Customer Reviews

What Is Thought? (A Bradford Book)

5 star:		(11)
4 star:		(3)
3 star:		(2)
2 star:		(3)
1 star:		(1)

 

 

Why can humans rapidly carry out tasks, such as learning to talk or recognizing an object, that seem intractable for computers?
According to Eric Baum, the human brain is much like a computer, but it runs programs that are different from the ones usually
written by human computer programmers. The programs run by the brain are insightful or ``compressed''; they have built in
a good deal of knowledge or ``understanding'' about the nature of the world. Human programmers have difficulty generating such efficient or compressed programs (except for limited special purposes), because to do so requires vast computing resources, far beyond what one can accomplish with pencil and paper or even with presently available computer assistance.

The key to understanding intelligence, according to Baum, is the theory of evolution; in the process that brought humans into being, evolution cycled through many billions of generations of organisms, in the course of which, in effect, vast computational resources were brought to bear on the problem of generating useful algorithms. The real secret to thought is thus stored in our DNA, which preprograms us with algorithms that
are more efficient and powerful than the ones usually available to computer scientists.

With this starting point, Baum proposes answers to many old riddles. Our sense of ``self'' reflects our origin in an evolutionary struggle for survival toward which all components of our biology are directed. ``Free will'' is a useful approximation because of the great complexity of our brains (and our limited knowledge about them) and the concommitant difficulty of predicting a person's behavior. Baum illustrates his arguments with numerous examples drawn from biology, psychology, and computer science; the material is generally quite interesting, though at times perhaps too detailed for a casual reader. His arguments are surprisingly persuasive, and, while certainly no expert, I suspect that Baum is closer to the mark than most of the old and new classic writers on these problems.

In the introduction to this handsomely bound book, the author suggests that it is an appropriate time for an explanation of how the dynamics of a human brain can be accounted for by computer science. His title is motivated by Erwin Schrödinger's enormously influential "What is life?" which launched the field of evolutionary biology by inducing both Francis Crick and James Watson to successfully seek the molecular basis of biological evolution, but the analogy is strained for several reasons.

Schrödinger's book is less than 100 pages in a current edition, while Baum's is about five times as long. In the context of Schrödinger's lifelong interest in biological problems and based on a series of three public lectures that he presented to the Irish intelligentsia in 1943 (as one of his statutory duties as the founding director of the Dublin Institute of Advanced Studies), "What is Life?" is a classic example of his exceptional expository skill---in a second language, no less---whereas Baum's book would have profited from another round of copy-editing. But the most striking difference between these two titles lies in the cogency of their respective contents.

Although Max Delbrück and his colleagues had used measurements of mutation rates of fruit flies under X-radiation to show that their genes were necessarily of molecular dimensions in the mid-1930s, the implications of these data were unnoticed by the literate world of the mid-1940s. Thus Schrödinger's public lectures were newsworthy, being favorably noted by Time magazine in the spring of 1943, and his subsequent book---after some difficulties with an Irish publisher and the Roman Catholic Church over the religious implications of his ideas---went on to sell over 100,000 copies for Cambridge University Press, with translations into seven languages. Is there a similar communications gap in our current understanding of the nature of thought?

Noting his background in computer science, one mightclassify Eric Baum among those who believe that ``our souls are software'', but this is not quite fair. Although he states that ``the obvious inability of present-day computer science to account for [the brain's behavior] is no reason at all for doubting that they can be accounted for by computer science,'' the intellectual perspectives of "What is Thought?" are broader than this assertion seems to suggest. The book begins with several interesting chapters on the nature of computation (I particularly liked the presentation of the traveling-salesman problem), which include discussions of the importance of making decisions at the level of semantics, the Turing test, properties of neural nets, hill climbing in a fitness landscape, among several other relevant topics. These discussions lead into the author's central thesis that the mind, like all efficient computer programs, is necessarily modular. In other words, each aspect of the brain's dynamics comprises several subroutines, which presumably can be further broken down into hierarchical structures of nested activities, and he discusses several permutations of this important concept. Curiously, Baum's otherwise comprehensive list of references does not include Donald Hebb's seminal and classic work, in which the notion of ``cell assemblies'' (which are dynamically self-sufficient modules of neurons) was first suggested over a half-century ago. As a psychologist, Hebb aimed to ``bridge the long gap between the facts of psychology and those of neurology,'' and coming at about the same time as the development of the digital computer, his formulation has provided the basis for many numerical studies starting in the 1950s and continuing to the present day which are in accord with a growing body of electrophysiological data. Setting this quibble aside, Baum offers compelling psychological evidence for the modular structure of mind and provides his readers with an interesting and informative account of how the structure of our thinking may have developed over the course of biological evolution, with particular attention paid to computational constraints on the development of learning mechanisms. Importantly, his perspectives are broader than those of many of his colleagues, as he asserts that the ``whole program'' of a brain's dynamics includes the ``complex society'' in which it is embedded. Indeed, the author's evident humility in the face of awesome intricacy of mental activity is, to me, one of the more appealing aspects of "What is Thought?"

