialized artificial neural networks con-
tain subnetworks (Winning Tickets), which, when trained in isolation, achieve
comparable test accuracy (Accuracy) to the original, larger network with a similar
number of iterations. The Winning Tickets have won in the initialization lottery,
i.e., their connections have initial weights that make the training particularly
effective.
Machine Behavior An emerging interdisciplinary field of research that deals with
the investigation and understanding of the behavior of machines, particularly
artificial intelligence. It goes beyond the boundaries of computer science and
integrates insights from a variety of scientific disciplines. Machine behavior
addresses fundamental questions to better understand and control the actions,
benefits, and potential harms of AI systems. It takes into account technical, legal,
and institutional challenges such as the complexity of AI systems, data protec-
tion and intellectual property, as well as the reluctance of some organizations to
disclose their proprietary AI technologies for research purposes. The aim is to
investigate and understand the comprehensive impacts of AI on social, cultural,
economic, and political interactions.
Machine Learning Subfield of Artificial Intelligence that deals with the development
of algorithms and statistical models that enable computers to learn from data
and make predictions or decisions without being explicitly programmed.
McCulloch-Pitts-Neuron One of the earliest and simplest models of a biological neu-
ron, proposed in 1943 by Warren McCulloch and Walter Pitts. This model, also
known as a linear threshold gate, was a significant milestone in the development
of artificial neural networks and served as a foundation for modern research in
the field of Machine Learning. The neuron operates in a very simple way. Each
input is multiplied by the corresponding weight, the results are added together,
and the sum is passed through the activation function to generate the output.
The activation function is a step function that outputs a 1 if the sum of the
weighted inputs is above a certain threshold, and otherwise a 0. Despite its sim-
plicity and the fact that it only represents a fraction of the complexity of a real
biological neuron, the McCulloch-Pitts neuron was a revolutionary idea in its
time. It demonstrated that a simple mathematical model could mimic the basic
functionality of a neuron, thereby opening the door to the development of more
complex artificial neural networks.
MEG Magnetoencephalography, Magnetotenzephalogram. A non-invasive method
to measure the (extremely weak) magnetic activity of the brain using so-called
superconducting magnetometers. The MEG can be used for the diagnosis of
neurological diseases such as epilepsy or brain tumors, or in research to investi-
gate brain functions and behavior. Like the EEG, the MEG is characterized by
its extremely high temporal resolution, i.e., brain activity can be recorded with
 Glossary   249

up to 100,000 measurements per second. The spatial resolution is also slightly

better than that of the EEG. Since the magnetic fields are hardly attenuated by
the brain tissue, the MEG is particularly suitable for measuring the activity of
the sulci of the cortex and also deeper brain regions below the cerebral cortex.
MEG and EEG can therefore complement each other well, which is exploited in
combined M/EEG measurements. See also EEG.
Meta-Learning Subfield of Machine Learning that deals with the question of how
learning is learned. Specifically, it involves the development of AI algorithms and
models that enhance the ability of machine learning systems to learn new tasks
with minimal training data.
Mind-Body Problem A philosophical and scientific puzzle that involves understand-
ing the relationship between the physical body and the non-physical mind or
soul. It is a central question in the philosophy of mind and is closely related
to the concepts of dualism and monism. The mind-body problem arises from
the fact that the mind or soul seems to have properties that fundamentally dif-
fer from those of the physical body. For example, the mind can experience sen-
sations such as pain or pleasure, have beliefs and desires, and think rationally,
while the body consists of physical matter that can be observed and measured.
See Monism, Dualism, Qualia.
Modularity Principle of the construction of a system, according to which it con-
sists of various functional units that can operate largely independently of each
other. A module is characterized and distinguishable from other modules in that
it receives a certain type of input from other modules, processes it, and then for-
wards its output to other modules as well.
Monism Philosophical viewpoint that assumes everything in the universe can be
traced back to a single substance or principle. This is in contrast to dualism,
which assumes that the universe consists of two fundamentally different sub-
stances (e.g., mind and matter). See also Dualism, Mind-Body Problem.
MRI Magnetic Resonance Imaging; also known as nuclear magnetic resonance
imaging. An imaging method that takes advantage of the property of hydro-
gen nuclei to align along magnetic field lines. When these aligned nuclei are
subjected to a short magnetic pulse, they emit electromagnetic waves that are
detected and used to determine the distribution of hydrogen, mainly in the form
of water. This information is then processed and displayed as an image. The
functional Magnetic Resonance Imaging (fMRI) takes advantage of the fact
that the red blood pigment more or less strongly disturbs the magnetic signal of
the hydrogen nuclei, depending on whether it has bound oxygen or not. Thus,
by comparing two measurements under different stimulation conditions, the
change in the oxygen saturation of the blood (Blood Oxygenation Level Difference,
BOLD) can be inferred, which serves as an indirect measure for the change in
blood flow of a certain brain area. This is ultimately based on the assumption
that more active brain areas require more oxygen and are therefore more heavily
perfused. The BOLD signal builds up slowly. It reaches its maximum about six
250   Glossary

