54

54 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 15

I
n the outside-in framework, the separation of perception from action is an sensory research today. e theoretical framework of a passively associating
important distinction. Perception-action separation is both intuitive and brain was a welcome strategy to experimenters because neuronal responses to
reasonable in the context of early anatomical studies of the nervous system. sensory stimuli could be registered also in anesthetized or paralyzed animals.
In the nineteenth century, François Magendie discovered that the motor When experiments were subsequently performed in waking animals, their
neuron bers that innervate skeletal muscles exit from the anterior root of the heads and even the eyes were xed to eliminate inconvenient behavior-induced
spinal cord, while the sensory neuron bers that carry information about touch variability in the recorded brain signals. However, building a complex percept
from the skin and the contraction states of the muscles from their own stretch- in the brain from its assumed elementary components turned out to be a formi-
detecting receptors (i.e., proprioceptors) enter through the dorsal root. 4 at dable hurdle given the astronomical realm of possible combinations.30 Luckily,
is, distinct sets of neurons deal with body sensation and muscle control. A few it turned out that natural visual scenes and sounds represent only a small subset
decades later, the Ukrainian neuroanatomist Vladimir A. Betz noted that the of all possible combinations, so sensory researchers recently began to use nat-
brain maintains the anterior-posterior segregation of the spinal cord: structures ural images and sounds. ey discovered that brains are “ ne-tuned” for nat-
dedicated to vision, hearing, taste, and somatosensation reside largely in the urally occurring stimuli, and thus many aspects of neural activity were missed
posterior half of the human brain, whereas areas behind the eye, called frontal with the previous arti cial stimuli. For example, even in a small brain, like that
structures, are mainly devoted to motor functions, 5 in keeping with the anterior of a frog, transmission of acoustic information is most e cient when the spec-
motor and posterior sensory organization of the spinal cord. However, while trum of sounds matches the spectrum of natural frog calls.31 But the nature of
anatomy supports the separation of sensation and motor commands, it has no the stimulus is only a technical problem. e bigger issue is that the brain is not
bearing on the temporal ordering of events, the connections between “sensory” a passive device that unconditionally associates signals.
and “motor” areas in the brain, or the postulated causal relationship between A tacit assumption of the outside-in32 framework is that attributes or features
perception and action. of objects are bound together by some brain mechanism. What makes an object
Furthermore, while sensation and perception are obviously related, they are an entity is the coherent fusion of its components, which separates it from its
di erent. Some authors use the terms interchangeably; others refer to sensa- surroundings. For example, a car has a particular shape with a body and wheels,
tion in lower animals and perception in higher animals. A more rigorous dis- color, sound, and a characteristic movement pattern. But there is a problem
tinction is that sensation is the instantaneous feeling that receptors are being with this philosophy, which is that the “attributes” of an object are not a xed
stimulated, whereas perception compares sensation with memories of similar set that reside in the object but are constructed in the brain of the observer.
experience to identify the evoking stimulus. Similarly, although “action” and Brain responses do not intrinsically “represent” stimuli until the experimenter
“motor output” belong to the same general category, “action” is a more general interprets the data. 33 Judgments of “similar” or “di erent” are subjective (i.e.,
term than “motor.” Outputs from the brain can not only a ect skeletal muscles observer-dependent). Yes, an experimenter can quantitatively de ne the rela-
tionship between stimuli and the spatiotemporal ring patterns of recorded
4. Jørgensen (2003); Bell (1811); Magendie (1822). neurons (Figure 1.2). However, this relationship does not mean that neurons in
5. Betz noticed that the anterior (motor) areas in humans contain large pyramidal cells in deep
layers, whereas in the posterior part the cells were smaller. is observation led to his best-
known contribution, deducing the function of the giant pyramidal neurons of the primary
motor cortex (known today as “Betz cells”), which are the origin of the corticospinal tract. “ e 30. An illuminating review on how methods in uence scienti c thinking is that by Evarts
sulcus of Rolando divides the cerebral surface into two parts; an anterior in which the large (1973). I o en wondered how many more recordings, monkeys, and cats would be needed to
pyramidal nerve cells predominate. . . . Undoubtedly these cells have all the attributes of so- explain vision with the outside-in strategy, even if all 30,000 neuroscientists were working on
called ‘motor cells’ and de nitely continue as cerebral nerve bres” (Betz, 1874, pp. 578–580). the visual system.
In addition to the human precentral gyrus, Betz found these cells in the same location in dogs, 31. Attneave (1954); Singh and eunissen (2003); Wang et al. (1995); Rieke et al. (1995).
chimpanzees, baboons, and other primates, and, based on the “brilliant physiological results” of
32. Related terms to the “outside-in” framework are “empiricist,” “associational,” and “repre-
Fritsch and Hitzig, he concluded that “these cells have all the attributes of the so-called ‘motor
sentational” strategies, and their various combinations. Principles of Neural Science, a 1,700-
cells’ and very de nitely continue as ‘cerebral nerve bres’ (for further details, see Kushchayev
page “bible” of neuroscience and the most widely used textbook for educating neuroscientists,
et al., 2012). e anterior cingulate cortex of hominids also contains such large neurons, and
also follows the outside-in framework (Kandel et al., 2012).
much speculation has been devoted to the signi cance of these “spindle cells,” also known as
von Economo neurons, in intelligence (Allman et al., 2001). 33. Skarda and Freeman (1987); Werner (1988).
16

16 THE BRAIN FROM INSIDE OUT

the brain of the experimental subject views the stimuli in the same way as the
experimenter, nor does it mean that the brain uses these recorded signals in the
same way or at all. is is the crux of the problem of the outside-in correlational
approach. It cannot yield adequate understanding of perception or neuronal 3
computation. I brie y summarize the outline of this argument here and return
to it in more detail in Chapter 3.
Perception from Action

ONLY ACTION CAN PROVIDE MEANING

TO BRAIN ACTIVITY

Like the philosophical issues that plagued physics through the twentieth cen-
tury, the issue in neuroscience is the “observer.” ere is a fundamental di er-
ence between an observer outside the brain, such as the experimenter, and an Our sensory systems are organized to give us detailed and accurate view
observer inside the brain, such as Descartes’s hypothetical homunculus residing of reality, exactly as we would expect if truth about the outside world
in the pineal gland. helps us navigate it more e ectively.
First, let’s take the perspective of an experimenter studying the visual —R T
system. Upon presentation of a visual stimulus, an image is converted to action e hardest thing is to do something which is close to nothing.
potentials by retinal ganglion cells and conveyed to the primary visual cortex —M A
via neurons in the thalamic lateral geniculate body. e experimenter can place
a recording device, such as an electrode array, inside the brain and monitor e eye obeys exactly the action of the mind. In enumeration the names
neural responses to di erent images presented to the eye. In this situation, the of persons or of countries, as France, Germany, Spain, Turkey, the eyes
experimenter has a privileged point of view: he has access to both the images wink at each new name.
(external world) and the neuronal responses in the brain (Figure 1.2). From —R W E ( )
this perspective, he can nd a reliable correlation between speci c features of
the presented images (input; the rose in Figure  1.2) and neuronal responses
(output for present purposes), which allows him to make a guess about the pos-
sible input–output transformation rules. A er establishing a correlation (o en
mistakenly called a neuronal code), he can recover stimulus properties from
the neuronal responses and even predict the properties of untested stimulus
patterns (e.g., another ower). In short, a particular constellation of neuronal
1. Robert Trivers, one of the most prominent evolutionary biologists and sociobiologists, made
spike patterns becomes information. But to whom is that information useful? this statement as recently as 2011 (Trivers, 2011). Many other prominent scientists share the
e recorded signal in the brain, combined with a priori knowledge about the view that our perception is veridical and is a window on truth, “veridicality is an essential char-
stimulus, can generate meaning only to the experimenter. e observed corre- acteristic of perception and cognition” (Pizlo et al., 2014). David Marr (1982) also expressed the
view that humans “de nitely do compute explicit properties of the real surfaces out there” but
lation between the image and neuronal response only means that something
he denied this ability in non-human species. Other theorists make such dichotomy even among
outside the brain was transformed into ring pattern changes. senses; for example, that vision is veridical but taste is not (Pizlo et al., 2014).
What about the internal observers, the brain circuits? e neurons that “re-
2. Abramovic (2016).
spond” to the image do not see the image, just patterns of neural spiking arriving
3. Chapter 5 on Behavior, in Emerson’s essay collection, e Conduct of Life (1899), is full of
from their peer neurons. Neurons in the visual cortex are blind to events that
outstanding observations on how human beings and their thoughts are exteriorized by eye
happen in the world and so is the rest of the brain or even a magical homun- movements. “ e eyes of man converse as much as the tongues, with the advantage that their
culus (i.e., the little human) watching the action of all visual cortical neurons ocular dialect needs no dictionary, but is understood all the world over” (p. 173).
52

Chapter 1. e Problem 17

(Figure 1.2). Imagine that the little human is the experimenter who, in a magic
experiment, can register the activity of all neurons in the brain. Without access
to information outside the skull, such knowledge cannot reveal what the brain is
sensing; it cannot see the rose. In the language of the neural code, the outside-in
framework has shown only that decoding of stimulus properties from neural ac-
tivity is possible in principle and only if an observer has the code book (i.e., the set
of stimulus–response correlations). However, the brain only has its own neural
“responses,” and the outside-in framework does not mention how such a code
would be generated or read by neural responses alone.
All any neuron in the brain ever “sees” is that some change occurred in the
ring patterns of its upstream peers. It cannot sense whether such change is
caused by an external perturbation or by the brain’s perpetual self-organized
activity (Chapters  7 and 8). us, neurons embedded in networks of other
neurons do not “know” what the brain’s sensors are sensing34; they simply re-
spond to their upstream inputs. In other words, unlike our ideal experimenter,
the neurons have no way of relating or comparing their spikes to anything else be-
cause they only receive retinal correspondences or processed “representations”
of the sensory input. But establishing correspondences without knowing the
rules by which those correspondences are constructed is like comparing Mansi
words with Khanty words when we understand neither language.35 Only a er
we have de ned the vocabulary of one language can we understand the corre-
sponding meaning of words in the other. Similarly, without further informa-
tion, sensory neurons can attach no meaning whatsoever to their spikes. Put
simply, the mind’s eye is blind.36
To explain my point di erently, let me invite you to do a thought experiment.
Let’s connect the output of a video camera (or the optic nerve of the retina)
to cultured neurons in a dish, derived from the visual cortex of the smartest
person in the world, so that a fraction of the neurons is stimulated by the

34. A classic paper on this topic is by Lettvin et al. (1959). See also Eggermont (2007).
35. is is not to say that everything is equal to a newborn brain. Phylogenetic experience can
provide the necessary cipher for certain patterns, such as human faces, for babies. But, for most
of the things in the world, we have to create a cipher through action-based grounding.
36. is “blindness,” in my view, applies also to the Global Workspace version (Baars, 1988;
Dehaene and Changeux, 2011) of the Cartesian theater metaphor of consciousness, since the
Global Workspace, even it if involves distributed processing in the entire brain, faces the exact
same problem: having no access to the external information. Realizing the shortcomings of the
Cartesian theater idea, Dennett and Kinsbourne (1992) suggested an alternative solution, what
they call the “Multiple Dra s” model, reminiscent of E. Roy John’s distributed consciousness
model (1976). However, this model is also a “passive” interpreter and has no access to outside
world information and, therefore, faces the same “grounding problem” (Harnard, 1990).
18

