Artificial intelligence and cognitive modeling have the same problem chapter2

Chapter 2
Artificial Intelligence and CognitiveModeling
Have the Same Problem
Nicholas L. Cassimatis
Department of Cognitive Science, Rensselaer Polytechnic Institute, 108 Carnegie, 110 8th
St. Troy, NY 12180
cassin@rpi.edu
Cognitive modelers attempting to explain human intelligence share a puzzle with artificial
intelligence researchers aiming to create computers that exhibit human-level intelligence:
how can a system composed of relatively unintelligent parts (such as neurons or transistors)
behave intelligently? I argue that although cognitive science has made significant progress
towards many of its goals, that solving the puzzle of intelligence requires special standards
and methods in addition to those already employed in cognitive science. To promote such
research, I suggest creating a subfield within cognitive science called intelligence science
and propose some guidelines for research addressing the intelligence puzzle.
2.1 The Intelligence Problem
Cognitive scientists attempting to fully understand human cognition share a puzzle with
artificial intelligence researchers aiming to create computers that exhibit human-level in-telligence:
how can a system composed of relatively unintelligent parts (say, neurons or
transistors) behave intelligently?
2.1.1 Naming the problem
I will call the problem of understanding how unintelligent components can combine to
generate human-level intelligence the intelligence problem; the endeavor to understand how
the human brain embodies a solution to this problem understanding human intelligence;
and the project of making computers with human-level intelligence human-level artificial
intelligence.
11

12 Theoretical Foundations of Artificial General Intelligence
When I say that a system exhibits human-level intelligence, I mean that it can deal
with the same set of situations that a human can with the same level of competence. For
example, I will say a system is a human-level conversationalist to the extent that it can
have the same kinds of conversations as a typical human. A caveat to this is that artificial
intelligence systems may not be able to perform in some situations, not for reasons of their
programming, but because of issues related to their physical manifestation. For example, it
would be difficult for a machine without hand gestures and facial expressions to converse
as well as a human in many situations because hand gestures and facial expressions are so
important to many conversations. In the long term, it may be necessary therefore to sort
out exactly which situations matter and which do not. However, the current abilities of
artificial systems are so far away from human-level that resolving this issue can generally
be postponed for some time. One point that does follow from these reflections, though,
is the inadequacy of the Turing Test. Just as the invention of the airplane was an advance
in artificial flight without convincing a single person that it was a bird, it is often irrele-vant
whether a major step-step towards human-intelligence cons observers into believing a
computer is a human.
2.1.2 Why the Intelligence Problem is Important
Why is the human-level intelligence problem important to cognitive science? The the-oretical
interest is that human intelligence poses a problem for a naturalistic worldview in-sofar
as our best theories about the laws governing the behavior of the physical world posit
processes that do not include creative problems solving, purposeful behavior and other fea-tures
of human-level cognition. Therefore, not understanding how the relatively simple and
“unintelligent” mechanisms of atoms and molecules combine to create intelligent behavior
is a major challenge for a naturalistic world view (upon which much cognitive science is
based). Perhaps it is the last major challenge. Surmounting the human-level intelligence
problem also has enormous technological benefits which are obvious enough.
2.1.3 The State of the Science
For these reasons, understanding how the human brain embodies a solution to the
human-level intelligence problem is an important goal of cognitive science. At least at
first glance, we are quite far from achieving this goal. There are no cognitive models that
can, for example, fully understand language or solve problems that are simple for a young
child. This paper evaluates the promise of applying existing methods and standards in