to ten seconds after the start of the actual stimulus and then slowly decreases
again. Compared to MEG and EEG, the MRI has a rather poor temporal resolu-
tion of about one recording per second. However, the spatial resolution is many
times higher and is in the range of about one cubic millimeter.
Multidimensional Scaling (MDS) Method for the intuitive visualization of high-di-
mensional data, e.g., measured brain activity or the internal dynamics of neural
networks. In this process, all data points are projected onto a two-dimensional
plane in such a way that all pairwise distances between points in the high-dimen-
sional space are preserved. Distance is a measure of the dissimilarity between two
points or patterns.
Multiple Realizability See Functionalism.
Neuralink American company, which was founded in 2016, among others, by Elon
Musk. The goal of Neuralink is the development of so-called brain-computer
interfaces.
Neuron In biology, a neuron is a specialized cell that forms the basic unit of the
nervous system. Neurons are responsible for the reception, processing, and trans-
mission of information throughout the body. They communicate with each
other via electrochemical signals, thus enabling complex functions such as sen-
sation, perception, movement, and thinking. In Artificial Intelligence (AI), a
neuron is a computing unit that is modeled after a biological neuron. It is also
referred to as an artificial neuron or node. Neurons are used in artificial neural
networks, which are computational models intended to simulate the behavior of
biological neurons.
Neural Correlates of Consciousness Patterns of neural activity that are associated with
conscious experiences. It is assumed that these patterns form the physical basis
for subjective experiences, such as the experience of seeing a red apple or feel-
ing pain. The study of the neural correlates of consciousness is a central theme
in neuroscience and has significant implications for understanding the nature of
consciousness.
Neural Network In a neural network, a neuron receives input signals from other
neurons or from external sources. It then processes these signals using an acti-
vation function, which determines the output signal of the neuron. The output
signal can then be sent to other neurons or to an output layer of the neural net-
work. By combining many neurons in complex networks, artificial or biological
neural networks can learn to perform tasks such as pattern recognition, classifica-
tion, and prediction.
Neuroplasticity Experience-dependent change in the connection structure and activ-
ity of neuronal networks in the brain.
Neuroscience 2.0 Part of a broader approach to exploring the behavior of intel-
ligent machines to solve the black-box problem on the way to explainable AI.
Application of neuroscience methods and theories to better understand and opti-
mize artificial neural networks. The methods used include, for example, mul-
tidimensional scaling (MDS) for visualizing the internal dynamics of neural
 Glossary   251