18 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 51

camera output. e part of the tissue culture that is stimulated by the camera that do not rely on cause-and- e ect arguments can also arrive at valid sci-
output can be called the sensory region. We may discover that repeated visual enti c conclusions. e concept of causation is especially problematic in
patterns, such as the picture of the rose, will induce somewhat similar neuronal self-organized systems with amplifying-damping feedback loops, such as the
responses. We may even detect some changes in synaptic connectivity among brain. Causes in such systems are o en circular or multidirectional; events
neurons as a result of repetition. It is also possible, though, that some members are not caused but emerge from the interaction of multiple elements.
of the cultured neuronal network will be preoccupied with each other and ig-
nore the stimulation altogether.37 e experimenter can establish a relationship
between the applied stimuli and the corresponding neuronal responses and de-
clare that some coding was identi ed. But the tissue culture does not “see” the
rose. So far, this is analogous to the passive, representational brain model and
the experimental approach used in many laboratories. Since the neurons in the
culture have no direct access to the outside world, they have no way of verifying
or “grounding” their patterns to the events outside. e term “grounding”
refers to the ability of the brain’s circuits to connect it to something meaningful
for the owner of the brain. It’s kind of a second opinion from a more reliable
source.38 Using our dictionary metaphor, if you know the meaning of words in
an English dictionary, you can ground the meaning of corresponding words in
a dictionary of any other language.
e situation in the dish may change dramatically if the spiking output of some
neurons is connected to a robot that can move the camera based on the spiking
pattern of these experimenter-designated “motor” neurons in the culture. Now
neurons in the dish have two functions. One function is generating an output
(move the camera) and the second is responding to the camera-supplied signals
(sense the inputs). Because the tissue culture circuit is supposedly well intercon-
nected, neurons in the “sensory” region will receive two types of inputs: input
from the camera signals and an additional input from their motor neuron part-
ners, which inform them that it was their action that induced movement of the
camera and, therefore, changes in the sensory input. Now, we have a closed-loop

37. It helps if the tissue culture is “awakened” by adding a cocktail of subcortical

neurotransmitters, mimicking the aroused state of the brain (Hinard et al., 2012).
38. Steve Harnad (1990) explains the need of “symbol grounding” in understanding. An e ec-
tive computer algorithm can discover correspondences between two sets of symbols without
understanding the meaning of any of them. Handheld translators have enormous vocabularies
and can convert commonly spoken words and phrases to dozens of other languages at an
amazing speed and even pronounce the words in the target language—all done by a few chips
and a microprocessor. However, the semantic “meaning” of symbols of one set (e.g., a new lan-
guage) can be constructed only if the translator’s knowledge is grounded to an already known
set of symbols (e.g., your mother tongue). Without grounding, a theorem remains discon-
nected from empirical phenomena. See also Fields (2014). Grounding is also related to the
“credit assignment problem”; that is, the determination of how collective action of neurons
relates to behavior.
50

50 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 19

perturbation methods should be combined whenever possible because they system in a dish, an output-driven sensory feedback device. e addition of the
represent complementary ways of analyzing the same phenomena. is is the robot provided the cultured neurons with the same advantage as the experi-
best we can do. menter: now the neurons can sense both the outside word and the internal com-
e di culty in applying causation as the sole tool for inference will putation and, therefore, can compare their joint impact. 39
o en reappear in this book. In the next chapter, we will examine the rela- Now let’s assign a goal to our closed-loop system—for example, to nd the
tionship between perception and action. e two are o en correlated. As we rose in a picture. Whenever the camera is focused on the rose by the chance
learned in this chapter, correlation is symmetric—in the mathematical sense. behavior of the cultured neurons, a magic potion (let’s call it dopamine) is
However, in searching for a cause with an assumed direction, we designate spritzed onto the culture to strengthen the connections between those neurons
one set of variables independent and the other dependent and make the re- that successfully moved the camera to the desired spot in the visual scene and
lationship asymmetric in our minds. To verify our intuition, we must vary the neurons that were activated by the rose. is simple modi cation (“plas-
one of the variables and quantify the consequence on the other variable. From ticity”) may increase the probability that such neuronal activity will happen
this perspective, there should be a distinction in whether we talk about a again and lead to pointing the camera at the rose more o en.40 us, out of
perception–action cycle (where perception is the critical variable for action) the large realm of possible neuronal network patterns, at least one unique event
or an action–perceptions cycle (where action is critical for perception). While acquired a meaning. By connecting a bunch of otherwise useless tissue culture
in everyday parlance such distinction may sound prosaic and hair-splitting, neurons with the world through an action-perception arc, we have just built a
I  submit that the failure to recognize the importance of this di erence is a machine with a purpose: a brain-like device.41 In fact, we may not need a magic
main contributor to why the outside-in framework continues to dominate potion if we assume that the brain is already prewired to some extent so that its
cognitive neuroscience. In an uncalibrated (naïve) brain, changing the sen- connectivity is not random but guided by some evolution-shaped statistical rule
sory inputs may not impact the motor response much (save phylogenetically (Chapter 11). e internal feedback from the action-inducing neurons to sen-
learned xed action patterns), whereas moving the sensors by actions brings sory neurons may be regarded as a second opinion, the needed grounding signal
about large changes in the activity of sensory structures. e next chapter that can attach meaning to input-induced neuronal responses (Chapter 3).
discusses how confusion between cause and e ect has impacted thinking in
perception research.
39. For an attempt to achieve such comparison by neurons, see Demarse et al. (2001). A com-
putational model using such building blocks is by Weng (2004); Choe et al. (2007).
SUMMARY 40. Arti cial intelligence (AI) devices using cameras give equal priority to all image pixels
at the earliest stage of processing. In contrast, the visual system prioritizes and isolates the
most relevant information based on previous knowledge and moves the eyes to attend those
In neuroscience, we o en compare our observational data against
aspects (Olshausen et  al., 1993). Inspired by the brain’s solution, recent AI architectures se-
association-based models, such as correlations, and plot our data as a re- rially take glimpses of the input image and update their internal state representations before
lationship between presumed independent and dependent variables. When sampling the next location. e goal of AI agents is to select the best outcome that results in
we identify some regularity between two variables ( nd a correlation), we maximal “reward” (Mnih et al., 2014). e main success of AI agents, in, for example, defeating
are tempted to regard one set of variables as the thing- to-be- explained and human players in Atari games, is that those games are also designed by humans using arti cial
algorithms. e current instantiations of AI agents are still pretty helpless when it comes to de-
the other set as the thing- that-explains. is is the basic logic of cause-and- veloping and solving temporally extended planning strategies.
e ect relationships. Interpreting correlations is a two-step process. e rst
41. William Grey Walter at the Burden Neurological Institute in Bristol, England built an elec-
step is the description of facts (the correlation), followed by a second, biased tronic “tortoise,” among the rst autonomous robots, which he called Machina speculatrix. His
step—an assumption of a cause-and-e ect relationship from a symmetric Machina had only two “neurons” (vacuum tubes) but could be assigned a goal, which was to
statistical relationship. nd its way to a recharging station when it ran low on battery power. Walter (1950) argued that
Causality is the most critical pillar of scienti c inference in the Western goal-directed action is a main characteristic of animals and that animated gadgets with a goal,
such as his robots, can emulate such behaviors. A  similar argument was made subsequently
world. Revealing a cause amounts to an explanation. However, other cultures
by Brooks (1991):  brains have no representations. Instead, “it is a collection of competing
behaviors. Out of the local chaos of their interactions there emerges, in the eyes of an observer,
a coherent pattern of behavior.”
20

20 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 49

In our tissue culture example, we arbitrarily designated two sets of of certain brain parts as a result of injury or disease has o ered numerous and
neurons and called them sensory input to where the stimuli from the camera powerful insights into the contribution of various brain structures to normal
were delivered and motor output because these neurons were connected to function, suggesting that di erent parts of the brain are specialized for di erent
the robot moving the camera. e tissue culture has no a priori bias toward functions. Complex functions, such as memory, emotions, and planning, have
these designations, and the two areas are reciprocally interconnected to found their homes in the neural substrates of the hippocampus, amygdala, and
each other. In real brains, the sensory and motor areas have somewhat dif- prefrontal cortex, respectively, with the o en explicit conclusion that neuronal
ferent internal connectivity but the similarities are more striking than the activity in these structures is causally related to the respective behaviors. In
di erences. e di erences are even more blurred when we compare their animal experiments, structures or their parts can be deliberately damaged,
partner structures, o en referred to as higher order sensory and supplemen- “turned on,” or “turned o ” as a “causal” manipulation. e ensuing behavioral,
tary motor areas.42 ese areas are reciprocally connected, and the tra c cognitive de cit can then be interpreted to indicate that, in the intact brain, the
moves not only from the sensory to the motor circuits but equally so from perturbed structure is responsible for that behavior.
the motor to the sensory areas. is latter motor- to-sensory projection has a But slam on the brakes. Such a conclusion is not always warranted. For ex-
fundamental impact on many levels of brain operations, as we will see in the ample, massive degeneration of substantia nigra neurons (loss of function of
subsequent chapters. dopamine-producing neurons) is followed by muscular rigidity, slowed move-
e picture I  just painted is unfortunately an oversimpli cation. I  did not ment (bradykinesia), and tremor. However, decreased dopamine is a neces-
mean to create the impression that our tissue culture thought experiment can sary but not always su cient condition for the symptoms to occur, and similar
explain the mechanisms of a thinking, feeling brain. It was only meant to il- symptoms may also occur in the absence of dopamine dysfunction. Similarly,
lustrate the minimum necessary requirements of a brain-like system. Random subcortical denervation of the hippocampus in rodents results in more severe
connections and plasticity are not enough (Chapter 12). A large part of brain cognitive de cits than damage to the hippocampus itself. is happens not only
structure is genetically determined, and self-organized activity is as important because the hippocampal circuits can no longer perform their physiological
for the development of brain dynamics as is sensory stimulation even in pri- jobs but also because the denervated hippocampus induces abnormal neuronal
mary sensory areas of the brain43 (Chapter 5). activity, adding insult to injury. Conversely, if behavior is una ected long a er
lesioning a particular brain structure or circuit, one cannot rule out its impor-
tant contribution to that behavior because the brain circuits have redundancies,
THE INSIDE- OUT, READER- CENTRIC FRAMEWORK degeneracies, and plasticity that can o en compensate for an acute impairment.
ese examples illustrate why, in a densely interconnected dynamical system
Can James’s Table of Contents and the outside-in strategy be renegotiated? I do like the brain, even minor and local perturbations may induce unexpected ac-
not think it would get us far. My point is not that measuring brain responses tivity in downstream and distant structures. Furthermore, functional recovery
may occur via alternative behavioral strategies and adaptive repurposing of
structures that are targets of the damaged area. e emerging nonphysiological
42. Mrganka Sur and colleagues (Sharma et al., 2000) “rewired” the brain of ferrets so that the activity in these downstream structures can be as damaging as the absence of
visual information was connected to the auditory system. Neurons in the auditory cortex of their upstream partners. ese considerations point out the caveats and limita-
these animals responded in many ways like the visual cortex of intact animals, including visual
orientation modules. Impressively, the ferrets used their newly wired brain to avoid visual tions of using causation in our search for physiological function in the complex
objects. labyrinth of brain dynamics.
43. Brain circuits can assemble themselves under genetic guidance to provide “proto- maps.”
In conclusion, statistical correlations are similar to associations. In neither
When the thalamic input to the neocortex is disrupted in Gbx-2 mutant mice, neocortical case can one rely on causal grounding, which requires active manipulations.
region-speci c gene expression and the main cortical layers and divisions develop normally In the strongly interconnected complex networks of the brain, perturba-
(Miyashita-Lin et al., 1999). Deletion of a single protein, Munc18-1, in mice abolishes neu- tion may bring about secondary and higher order changes, which need to be
rotransmitter secretion from synaptic vesicles throughout development. However, this does
separated from the primary e ects. is is rarely straightforward as complex
not prevent an apparently normal initial assembly of the brain. However, persistent absence
of synaptic activity leads to synapse degeneration and the mice die (Verhage et al., 2000). e networks o en respond to challenges in an unpredictable manner. Since there
constellation of ion channels in individual neurons is also driven by their activity. is no simple recipe for how to deal with emergent systems, correlational and
48