Artificial Intelligence and Cognitive Modeling Have the Same Problem 13
cognitive science to solve this problem and ultimately proposes establishing a new subfield
within cognitive science, called Intelligence Science1, and outlines some guiding principles
for that field.
Before discussing how effective the methods and standards of cognitive science are
in solving the intelligence problem, it is helpful to list some of the problems or questions
intelligence sciencemust answer. The elements of this list are not original (see (Cassimatis,
2010) and (Shanahan, 1997)) or exhaustive. They are merely illustrative examples:
Qualification problem How does the mind retrieve or infer in so short a time the excep-tions
to its knowledge? For example, a hill symbol on a map means there is a hill in the
corresponding location in the real world except if: the mapmaker was deceptive, the hill
was leveled during real estate development after the map was made, or the map is of shift-ing
sand dunes. Even the exceptions have exceptions. The sand dunes could be part of a
historical site and be carefully preserved or the map could be based on constantly updated
satellite images. In these exceptions to the exceptions, a hill symbol does mean there is
a hill there now. It is impossible to have foreseen or been taught all these exceptions in
advance, yet we recognize them as exceptions almost instantly.
Relevance problem Of the enormous amount of knowledge people have, how do they
manage to retrieve the relevant aspects of it, often in less than a second, to sort from many
of the possible interpretations of a verbal utterance or perceived set of events?
Integration problem How does the mind solve problems that require, say, probabilis-tic,
memory-based and logical inferences when the best current models of each form of
inference are based on such different computational methods?
Is it merely a matter of time before cognitive science as it is currently practiced answers
questions like these or will it require new methods and standards to achieve the intelligence
problem?
2.2 Existing Methods and Standards are not Sufficient
Historically, AI and cognitive science were driven in part by the goal of understand-ing
and engineering human-level intelligence. There are many goals in cognitive science
and, although momentous for several reasons, human-level intelligence is just one of them.
Some other goals are to generate models or theories that predict and explain empirical data,
11Obviously, for lack of a better name.

to develop formal theories to predict human grammatically judgments and to associate cer-tain
kinds of cognitive processes with brain regions. Methods used today in cognitive
science are very successful at achieving these goals and show every indication of contin-uing
to do so. In this paper, I argue that these methods are not adequate to the task of
understanding human-level intelligence.
Put another way, it is possible to do good research by the current standards and goals
of cognitive science and still not make much progress towards understanding human intel-ligence.
Just to underline the point, the goal of this paper is not to argue that “cognitive science
is on the wrong track”, but that despite great overall success on many of its goals, progress
towards one of its goals, understanding human-level intelligence, requires methodological
innovation.
2.2.1 Formal linguistics
The goal of many formal grammarians is to create a formal theory that predicts whether
a given set of sentences is judged by people to be grammatical or not. Within this frame-work,
whether elements of the theory correspond to amechanismhumans use to understand
language is generally not a major issue. For example, at various times during the devel-opment
of Chomsky and his students’ formal syntax, their grammar generated enormous
numbers of syntactic trees and relied on grammatical principles to rule out ungrammatical
trees. These researchers never considered it very relevant to criticize their framework by
arguing that it was implausible to suppose that humans could generate and sort through this
many trees in the second or two it takes them to understand most sentences. That was the
province of what they call “performance” (the mechanisms the mind uses) not competence
(what the mind, in some sense, knows, independent of how it uses this knowledge). It is
possible therefore to do great linguistics without addressing the computational problems
(e.g. the relevance problem from the last section) involved in human-level language use.
2.2.2 Neuroscience
The field of neuroscience is so vast that it is difficult to even pretend to discuss it in
total. I will confine my remarks to the two most relevant subfields of neuroscience. First,
“cognitive neuroscience” is probably the subfield that most closely addresses mechanisms
relevant to understanding human intelligence. What often counts as a result in this field is a
demonstration that certain regions of the brain are active during certain forms of cognition.