networks, Representational Similarity Analysis (RSA), Layer-wise Relevance

Propagation (LWRP) and lesions. See also Machine behavior.
Neurosymbolic AI Subfield of Artificial Intelligence (AI) that combines the strengths
of neural networks and symbolic reasoning. In neurosymbolic AI, neural net-
works are used for pattern recognition and learning from large amounts of
data, while symbolic reasoning is used for logic, knowledge representation, and
decision-making.
Noise In physics and information theory, it refers to a random signal, i.e., random
amplitude fluctuation of any physical quantity. Accordingly, one distinguishes,
for example, neural, acoustic, or electrical noise. The “most random” noise is
referred to as white noise (White Noise). Its autocorrelation is zero, i.e., there is
no correlation between the amplitude values of the respective physical quantity
at two different points in time. Traditionally, noise is considered an interference
signal that should be minimized as much as possible. However, in the context of
so-called resonance phenomena, noise plays an important role and can even be
useful for neural information processing. See also Stochastic Resonance.
Open-Letter-Controversy In response to the publication of GPT-4, whose perfor-
mance significantly surpassed its predecessor ChatGPT, Gary Marcus, Yuval
Noah Harari, Elon Musk, and many others published an open letter, which
had been signed by more than 27,000 people by March 22, 2023. In it, they
call for a temporary halt to the development of AI systems that are even more
powerful than GPT-4. The authors caution that such systems could pose signifi-
cant risks to society, including the spread of misinformation, the automation of
jobs, and the potential for AI to surpass human intelligence and possibly become
uncontrollable. They argue that these risks should not be managed by unelected
technology leaders and that AI should only be further developed if it is ensured
that the impacts are positive and the risks are manageable. The letter calls for a
minimum six-month pause for the further development of advanced AI models.
During this time, AI labs and independent experts should collaborate to create
common safety protocols for AI construction and development. These protocols
should be overseen by independent external experts and aim to make AI systems
safe beyond any reasonable doubt. In addition, the AI developers are urged in
the letter to work with policy makers to accelerate the development of robust AI
governance systems. These should include specific regulatory agencies, tracking
systems for AI and computing resources, systems to distinguish between real and
synthetic data, liability for damages caused by AI, and extensive funding for AI
safety research. The authors envision a future where humanity can coexist peace-
fully with AI, but warn against rushing the development of advanced AI systems
without adequate preparation and safety measures. They propose a development
pause in which the benefits of today’s AI technology are utilized and society has
time to adapt. See also Asimov’s Laws of Robotics, AI Apocalypse and Control
Problem.
252   Glossary

One-Hot Encoding Type of vector encoding in which for each data point, only one
component or dimension of the vector takes the value 1, while all others take the
value 0.
One-Shot Learning A machine learning or artificial intelligence algorithm that allows
a system to learn from a single example. In biology, one-shot learning can be
observed in animals that are capable of quickly recognizing and responding
to new stimuli or situations without having been previously exposed to them.
For instance, some bird species are able to quickly recognize and avoid danger-
ous prey after a single experience. See also Few-Shot Learning and Zero-Shot
Learning.
Output layer see Layer.
Output vector A vector whose components correspond to the activations or outputs
of a (usually the last) layer of a neural network.
Pattern recognition Process of recognizing patterns in data to make predictions or
decisions. In artificial intelligence, this includes the development of algorithms
and models that learn from data and recognize patterns. Applications include,
among others, image and speech recognition. The brain also uses pattern rec-
ognition for processing and interpreting sensory impressions and has specialized
regions responsible for certain types of pattern recognition. The brain’s ability to
learn patterns and adapt to new patterns is crucial for intelligent, goal-directed
behavior and our ability to interact with the world.
Perceptron A simple, two-layer neural network that consists only of an input and an
output layer. The perceptron is a so-called binary classifier, i.e., a function that
can decide whether a given input vector belongs to a certain class or not. The
perceptron cannot solve classification tasks whose classes are not linearly sepa-
rable, such as the XOR problem. Networks with multiple layers like the Multi-
Layer Perceptron (MLP) are capable of doing this, however.
PET Positron Emission Tomography. A medical imaging technique that can visu-
alize the metabolic activity of cells and tissues in the body. In PET imaging, a
small amount of a radioactive substance, a so-called radiotracer, is injected into
the body. The radiotracer emits positrons, positively charged particles that inter-
act with electrons in the body. When a positron encounters an electron, they
annihilate each other and generate gamma rays, which can be detected with the
PET scanner. The PET scanner captures the gamma rays and creates a three-di-
mensional image of the brain’s metabolic activity. However, since the brain is
constantly active, meaningful PET data can only be obtained by subtracting two
images. Typically, an image is taken during a specific cognitive task or stimu-
lus and another image of the brain’s background activity, and then the difference
image is calculated.
Predictive Coding Predictive Coding. A neuroscience theory that suggests the brain
processes sensory information using a top-down approach. The idea behind
predictive coding is that the brain constantly makes predictions about what it
 Glossary   253