48 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 21

Just because two sets of data move together over time, that does not necessarily to stimuli yields no valuable data or insight. In fact, a primary slogan in my
mean they are connected by a causal link. Correlations are not causation, as con- laboratory is that “it is impossible to nd nothing” in the brain. An enormous
ventional wisdom rightly says. erefore, cause does not reduce to correlation. amount of precious knowledge has been obtained by identifying neuronal
To understand a machine, it is not enough to look at it carefully. We must take “correspondences” and “representations” of the sensory world over the past
it apart and put it together or remove parts from it and see what happens. Of decades. Until recently, the empiricist outside-in approach has generated most
course, in a closed system, such as a car, if all parts and all their linear interactions of the available knowledge about sensory coding, and it continues to be an im-
are known, in principle, the state of the system can be described at any time. portant strategy. However, when studying the more central parts of the brain
However, in open systems, such as the brain with its complex dynamics, such involved in cognition, the limitations of a strict stimulus–response strategy be-
complete description is rarely possible, mainly because the interactions are non- come apparent.44 e ultimate goal of brain research cannot just be to uncover
linear and hard to predict. correspondences between external signals and neuronal responses. Building
a dictionary of representations from scratch would be an enormous e ort,
but it is easy to see why such a program has limited success (Chapter 4). e
Challenging Correlations outside-in approach interrogates brain circuits by asking what those circuits
can potentially do, rather than what they actually do, which produces an illu-
Perturbation, a targeted challenge to the system, is a powerful method for sory understanding of brain computation. Overall, I suggest that searching for
investigating a system provided that perturbation is done using a correct design alternative approaches is warranted.
and the resultant changes are interpreted properly.30 In neurology and brain sci- e essence of the complementary strategy I  suggest is simple:  under-
ences, the perturbation method is one of the oldest. Damage or degeneration stand the brain from within. In this inside-out framework, the key issue is the
mechanisms by which stimuli and situations become meaningful percepts and
parameter” is the synergy or simultaneous action of emergence. In synergetic systems, up- experiences for the brain.45 e main emphasis is on how the brain’s outputs,
ward (local-to-global) and downward (global-to-local) causations are simultaneous. us, the re ected by the animal’s actions, in uence incoming signals. By linking an oth-
“cause” is not one or the other, but it is the con guration of relations. One can call such causa- erwise meaningless brain pattern to action, that pattern can gain meaning and
tion reciprocal or circular, terms that are forbidden by Aristotelian and Humean logic (circulus signi cance to the organism. Within the brain, the emphasis is on how down-
vitiosus). However, in complex systems, circular causation is neither vicious nor virtuous but
a norm. Reciprocal interactions weave the correlations. An excellent discussion of synergetics
stream networks make use of the messages obtained from their upstream part-
in neuroscience is by Kelso (1995), with an abbreviated version in Bressler and Kelso (2001). ners. In this framework, the goal of the brain is to explore the world and register
30. Controlling neuronal activity in real time was only a dream for neuroscientists not so long
the consequences of successful exploratory actions to improve the e cacy of
ago. Since the introduction of optogenetics into mainstream neuroscience by Karl Deisseroth future actions. us, an action–perception loop learns to make sense of sensory
and Ed Boyden (Boyden et  al., 2005), we now have the ability to turn on or o genetically inputs. Perception is what we do (Chapter 3).46
identi ed neurons at will. is revolutionary tool allows us to challenge long-held views and
generate new advances in our understanding of brain function in both health and disease.
Unfortunately, when new techniques are invented, optimism and hype o en trump humility.
Optogenetic techniques are too o en advertised as causal tools to discover the relationship be- 44. Rieke et al. (1997); Friston (2010, 2012).
tween activity patterns in speci c neuronal circuits and behavior. However, because neurons 45. e outside-in versus inside-out dichotomy should not be confused with the top-down
are both embedded in circuits and contribute to circuit function, their perturbation can bring and bottom-up distinctions. e former refers to the relationship between the world and the
about secondary and higher order changes, which need to be separated from the primary ac- brain, whereas the latter refers mainly to the spread of activity in anatomical space. It is also
tion of optical stimulation (Miesenbock, 2009). is separation is rarely straightforward be- distinct from Paul MacLean’s (1970) concept of the triune brain, a division into reptilian, paleo-
cause the network may change in an unpredictable manner. For this reason, it is not easy to mammalian, and neo-mammalian brain. e relationship of the inside-out framework to
tease out whether the observed variable is caused by silencing of a speci c group of neurons or embodied cognition and predictive coding will be addressed in Chapters 3 and 5.
is the unseen consequences of novel patterns in their targets. Without knowledge of the precise
e ect, the assumed cause ceases to be a true cause. A potential improvement is the combina- 46. In my younger days, I thought these ideas were solely mine. But, with time, I discovered
tion of correlation and perturbation methods. To fully exploit the advantages of a perturbation that many people, o en independently from each other, have come to a similar conclusion (e.g.,
technique for brain circuit analysis, one should try to limit the perturbation to a small number Merleau-Ponty, 1945/2005; Mackay, 1967; Bach-y-Rita, 1983; Held and Hein, 1983; Paillard,
of neurons that are continuously monitored together with their neighbors and upstream targets 1991; Varela et al., 1991; Berthoz, 1997; Bialeket al., 1999; Järvilehto, 1999; O’Regan and Noë,
(Buzsáki et al., 2015; Wol and Ölveczky, 2018). 2001; Llinás, 2002; Noë, 2004, 2009; Choe et  al., 2007; Chemero, 2009; Scharnowski et  al.,
2

22 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 47

Nearly all chapters of this book revolve around an action-centered brain MUTUAL AND SELF- CAUSATION
because I  want to make it as clear as possible that no meaning or advan-
tage emerges for the brain without the ability to calibrate neural patterns by Probabilistic causation appears more attractive and perhaps more appropriate
behavior-induced consequences. Without generating outputs to move and for describing interactions in complex systems such as the brain. A er all, it
optimize the sensors, and calibrate brain circuits at some point in our lives, is only in interactions that nature a ects the world. One may wonder whether
I claim, no perception or cognition exists.47 transforming the philosophical problem to a precise mathematical descrip-
A major advantage of inside- out approach is that it is free of philosoph- tion makes the word “causation” more valid. is issue is especially relevant
ical connotations. It takes brain mechanisms as independent variables, as in mutually interconnected and strongly recurrent complex systems with
opposed to attempting to nd correspondences between subjectively de- amplifying-damping feedback loops and emerging properties. Brain networks
rived categories and brain responses. Brains largely organize themselves, are robust and balanced, so sometimes even very strong stimuli may fail to af-
rather than being enslaved to input signals. I am interested in how internally fect the brain. Other times, minute perturbations can have a large impact on
generated, self-organized patterns in the brain acquire “meaning” through neuronal activity depending on the state of these networks. e most important
action, which becomes what we call “experience.” Instead of looking for feature of complex systems is perpetual activity supported by self-organized
correlations between experimenter-selected signals and activity patterns of dynamics (Chapter 5). Ever since an EEG signal was rst detected from the sur-
neurons, we should ask a more important question:  How do those ring face of the brain and scalp, its ever-changing electric landscape has been called
patterns a ect downstream “reader” neurons? We can call this strategy the “spontaneous.” Related terms, such as endogenous, autogenous, autochthonous,
neural observer, or reader-centric, view. Before you accuse me of simply autopoietic, autocatakinetic, or self-assembled are more popular in other com-
replacing the homunculus with the “reader,” please consult Chapters 4 and plex systems, but they refer to the same idea.28 ese concepts have proved to be
5.  e term “reader/ observer” is meant to be an engineering term, referring di cult to tackle because, by de nition, the generation of spontaneous activity
to an actuator (such as a robot, muscle, or simply a downstream neuron) occurs independently of outside in uences. us it must be induced by some
that responds to a particular pattern of upstream neuronal activity but not self-cause, not by external causes, although the latter are much easier to study.
to others; only a particular key of a bunch on a keychain opens a particular How can we separate causes from e ects in self-organized brain networks
lock. In the subsequent chapters, I will give you a taste of how this is done in with multiple parallel and interacting feedback loops? In many cases, brain
the brain by discussing a set of recent ndings that illustrate progress toward patterns, such as neuronal oscillations, are not caused but are released or
an inside-out research program. permitted. For example, sleep spindles in the thalamocortical system can be
induced occasionally by some sensory stimuli but most o en they just occur
because the decreased activity of subcortical neuromodulators during sleep do
not prevent their occurrence. Spontaneous patterns emerge and evolve without
2013). Interaction between the body and mind was natural to the Greek philosophers. Aristotle obvious trigger events. A pertinent analogy is the emergence of a tune from the
might have said that we learn about the world by our actions on it. Only a er the concept of interaction of jazz musicians. When they nd a coherent melody, it decreases
the soul fused with the mind came the full separation, especially with Descartes’s declaration the degree of freedom of each player and forces them to time their actions to
of the independence of res extensa (corporeal substance or the external stu ) and the God- the emergent tune. But—and this is where it gets interesting—emergence is not
given res cogitans (the thinking thing or soul). Perhaps Henri Poincaré was the rst thinker
who speculated that the only way sensations become relevant and turn into an experience is by
the sum of multiple decomposable physical causes but a result of a multitude of
relating them to the body and its movements. “To localize an object simply means to represent interactions. Emergence is not a “thing” that a ects neurons or humans. 29
oneself the movements that would be necessary to reach it” (Poincaré, 1905; p. 47). e volume
by Prinz et al. (2013) is an excellent update on the renewed interest in action science. Tolman’s
cognitive map theory (1948) is another important and major departure from the outside-in 28. e term “self-organization” was introduced by the British psychiatrist W.  Ross Ashby
framework but it is not action-based. (Ashby, 1947). Increasing order in self-organized systems is discussed in Kampis (1991), an
accessible text on self-organization. For a more complex treatment of the subject, see Maturana
47. Once a brain pattern acquires meaning (i.e., it has been calibrated through action), it re-
and Varela (1980).
mains meaningful as long as memory is intact, without further action (Chapters  5 and 8).
Learning through action also applies to species knowledge acquired through evolutionary se- 29. Hermann Haken (1984) introduced the term “synergetics” to explain emergence and
lection (Watson and Szathmáry, 2016). self-organized patterns in open systems far from thermodynamic equilibrium. His “order
46