The often suggested possibilities for quantum computation are discussed in some detail, along with an analysis of the widely noted example of ``Schrödinger's cat'' which was originally proposed to emphasize the difficulties of applying ideas developed for atomic dynamics to complex macroscopic systems. Considering that a quantum computer---if it is at all possible to construct one---must be carefully isolated from structural irregularities and operated near absolute zero of temperature, Baum joins the majority of physical scientists in concluding that it is ``highly unlikely that quantum computation is relevant to the mind.''

Eric Baum has a dog, and---like most of us dog owners---he is convinced that his pet is conscious, but he goes on to assert that ``we do not need to posit new qualitative modes of thinking to explain human advance over animals. To my mind, the difference between human intelligence and animal intelligence is straightforwardly explainable by cumulative progress once there is the ability to communicate programs.'' Here, again, Baum could profit from reading Hebb's book, which contains but a single mathematical expression, namely A/S. This parameter represents the ratio of the associative area (A) of a mammalian neocortex to its sensory area (S), and it becomes greater as one progresses from rats through dogs to humans. A related physiological parameter---with profound significance for the ease and rate at which modules (or cell assemblies) can switch on and off---is the percentage of inhibitory intercortical neurons, varying as follows: rabbit (31%), cat (35%), monkey (45%), human (75%) [6]. Of course, these relative differences may be examples of the ``cumulative progress'' to which Baum refers.

In a penultimate section, Baum discusses the question of free will, noting that ``our decisions look, from any reasonable perspective short of knowing the exact state of our brains and simulating them in detail, like they are introducing genuinely new information.'' In reaching this conclusion, he may be confused by the continuing tendency of many scientists to overlook a phenomenon called ``sensitive dependence on initial conditions'' first studied by the eminent French mathematician Henri Poincaré and widely observed nowadays by those who study nonlinear dynamic phenomena (chaos theory). As Poincare` famously put it over a century ago:

"If we knew exactly the laws of nature and the situation of the universe at the initial moment, we could predict exactly the situation of that same universe at a succeeding moment, but even if it were the case that the natural laws had no longer any secret for us, we could still only know the initial situation approximately. If that enabled us to predict the succeeding situation with the same approximation, that is all we require, and we should say that the phenomenon had been predicted, that it is governed by laws. But it is not always so; it may happen that small differences in the initial conditions produce very great ones in the final phenomena. A small error in the former will produce an enormous error in the latter. Prediction becomes impossible, and we have the fortuitous phenomenon."

For an author who bases many of his conclusions on close mathematical reasoning and offers a theory that purports to be ``capable of explaining everything,'' the implications of these ``fortuitious phenomena'' should be carefully digested.

Alwyn Scott
http://personal.riverusers.com/~rover/

The first half of this book is an overview of the field of artificial intelligence that might be one of the best available introductions for people who are new to the subject, but which seemed fairly slow and only mildly interesting to me.
The parts of the book that are excellent for both amateurs and experts are chapters 11 through 13, dealing with how human intelligence evolved.
He presents strong, although not conclusive, arguments that the evolution of language did not involve dramatic new modes of thought except to the extent that improved communication improved learning, and that small catalysts created by humans might well be enough to spark the evolution of human-like language in other apes.
His recasting of the nature versus nurture debate in terms of biases that guide learning is likely to prove more valuable at resisting the distortions of ideologues than more conventional versions (e.g. Pinker's).
His arguments have important implications for how AI will progress. He convinced me that it will be less sudden than I previously thought, by convincing me that truly general-purpose learning machines won't work, and that much of intelligence involves using large quantities of data about the real world to choose good biases with which to guide our learning.