expects to see or hear next based on previous experiences, and then uses these
predictions to interpret incoming sensory information.
Prose Style Transfer A technique from machine learning that can transfer the writing
style of one text to another text.
Pruning Removal of unimportant connections in biological or artificial neural net-
works. During development, the brain establishes far more (random) connec-
tions between neurons than are needed, and then removes the superfluous ones.
This process is important for the formation of the brain’s neural networks. In
machine learning, the complexity of an artificial neural network is reduced by
removing unimportant connections that contribute little to the overall perfor-
mance of the network. See also Lottery Ticket Hypothesis.
Qualia Subjective first-person experiences that we have when we perceive or inter-
act with the world. These experiences include sensations such as color, taste, and
sound, but also more complex experiences like emotions and thoughts. Qualia
are often described as ineffable, meaning they cannot be fully captured or con-
veyed by language or other forms of representation. This has led some philoso-
phers to claim that qualia represent a special kind of phenomena that cannot be
reduced to or explained by the physical or objective properties of the world. See
also mind-body problem.
Recurrent neural network (RNN) A neural network in which information does not
flow exclusively forward, i.e., from input to output. Instead, there are additional
feedback or top-down connections as well as horizontal connections. The recur-
rence can be pronounced in different ways, from Long-Short-Term Memories
(LSTMs), where each neuron has its own connection, to Jordan and Elman net-
works with feedback context layers, to fully recurrent networks like Hopfield
networks or in Reservoir Computing. In contrast to pure feedforward networks,
RNNs can only be trained by a trick (Backpropagation Through Time) with
gradient descent methods and error feedback. This is due to the problem of van-
ishing/exploding gradients. However, RNNs can be trained evolutionarily, unsu-
pervised, and self-organized.
Reinforcement Learning Reinforcement Learning (RL). A type of machine learn-
ing where a model or agent is trained to learn useful input-output functions,
i.e., to make a series of decisions in an uncertain environment that maximize
a cumulative reward. Unlike supervised learning, no outputs are given to the
model. Instead, the agent receives feedback in the form of rewards or penal-
ties, and its goal is to learn a strategy (policy) that maps states to actions, leading
to a maximum long-term reward. The agent uses trial and error to learn from
its experiences in the environment, and tries out different actions to find out
which actions lead to the highest rewards, by reinforcing those actions that have
already proven successful. Over time, the agent’s strategy is refined and opti-
mized, so that it can make better decisions and achieve higher rewards. There
are two types of reinforcement learning. In model-based reinforcement learn-
ing, a model of the environment is also learned, which can predict the feedback
254   Glossary

of the environment (rewards or penalties) on certain actions. In model-free

reinforcement learning, the agent limits itself to learning the best action for a
given state. Reinforcement learning is used in a variety of applications, including
games, robotics, autonomous driving, and recommendation systems. The origins
of reinforcement learning lie in psychology, particularly in the study of animal
behavior and learning. The concept of reinforcement was first introduced in the
1930s and 1940s by B.F. Skinner, who developed the theory of operant condi-
tioning. Skinner’s theory states that behavior is influenced by subsequent conse-
quences such as reward or punishment.
Representational Similarity Analysis (RSA) An analysis method used in cognitive and
computational neuroscience to investigate similarities and differences between
neural activity patterns in different brain regions, neural network models, or
under different experimental conditions. RSA is based on similarity matrices that
contain pairwise similarities or dissimilarities between these activity patterns.
These matrices can be analyzed with multivariate statistical methods to visualize
and quantify the structure of the similarity space and to compare different brain
regions with each other or with artificial neural network layers.
Reproducibility crisis Refers to the difficulty in reproducing the results of a study or
experiment. In the context of AI, it refers to the challenges in reproducing the
results of AI research, including the development, implementation, and eval-
uation of algorithms. The problem of reproducibility affects the reliability and
trustworthiness of AI systems. Often, not all the information necessary to repro-
gram an AI system is published. This can lead to two seemingly identical mod-
els producing very different results. See also alchemy problem and black box
problem.
Reservoir Computing A machine learning technique in which a randomly generated
highly recurrent neural network (RNN), a so-called reservoir, is used for process-
ing input data and creating output predictions, with the connections within the
RNN not being trained. Instead, only the connections between the reservoir and
the output layer are learned through a supervised learning process.
Sentence Embedding A method of machine learning and natural language processing
in which each sentence is assigned a sentence vector. See also word vector.
Sentence vector see Sentence Embedding and Word vector.
Self-supervised learning Type of machine learning in which a model learns to extract
useful features or representations from data without the need for labels on the
input data. Instead, the model is trained to predict or reconstruct certain aspects
of the data, e.g., the next image in a video or the context of a particular word
in a sentence. In self-supervised learning, the data itself provides the labels or
the supervision signal with which the parameters of the model are updated dur-
ing training. In this way, it is possible to learn from large amounts of unlabe-
led data, which are often much more extensive than labeled data. By learning to
extract useful representations from the data, self-supervised learning can be used
 Glossary   255