46 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 23

prediction is also common in neuroscience data. Yet correlating, explaining, BLANK PAP ER OR PREEXISTING CONSTRAINTS?
predicting, and causing are di erent.
Like causality, the notion of probability is o en entangled in metaphys- e choice of outside-in and inside-out, action-centric frameworks also shapes
ical disputes, and its interpretation appeals to mysterious properties. 25 Such our ideas about fundamental brain operations. One of the oldest outside-in
considerations led Bertrand Russell to suggest that we remove causation views was formulated by Aristotle, who suggested that we are born with a
from scienti c thinking: “In the motions of mutually gravitating bodies, there blank slate (or tabula rasa) on which experiences are written.48 e tabula rasa
is nothing that can be called a cause, and nothing that can be called an ef- view is an almost inevitable assumption of the outside-in framework because,
fect; there is merely a formula.”26 In the mathematical equations of physics, under the empiricist view, the goal of the brain is to learn and represent the
there is no room or need for causes. Do we need them for neuroscience? truth with its veridical details. is view has in uenced thinking in Christian
Paradoxically, Russell’s quantitative probabilistic formulation managed to save and Persian philosophies, British empiricism, and the Marxist doctrine, and
rather than eliminate causation as the tool of scienti c explanation. e con- it has become the leading school of thought in cognitive and social sciences.49
cept of probabilistic causation replaced deterministic substrate-dependent Similar to the free will idea that humans can do anything, tabula rasa thinking
causation.27 Spurious regularities, a headache for deterministic causality, had implies that anything can be written on a blank brain. Although I have never
become quanti able and therefore could be contrasted with “true” regularities. met a neuroscientist colleague who would openly subscribe to the tabula rasa
However, making causal inferences from correlational data is harder than is view, many experiments and modeling studies today are still performed ac-
usually thought, mainly because probabilistic causation is a statistical inference, cording to this associationist philosophy-based framework, or its modern-day
whereas causation should involve physical substrates (“enhanced neuronal syn- version: “connectionism.” e main idea of connectionism, especially in cog-
chrony is the cause of epilepsy”). e correlation between various genes and nitive psychology and arti cial intelligence research,50 is that perception and
schizophrenia leaves no doubt that aberrations in such genes increase the cognition can be transcribed into networks of simple and uniform neurons,
probability of the disease. On the other hand, claiming that my friend’s schiz- which are interconnected relatively randomly with largely similar synaptic
ophrenia was caused by particular genes is hard to support. Correlations need connections. Nothing is further from the facts, as I explain in Chapter 12 and
to be “challenged” to improve their explanatory power. Adding true meaning brie y summarize here.
to correlations requires a second validation step, a hypothesis-testing practice Early behavioral observations already argued against the idea of the brain
based on some perturbation of the explanatory independent variable. For this as a blank slate. Researchers demonstrated repeatedly that animals do not as-
reason, we need to do many independent experiments and worry more about sociate everything equally and cannot be trained to do all tricks the exper-
replicability than causal explanation. imenter expects them to do. Behaviors that relate to the animal’s ecological
niche can be trained easily because the brain is predisposed or “prepared’ to
do things that have survival and reproductive advantage. For example, “spon-
taneous alternation,” the tendency in rodents to choose di erent paths during
25. Despite the solid mathematical bases of probability and statistical inference, the connec- foraging, is an instance of biological preparedness for the rapid acquisition
tion between probability relations and correlational dependencies is o en vague.
26. Russell et al. (1992). e cause (C) has a causal in uence on the e ect (E) if conditional
probability (E ı C) for E reliably changes under free variation of C. In an open, linear system,
48. e tabula rasa concept was more explicitly developed by the Persian Muslim Avicenna
this requirement is not an issue, but in complex interconnected systems, such as the brain, the
(Abu Ali al-Hussain Ibn Sina). Avicenna was an important precursor of British empiricists.
assumption of free variation of C is most o en not tenable.
His syllogistic, deductive method of reasoning in uenced Francis Bacon, John Locke, and John
27. Moyal (1949); Churchland (2002); Mannino and Bressler (2015). While the term “probabi- Stuart Mill and continues to in uence scienti c reasoning even today. Steven Pinker’s Blank
listic” is a useful theoretical construct, its use in real life may not be always practical. Consider Slate (2003) is an excellent overview on this old philosophical problem.
Richard Feynman’s explanation of the increase of entropy as de ned by the second law of ther-
49. Read, for example, Popper (1959).
modynamics, which says that entropy is always increasing. “ ings are irreversible only in a
sense that going one way is likely, but going the other way, although it is possible and is ac- 50. e classic works on connectionist networks have been summarized in Rumelhart et al.
cording to the laws of physics, would not happen in a million years” (Feynman, 1965, p. 112). (1986) and McClelland et al. (1986). e much more sophisticated recent neuronal networks
Similarly, in probabilistic causation, everything can be explained by singular causes in prin- also fall under the tabula rasa framework. e leading author on connectionism in cognitive
ciple, but most of them may be statistically and practically improbable. psychology is Donald Hebb (1949).
24

24 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 45

of species-speci c learning. Returning to the same location for food within a springtime brings owers and higher birthrates. Nobody argues against this
limited time window is not an e cient strategy because choosing an alternate reliable relationship. However, even if this relationship is acknowledged, the
route will more likely lead to reward. In contrast, associations that would be issue of causality remains ambiguous because the likely cause is not that spring
detrimental to survival are called “contraprepared.” For example, it is virtually induces labor, but that couples spend more time together during the long dark
impossible to train a rat to rear on its hindlimbs to avoid an unpleasant elec- nights of winter. Similarly, there is a reliable correlation between the blooming
tric shock to the feet since rearing is an exploratory action and incompatible of owers and birthrate increases. If we are not aware of the nonrecorded
with the hiding and freezing behaviors deployed in case of danger.51 A prac- “third-party” variable (spring), we may suspect a causal relationship between
tical message here is that if training takes weeks or longer, the experimenter these two variables as well. is example may seem silly for an easily relatable
should seriously consider that the brain signals associated with such extended relationship such as owers and birth rate, but when we are observing variables
shaping of behavior may not reveal much about the intended question because for which we do not have the luxury of everyday intuition, the reader can see
each animal may use di erent tricks to solve the task, and these may remain how easy it would be to mistake correlation for causation. Hidden or permis-
hidden for the experimenter. sive common causes are the most frequent sources of misinterpretations based
Neuroscience has also accumulated much experimental evidence against the on correlational arguments.24 For these reasons, deterministic causality has
tabula rasa model. e most important is the recognition that a great deal of been challenged in several other areas of science as well.
activity in the brain is self-organized instead of being driven by outside signals.
is self-generated persistent activity is supported by the numerous local and
brain-wide neuronal rhythms, which we will discuss in Chapter 6. For now, it Probabilistic Causation
is su cient to note that these oscillations are not only part of and help to stabi-
lize neuronal dynamics but also o er a substrate for syntactical organization of Events can follow each other with high probability, very low probability, or
neuronal messages. Each period of an oscillation can be conceived as a frame anything in between, so probabilistic in uences uctuate over a wide range
that contains a particular constellation of spiking neurons. Metaphorically, we from 0, meaning chance, to 1, meaning certainty. However, it is not obvious
can call it a “neuronal letter.” In turn, the cycles of the many simultaneously which aspects of the distribution are “causal” and which are not. Probability
acting oscillations can concatenate neuronal letters to compose neural words is an abstract mathematical concept and not a “mover” needed for causation
and sentences in a virtually in nite number of ways. is brain syntax is the in Newtonian physics. Determining probabilistic causation is a two-step pro-
main focus of the discussion in Chapters 6 and 7. cess:  the rst step is the description of patterns and probabilistic facts (e.g.,
the substrate-free illustration of Figure 2.1) and the second is to draw conclu-
sion from the measured probabilities. is second step is not inherent in the
observed facts, however, but is an interpretation of the experimenter. When
we observe a regular succession between things, we feel compelled to assume
a cause. But how likely does succession have to be to reach a threshold to be
51. Each of these telegraphic statements represents large chapters of cognitive and behavioral
psychology. Skinner believed that any behavior could be shaped by his method of “successive perceived as “causal”? Might your threshold be di erent from mine? We single
approximations” and was able to teach rodents and other animals to do some extraordinary out high- and low- probability events from a continuous distribution and call
things. However, his students (Breland and Breland, 1961) countered that animals o en mis- them cause and chance. Confusion about chance, low probability, and high
behave and develop “autoshaped” or “superstitious rituals” (e.g., Catania and Cutts, 1963). e probability is a major contributor to wishfully interpreted “causal” relationships
literature on autoshaping is a rich source of information for the expected surprises that today’s
users of the increasingly popular “virtual reality” tasks will have to face (Buzsáki, 1982). e pre-
between gender and performance, genes and intelligence quotient (IQ), and
pared and contraprepared spectrum of behaviors, developed by Martin Seligman, had a great religion and moral values. Con ation between explanation (postdiction) and
in uence on human phobia research (Seligman, 1971, 1975). Phobias (e.g., to spiders, snakes,
and heights) can be viewed as a prominent prepared category because they are commonly asso-
ciated with objects and situations that have threatened us throughout our evolutionary history. 24. Calcott (2017) discusses the distinction between permissive and instructive factors. ese
Joe LeDoux, the leading authority on threat behavior, ampli es this view further by explaining factors are not really causal, yet their presence may o en be critical. Permissiveness can be
that brain circuits that detect and respond to threats in rodents are not necessarily the same as viewed as a gating function that allows the cause to exert its e ect. Instructive factors may be a
those that produce conscious fear in humans (LeDoux, 2015). prerequisite but can induce a nal e ect only in conjunction with other factors.
4