The main thesis of this book, asserted repetitively, is that the mind is a computer program. Once this is borne in mind, pardon the alliteration, most of the book is reduced to an argument in its favour, rather than an investigation into its credibility. The book often reaches for blunt assertions to support its positions and only afterwards begins a slight retracing of steps. For example, we are told that inductive bias and learning algorithms are coded into the genome. It is obvious, bit of speculation on DNA, evolution and algorithms and out comes the result!

In his observance of Occam's Razor, the author confuses the appeal of the simplest explanatory hypothesis with the belief that he has found such. The discussion of neural networks leaves aside recurrent networks, which are probably more biologically plausible than competitors.

Likewise the idea that the brain essentially 'runs' compressed programs due to evolutionary endowments is unconvincing and philosophically leaky.

I don't want to be over critical of the book as it has brought together many interesting strands of work, but it just has not woven them into anything interesting. There is little new here, whether from modularity or evolutionary programming constraints on neural activity. A lot of it is speculative and several of the key themes are discordant due to under analysis of their assumptions.

Several of the elaborations verge on the frivolous. For example, there is a particularly woolly argument linking the learning of Scheme to "what goes on in constructing our understanding of the world" (p. 222). Likewsie in discussing awareness and consciousness, the author relies on the use of 'main' in C to metaphorically explain how information might come together in the brain (p. 413-415). All kinds of reification fallacies come to mind, leaving aside the thinnes of the argument.

The bottom line is that the book pursues a strong cognitivist program (the brain is a computer) without convincingly examining various sides of the argument. I was certainly no wiser off at the end of it.

Baum's book aims -- and in my estimation succeeds brilliantly -- at illuminating what we know and don't know about computation and the modeling of mind: memory, learning, perception, reasoning, etc. Baum summarizes the main perspectives of various schools of thought on the topic, notably including both proponents of the artificial intelligence enterprise as well as critics, plus neural, sociobiological, psychological and philosophical points of view. He summarizes the main results of computer science and shows their relevance to mind. Best of all, the book is very well-written, and despite the fact that it includes considerable technical depth, it does not presuppose prior knowledge of the subject and should therefore be accessible to a broad audience.

Eric Baum's recent book "What is Thought" is a must-read for anyone interested in artificial intelligence or cognitive science and neuroscience. In the highly saturated area of "consciousness books" this one stands out as one likely to be remembered and referenced much longer than the others. One reason for this is the absolute clarity with which he argues the hard AI position, that the mind is the result of the computer program that is not just run by the brain, but a result of the brain's very architecture produced by several billion years of evolution, the original and ultimate genetic program. The major thesis of the book is that "meaning" should be considered to be identical to a compact description of the data, such as the sensory input from the external world. One example he gives is the compact description of a set of data as falling on a line. This is, of course, a completely operational definition of semantics, but I think a useful one. This leads to the conclusion that meaning is intrinsically determined by the interaction of the world with the architecture of our 100 billion neuron brain as produced by the action of a mere 30,000 genes in generating its architecture. He does not ignore learning and culture, of course, but the point is that, at least at this point in our evolution, most of the compaction is already in the structure. Baum's credentials for many of these speculations come from his solutions to several classical AI problems, such as "Blocks World" using genetic programming techniques. The most successful of these are embodied in an artificial economy model call "Hayek" that solves the credit assignment problem well enough to have advanced solutions to such complex problems considerably. The description of the Hayek system is worth reading in its own right for those interested in various AI approaches to these classical problems, although I found these sections somewhat sparse in details for trying to implement the code. What Baum is very clear about is the formidable challenge of producing, in any current computer system, an equivalent compact description of data similar to that for which humans have evolved. Thus, from first principles, we cannot expect any current AI system to display anything like the ability to generate common sense meaning from the world that has been produced by the great genetic program that is the evolution of the human brain on earth, because the number of equivalent learning cycles (on the order of 4 billion years times the number of example animals) is so many orders of magnitude greater for biological brains than artificial ones. But there is hope in the future from Moore's law of the continued increase in computer power. If you accept these arguments about the vast computational power embodied in our brain's structure, then our inability to comprehend issues such as "qualia" and the feeling of having free will are to be attributed to simple ignorance, a quantitative difference, rather than to more mystical qualitative boundaries. This is consonant with arguments previously eloquently made by the philosopher Dennett, among others. Whether you are for or against such a hard AI position, this book makes its case more honestly, eloquently, and in more detail than any other I have read. Besides the lack of detail for implementation in the discussion of the Hayek system for solving classic AI problems such as Blocks World, one other complaint I have is the lack of reference to some previous work. For example, although Baum does not borrow in any direct way from the CopyCat work of Hofstadter and Mitchell, in spirit, at least, the set of autonomous agents in Hayek sound a lot like codelets and other elements in the CopyCat system, and I don't see why Baum could not have referenced that. I also believe that the reduction of data to a compact description as being equivalent to meaning is slightly incomplete. I think such a compact description is equivalent to an instinct, or an intuition. The embodiment of the compact description that can be manipulated within a system of such descriptions is what actually generates meaning, and the equivalent of thought.