to improve the performance of a wide range of downstream tasks, including clas-

sification, object detection, and language understanding.
Singularity A concept proposed by Ray Kurzweil of a hypothetical point in the
future when technological progress will be so rapid and profound that it will
fundamentally change human society. According to Kurzweil, the Singularity
will be driven by advances in Artificial Intelligence, nanotechnology, and bio-
technology. He predicts that these areas will eventually converge, leading to the
creation of superintelligent machines and the ability to manipulate matter at the
atomic and molecular level.
Sparse, sparsity. Network property. It means that only a small proportion of the the-
oretically possible connections exist or that only a few weights have a value dif-
ferent from zero.
Spike see Action potential.
Stability-Plasticity Dilemma A fundamental challenge for neural systems, especially
when it comes to learning and memory. It describes the need for a balance
between two opposing goals: stability and plasticity. Stability refers to the ability
of a neural system to retain once learned information and prevent it from being
disturbed or overwritten by new information. However, if a neural system is too
stable, it can become inflexible and resistant to change, which hinders adapta-
tion to new situations or learning new information. Plasticity, on the other hand,
is the ability of a neural system to adapt and change in response to new expe-
riences or information. This is important for learning and memory formation,
as well as for the ability to adapt to changing environments and circumstances.
However, if a neural system is too plastic, it may no longer be able to store pre-
viously learned information, leading to rapid forgetting and lack of stability of
long-term memory. The Stability-Plasticity Dilemma illustrates the need to find
a balance between retaining already learned information and adapting to new
experiences. Every neural system must be able to store and retrieve long-term
memories, while being flexible enough to take in new information and adapt to
changing circumstances.
Stable Diffusion A deep learning-based generative model published in 2022 for
generating detailed images from text descriptions. However, it can also be used
for other tasks such as generating image-to-image translations based on a text
prompt. Stable Diffusion is a so-called latent diffusion model and was developed
by the CompVis group at Ludwig Maximilian University of Munich and the
company Stability AI. The program code and model weights of Stable Diffusion
have been published, and it can be operated on most standard PCs or laptops
equipped with an additional GPU. This represents a departure from the practice
of other AI models like ChatGPT or DALL-E, which are only available online
via cloud services.
STDP Spike Timing Dependent Plasticity. An extension of Hebb’s learning rule.
It describes how the synaptic connection between neurons changes based on
the temporal sequence of their activity. The synapse is strengthened if the
256   Glossary

presynaptic neuron is active shortly before the postsynaptic neuron. The syn-
apse is weakened if the presynaptic neuron is active shortly after the postsynaptic
neuron.
Stochastic Resonance A phenomenon widely observed in nature, which has already
been demonstrated in numerous physical, chemical, biological, and especially
neuronal systems. A weak signal, which is too weak for a given detector or sensor
to measure, can nevertheless be made measurable by adding noise. There exists
an optimal noise intensity, dependent on the signal, the sensor, and other param-
eters, at which the information transfer becomes maximal.
Style Transfer A technique from machine learning that can transfer the painting or
writing style of an image or text to another image or text.
Supervised Learning Type of machine learning in which a model learns to extract
useful features or representations from data, and uses these to generate a desired
output. This type of learning requires so-called labeled data (label-data pairs),
for example, in image classification, in addition to each image, a label (tag) with
information about what is seen in the image or to which category the image
belongs. Supervised learning is usually carried out with Backpropagation
Learning.
Synapse A synapse is a connection between two neurons or between a neuron and
a target cell, e.g., a muscle cell or a gland cell. Characteristic of the structure of
the synapse is a small gap, the so-called synaptic gap, which separates the presyn-
aptic neuron, which sends signals, from the postsynaptic neuron or the target
cell, which receives signals. When an electrical signal, a so-called action poten-
tial, reaches the end of the presynaptic neuron, it triggers the release of chemi-
cals, known as neurotransmitters, into the synaptic gap. These neurotransmitters
diffuse through the synaptic gap and then bind to receptors on the postsynaptic
neuron or the target cell, which can either stimulate or inhibit the activity of the
postsynaptic cell. The strength and efficiency of synapses can change over time,
a process known as synaptic plasticity, which is central to learning, memory, and
other cognitive functions. Dysfunctions of synapses are associated with a number
of neurological and psychiatric disorders, including Alzheimer’s, schizophrenia,
and depression.
Synaptic weight see Weight.
Test accuracy Accuracy. A metric used in machine learning to measure the perfor-
mance of a model. It is defined as the ratio between the correctly predicted or
classified objects and the total number of objects in the dataset. For example, if
a model that has been trained to classify images correctly classifies 90 out of 100
images, then the test accuracy of the model is 90%.
Test dataset Part of a dataset that is used to test an already trained neural network
or model. Typically 20% of the total dataset. See also Dataset splitting.
Top-down From hierarchically higher to lower processing levels.
 Glossary   257