A simplistic, but not wholly inaccurate way of describing how this methodology would
apply to understanding intelligence would be to say that the field is more concerned with
what parts of the brain embody a solution to the intelligence problem, not how they actually
solve the problem. It is thus possible to be a highly successful cognitive neuroscientist
without making progress towards solving the intelligence problem.
Computational neuroscience is concerned with explaining complex computation in
terms of the interaction of less complex parts (i.e., neurons) obviously relevant to this dis-cussion.
Much of what I say about cognitive modeling below also applies to computational
neuroscience.
2.2.3 Artificial intelligence
An important aim of this paper is that cognitive science’s attempt to solve the intelli-gence
problem is also an AI project and in later sections I will describe how this has and
can still help cognitive science. There are, however, some ways AI practice can distract
from that aim, too. Much AI research has been driven in part by at least one of these two
goals.
(1) A formal or empirical demonstration that an algorithm is consistent with, approximates,
or converges on some normative standard. Examples include proving that a Bayes network
belief propagation algorithm converges on a probability distribution dictated by probability
theory or proving that a theorem prover is sound and complete with respect to a semantics
for some logic. Although there are many theoretical and practical reasons for seeking these
results (I would like nuclear power plant software to be correct as much as anyone), they
do not necessarily constitute progress towards solving the intelligence problem. For exam-ple,
establishing that a Bayes Network belief propagation algorithm converges relatively
quickly towards a normatively correct probability distribution given observed states of the
world does not in any way indicate that solving such problems is part of human-level intel-ligence,
nor is there any professional incentive or standard requiring researchers to argue
for this. There is in fact extensive evidence that humans are not normatively correct rea-soners.
It may even be that some flaws in human reasoning are a tradeoff required of any
computational system that solves the problems humans do.
(2) Demonstrating with respect to some metric that an algorithm or system is faster, con-sumes
fewer resources and/or is more accurate than some alternative(s). As with proving
theorems, one can derive great professionalmileage creating a more accurate part of speech

tagger or faster STRIPS planner without needing to demonstrate in any way that their so-lution
is consistent with or contributes to the goal of achieving human-level intelligence.
2.2.4 Experimental psychology
Cognitive psychologists generally develop theories about how some cognitive process
operates and run experiments to confirm these theories. There is nothing specifically in
this methodology that focuses the field on solving the intelligence problem. The field’s
standards mainly regard the accuracy and precision of theories, not the level of intelli-gence
they help explain. A set of experiments discovering and explaining a surprising new
phenomenon in (mammalian-level) place memory in humans will typically receive more
plaudits than another humdrum experiment in high-level human reasoning. To the extent
that the goal of the field is solely to find accurate theories of cognitive processes, this makes
sense. But it also illustrates the lack of an impetuous towards understanding human-level
intelligence. In addition to this point, many of Newell’s (Newell, 1973) themes apply to
the project of understanding human-level intelligence with experimental psychology alone
and will not be repeated here.
A subfield of cognitive psychology, cognitive modeling, does, at its best, avoid many
of the mistakes Newell cautions against and I believe understanding human cognition is
ultimately a cognitive modeling problem. I will therefore address cognitive modeling ex-tensively
in the rest of this paper.
2.3 CognitiveModeling: The Model Fit Imperative
Cognitive modeling is indispensable to the project of understanding human-level intel-ligence.
Ultimately, you cannot say for sure that you have understood how the human brain
embodies a solution to the intelligence problem unless you have (1) a computationalmodel
that behaves as intelligently as a human and (2) some way of knowing that the mechanisms
of that model, or at least its behavior, reflect what is going on in humans. Creating com-puter
models to behave like humans and showing that the model’s mechanisms at some
level correspond to mechanism underlying human cognition is a big part of what most cog-nitive
modelers aim to do today. Understanding how the human brain embodies a solution
to the intelligence problem is thus in part a cognitive modeling problem.
This section describes why I think some of the practices and standards of the cognitive
modeling community,while beingwell-suited for understanding many aspects of cognition,

