LINGUIST List 15.1716

Fri Jun 4 2004

Review: Cognitive Science: Baum (2004)

Editor for this issue: Naomi Ogasawara <naomilinguistlist.org>


What follows is a review or discussion note contributed to our Book Discussion Forum. We expect discussions to be informal and interactive; and the author of the book discussed is cordially invited to join in. If you are interested in leading a book discussion, look for books announced on LINGUIST as "available for review." Then contact Sheila Dooley Collberg at collberglinguistlist.org.

Directory

  1. Eduard Barbu, What is Thought?

Message 1: What is Thought?

Date: Fri, 4 Jun 2004 16:55:58 -0400 (EDT)
From: Eduard Barbu <eduardracai.ro>
Subject: What is Thought?

AUTHOR: Baum, Eric
TITLE: What Is Thought?
SERIES: Bradford Books
PUBLISHER: MIT Press
YEAR: 2004
Announced at http://linguistlist.org/issues/15/15-196.html


Eduard Barbu
Institute for Artificial Intelligence, Romanian Academy

INTRODUCTION

Written by an advocate of Strong Artificial Intelligence point of
view, the book has every reason to be considered a controversial
one. Eric Baum proposes a computational model that is meant to explain
everything (mind, evolution, language etc.). When the ambition is so
big it is no wonder that there will be voices rising against it. My
review will have the following structure. In this introduction I will
try to put the Baum's work in a global research context for a
scientific theory of consciousness. I will present then the main
ideas in every chapter and I will address what I think are
inconsistencies and what are week points that should be
elaborated. Finally I will question Baum picture of the mind.

Some years ago philosopher John Searle (1980) distinguished two points
of view in Artificial Intelligence (AI): a. The Strong AI (SAI) thesis
is that human have cognitive states by implementing the right kind of
computation. Its followers' hope is that computer scientists will find
the program of ''human mind''. b. The Weak AI thesis is that the only
significant contribution of AI to the understanding of human
psychology is by building useful tools for testing scientific
hypothesis.

For a better understanding of Baum enterprise I will use Penrose's
classification of theories about mind and Sloman's analyses of the
field of SAI. According to Penrose (1990) there are four viewpoints on
the nature of mind:

A. All thinking, all sensations, feelings, etc. are the result of
implementing the appropriate computation. This is SAI.

B. Consciousness is the result of the physical actions of the
brain. The computation just simulates cognitive states but cannot
actually produce them.

C. Consciousness is a characteristic of the brain but the actions of
the brain cannot be simulated computationally.

D. Science will never explain consciousness. This is the mystical
viewpoint on the nature of consciousness.

SAI itself is not compact and easy to classify. The different
positions in the field are marked by different views on what can be a
theory of implementation (see Sloman 1992 for a detailed
analysis). Because Baum thinks that Searle can implement the algorithm
which supposedly describes the language competence of a Chinese
speaker, he is an adept of Sloman's T2 thesis which requires that
''computation be causally related to an explicit program'' . By
relying on Occam's razor for finding the best theory compatible with
the data, Baum escapes a well-known objection to SAI. The objection
states that if the behavior alone is sufficient for the possession of
cognitive states then a program implemented as a giant lookup table,
whose behavior is equivalent with the behavior produced by an
intelligent system, will posses mentality.

The main tenet of this book is that mind is an evolved program encoded
in a compact code (about 10 Mega of DNA). However this code cannot be
replicated yet because we have not the computational resources the
evolution had. But with the aid of evolutionary programming and using
Occam's razor we will succeed in the future to find the program of
mind.

SUMMARY AND COMMENTS

Chapter 1. Introduction

The introductory chapter has two parts. In the first part the author
lays out in short all his chief ideas. The second part is just a road
map to the rest of the book. Baum starts by stating that his purpose
is to present a picture of the mind which is consistent with all our
present knowledge. He continues by rejecting the mystical conception
of the mind and then he presents SAI as if it were the only
alternative to mysticism. As we saw this is not true, there are two
other theories labeled B and C above, which are worth considering. It
is necessary at least to present the other scientifically compatible
points of view and try to refute them.

Then he presents the main ideas that he tries to sustain and defend in
the rest of the book:

1. Mind is a computer program. More precisely mind is a modular
program with dedicated subroutines. These subroutines are used in
different contexts, thus facilitating learning. The analogy and the
metaphor are explainable by code reuse.

2. ''Thought is all about semantics.'' For an entity to have meaning
is to have the capacity to capture and exploit ''the compact structure
of the world''.

3. The best theory is the simplest one. If there are many theories (in
our case programs) which explain the same facts and made accurate
predications, to choose the right one we should use Occam's razor.