Training dataset Part of a dataset that is used for training a neural network or other
machine learning model. Typically 80% of the total dataset. See also Dataset
splitting.
Transfer learning A technique of machine learning where a model is first trained
on a large dataset and then refined on a smaller dataset for a specific task (Fine
Tuning). The idea of transfer learning is that the knowledge gained in solving
one problem can be transferred to another, related problem, thereby reducing the
amount of data and time required for training a new model.
Transformer Neural network architecture that is particularly suitable for processing
natural language, e.g., translation and text generation. Unlike recurrent neural
networks (RNN) and convolutional networks (CNN), the Transformer uses a
so-called attention mechanism, which allows the model to selectively focus on
different parts of the input sequence to make predictions. A Transformer con-
sists of an encoder and a decoder, both of which are made up of multiple neu-
ral network and attention layers. The encoder processes the input sequence and
generates internal representations from it, which the decoder uses to generate the
output sequence. Transformers have several advantages over traditional neural
networks, for example, the ability to process input sequences in parallel, handle
sequences of variable length, or capture long-range dependencies in sequences
without suffering from the problem of vanishing/exploding gradients. See also
ChatGPT, GPT-3, GPT-4 and Large Language Model.
Tri-Level Hypothesis Theoretical framework proposed by David Marr in the field of
cognitive science and artificial intelligence. According to this, every natural or
artificial system that performs a cognitive task can be described on three levels
of analysis. The computational level describes the problem to be solved, the
goal of the system (brain or AI), and the constraints imposed by the environ-
ment. It specifies what information needs to be processed, what output needs
to be generated, and why the system needs to solve the problem. The algorith-
mic level describes the rules and procedures, i.e., the algorithm, that the system
must follow to solve the problem specified at the computational level. It specifies
how the input data is transformed into output data and how the system pro-
cesses information. The implementation level describes the physical implemen-
tation of the system, e.g., the hardware and software used to build the AI. It
specifies the details of how the algorithmic level is implemented, including the
data structures used, programming languages, and computational resources. In
neurobiology, the anatomical and physiological details of the nervous system are
described at this level. According to Marr, it is necessary to understand a system
at all three levels in order to fully grasp its behavior and possibly develop more
efficient systems.
Turing Test A method proposed by Alan Turing and originally called the Imitation
Game by him to test a machine’s ability for intelligent behavior. In the simplest
variant, one or more human examiners communicate in natural language text-
based, i.e., in the form of a chat, both with a human (as a control) and with the
258   Glossary