44 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 25

But many situations common in neuroscience will result in Granger’s for- Wide- Range Distributions: Good Enough
malism giving the incorrect answer. Consider the following real-life observation. and Precise Solutions
In recording from groups of neurons in area A and area B in the brain, we nd
a reliable correlation between the peaks of spiking activity of the populations, As an alternative to the empiricist outside-in tabula rasa view, I raise the pos-
with neuronal spikes in area A, on average, consistently and reliably preceding sibility that the brain already starts out as a nonsensical dictionary. It comes
the peak spiking activity in area B. Logic and Granger’s formula dictate that it with evolutionarily preserved, precon gured internal syntactical rules that can
is likely that activity of neurons in area A causes the increased activity in area generate a huge repertoire of neuronal patterns. ese patterns are regarded as
B. But we may be surprised to learn that axons of neurons in area B innervate initially nonsense neuronal words which can acquire meaning through experi-
neurons in area A unidirectionally; in short, area B is upstream to area A. To ence. Under this hypothesis, learning does not create brand new brain activity
explain the contradiction between the anatomical ground truth and the physio- patterns from scratch but is instead a “ tting process” of the experience onto a
logical observations, let’s assume that neurons in area A respond to inputs from preexisting neuronal pattern.
B with a low threshold (i.e., before neurons in area B reach their population e precon gured brain is supported by a mechanism I refer to as “high-
peak activity) and stop ring a er their rst spike. is results in peak spiking diversity” organization (Chapters 12 and 13). Diversity and large variations are
of the population in area A earlier than the peak in area B! Neuronal patterns the norm in the brain, whether we are analyzing its microscopic or macroscopic
o en have such properties. In fact, it is a basic feature of a common feedforward connectivity or dealing with its dynamics at a small or a large scale. Neuronal
inhibition motif found throughout the brain. Neurons can stop ring under ring rates, connection strength between neurons, and the magnitude of their
continuous and even increasing drive because intracellular mechanisms may concerted action can vary by three to four orders of magnitude. I speculate that
prevent sustained ring in neurons of area A (such response is known as spike this diversity, which respects a logarithmic rule (Chapter  12), is the essential
frequency adaptation) or because strong feedback from inhibitory neurons backbone that provides stability, resilience, and robustness to brain networks
prevents the occurrence of further spiking. Without knowing the anatomical against competing requirements for wide dynamic range, plasticity, and redun-
connections and biophysical/circuit properties of neurons, Granger’s formula dancy. ere are many advantages of a precon gured brain52 over the blank
gives the false conclusion that area A causes activity in area B. is example slate model, the most important of which is stability of brain dynamics because
illustrates just one of the many pitfalls that have to be considered in any “causal” adding new experience a er each learning event does not perturb much the
analysis.23 Of course, this example does not disqualify the existence and impor- overall state of neuronal networks. I will discuss that even the inexperienced
tance of cause-and-e ect relationships; it only reveals that the method is not brain has a huge reservoir of unique neuronal patterns (Chapter  13), each
always appropriate. of which has the potential to acquire signi cance through the experience of
Hidden common causes are the most frequent sources of misinterpretations unique events or situations.
based on correlational arguments. Take as an example the observation that e tails of these wide-range distributions, such as neurons with high and
slow ring rates and strong and weak synaptic connections, can support seem-
ingly opposing functions, such as judging a situation as familiar or novel.
(Shafer, 1996), propensity scores (Rosenbaum and Rubin, 1983) and convergent cross mapping Neural circuits involving a minority of highly interactive and well-connected
(Sugihara et al., 2012). Applying these methods to the brain o en requires the assumption that
the brain is a deterministic, nonlinear dynamic system, although this assumption is question- neurons may appear as a constraint. However, the bene t of such apparent
able (Friston et al., 2012). constraint is that the collective action of such oligarchic circuits with skewed
distribution of resources allows the brain to generalize across situations and
23. Another striking illusion of causality has been described in the entorhinal-hippocampal
circuits. While super cial layer neurons of the entorhinal cortex are known to innervate
o er its “best guess” under any circumstances, so that no situation is regarded
hippocampal neurons but not vice versa, population activity in the downstream hippocampal
neurons within theta oscillations cycles peak earlier than peak activity in the upstream
entorhinal neurons (Mizuseki et  al., 2009). e explanation in this case is that it is the fast 52. “Precon gured” usually means experience-independent. e backbone of brain connec-
spiking minority of neurons that re early in the theta phase, and they have a much stronger im- tivity and its emerging dynamics are genetically de ned (Chapter 12). In a broader sense, the
pact on their target interneurons than the majority of slow ring cells. We discuss in Chapter 12 term “precon gured” or “preexisting” is also o en used to refer to a brain with an existing
that skewed distributions are the main reason that mean population-based calculations are knowledge base, such as mine at this instant of time a er numerous years of action-based
o en misleading. calibration.
26

26 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 43

as completely unknown. I call this aspect of neuronal organization the “good- knowledge of the real world is limited by inevitably noisy measurements with
enough” brain. e good-enough brain uses only a small fraction of its re- limited precision, and such wisdom does not o er much solace. Even if we did
sources, but its member neurons are always on the alert. On the other hand, have complete knowledge of the universe’s past, the time needed to compute
the remaining majority of slow- ring member neurons with weak synaptic the future might be as long as the elapsed past.
connections have complementary properties, such as high plasticity and slow
but precise and reliable processing. is other extreme of network organization
comprises the “precision” brain, whose main job is to rede ne a situation or Troubles with Causation
an event and determine its distinctness in high detail. e “good enough” and
“precision” aspects of brain organization complement, and in fact, assume, each Problems with causality surfaced early in contemporary physics with the birth of
other (Chapters 12 and 13). quantum physics, nonlinear dynamics, chaos theory, and quantum entanglement.
In the precon gured brain model, learning is a matching process, in which Quantum physics underwent a multitude of interpretations, including determin-
preexisting neuronal patterns, initially nonsensical to the organism, acquire istic causality. On the other hand, in the framework of general relativity, time be-
meaning with the help of experience. As discussed earlier, attaching meaning came symmetric and the foundations of cause–e ect logic were shaken.
or signi cance to any neuronal pattern requires grounding, which can be To understand complex economic data, Clive Granger reintroduced the
provided only by action, the ultimate source of knowledge (Chapter 3). In time-asymmetry axiom (Chapter 10) and a novel statistical method for testing
contrast to the tabula rasa model, in which the complexity of neuronal dy- causality. “ e past and present may cause the future, but the future cannot
namics scales with the amount of experience, in the precon gured brain cause the past.” His method provides predictability, as opposed to correlation,
network homeostasis is only minimally and transiently a ected by new to identify the cause–consequence relationship with the key requirement of
learning because nothing is added to the circuitry; only rearrangements separability; that is, that information about a causative factor is independently
take place. e total magnitude of plasticity in brain circuits and the overall unique from the variable. Strictly speaking, Granger’s method does not present
spiking activity remain constant over time, at least in the healthy adult brain a de nition of causality; it only de nes a metric by which we can measure how
(Chapter 12). I make no claim that my strategy is de nitely right, but the al- much knowledge of one signal can enable you to predict another signal at a
ternative formulation that I am suggesting might provide di erent insights later time. It is ambiguous to the underlying source of predictability.
than the current representational framework. Yet the absence of the cause and e ect between activities in brain regions does
In contrast to the precon gured brain, the outside-in framework assumes not prevent neuroscientists from talking comfortably about causal inferences in
that there are obvious things in the outside world, which we o en call in- correlational data from the brain-derived data all the time. Granger’s method
formation, and the brain’s job is to incorporate or absorb that information. has become popular in neuroscience in attempts to interpret cause-and-e ect
But how? relationships, especially in complex functional magnetic resonance imaging
(fMRI), electroencephalograms (EEG), and magnetoencephalographic (MEG)
data.22
THE BRAIN “CODES INFORMATION”: OR DOES IT ?
is o en-heard statement should communicate a lot of content to everyone. 22. Statistical models used for testing causality are almost always association-based models
A er all, we live in an Information Age in which everyone is familiar with applied to observational data. e theory itself is expected to provide the proof of causality,
such terms as “information” and “coding.” But when we attempt to explain but, in practice, the ground truth is o en absent (see Granger, 1969; Wiener, 1956; Mannino
and de ne information coding in the brain, we pause because we do not and Bressler, 2015). Causality as a tool of inference has grown from a nebulous concept into
a mathematical theory with signi cant applications in many elds of sciences (Pearl, 1995).
have a widely accepted, disciplined de nition. When terms are not de ned
In addition to Granger’s causality measures, other methods, such as directed coherence, par-
precisely, they o en acquire many disguised interpretations and, therefore, tial coherence, transfer entropy, and dynamic causal modeling, have been used to assess the
convey di erent meanings to di erent individuals. “Is my grant proposal direction of activity ow in the brain (Barnett et  al., 2009; Seth 2005). “Proper” statistical
rejected or approved?” From the perspective of information theory, the methodology for testing causality has been advanced, such as specially designed experiments
or association-based, specialized causal inference methods for observational data, including
causal diagrams (Pearl, 1995), discovery algorithms (Spirtes at al., 2000), probability trees
42

42 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 27

the two. Hume thought that we do not experience the cause itself; we simply answer represents a single bit of information (yes or no). 53 But for the receiver
infer it from the regular appearance of the cause and e ect. Immanuel Kant of such information (me), it induces profound and long-lasting changes in
countered Hume’s interpretation by arguing that there is a universal principle my brain.
of causality that cannot be based on experience. Kant promoted the idea that While the mathematical formalism of information theory is neutral to in-
the realm of things in themselves (das Ding an sich) is the cause of the realm formation “content,” the same cannot be said for the brain. For a behaving
of appearances (phenomena).19 Our experience is shaped by a priori causality organism, “decoding” information cannot be divorced from “meaning”:  the
because the mind interprets the sensory inputs arriving from an objective, experience-based prediction of potential consequences of that informa-
observer-independent world. tion. ere is no such thing as invariant information content in a stimulus. 54
Hume’s idea of representation has shaped the thinking of cognitive “Features” or attributes of a stimulus are not objective physical characteristics
psychologists and neuroscientists alike and became the de facto scienti c ex- of the stimulus. Instead, the relevance of the stimulus depends on changing
planation of how “attributes” of objects become neuronal “representations.” internal states and the brain’s history with similar situations. In the following
His subjectivism has been modernized to a subjective realism in which the chapters, I hope to show you that this situational relevance is not a small matter
mind represents, but does not faithfully mirror, the outside reality. erefore, but has a profound and observable e ect on the nature of neural activity and
my representation of truths and judgments can only become similar to your its relationship to the sensory environment. e same visual stimulus (e.g.,
representations through a mutually agreed consensus. Each brain in isolation the rejection letter) evokes very di erent neuronal patterns in a frog’s brain,
would generate a somewhat di erent representation of the same object, color, my brain, or the brains of other humans. e induced pattern even in my
or sound. If objective and subjective reality are separate, then our mental in u- brain depends on its current state. If the letter arrived the same week when
ence of causation does not necessarily apply to the relationships present in the I received the news about other rejections, it would induce a di erent e ect
physical world. compared to an alternative situation when the letter arrived a day a er I won
In a deterministic system with one-way interactions among the constituents, the lottery jackpot.
cause-and-e ect relationships can be reliably described by Aristotelian–
Humean arguments because the cause regularly precedes and induces the
e ect.20 Determinism heavily rests on causality. In the Newtonian world, cause-
and-e ect relationships seem to work well since nothing moves unless it is 53. Once agreed by the sender and receiver agents, everything can be represented by symbols;
for example, the head or tail outcome of coin ipping by 1 and 0. Each ip produces one bit of
being moved.21 If the initial conditions and all variations are known in a system,
information, which can be concatenated by some agreed rules. In the American Standard Code
in principle, all consequences can be accounted for. However, in practice, our for Information Interchange (ASCII), seven bits represents a character; for example, 1100101 is
code for the letter “e.” Each block representing a single letter is separated by an agreed SPACE
code (0100000). As in Morse code or any other coding scheme, spaces are the essential part
19. Interestingly, Muslim theology makes a similar distinction between batin (things-in- of packaging messages (Chapter 6). Without space codes, the information may become com-
themselves or noumena) and zahir (phenomena). Muslim scholars believe that categories of pletely undecodable. Physics uses information as a conceptual tool to understand interactions
causality and existence apply only to the noumenal realm, but not to zahir. Muhammad exists, and relationships. Information can be mathematically de ned as the negative of the logarithm
but we do not know how (El-Bizri, 2000); we just have to believe it. For a concise exposure to of the probability distribution (I = log2 N). When n = 2, information I is 1, thus the unit of infor-
fresh thinking about causation, see Scha er (2016). mation is a choice between two possibilities, measured in “bits” (Shannon, 1948). e meaning
of the observed events does not matter in this de nition of information.
20. e concept of cause, as usually presupposed in standard causal modeling, is a “mover” and
cannot be used to reduce probability. Can the absence or reduction of something be a cause? 54. I accept that one solution would be to stick to the framework of Shannonian theory or en-
Inhibition in the brain is such a case. e primary role of a special, dedicated group of neurons tropy (Shannon, 1948) and use it to de ne relationships between world events and brain events.
in the brain is to inhibit excitatory neurons and each other. us, elevation of ring rates of ex- A connection to information theory is free energy that quanti es the amount of prediction error
citatory cells can be brought about by excitation via other excitatory neurons or by reduction of or surprise. Prediction error is the di erence between the representations encoded at one level
inhibition (“disinhibition cause?”). and the top-down predictions generated by the brain’s internal model: a comparison. More for-
mally, it is a variational approximation to negative log likelihood of some data given an internal
21. Nearly every tool and machine is based on the logic of cause-and-e ect relationship. model of those data (Friston, 2010). However, in this de nition, we should abandon the wider
Machines may represent the exteriorized re ection of this logic (Chapter  9), or, conversely, use of information as it is typically applied in neuroscience.
the functioning of the tool and machine components could have led to internalization of the
observed logic of things (Chapter 5).
28