Unfortunately this chapter contains some inconsistencies: For example,
at page 3 Baum says that: ''The execution of a computer program is
always equivalent to pure syntax.'' On any construal this statement is
false. If a program in execution (a process) is just syntax and the
mind is a computer program how can mind have semantics? A follower of
SAI would say that a process, if it were to be considered intelligent,
would have semantics.

At the same page the author claims that ''mind typically produces a
computer program capable of behaving'', which immediately raises the
question: Is the mind a computer program or does the mind produces a
program?

Chapter 2. The Mind Is a Computer Program.

The chapter builds in his most part on historical considerations. The
author introduces important concepts for understanding the
computational picture of the mind such as Turing machine, universal
Turing machine and the self-reproducing automaton of John von Neumann.

Firstly Baum stresses that the creation of the Turing machine was the
consequence of the attempt to answer one of the problems put by
Hilbert: is there an effective procedure for solving all the problems
of mathematics (a well define class of problems as Diophantian
equations, to be more precise)? The problem was independently and
negatively solved by Turing who developed the concept of Turing
machines and by Church who developed the lambda calculus. Turing
machines, lambda calculus and Emil Post production systems formalize
what an algorithm is. But at this point someone might ask: what is the
link between a Turing machine and the thinking process? The only
answer that Baum gives is that a mathematician mind solving a certain
problem is equivalent to the lookup table of the Turing machine and so
the states of the mind of the mathematicians are computational
states. But there are at least two questions that should be answered:

1. Mathematical thinking is just a small part of what can be titled as
thinking. Even admitting that the concept of computation captures the
mathematical thinking how about the rest of thinking?

2. Is mathematical thinking computational? Many authors, notably Lucas
and Penrose, believe it is not. They use in their support the
celebrated Kurt Godel's theorem, also mentioned by Baum in this
chapter.

Moreover, the account Baum gives to the Hilbert problem is a little
bit misleading. He says at page 50 that Godel answered the problem
''(mentioned earlier in this chapter) in the negative: there is not
effective procedure that can prove all the true theorems of
mathematics''. Instead of being concerned with this problem, Godel was
concerned with other problem posed by Hilbert, namely to give an
absolute consistency proof and also completeness proof for the
mathematics. Godel showed that a system equivalent to the
Russell-Whitehead Principia Mathematica contains undecidable
propositions and one of them is the consistency of the system.

In the second part of chapter 2 the author describes ''the
computational process that is life'', a process that creates and
maintains us. He gives us some background from biochemistry and argues
that the program of life is isomorphic with a Post Production system.

Chapter 3. The Turing Test, the Chinese Room, and What Computers Can't Do

This chapter is Baum's first attempt to answer some critics of SAI
point of view. He is addressing in principal the problems of qualia
and understanding. The first problem that Baum takes on is the
important problem of experience or qualia. How is that we can feel,
smell and so on if our minds are just computer programs? Unfortunately
Baum's answer will be accepted as a valid one only by SAI adepts.
Basically, he says that the fact that we cannot accept that computers
can have qualia is just a failure of our imagination. Moreover, he
claims that we will accept this when we have a chat with a computer
that will insist that it can have experiences (p. 67).

Baum then skips to the other problem, the problem of understanding. In
this context he presents the Turing test and what is considered by
most people to be the most powerful argument against SAI, Searle's
Chinese room argument (CRA). The way Baum tries to refute the CRA is
not new and it is just a variant of the ''System Reply''. Searle's
answer to the ''System Reply'' is that he can internalize all the
elements of the system, he can answer any question an external
observer asks, but he will fail again to understand Chinese. Baum
rejects Searle's conclusion that he doesn't understand Chinese and
sustains that, by internalizing a Turing machine that can answer
Chinese questions, Searle does indeed understand Chinese. But that is
to misunderstand Searle's proof. The thought experiment that Searle
proposes has as its point to show that someone can pass a Turing test
without understanding anything, that human understanding is different
from computer understanding in that we have intentionality, which
computers lack. Baum puzzles us further when (pp. 78) he misidentifies
intention with the intentionality: ''This concern that intention is
something that only human beings can have is still reverberating in
philosophical literature.''

Chapter 4. Occam's Razor and Understanding

The theme of this chapter is how symbols in computer programs can mean
something.

Refuting the CRA and accepting the Turing Test as a measurement for
understanding, it is clear that the answer to the problem of how the
symbols of a computer can have meaning will be based on the capacity
of programs to explain a series of facts and make predication. With
this meaning of ''meaning'' in mind Baum shows us how the external
data can be fit by a good theory. The chapter is a plea for Occam's
razor, which is seen by the author to be at the heart of science.

The author presents three formalizations of Occam's razor. The first
formalization uses Vapnik-Chervonenkis dimension, the second is based
on the description length principle, and the third one is ground on
Bayesian probability.