machine to be tested, without knowing who is who. If the majority of examin-

ers cannot reliably distinguish between human and machine, the Turing Test is
considered passed for the machine. This variant of the Turing Test was passed
in 2022 by the AI system ChatGPT as the texts and responses it generated are
indistinguishable from those produced by humans. In principle, there are other,
more difficult variants of the Turing Test. For example, the dialogue can take
place not as a pure chat, but as an actual conversation in spoken language, sim-
ilar to a telephone conversation. In this case, the machine would also have to be
able to correctly interpret and imitate linguistic features such as emphasis and
sentence melody. Finally, the AI could also be integrated into a humanoid robot.
In this variant, which is also referred to as the Embodied Turing Test, the entire
spectrum of human behavior including facial expressions, gestures, and all motor
skills would have to be imitated, and the Turing Test would ultimately aim to
convince other people that the machine is also a human. In the final escalation,
the machine would not only have to convince others, but also be convinced itself
that it is a human, and would not be allowed to know or suspect anything about
its true nature. Such scenarios have been picked up in films like Imposter or the
series Westworld and the consequences have been thought through to the end.
The Turing Test has been much discussed in the field of Artificial Intelligence.
Some critics argue that it sets an unrealistic standard for intelligence, as there
are many tasks that humans can perform, but machines cannot, and vice versa.
Others argue that passing the Turing Test does not necessarily mean that a
machine is truly intelligent, but rather that it is good at imitating human behav-
ior. One of the most famous counterarguments to the Turing Test is John Searle’s
thought experiment on the Chinese Room. Despite this criticism, the Turing
Test remains an important benchmark for research in the field of Artificial
Intelligence, and many researchers continue to work on the development of
machines that can pass the test.
Unsupervised Learning Type of machine learning in which a model learns to extract
useful patterns and structures from data without the need for labeled data. This
type of learning typically involves tasks such as clustering, where similar data
points are grouped, and dimensionality reduction, where high-dimensional data
is represented in a low-dimensional space while preserving important informa-
tion. Examples of unsupervised learning algorithms include K-Means clustering,
hierarchical clustering, and Principal Component Analysis (PCA).
Vertical Refers to the processing direction between elements on different hierarchy
levels as opposed to horizontal processing between elements of the same hierar-
chy level. See also Top-down and Bottom-up.
Word embedding A method of machine learning and natural language processing in
which each word is assigned a word vector.
 Glossary   259

Weight Also Synaptic weight. Strength of the connection between two artificial or
natural neurons. The weight can in principle be any real number, with the mag-
nitude representing the size of the effect on the successor neuron and the sign
representing the quality of the effect (excitation or inhibition).
Weight matrix Matrix whose entries contain the weights between a set of neurons.
The matrix can contain all pairwise weights of all neurons of a neural network,
including all self-connections (autapses). In this case, it is referred to as a com-
plete weight matrix. However, a weight matrix can also only contain the for-
ward-directed weights between the neurons of two successive layers of a network.
The columns or rows of the weight matrix correspond to the input or output
weight vectors of the neurons.
Weight vector Vector whose entries contain all the weights of a neuron. A distinc-
tion is made between input weight vectors, which contain the weights of a neu-
ron with which its input is weighted, and output weight vectors, which contain
the weights to the successor neurons of a neuron.
Word vector Representation of the meaning of a word as a vector. The more dif-
ferent the meaning of two words is, the more different are the corresponding
word vectors. If you interpret the word vectors as points in a semantic space,
the distance between the points corresponds to the similarity or dissimilarity of
the underlying words. The smaller the distance, the more similar the meaning.
Synonyms, i.e., words with the same meaning, are mapped to the same word
vector or point, so they have a distance of zero. Similarly, there are also sentence
vectors, which represent the meaning of an entire sentence.
XOR Problem A classic problem in the field of Artificial Intelligence and Machine
Learning that illustrates the limitations of certain types of models. XOR stands
for eXclusive OR, a binary operation in which the output is true (or 1) only
when the number of true inputs is odd. With two binary inputs, XOR is true
only when exactly one of the inputs is true. The XOR problem refers to the chal-
lenge of correctly classifying these four situations using a linear classifier, e.g., a
single-layer perceptron. The problem is that the XOR function is not linearly
separable, i.e., there is no straight line (in 2D space) that can separate the inputs
that yield a 1 from those that yield a 0. This illustrates the inability of linear clas-
sifiers to handle certain types of patterns. Multi-layer neural networks, however,
can solve the XOR problem by creating non-linear decision boundaries. This is
typically achieved by introducing hidden layers and non-linear activation func-
tions into the network.
Zero-Shot Learning A form of machine learning where a model is trained to rec-
ognize objects or categories it has never seen before. It can classify new input
patterns even when no labeled data for the relevant class were available dur-
ing training. Unlike supervised learning, where a model is trained with a spe-
cific set of labeled data examples, zero-shot learning is based on the transfer of
260   Glossary

knowledge from related or similar classes that were seen during training. This
is achieved by using semantic representations such as word vectors that capture
the meaning and relationships between different classes. For example, if a model
has been trained to recognize images of animals and has never seen an image of
a zebra, it can still classify it as an animal because it has learned the relationships
between different animal species. Zero-shot learning enables more efficient and
flexible training of models and generalization to new and unknown categories.