are not sufficient to, and sometimes even impede progress towards, understanding human-level
intelligence.
The main approach to modeling today is to create a model of human cognition in a task
that fits existing data regarding their behavior in that task and, ideally, predicts behavior in
other versions of the task or other tasks altogether. When a single model with a few param-eters
predicts behavior in many variations of a task or in many different tasks, that is good
evidence that the mechanisms posited by the model correspond, at least approximately, to
actual mechanisms of human cognition. I will call the drive to do this kind of work the
model fit imperative.
What this approach does not guarantee is that the mechanisms uncovered are impor-tant
to understanding human-level intelligence. Nor does it do impel researchers to find
important problems or mechanisms that have not yet been addressed, but which are key to
understanding human-level intelligence.
An analogy with understanding and synthesizing flight will illustrate these points2. Let
us call the project of understanding birds aviary science; the project of creating computa-tionalmodels
of birds aviary modeling and the project of making machines that fly artificial
flight. We call the problem of how a system that is composed of parts that individually suc-cumb
to gravity can combine to defy gravity the flight problem; and we call the project of
understanding how birds embody a solution to this problem understanding bird flight.
You can clearly do great aviary science, i.e., work that advances the understanding of
birds, without addressing the flight problem. You can create predictive models of bird
mating patterns that can tell you something about how birds are constructed, but they will
tell you nothing about how birds manage to fly. You can create models that predict the
flapping rate of a bird’s wings and how that varies with the bird’s velocity, its mass, etc.
While this work studies something related to bird flight, it does not give you any idea of
how birds actually manage to fly. Thus, just because aviary science and aviary modeling
are good at understanding many aspects of birds, it does not mean they are anywhere near
understanding bird flight. If the only standard of their field is to develop predictive models
of bird behavior, they can operate with great success without ever understanding how birds
solve the flight problem and manage to fly.
I suggest that the model fit imperative in cognitive modeling alone is about as likely to
lead to an understanding of human intelligence as it would be likely to drive aviary science
towards understanding how birds fly. It is possible to collect data about human cognition,
2I have been told that David Marr has also made an analogy between cognitive science and aeronautics, but I
have been unable to find the reference.

build fine models that fit the data and accurately predict new observations – it is possible
to do all this without actually helping to understand human intelligence. Two examples
of what I consider the best cognitive modeling I know of illustrate this point. (Lewis &
Vasishth, 2005) have developed a great model of some mechanisms involved in sentence
understanding, but this and a dozen more fine pieces of cognitive modeling could be done
and we would still not have a much better idea of how people actually mange to solve
all of the inferential problems in having a conversation, how they sort from among all the
various interpretations of a sentence, how they manage to fill in information not literally
appearing in a sentence to understand the speaker’s intent. Likewise, Anderson’s (Ander-son,
2005) work modeling brain activity during algebraic problem solving is a big advance
in confirming that specific mechanisms in ACT-R models of cognition actually reflect real,
identifiable, brain mechanisms. But, as Anderson himself claimed3, these models only
shed light on behavior where there is a preordained set of steps to take, not where people
actually have to intelligently figure out a solution to the problem on their own.
The point of these examples is not that they are failures. These projects are great suc-cesses.
They actually achieved the goals of the researchers involved and the cognitive mod-eling
community. That they did so without greatly advancing the project of understanding
human intelligence is the point. The model fit imperative is geared towards understanding
cognition, but not specifically towards making sure that human-level intelligence is part
of the cognition we understand. To put the matter more concretely, there is nothing about
the model fit imperative that forces, say, someone making a cognitive model of memory to
figure out how their model explains how humans solve the qualification and relevance prob-lems.
When one’s goal is to confirm that a model of a cognitive process actually reflects
how the mind implements that process, the model fit imperative can be very useful. When
one has the additional goal of explaining human-level intelligence, then some additional
standard is necessary to show that this model is powerful enough to explain human-level
performance.
Further, I suggest that the model fit imperative can actually impeded progress towards
understanding human intelligence. Extending the analogy with the flight problem will help
illustrate this point. Let us say the Wright Brothers decided for whatever reason to subject
themselves to the standards of our hypothetical aviary modeling community. Their initial
plane at Kitty Hawk was not based on detailed data on bird flight and made no predictions
about it. Not only could their plane not predict bird wing flapping frequencies, its wings
3In a talk at RPI.