28 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 41

Computation performed by computer programs or machines is referred to, Although Aristotle’s formulation of the cause-and-e ect relationship has
in general, as “information processing.”55 In reality, the information is not in undergone extensive revisions over the past millennia, it has remained the
the processing. It becomes information only when interpreted by an observer main tool of reasoning in the laboratory, likely due to its simplicity. However,
that recognizes its meaning and signi cance, be it a human interpreter or a me- even in everyday life, one can o en see how such logic fails. My shadow follows
chanical actuator. A computer program can e ectively control a robot or other me, and without me it never occurs. My shadow and I (o en) correlate. I am
man-made machines, thus giving rise to the illusion that information resides the necessary (but not su cient) condition because my shadow is not always
in the program. However, the process is based on a human-designed solution detectible. In darkness, the suspected causal rule is violated because the occur-
even if it involves a complex trial-and-error learning process, including “deep rence of my shadow is conditional upon the presence of light. However, light is
learning” in arti cial intelligence programs.56 In short, information is not in- not a cause and not even a co-contributing component but a condition.18
herent in the computation (in machines or brains) but becomes such when it Similar violations are found in neuroscience. A weak stimulus, such as an ant
is interpreted.57 climbing on my leg, can sometimes induce reliable ring of neurons in multiple
In neuroscience, information is o en understood as acquired knowledge, a brain circuits and is prominently noticed (an e ect). At other times, the same
broader concept than in information theory. Claude Shannon warned us that cause (the ant) may not induce a noticeable e ect at all (for example, when I am
information theory is “aimed in a very speci c direction, a direction that is not intensely preoccupied with writing this book chapter). Again, the change in
necessarily relevant to such elds as psychology.”58 In the forthcoming chapters, brain state is not considered a co-cause, but it is a necessary condition for the
I  discuss why the brain is not an information-absorbing, perpetual coding detectable cause–e ect relationship. Later, I highlight some of the pitfalls of our
device, as it is o en portrayed, but a venture-seeking explorer, an action- everyday reliance on causation as a scienti c rule. My goal, of course, is not to
obsessed agent constantly controlling the body’s actuators and sensors to test disqualify formalized reasoning but to call attention to why the simple recipe
its hypotheses. e brain ceaselessly interacts with, rather than just detects, of logic fails in neuroscience, especially when things become really interesting.
the external world in order to remain itself. It is through such exploration Because of the omnipresence of logical inference and causation in sci-
that correlations and interactions acquire meaning and become information ence, both philosophers and scientists have undertaken multiple rounds of
(Chapter 5). Brains do not process information: they create it. explanations and attempts to re ne them. Each time, these concepts were
Now let me try to give a broad de nition of coding as it applies to the brain. approached with a particular point of view and purpose, and, not unexpect-
In essence, coding is an agreement for communication between the sender and edly, the resulting explanations vary. Philosophers are primarily interested in
receiver, making the content of the encoded information a secret to outsiders. the ontological question of whether causation is an inherent property of the
physical world independent of the human mind or whether it has epistemic
limits. As we learned in Chapter 1, David Hume’s subjectivism separated the
55. e term “processing” has an interpretative connotation. It refers to a procedure that takes objective world from our rst-person, subjective experiences and introduced
something and turns it into something else, a sort of format transformation. is is what is usu-
ally referred to as “representation.” e o en used “information-processing” expression relies
the idea that our mind transforms and “represents” the objective reality. He
on this view and tacitly implies that representation is the nal product. However, representa- criticized rationalists who assumed that we can acquire knowledge of the world
tion is not a thing; it is a process or an event (Brette, 2015, 2017). erefore, representation (and by trusting our minds. For instance, we observe a regular relationship between
perception) cannot be described as a transformation of one thing into another thing. See also lightning and thunder, but the inference that lightning causes thunder is only
Jeannerod (1994).
a subjective mental operation, so there may be no objective causality between
56. Deep learning (or hierarchical deep machine learning or deep structured learning) is
part of a broader family of machine-learning algorithms that attempts to model high-level
abstractions in data by using a “deep graph” with multiple processing layers, akin to the asso-
ciational networks sandwiched between sensory and motor areas in the brain. For the origin 18. In Boolean logic, it could be called a “gate.” In algebra, Boolean logic follows Aristotelian
of this highly successful and rapidly expanding eld, see Hinton et al. (1995) and Sejnowski logic in a formalized manner, where all values are reduced to a binary true (1)  or false (0).
(2018). Everything in the middle is excluded or forced to belong to the true or false category (you either
perceive something or not). e binary nature of Boolean algebra is what makes it so suitable
57. e thoughtful paper by Chiel and Beer (1997) outlines a broad program for brain-inspired
for computer science, using the true and false statements with the fundamental operators And,
robot design. See also König and Luksch (1998).
Or, Xor and Not. Boole’s symbolic system makes computation and logic inseparable. Early brain
58. Shannon (1956). models assumed that brain operations follow a Boolean logic (McCulloch and Pitts, 1943).
40

40 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 29

thus eliminating potential causes rather than proving the truth. Scienti c If the code is not known to the receiver, the packets carry no inherent infor-
knowledge is tentative and not de nitive; it is progressive but not absolute. mation (Chapter 7). ere are many forms of coding (e.g., Morse code or the
genetic code), but the fundamental features remain the same. e encoding–
decoding process is done via a cipher, 59 a transforming tool that encrypts
Cause and Effect messages by mutually agreed syntactical rules, so the information looks like
nonsense to the uninitiated. Encryption is not always intentional, of course;
Causality is a subtle metaphysical notion, an abstraction that indicates how sometimes it just happens. Evolution has no aim of secrecy, yet biological codes
events in the world progress. e common notion of causation can be traced are o en mysterious to us—that is, until we uncover the cipher. In the outside-
back to Aristotle and is a statement about how objects in uence each other in philosophy, the words in James’s list are the senders and the brain is the re-
across space and time. From the correlation graph (Figure 2.1), we cannot make ceiver. However, terms such as “emotion,” “will,” “attention,” and others are not
the claim that x is the cause of y without extra manipulations. 16 Causation is de- senders of anything and possess no inherent information. Space and time are
ned as a chain of dependency, where the cause always precedes the e ect: the not sensed directly by the brain either (Chapter 10), and no messages are being
cause is in the past of a present event. An event, A, is a cause of event B if the sent by them. If we want to understand information, we should start with its
presence of A is necessary for B to happen. For A to have any e ect on B, event creator: the brain.
A has to happen rst.17 David Hume remarked that “if the rst object had not
been, the second never had existed.” In a deterministic framework, the pres-
ence of the cause is a necessary and su cient condition for an e ect to happen. CODE- BREAKING EXERCISES
Causality and determinism are o en intertwined. Of course, an e ect can have
many co-occurring causes, but we are compelled to believe that we should be e American writer, Edgar Allen Poe, whose hobby was cryptography, believed
able to identify at least one. For example, when an individual gene alone cannot that “human ingenuity cannot concoct a cipher which human ingenuity cannot
account for a mental disease, it can still be viewed as a conditional cause when resolve.” But modern science claims that it is possible to encrypt messages that
co-occurring with other causes. When enough necessary components occur no human or machine can break.60 Does the cipher of brain dynamics fall into
together, they may be su cient to induce the disease. Under this framework, this highly secured category, or can it be outsmarted?
the triad of correlations and necessary and su cient conditions complement As we search for ciphers in the brain, we may learn from previous deciphering
each other. In another formulation, one can distinguish an immediate or prox- successes. e Englishman omas Young and the Frenchman Jean-François
imal cause (also called precipitating cause:  a stroke in the motor area caused Champollion are credited with deciphering Egyptian hieroglyphics by using
paralysis), predisposing cause (she got upset), and sustaining cause (she had high a written cipher:  the Rosetta Stone. is most famous artifact of the British
blood pressure). Neither the predisposing nor the sustaining course alone could Museum contains the same script in three languages (Greek, Demotic, and
be designated as necessary and su cient condition. Yet the proximal cause may hieroglyphics). As a polymath, Young’s strategy was to nd correspondences of
not have happened without the sustaining and predisposing causes. well-known names of gods and royalties, such as Ptolemy and Cleopatra, be-
tween the words in the Greek text and the glyphs. is approach provided fast
initial progress but could not uncover the generative rules of hieroglyphic lan-
16. One step forward in interpreting a graph is to vary the independent variable. In this case,
the designation of the independent and dependent variables is critical because regression of x guage. at task needed Champollion’s genius and his understanding of syntactic
on y is not the same as regression of y on x.
17. is logic stems from a Newtonian framework, which has a “time arrow.” When this funda-
59. From the Arabic sifr, meaning “zip” or “blank.” Deciphering is the process of breaking the
mental requirement of causation (i.e., the notion of time) is removed from its de nition, things
secret (i.e., decoding). A er reading extensively about ciphers, I realized that brain research
become rather complicated. In quantum-mechanical examples—such as entanglement—two
badly needs such concepts. Yet very little speci c research has tried to uncover and understand
particles can have interdependent or “entangled” properties. e usefulness of “classical” causa-
such mechanisms, even though we pay extensive lip service to “breaking the brain’s code.”
tion and the role of time as part of causality are intensely debated in both general relativity and
quantum mechanics (Hardy, 2007; Brukner, 2014; Ball, 2017). As logic is timeless, it is not clear 60. Poe (1841). An accessible review on the current state of ciphers is by Eckert and Renner
why time should be part of causation. We visit the nature of time and its alleged role in brain (2014). e key is to make choices that are independent of everything preexisting and are hence
operations in Chapter 10. unpredictable: a device-independent cryptogaphy.
30

30 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 39

rules in Hebrew, Syriac, Greek, Arabic, Aramaic, Persian, and, critically, Coptic without accepting such causality rules. e idea of causation is quite di erent in
and Demotic. He hypothesized that phonetic signs were not just used for the Asian cultures than in European cultures. For example, in Buddhist philosophy,
names of royals but also represented a general rule. is insight suggested that nothing exists as a singular and independent entity because things and events
Egyptian hieroglyphs could be read and understood by transforming the glyphs depend on multiple, co-arising conditions.12 A Buddhist monk or teacher lights
to sounds and concatenating them.61 us, the nonsensical messages of a for- three incense sticks leaning against each other, as a demonstration that if any of
gotten language became information. However, without the Rosetta stone, the the sticks falls, the others will fall as well. Similarly, Confucians approached cau-
secrets of the Egyptian hieroglyphs might well have remained forever hidden. sation in an all-encompassing manner, searching for a network of connections.
My second example of a cipher is connected to the most celebrated (post- Historical changes are explained by coincidences of events instead of a partic-
humously) code breaker, Alan Turing, a genius by many accounts. He and his ular cause. In Muslim philosophy, causal relations have no inherent necessity
team at Bletchley Park, north of London, are credited with breaking the code of at all. Instead, they are just expressions of what God does. 13 erefore, it should
the secret messages of the German Navy in World War II.62 With all my admi- not surprise us that people from di erent cultures—yet with the same kind
ration and respect for Turing’s contributions, I must make a correction: neither of brains—judge the a airs of the world surrounding us di erently.14 So why
he nor his team “broke the code.” ey received the cipher, the key syntactical do we assume that the European philosophy of science is the superior or only
tool of encryption, in the form of Enigma machines,63 from the Polish Cipher way to view the world and ourselves, especially given that many other cultures
Bureau early during the war.64 Without this access to the syntactical rules of the have achieved great scienti c discoveries? Is it really true that when the cause
cipher, the secret messages sent to the German U-boats by the Kriegsmarine is identi ed, the truth is found? Even Descartes hesitated to accept such a link.
Headquarters would likely have remained gobbledygook to the allies. Alan First, he states: “If a cause allows all the phenomena to be clearly deduced from
Turing’s Rosetta Stone was the Enigma’s hardware. it, then it is virtually impossible that it should not be true,” but then he imme-
Is deciphering a secret code possible without (some) a priori knowledge of diately backtracks and adds:  “Nevertheless, I  want causes that I  shall set out
the syntax? Our third example, discovery of the genetic code and the trans- here to be regarded simply as hypotheses” because we o en make “assumptions
lation of DNA to RNA to protein, was such a case. e cipher in this case which are agreed to be false.”15 e modern meaning of the term “hypothesis”
was the double spiral. e two strands of the helix consist of polynucleotides refers to a conjecture which can be rejected by observations and experiments,
made up from simpler units of nucleotides. Each nucleotide is composed of
a nitrogen-containing nucleobase—cytosine (C), guanine (G), adenine (A),

12. “Origination” (or pratityasamutpada in Sanskrit) refers to the principle of interdependent

61. Silvestre deSacy, Champollion’s master, had already found a relationship between pro- causation and explains the twelvefold chain that describes the endless rebirth of life (Dalai
nounced sounds of Demotic symbols and their meaning (Allen, 1960). Champollion is o en Lama and Chodron, 2017). In metaphysics, accidentalism denies the doctrine that everything
criticized for not giving enough credit to deSacy and Young. Eventually, 24 glyphs, representing occurs or results from a de nite cause.
single consonants, were identi ed, similar to an alphabet.
13. Leaman (1985).
62. Turing’s short and complex life continues to inspire generations of creative people. e
14. Carl Jung also rebelled against the Kantian axiom that “nothing can happen without being
movie e Imitation Game (2014) is perhaps the best known, albeit heavily romanticized, ac-
caused.” Instead, he argued for “an acausal connecting principle,” or the principle of “mean-
count of his code-breaking years and personal life.
ingful coincidences,” based on Eastern philosophies. He and the physicist Wolfgang Pauli
63. If the letters are shu ed properly, as can be achieved today with fast computers, the cipher coined the term “synchronicity” to replace causality as a principle of explanation. Explanation
is practically unbreakable. “cannot be a question of cause and e ect, but of a falling together in time, a kind of simulta-
neity” (Jung, 1973). Jung explained everything with synchronicity, including telepathy. For a
64. In a book about the history of British Intelligence in World War II, mathematician Marian
statistician, coincidence is not a surprising, meaningful event because “with a large enough
Rejewski noted, “we quickly found the [wirings] within the [new rotors], but [their] introduc-
sample, any outrageous thing is likely to happen with no apparent causal connection” (Diaconis
tion . . . raised the number of possible sequences of [rotors] from 6 to 60 . . . and hence also
and Mostelle, 1989). ese issues are eloquently exposed in Arthur Koestler’s book e Roots of
raised tenfold the work of nding the keys. us, the change was not qualitative but quan-
Coincidence (1973). [NeuroNote: Koestler attempted suicide in Lisbon in 1940 a er he failed to
titative” (as quoted in the appendix of Kozaczuk, 1984)  Another important predecessor of
return to England, but he survived. A er he developed Parkinson’s disease in his seventies, he
Turing’s decoding venture was the work of another British cryptographer, Dillwyn Knox, who
and his wife committed double-suicide.]
worked on deciphering Enigma-encrypted messages of the Nationalists in the Spanish Civil
War (Batey, 2009). 15. Descartes (1984).
38

38 THE BRAIN FROM INSIDE OUT Chapter 1. e Problem 31

vary the independent values experimentally and examine how the dependent or thymine (T)—as well as a monosaccharide sugar called deoxyribose and a
variable follows such changes. 10 Direct manipulation of physical objects is the phosphate group. e syntax is elegantly simple. A is always paired with T, and
most basic way for arguing the persuasive rule of cause and e ect. When a C paired with G, providing the rule for making the double-helix DNA: the hy-
child discovers that ipping a light switch can make the room appear and disap- drogen bonds between these matched pairs of nucleotides bind the nitrogenous
pear, he repeatedly practices it to learn that he is the sole causative agent of the bases of the two separate polynucleotide strands. is DNA sequence is ulti-
consequences. Similarly, when a rat presses a lever that delivers food, it repeat- mately translated into a string of amino acids, the building blocks that make
edly tests the consequences of its action. It seems that functioning in the world up proteins. is is done via RNA, which transcribes a particular set of three
requires animals to develop the concept of cause–e ect relationships. nucleotides into a particular amino acid.65 Using this messenger RNA as a tem-
Both mathematics and scienti c thinking require logic. All mathematical plate, the protein molecule is then built by linking neighboring amino acids.
theories can be reduced to some collection of axioms or set of rules on which ese examples of code-breaking practices and their players o er an impor-
an internally consistent system is built. ese axioms should not be violated. tant lesson for brain scientists in our search for a neuronal “code.” When the
But science is supposedly di erent. It has no a priori rules. We simply ob- syntax is known, breaking a code is a hopeful exercise. Uncovering the meanings
serve regularities in nature and call them “laws.” 11 When we interfere with of the hieroglyphs took a few thinkers and a few decades of hard work. A few
those regularities, we become part of the relationship and therefore may alter years of nerve-wracking search for correspondences, with several hundred
it. As Heisenberg, the father of uncertainty in quantum mechanics, famously people working together, were required to decipher the messages encrypted by
stated:  “What we observe is not nature, but nature exposed to our method the Enigma machines, even though the cipher was available. In the case of the
of questioning.” is healthy skepticism is important because our measuring genetic code, with well-understood goals (a combination of four elements to
instruments may a ect the cause–e ect relationship. code in triplets for 20 known amino acids), the e orts of several hundreds of
Every action we take has an e ect on the system we wish to observe. scholars over the span of a few decades provided a mechanistic explanation for
However, this cause-and-e ect situation runs deeper than just a di culty in a “miracle” of nature, the hypothetical force that produces complex organisms
experimental science—it forms the root of scienti c inference in the Western from simple seeds. So what should be our strategy for identifying the brain’s
world: once the cause is revealed, the problem is “explained,” the truth is found. cipher? Collecting correspondences and correlations is not enough; we also
In the traditional wisdom, there are no e ects without causes, and no logical need to understand the neuronal syntax (Chapters 6 to 8) and to “ground” our
argument can be delivered without causation. Western philosophy is built on a correlational observations with a rmer “second opinion” (Chapters 3 and 5).
foundation of causality. Yet much of the world has moved forward and onward Only with such independent knowledge can we hope to gure out the secret
meanings of the brain’s vocabulary.
At this point, the reader may rightly ask: Why does it matter to logically pro-
astronomer would disagree and claim that the impact is unpredictable given the complex ceed from inside out rather than the other way around? Can we not understand
relationships among the celestial bodies (and many more such perturbations, such as removal relationships either way? Answering this will require discussion of the complex
of other planets or perhaps galaxies, are needed to get a rm answer). ese are the types of problem of causality in neuroscience, the topic we are going to explore in the
questions that neuroscience constantly faces.
next chapter.
10. Without manipulating the system experimentally, one can add or leave out variables to “re-
gress out” noncausal agents. It is never foolproof, but o en this is the best one can do without
interfering with the system. Descriptive statistics that reveal correlations are a bit like passive
65. James Watson, Francis Crick, and Maurice Wilkins were acknowledged for their discovery
associations. In neither case can one make causal grounding. e source of knowledge is in the
of the DNA code with the Nobel Prize in 1962. e Russian physicist Georgiy Gamow had
manipulation.
already calculated, using mathematical modeling, that a three-letter nucleotide code could de-
11. Laws, as organization tools of scienti c thinking, have speci c theological associations. ne all 20 amino acids. He formed the “RNA Tie Club” with 20 regular members (one for each
Physicists, including Newton, who introduced these laws have assumed that things are “ruled” amino acid), and four honorary members (one for each nucleotide; Gamow, Watson, Crick,
or “governed” by God, who was regarded as changeless and eternal. In Christian and many other and Sydney Brenner) with the goal of wide cooperation in understanding how RNAs can build
religions everything must have a cause and laws and Order of Nature are assumed to be xed proteins. Yet the breakthrough came from outsiders, Marshall W.  Nirenberg, Har Gobind
in their forms (Toulmin and Good eld, 1965). However, laws are substrate-free. ey are not Khorana, and Robert W. Holley, who were awarded the Nobel Prize “for their interpretation of
physical things, so they cannot exert an e ect on anything else, short of divine interventions. the genetic code and its function in protein synthesis” in 1968. See http://www.nobelprize.org/
ey simply re ect regularities—the earth orbits around the sun. educational/medicine/gene-code/history.html.
32

32 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 37

SUMMARY CAUSALITY AND SELF- CAUSATION IN TH E BRAIN

In this introductory chapter, I presented some contrasting views and research In Figure 2.1, we identi ed some regularity between two variables: one is the
strategies that can provide complementary knowledge about brain operations. thing-to-be-explained (explanandum), called the dependent (y) variable, and
e rst comparison was between “outside-in” and “inside-out” strategies. the other is the thing-that-explains (explanans), called the independent (x)
e outside-in or perception-action framework has deep philosophical roots, variable.7 In more general terms, we can also call them the “de ned concept”
starting with Aristotle and most explicitly elaborated by the empiricist phi- and the “de ning concept.” Of course, our strongest desire is to conclude that
losopher David Hume. According to this framework, all our knowledge one variable predicts or causes the other. Ideally, the de ning concept is more
arises from perceptual associations and inductive reasoning of cause-and- general than the de ned concept, with some asymmetry between them. “ e
e ect relationships. is philosophy is the foundation of Western science and brain is a machine”; however, the other way around— “the machine is a brain”—
continues to in uence thinking in contemporary neuroscience. An inevitable does not apply. Accordingly, our choice to call one variable independent and
consequence of this perceptual representation-centered view is the assumption the other dependent is o en arbitrary and may re ect our preconceived bias.
of a hidden homunculus that decides whether to respond or not. In contrast, In some cases, we are able to manipulate the independent variable experimen-
my recommended inside-out framework takes action as the primary source of tally and observe its e ect on the dependent variable. In neuroscience we o en
knowledge. Action validates the meaning and signi cance of sensory signals observe quantities which we cannot manipulate directly or for which there
by providing a second opinion. Without such “grounding,” information cannot is no way to “independently” manipulate a single variable without indirectly
emerge. manipulating many more properties of the system. In this case, the choice of
Next, we discussed the relationship between the tabula rasa (blank slate) calling one variable independent and the other dependent is o en arbitrary
and precon gured brain models. In the empiricist outside-in model, the brain and re ects a bias. In fact, the plot itself is symmetric and its statistical descrip-
starts out as blank paper onto which new information is cumulatively written. tion would be the same if we reversed the x and y coordinates: the ratios would
Modi cation of brain circuits scales with the amount of newly learned know- not change. Yet, in neuroscience, we place behavioral observations or stimulus
ledge by juxtaposition and superposition. A contrasting view is that the brain is features on the x axis (in this instance, the distance walked) and brain meas-
a dictionary with preexisting internal dynamics and syntactical rules but lled ures on the y axis. is tradition takes its origin from the outside-in framework
with initially nonsense neuronal words. A large reservoir of unique neuronal as discussed in Chapter  1. Putting aside this historical issue, may we be able
patterns has the potential to acquire signi cance for the animal through ex- to make a causal inference from such data by plotting and evaluating the cor-
ploratory action and represents a distinct event or situation. In this alternative relational measurements in some other way?8 Equally important, we should
model, the diversity of brain components, such as ring rates, synaptic con- ask about our con dence in the prediction:  the probability of being right or
nection strengths, and the magnitude of collective behavior of neurons, leads wrong. Are these observational data enough, or do we need to do additional
to wide distributions. e two tails of this distribution o er complementary experiments? 9 To show convincingly that one variable causes the other, we must
advantages: the “good-enough” brain can generalize and act fast; the “precision”
brain is slow but careful and o ers needed details in many situations. 7. Hempel and Oppenheim (1948).
8. Our expectations o en in uence our interpretation. From the outside-in framework, one
can assume that the brain “measures” distances in the environment and represents them by
temporal o sets between the representing neurons. However, experiments designed to chal-
lenge such correlations showed that the temporal o sets between hippocampal neurons
are determined by internal rules of brain dynamics that remain the same across increasing
size environments. As a result, the spatial resolution of the neurons decreases in larger
environments. Distance representations in the hippocampus scale with the size of the sur-
rounding world (Diba and Buzsáki, 2008). us, brain network dynamics is the independent
variable that explains distance, not the other way around.
9. What if we remove Jupiter from the solar system? One astronomer would say that we can
calculate the consequences from the available data (no experiment is needed). But another
36

36 THE BRAIN FROM INSIDE OUT

any relationship and are thus substrate-free. For example, they could represent
an observed relationship between emotional magnitude and neuronal ring.
In contrast, the interpretation of the relationship always depends on their sub-
strate. Real things have a physical substance. Because the substrates of the 2
variables are missing from the graph, the experimenter adds them in his or her
interpretation to deduce a substrate-dependent mechanism that supposedly
Causation and Logic in Neuroscience
explains the observed relationships. is interpretational desire is the reason
we give various names, such as contingency, dependence, interdependence, in-
dication, anticipation, prediction, or forecasting (I might add fortune-telling or
crystal gazing), to such simple x-y relationships.6
In our example in Figure 2.1, we took distance as the independent variable
in the tradition of the outside-in framework. e distance between the two
place elds in the maze can be measured objectively by a ruler, and the goal is
to learn how distance is “represented” in the brain (e.g., by time o sets). In a Nihil in terra sine causa t.
sense, we are asking how distance predicts duration. However, as you will see in —T B
Chapter 7, the brain internally generates time o sets between neuronal spikes, Everything connects to everything else.
and, with the help of the locomotion speed of the animal, it calculates the —L V
distances. In the inside-out framework, the cause is a brain mechanism and the
consequence is distance estimation. Distances are meaningless to a naïve brain e law of causation . . . is a relic of a bygone age, surviving, like the mon-
without calibrating the visual cues with action (Chapter 3). us, depending on archy, only because it is erroneously supposed to do no harm.
the framework, the cause-and-e ect relationships can change. —B R ( )
However, another experimenter may disagree with our interpretation of the
x–y relationship. At the ends of the clouds, there are fewer data points, but
those seem to lose the linear relationship characteristic of the other data points.

A
Perhaps a sigmoid function (S-shape curve, in gray) is a better t. To a third ob- er coming to America as a postdoctoral fellow, I  gradually learned
server, the cloud of dots may appear to be split (Figure 2.1B). When all the dots how strongly our upbringing a ects our interpretation of the world.
are lumped together in our mind, the single-cloud interpretation justi es the Di erent cultures shape the brain di erently, making it di cult to em-
use of the term “integration” or “pattern completion,” by analogy to completing pathize with a new society that holds viewpoints that are distinct from the so-
a picture or a sentence with missing details. But the more experienced eye of the cial milieu where you spent your formative years. Shaking hands while looking
third observer suggests that there are actually two clouds, and, therefore, two at another’s eyes induces di erent emotions than bowing and expressing sub-
relationships exist, a case for segregation or pattern separation. Do the points mission. Most people agree that our upbringing is linked to our prejudices.
in the graph re ect a single integrated relationship or two segregated ones? e I  o en wondered whether cultural di erences can appear not only in moral
regression problem then becomes a classi cation problem. Our scienti c vo- and emotional issues but also in scienti c thinking. Even practices that appear
cabulary and views are o en formed by data plots even if they are more so- trivial, such as reading and writing from le to right or vice versa, may in u-
phisticated than those shown in this simple gure. Without an arbitrator or ence the workings of the brain.2
supervisor, we cannot tell who is right and who is wrong.
1. Brulent and Wamsley (2008).
2. Scholz et  al. (2009). Learning a second language can induce structural changes in white
6. Francis Bacon declared that experiments and observations are su cient to obtain reliable matter (Schlegel et  al., 2012). e brains of musicians are favorite targets of researchers
ideas about the world. In contrast, René Descartes, although he had faith in mathematics, examining practice-induced structural changes in the brain. Use-dependent brain plasticity
argued that bare facts are never self-explanatory: they require interpreting. A graph will never has been consistently demonstrated in the brains of professional musicians and in children re-
tell how thing happened, only how they could have happened. ceiving instrumental musical training for several months (Hyde et al., 2009; Lutz, 2009).
34

34 THE BRAIN FROM INSIDE OUT Chapter 2. Causation and Logic in Neuroscience 35

(a) (b) suggests a reliable relationship: longer distances correspond to larger intervals

between spikes. A simple way to quantify this relationship is to t a “regression
line” over the values (the solid black line). e line is a great simpli cation, an
assumption that all measurements are “represented” by the regression line. e
slope of the line and the magnitude of the variance of the values on both sides of
the line indicate the reliability and strength of the relationship. is regularity
is a “co-relationship” or correlation, in which approximately the same number
of values lie above and below the regression line.4
While a plot such as that in Figure 2.1 shows only the values of observations,
we o en wish to read much more from a static cloud of data points. For ex-
Figure 2.1. Interpreting graphs. A: Each dot corresponds to the measured distance the ample, there are many empty areas without measured values on the graph. We
animal walked in an experimental apparatus plotted against time di erence between
can simply assume that those values exist so that, if we had more time and
the spikes of a neuron pair. e black line is a linear regression of all values. Taking into
consideration the tails of the distribution, the sigmoid (gray line) is a better t. B: An
more funding to generate more data, those new points would ll in the blanks
alternative interpretation is that there are actually two clouds (black and gray), which between the already-measured ones. is operation, performed intuitively or
have di erent relationships between distance representation and the time lag of spikes. with mathematical rigor, is called interpolation (the arrow in Figure 2.1), but we
can also call it explanation or deduction. Postdiction would be another possible
term because interpolation is done based on data collected in the past. If we feel
Can we somehow put ourselves outside our subjective world and use objec-
comfortable lling the graph with imagined data between the dots, perhaps it
tive methods to make valid judgments without being in uenced by our pasts?
is not much of a jump to imagine more data that extend the cloud, again fol-
If my reasoning has some merit, then examining the same problems from mul-
lowing the trend in the observations. is operation is called extrapolation or
tiple angles may be useful in understanding the brain. Let’s start by evaluating
induction. Alternatively, we can call it prediction or guessing from the currently
our everyday tools for reasoning in the Western world: logic and correlations
available data, an act of “logical inference.”5 It is an abstraction from the pre-
and the necessary and su cient conditions of causation.
sent, which amounts to predicting the future. How far can we go in our extrap-
olation of the graph? Just because two sets of variables correlate now, there is no
guarantee that this correlation will continue into the future.
OBSERVATIONS, CO- OCCURRENCES,
Interpolation and extrapolation, as mathematical abstractions, are inde-
AND CORRELATIONS
pendent of the objects or observations being represented. e graph itself is an
abstract relationship. e dots have lost their physical correspondences and are
Hypotheses in the neurosciences are most o en tested by models of associations
now embedded in a mathematical graph. In fact, the same dots could represent
between observed properties of the world. Figure 2.1 shows a generic example of
a relationship between two observed variables: in this illustration, between the
distances walked by an animal and the temporal relationships between spikes of 4. e coordinates are both rotationally and translationally symmetric. ere is no special
neuron pairs corresponding to those distances. 3 Inspection of the cloud of dots need for the y and x axis to point north and east. e graph can be oriented in any direction
as long as the relationships between x and y values are unchanged. is symmetry applies to
Euclidian space. In Minkowski spacetime, the symmetry applies to time as well (Chapter 10).
3. ese data are modi ed from Diba and Buzsáki (2008), showing the distance measure be- 5. Con ation among expressions such as “explanation,” “prediction,” “postdiction,” and “cause”
tween place eld peaks of a pair of hippocampal place neurons (x coordinate) while the rat is common in interpreting statistical data, but distinguishing them is important for advancing
ran on a linear track and the time di erence between the spikes of the same neuron pair. e scienti c knowledge. e statistical literature o en lacks a thorough discussion of the many
physiological explanation of the at ends of the sigmoid curve is that the distance-to-duration di erences that arise in the process of modeling explanatory versus predictive processes.
conversion is brought about by theta oscillation dynamics. e duration of the theta cycle is 140 Explaining (the past) and predicting (the future) are logically di erent (Shmueli, 2010). To
ms (±70 ms in the gure), which limits the available distance representation to approximately “predict” (from the Latin pre- in front of, and dicere -to speak) means to verbally declare what
1 meter in the dorsal hippocampus. To resolve longer distances, the animal must move in the is going to happen in the future. Postdiction is an inference. us, statistics is not free of sub-
environment (see further discussion in Chapter 7). jective interpretation.