Noting his background in computer science, some classify Eric Baum among those who believe that ``our souls are software'', but this is not quite fair. Although he states that ``the obvious inability of present-day computer science to account for [the brain's behavior] is no reason at all for doubting that they can be accounted for by computer science,'' the intellectual perspectives of What is Thought? are broader than this assertion seems to suggest. The book begins with several interesting chapters on the nature of computation (I particularly liked the presentation of the traveling-salesman problem), which include discussions of the importance of making decisions at the level of semantics, the Turing test, properties of neural nets, hill climbing in a fitness landscape, among several other relevant topics. These discussions lead into the author's central thesis that the mind, like all efficient computer programs, is necessarily modular. In other words, each aspect of the brain's dynamics comprises several subroutines, which presumably can be further broken down into hierarchical structures of nested activities, and he discusses several permutations of this important concept. Curiously, Baum's otherwise comprehensive list of references does not include Donald Hebb's seminal and classic work, in which the notion of ``cell assemblies'' (which are dynamically self-sufficient modules of neurons) was first suggested over a half-century ago. As a psychologist, Hebb aimed to ``bridge the long gap between the facts of psychology and those of neurology,'' and coming at about the same time as the development of the digital computer, his formulation has provided the basis for many numerical studies starting in the 1950s and continuing to the present day which are in accord with a growing body of electrophysiological data. Setting this quibble aside, Baum offers compelling psychological evidence for the modular structure of mind and provides his readers with an interesting and informative account of how the structure of our thinking may have developed over the course of biological evolution, with particular attention paid to computational constraints on the development of learning mechanisms. Importantly, his perspectives are broader than those of many of his colleagues, as he asserts that the ``whole program'' of a brain's dynamics includes the ``complex society'' in which it is embedded. Indeed, the author's evident humility in the face of awesome intricacy of mental activity is, to me, one of the more appealing aspects of What is Thought?

The often suggested possibilities for quantum computation are discussed in some detail, along with an analysis of the widely noted example of ``Schrödinger's cat'' which was originally proposed to emphasize the difficulties of applying ideas developed for atomic dynamics to complex macroscopic systems. Considering that a quantum computer---if it is at all possible to construct one---must be carefully isolated from structural irregularities and operated near absolute zero of temperature, Baum joins the majority of physical scientists in concluding that it is ``highly unlikely that quantum computation is relevant to the mind.''

Eric Baum has a dog, and---like most of us dog owners---he is convinced that his pet is conscious, but he goes on to assert that ``we do not need to posit new qualitative modes of thinking to explain human advance over animals. To my mind, the difference between human intelligence and animal intelligence is straightforwardly explainable by cumulative progress once there is the ability to communicate programs.'' Here, again, Baum could profit from reading Hebb's book, which contains but a single mathematical expression, namely A/S. This parameter represents the ratio of the associative area (A) of a mammalian neocortex to its sensory area (S), and it becomes greater as one progresses from rats through dogs to humans. Of course, these relative differences may be examples of the ``cumulative progress'' to which Baum refers.