did not flap at all. Thus, while perhaps a technological marvel, their plane was not much
of an achievement by the aviary modeling community’s model fit imperative. If they and
the rest of that community had instead decided to measure bird wing flapping rates and
create a plane whose wings flapped, they may have gone through a multi-decade diversion
into understanding all the factors that contribute to wing flapping rates (not to mention
the engineering challenge of making plane whose wings flaps) before they got back to the
nub of the problem, to discover the aerodynamic principles and control structures that can
enable flight and thereby solve the flight problem. The Wright Flyer demonstrated that
these principles were enough to generate flight. Without it, we would not be confident that
what we know about bird flight is enough to fully explain how they fly. Thus, by adhering
to the model fit imperative, aviary science would have taken a lot longer to solve the flight
problem in birds.
I suggest that, just as it would in aviary science, the model fit imperative can retard
progress towards understanding how the human brain embodies a solution to the intelli-gence
problem. There are several reasons for this, which an example will illustrate. Imag-ine
that someone has created a system that was able to have productive conversations about,
say, managing one’s schedule. The system incorporates new information and answer ques-tions
as good as a human assistant can. When it is uncertain about a statement or question
it can engage in a dialog to correct the situation. Such a system would be a tremendous
advance in solving the intelligence problem. The researchers who designed it would have
had to find a way, which has so far eluded cognitive science and AI researchers, to integrate
multiple forms of information (acoustic, syntactic, semantic, social, etc.) within millisec-onds
to sort through the many ambiguous and incomplete utterance people make. Of the
millions of pieces of knowledge about this task, about the conversants and about whatever
the conversants could refer to, the system must find just the right knowledge, again, within
a fraction of a second. No AI researchers have to this point been able to solve these prob-lems.
Cognitive scientists have not determined how people solve these problems in actual
conversation. Thus, this work is very likely to contain some new, very powerful ideas that
would help AI and cognitive science greatly.
Would we seriously tell these researchers that their work is not progress towards un-derstanding
the mind because their system’s reaction times or error rates (for example) do
not quite match up with those of people in such conversations? If so, and these researchers
for some reason wanted our approval, what would it have meant for their research? Would
they have for each component of their model run experiments to collect data about that

component and calibrate the component to that data? What if their system had dozens of
components, would they have had to spend years running these studies? If so, how would
they have had the confidence that the set of components they were studying was important
to human-level conversation and that they were not leaving out components whose impor-tance
they did not initially anticipate? Thus, the data fit model of research would either
have forced these researchers to go down a long experimental path that they had little con-fidence
would address the right issues or they would have had to postpone announcing,
getting credit for and disseminating to the community the ideas underlying their system.
For all these reasons, I conclude that the model fit imperative in cognitive modeling
does not adequately drive the field towards achieving an understanding of human intelli-gence
and that it can even potentially impede progress towards that goal.
Does all this mean that cognitive science is somehow exceptional, that in every other
part of science, the notion of creating a model, fitting it to known data and accurately
predicting new observations does not apply to understanding human-level intelligence?
Not at all. There are different levels of detail and granularity in data. Most cognitive
modeling involves tasks where there is more than one possible computer program known
that can perform in that task. For example, the problem of solving algebraic equations
can be achieved by many kinds of computer programs (e.g., Mathematica and production
systems). The task in that community is to see which program the brain uses and to select
a program that exhibits the same reaction times and error rates as humans is a good way
to go about this. However, in the case of human-level intelligence, there are no known
programs that exhibit human-level intelligence. Thus, before we can get to the level of
detail of traditional cognitive modeling, that is, before we can worry about fitting data
at the reaction time and error rate level of detail, we need to explain and predict the most
fundamental datum: people are intelligent. Once we have a model that explains this, we can
fit the next level of detail and knowthat the mechanisms whose existence we are confirming
are powerful enough to explain human intelligence.
Creating models that predict that people are intelligent means writing computer pro-grams
that behave intelligently. This is also a goal of artificial intelligence. Understanding
human intelligence is therefore a kind of AI problem.

2.4 Artificial Intelligence and CognitiveModeling Can Help Each Other
I have so far argued that existing standards and practices in the cognitive sciences do
not adequately drive the field towards understanding human intelligence. The main prob-lems
are that (1) each field’s standards make it possible to reward work that is not highly
relevant to understanding human intelligence; (2) there is nothing in these standards to en-courage
researchers to discover each field’s gaps in its explanation of human intelligence;
and (3) that these standards can actually make it difficult for significant advances towards
understanding human-intelligence to gain support and recognition. This section suggests
some guidelines for cognitive science research into human intelligence.
Understanding human-intelligence should be its own subfield Research towards un-derstanding
human intelligence needs to be its own subfield, intelligence science, within
cognitive science. It needs its own scientific standards and funding mechanisms. This
is not to say that the other cognitive sciences are not important for understanding human
intelligence; they are in fact indispensable. However, it will always be easier to prove the-orems,
fit reaction time data, refine formal grammars or measure brain activity if solving
the intelligence problem is not a major concern. Researchers in an environment where
those are the principal standards will always be at a disadvantage professionally if they
are also trying to solve the intelligence problem. Unless there is a field that specifically
demands and rewards research that makes progress towards understanding how the brain
solves the intelligence problem, it will normally be, at least from a professional point of
view, more prudent to tackle another problem. Just as it is impossible to seriously propose
a comprehensive grammatically theory without addressing verb use, we need a field where
it is impossible to propose a comprehensive theory of cognition or cognitive architecture
without at least addressing the qualification, relevance, integration and other problems of
human-level intelligence.
Model the right data I argued earlier and elsewhere (Cassimatis et al., 2008) that the
most important datum for intelligence scientists to model is that humans are intelligent.
With respect to the human-level intelligence problem, for example, to worry about whether,
say, language learning follows a power or logarithmic law before actually discovering how
the learning is even possible is akin to trying to model bird flap frequency before under-standing
how wings contribute to flight.
The goal of building a model that behaves intelligently, instead of merely modeling
mechanisms such as memory and attention implicated in intelligent cognition, assures that

the field addresses the hard problems involved in solving the intelligence problem. It is hard
to avoid a hard problem or ignore an important mechanisms if, say, it is critical to human-level
physical cognition and building a system that makes the same physical inferences that
humans can is key to being published or getting a grant renewed.
A significant part of motivating and evaluating a research project in intelligence science
should be its relevance for (making progress towards) answering problems such as the
qualification, relevance and integration problems.
Take AI Seriously Since there are zero candidate cognitive models that exhibit human-level
intelligence, researchers in intelligence science are in the same position as AI re-searchers
aiming for human-level AI: they are both in need of and searching for compu-tational
mechanisms that exhibit a human-level of intelligence. Further, the history of AI
confirms its relevance to cognitive science. Before AI many philosophers and psycholo-gists
did not trust themselves or their colleagues to posit internal mental representations
without implicitly smuggling in some form of mysticism or homunculus. On a technical
level, search, neural networks, Bayesian networks, production rules, etc. were all in part
ideas developed by AI researchers but which play an important role in cognitive modeling
today.
Chess-playing programs are often used as examples of how AI can succeed with brute-force
methods that do not illuminate human intelligence. Note, however, that chess pro-grams
are very narrow in their functionality. They only play chess. Humans can play many
forms of games and can learn to play these rather quickly. Humans can draw on skills in
playing one game to play another. If the next goal after making computer programs chess
masters was not to make them grandmasters, but to make them learn, play new games and
transfer their knowledge to other games, brute force methods would not have been suffi-cient
and researchers would have had to develop new ideas, many of which would probably
bear on human-level intelligence.
Have a success Many AI researchers have retreated from trying to achieve human-level
AI. The lesson many have taken from this is that one should work on more tractable prob-lems
or more practical applications. This attitude is tantamount to surrendering the goal
of solving the human intelligence problem in our lifetimes. The field needs a success to
show that real progress is capable soon. One obstacle to such a success is that the bar, espe-cially
in AI, has been raised so high that anything short of an outright demonstration of full
human-level AI is considered by many to be hype. For a merely very important advance

towards human-level intelligence that has no immediate application, there is no good way
to undeniably confirm that importance. We thus need metrics that push the state of the art
but are at the same time realistic.
Develop realistic metrics Developing realistic methods for measuring a system’s intelli-gence
would make it possible to confirm that the ideas underlying it are an important part
of solving the intelligence problem. Such metrics would also increase confidence in the
prospects of intelligence science enabling quicker demonstrations of progress. My work
on a model of physical cognition has illustrated the value of such metrics (Cassimatis, in
press). I have so far tested this model by presenting it with sequences of partially occluded
physical events that I have partly borrowed from the developmental psychology literature
and have partly crafted myself. My strategy has been to continually find new classes of
scenarios that require different forms of reasoning (e.g., probabilistic, logical, defeasible,
etc.) and update my model so that it could reason about each class of scenarios. Using
superficially simple physical reasoning problems in this way has had several properties that
illustrate the value of the right metric.
Difficulty Challenge problems should be difficult enough so that a solution to them re-quires
a significant advance in the level of intelligence it is possible to model. Human-level
intelligence in the physical cognition domain requires advances towards understanding the
frame problem, defeasible reasoning and how to integrate perpetual and cognitive models
based on very different algorithms and data structures.
Ease While being difficult enough to require a real advance, challenge problem should
be as simple as possible so that real progress is made while avoiding extraneous issues and
tasks. One benefit of the physical cognition domain over, for example, Middle East politics
is the smaller amount of knowledge required for a system to have before it can actually
demonstrate intelligent reasoning.
Incremental It should be possible to demonstrate advances towards the goal short of
actually achieving it. For example, it is possible to show progress in the physical cognition
domain without actually providing a complete solution by showing that an addition to the
model enables and explains reasoning in a significantly wider, but still not complete, set of
scenarios.
General The extent to which a challenge problem involves issues that underlie cognition
in many domains makes progress towards solving that problem more important. For exam-

ple, I have shown (Cassimatis, 2004) how syntactic parsing can be mapped onto a physical
reasoning problem. Thus, progress towards understanding physical cognition amounts to
progress in two domains.
2.5 Conclusions
I have argued that cognitive scientists attempting to understand human intelligence can
be impeded by the standards of the cognitive sciences, that understanding human intelli-gence
will require its own subfield, intelligence science, and that much of the work in this
subfield will assume many of the characteristics of good human-level AI research. I have
outlined some principles for guiding intelligence science that I suggest would support and
motivate work towards solving the intelligence problem and understanding how the human
brain embodies a solution to the intelligence problem.
In only half a century we have made great progress towards understanding intelligence
within fields that, with occasional exceptions, have not been specifically and wholly di-rected
towards solving the intelligence problem. We have yet to see the progress that can
happen when large numbers of individuals and institutions make this their overriding goal.
Bibliography
[1] Anderson, J.R. (2005). Human symbol manipulation within an integrated cognitive architec-ture.
Cognitive Science, 313–341.
[2] Cassimatis, N.L. (2004). Grammatical Processing Using the Mechanisms of Physical Infer-ences.
Paper presented at the Twentieth-Sixth Annual Conference of the Cognitive Science
Society.
[3] Cassimatis, N.L., & Bignoli, P. (in press). Testing Common Sense Reasoning Abilities. Journal
of Theoretical and Experimental Artificial Intelligence.
[4] Cassimatis, N.L., Bello, P., & Langley, P. (2008). Ability, Parsimony and Breadth in Models of
Higher-Order Cognition. Cognitive Science, 33 (8), 1304–1322.
[5] Cassimatis, N., Bignoli, P., Bugajska, M., Dugas, S., Kurup, U., Murugesan, A., & Bello, P.
(2010). An Architecture for Adaptive Algorithmic Hybrids. IEEE Transactions on Systems,
Man, and Cybernetics, Part B, 4 (3), 903–914.
[6] Lewis, R.L.,& Vasishth, S. (2005). An activation-based model of sentence processing as skilled
memory retrieval. Cognitive Science, 29, 375-419.
[7] Newell, A. (1973). You can’t play 20 questions with nature and win. In W.G. Chase (Ed.),
Visual Information Processing, Academic Press.
[8] Shanahan, M. (1997). Solving the frame problem: a mathematical investigation of the common
sense law of inertia. MIT Press. Cambridge, MA.

http://www.springer.com/978-94-91216-61-9

Artificial intelligence and cognitive modeling have the same problem chapter2

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Artificial intelligence and cognitive modeling have the same problem chapter2 (20)

Recently uploaded (20)

Artificial intelligence and cognitive modeling have the same problem chapter2