Baum says very few words about neural nets, and this is perhaps the
greatest drawback of the book. In addition, the presentation of neural
nets misguides the uniformed reader in that it can create the false
impression that neural nets are ''complicated models of brain
circuits''. In fact, brain circuits without being their models only
inspire neural nets. The author will want perhaps to discuss this and
to correct this drawback in a next edition of the book.

Chapter 5. Optimization

If the chapter before was concerned with the relation between compact
programs and the data they are the description of, this chapter deals
with the heuristic of finding the best program consistent with some
given data. The author rules out the prospect of searching through
all the possibilities and then finding the most compact description
consistent with the data due to the fact that such an algorithm has an
exponential complexity. He also stresses that we don't need the
smaller possible representation in order to extract semantics, but one
sufficiently smaller than the data. Baum argues that the solution to
this problem could be a general optimization technique known as hill
climbing. The hill climbing and its advantages are nicely exemplified
with the Traveling Salesman Problem. The author speculates that a
similar technique was used by evolution in its searching for
meaningful possibilities.

Chapter 6. Remarks on Occam's Razor

Baum starts the most substantive part of the chapter 6 with a
discussion of a critique of neural networks. The critique is that
someone who looks at the inner structure of the neural networks cannot
understand what the net is doing. Then the author elaborates this
point of view, but not sufficiently because some things remain
unclear. Because Baum doesn't make explicit his point of view in the
so- called ''Systematicity debate'' launched in 1988 by Fodor and
Pylyshyn (1988) I cannot tell what Baum thinks about neural
networks. However, my opinion is that Baum thinks that connectionism
can account for higher cognitive functions only but implementing the
classical model. Baum confuses me further when he writes: '' Neural
nets are not sufficiently powerful to describe minds. One must talk
instead about more powerful programming languages.'' (pp 133)

Does Baum think that neural networks are programming languages!? If
not, what else does this mean?

Then the author compares DNA with the source code and the mind with
the executable. Baum extends the analogy by comparing commentaries in
programming languages with the base pairs in the DNA which are not
read during development. The chapter ends with a discussion of
generalization in neural networks and with a sketch of the proof of
the lower bound theorem.

Chapter 7. Reinforcement Learning

Baum argues that a passive framework, which has as a goal just the
data classification, is not sufficient to account for
consciousness. Instead we need a more powerful model which can account
for the interaction between robots (us) and the world. The author
discusses some reinforcement learning techniques and argues that
neural nets are too weak representations.

Chapter 8. Exploiting Structure

The author draws the difference between recognizing structure and
exploiting structure. He correctly argues that a theory that only
accounts for data classification is fundamentally incomplete. Three
problems which involve structure exploitation are presented (Blocks
World problem, the game of Chess and Go). Baum shows the
insufficiencies of already tried computational approaches for solving
these problems. He sustains that evolution has ''trained'' the mind on
vast number of problems. Now the mind has the capacity to generalize
and to solve problems it was not trained for such as the
above-mentioned games. In these cases the program of mind is better
because he has the capacity to: ''analyze new problems such as chess
into a collection of localized objects that interact causally
... ''(pp.196). Unfortunately the author does not elaborate this
point and I cannot understand what notion of causality he has in
mind. Moreover, he should further explain why this capacity cannot be
captured by present computer programs.

Chapter 9. Modules and Metaphors

The chapter is a plea for the modular structure of the mind and for
the metaphorical nature of thought. In support of the former the
author brings evidence from cognitive science biology and
psychology. After succinctly presenting the ideas of Lakoff and
Johnson he explains the metaphoric nature of thinking by code reuse.

Chapter 10. Evolutionary Programming

Some computational experiments for evolving code are discussed. The
author together with his colleague tries to solve Blocks World problem
by using evolutionary programming. They succeeded to evolve a program
that implements the same algorithm that we use when trying to solve
this problem. However the resulted code is not superior to a code not
written by this method.

Chapter 11. Intractability

A more complete and detailed analysis of the techniques used by
computer scientists to solve general classes of problems is given,
along with examples of polynomial time mapping of instances of NP
complete problems. There are also presented some experiments with
evolutionary robotics and it is argued that constraints propagation
allows for solutions for ''intractable problems''.

Chapter 12. The Evolution of Learning

A parallel between learning and development is drawn. It is postulated
that the learning process is largely dependent on the inductive biases
that evolution produced. Moreover, learning affects evolution by
''Baldwin effect'' (the ability of individuals to learn can guide the
evolutionary process) and culture. However, for Baum, the concept
''culture'' has not its usual meaning, but it means passing the
information through ''parental instruction''. Perhaps for Baum culture
means the set of acquired behaviors? An example of cultural
interaction would be the parental instructions given by mother bear to
her children (p. 335). The author believes that most of our concepts
are innate. Baum is a supporter of Chomsky's ideas and he states that
from the perspective of evolutionary programming: ''Chomsky's proposal
that there is some universal grammar wired into genome is
tautological'' (pp. 343).