Alwyn Scott
http://personal.riverusers.com/~rover/

Baum's book is always stimulating and in some ways admirable, especially in its insistence that there is nothing magical in the brain. But he's wrong in several crucial ways, the same ways that Pinker gets wrong (for example, in "The Slate's Last Stand").
1. Despite his neural network background, Baum fatally underestimates the power of unsupervised learning. While he's right that complex networks cannot be explicitly trained without astronomically numerous examples, it's now clear that unsupervised learning (where the number of examples is quite literally astronomical) combined with the rather regular (albeit complex) structure of the world, can do most of the heavy lifting, with supervision filling in details. Explaining unsupervised learning to a lay audience is not easy (I know of no successful attempts) but cannot be shirked.
2. Because of his background, Baum fatally overestimates the power of Darwinian evolution. For example, he completely omits the Eigen error threshold problem, he does not take seriously the gap between the information content of genomes and brains, and he seems to think that adding one bit per generation (which is all evolution can do) is a powerful learning procedure.
3. He's hopelessly starry-eyed about the ability of Darwinian evolution to find "compressed descriptions" (though he's spot on in his emphasis on compression). Both evolution and learning are algorithms for adapting, and Baum completely overlooks the possibility that brains can implement the Darwinian algorithm in a different physical medium (synapses instead of nucleotides). To validly draw the conclusions he jumps to, he would have to prove that either the Darwinian algorithm cannot be implemented neurally, or that it would be far too slow (while the evidence suggests that the basic update can be done neurally a billion times faster neurally than genetically). As Dawkins has emphasised, Darwinism is the only way to get intelligence, but this does NOT mean that only DNA can do it.
In sum, a book for the beach, not for eternity.

Those who are not yet convinced that the brain is a computing mechanism, or who believe that mysticism is required to explain thought, will find quite a bit of value in this book. The book surveys numerous areas of Computer Science, AI, and even a bit of biology, in an attempt to build a case for the brain as a computing mechanism. The book also wades into evolution to try to explain how it came to be so. The scope of the book is ambitious.

Anyone with a background in AI or Cognitive Science will likely find "What is Thought" disappointing as it has little new to say. I fall into this category, and I find a number of aspects of this book unsatisfying.

This is a long book in which there is a short book struggling to get out. The author's main thesis, that the brain is a modular computing mechanism that is the result of evolution, is repeated numerous times at considerable length to the point of tedium. While the author shows his thesis to be consistent with numerous observations, it is never developed to any greater depth. In fact, one of the author's conclusions is that we may never understand the inner workings of the brains "subroutines" because, as a result of evolution, they are now so "compressed".

The author rarely defines his terms. Merely replacing the words "compressed" and "compact" by the word "concise" would enhance the clarity of this book considerably. The author also seems to be of the opinion that generalization, which is the result of "compressed" representations, is the essence of understanding. This view is inadequate for explaining our abilities to plan our own actions and predict the actions of other agents, for example.

Because of the informal, breezy style, the book comes across as an introduction for novices or a position paper rather than a scholarly work. While some may enjoy this style, I find it lacks a certain satisfying clarity and crispness needed for a convincing presentation of such an abstract topic.

The review published in Nature does a better job than I could,
so I'll excerpt it.

"'What is thought?' is not a new question. For Aristotle, thought was what the soul does, and for Descartes it was the unequivocal evidence of one's existence. For Eric Baum, a US expert in machine learning, it is a computer program. This is not a superficial assertion: Baum pursues the idea with elegance, clarity, and considerable pursuasion...

"It is important not to treat the idea that thought is a program in too superficial a fashion. Popular texts often include explanations such as `brain is like hardware and mind is like software.' Baum intends a level of sophistication far above this. Thought for him is the process that 'understands' the complexities of the world...

"Baum's central point is that it is possible for programs to evolve, adapt, and learn, making them more powerful than anything that a programmer can concoct...

"Baum gives a reasoned response to John Searle's claim that no program can 'understand' the world, and to Roger Penrose's contention that conscious insight lies outside the logic that can be achieved by computation...

"this is a splendid book for discovering what is new. It will enthrall some computer scientists and provoke some philosophers. And it should engage general readers who wish to enjoy a clear, understandable description of many advanced principles of computer science."