Chapter 13. Language and Evolution of Thought

Language has a purely communication function. Language is what
differentiates animals and humans. Baum proposes a theory of language
where language words are seen as labels for computational
modules. Grammar, which has a nonstandard definition here, is seen as
a mapping function from combinations of words in a sentence to
corresponding code. It is argued that thought has nothing to do with
language because concepts predate language. Learning new words involve
attaching labels to already existing computational modules.

Chapter 14. The Evolution of Consciousness

This chapter deals primarily with consciousness, awareness, qualia and
free will. In the author's view, evolution produced mind and finally
consciousness. Baum tries to give an account for the notion of
self. Even if he thinks that the mind is a distributed program with
many subroutines, the fact that the subroutines are working toward the
same end build the self. The author stresses also that we are not
aware of most of our computation and that our awareness is just a
module of mind which concentrates the results of vast amounts of
unconsciousness computation. One good argument against physicalism is
that it cannot account for sensations and feelings. Baum argues
against this on the ground that sensations are necessary for the
evolution process and are built into our program at a ''fundamental
level''. But this doesn't address the question in that he presupposes
the existence of sensations, and does not show how the execution of a
computer code gives rise to sensations. The author thinks that free
will is entirely consistent with physicalism. He says that free will
does not exist but it is a useful concept in predicting the behavior
of others.

FINAL REMARKS

Baum's book raises many questions. His simple system of explaining
everything could be extensively questioned. In what follows I will
raise some possible objections. However, not being competent in
genetics, I will not question the validity of Baum's hypotheses from
this point of view.

1. Baum states that he offers a theory of mind compatible with our
present knowledge. But it can turn out that our current knowledge is
not sufficient for explaining mind. For instance, a theory of physics
compatible with the 17- century knowledge would be plainly false.

2. Baum does not give a theory of implementation. This makes his
construction vulnerable at Putnam's (1988) or Searle's (1990)
arguments. Putnam for example argued that if we are allowed to
consider arbitrary disjunction of physical states as realization of
the formal states of an automaton then the result we will obtain is
that any open physical system implements any finite state automaton
(FSA). Similarly, Searle argues that because syntax is an observer
relative notion, then a wall or stomach can be seen as implementing
any computation.

3. Surprisingly for a theory that tries to explain thought, Baum
doesn't refer to propositional attitudes and doesn't try a
naturalistic reduction of Intentionality. Without this his model has
not force and cannot explain anything.

4. He claims (chapter 8) that he solved the old philosophical problem
of whether the world exists or is just an illusion but he is just
begging the question. He presupposes the existence of physical objects
and of a code as the compact representation of the world and then
tries to prove that the world really exists.

5.His theory of language is roughly this:

a. Every concept corresponds to a piece of code (a subroutine).
b. The most part of concepts is innate.
c. The meaning of expressions is compositional and it is
 obtained by a module calling other modules.
d. Learning new words means to attach labels to presently
 existing modules.

This raises at least the following questions: If the most part of
concepts is innate how can they fit in 10 MG of code in DNA? How can a
speaker choose the right sense of a word when the word is ambiguous
(In Baum's formulation, how a subroutine knows what subroutine to
call)? How does this model treat pragmatics? How is the new code added
and how is it compiled when we learn new concepts?

6.Baum seems to think that evolution itself is a computational process
(note that this is different from saying that the evolution can be
simulated computationally). What algorithm does evolution implement?
For example Baum says: ''evolution effectively searched over
combinations of meaningful macros. Add long legs, and see if that
helps''. If evolution tried many possibilities where is the evidence?

Despite these problems, Baum's book was a pleasure to read. The
author manages to explain hard concepts like self- reproducing
automata and NP completeness to those less familiar with
them. Moreover, he has a keen sense of humor that adds joy to the
reading.

REFERENCES

Fodor, J. and Z. Pylyshyn (1988) Connectionism and Cognitive
Architecture: a Critical Analysis, Cognition: 28.

Penrose, R. (1990). The Emperor's New Mind. Oxford University Press.

Putnam, H. (1988) Representation and Reality, The MIT Press,
Cambridge, Mass.

Searle, J.R. (1980) Minds, brains and programs, Behavioral and Brain
Sciences: 3.

Searle, J.R. (1992) The Rediscovery of the Mind, Cambridge, MIT Press.

Sloman, A. (1992) The emperor's real mind: review of Roger Penrose's
The Emperor's New Mind, Artificial Intelligence.

ABOUT THE REVIEWER

Eduard Barbu is a researcher at the Romanian Institute for Artificial
Intelligence. He was involved in several European projects. His
interests are: formal and lexical ontology, cognitive science,
philosophy of language and mind. He is presently working at a
dictionary of Philosophy of Mind.
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue