Brains, Machines, and Mathematics

Transcription

1 Brains, Machines, and Mathematics

2 Michael A. Arbib Brains, Machines, and Mathematics Second Edition With 63 Illustrations Springer-Verlag New York Berlin Heidelberg London Paris Tokyo

3 Michael A. Arbib Departments of Biomedical Engineering, Computer Science, Electrical Engineering, Neurobiology, Physiology, and Psychology University of Southern California Los Angeles, CA USA Library of Congress Cataloging-in-Publication Data Arbib, Michael A. Brains, machines, and mathematics. Bibliography: p. Includes index. 1. Cybernetics. 2. Machine theory. 1. Title. Q31O.A Previous edition: M.A. Arbib, Brains, Machines, and Mathematics by McGraw-Hill, Inc by Springer-Verlag New York Inc. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer-Verlag, 175 Fifth Avenue, New York, New York 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use of general descriptive names, trade names, trademarks, etc. in this publication, even if the former are not especially identified, is not to be taken as a sign that such names, as understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone. Typeset by Asco Trade Typesetting Ltd., Hong Kong. Softcover reprint of the hardcover 1st edition I ISBN-13: l53-4 e-isbn-13: DOl: /

4 To Fred Pollock and Rhys Jones for teaching wisely, wittily, and well

5 Preface to the Second Edition This is a book whose time has come-again. The first edition (published by McGraw-Hill in 1964) was written in 1962, and it celebrated a number of approaches to developing an automata theory that could provide insights into the processing of information in brainlike machines, making it accessible to readers with no more than a college freshman's knowledge of mathematics. The book introduced many readers to aspects of cybernetics-the study of computation and control in animal and machine. But by the mid-1960s, many workers abandoned the integrated study of brains and machines to pursue artificial intelligence (AI) as an end in itself-the programming of computers to exhibit some aspects of human intelligence, but with the emphasis on achieving some benchmark of performance rather than on capturing the mechanisms by which humans were themselves intelligent. Some workers tried to use concepts from AI to model human cognition using computer programs, but were so dominated by the metaphor "the mind is a computer" that many argued that the mind must share with the computers of the 1960s the property of being serial, of executing a series of operations one at a time. As the 1960s became the 1970s, this trend continued. Meanwhile, experimental neuroscience saw an exploration of new data on the anatomy and physiology of neural circuitry, but little of this research placed these circuits in the context of overall behavior, and little was informed by theoretical concepts beyond feedback mechanisms and feature detectors. By contrast, I have always insisted that the metaphor "the brain is a computer" must not be read as reducing the brain to the level of current computer technology, but rather must expand our concepts of computation to embrace the style of the brain, depending on the constant interaction of many concurrently active systems,

6 viii Preface to the Second Edition many of which express their activity in the interplay of spatiotemporal patterns in diverse layers of neurons. However, many computer scientists are now abandoning the serial computer paradigm, and are seeking to understand how problem-solving can be distributed across a network of interacting concurrently active processors, while within AI a whole new field of "connectionism" or PDP (parallel distributed processing) has arisen, which combines the insight of studies of "learning networks" made in the 1950s and 1960s with the sophistication gained from more than two decades of studies in AI. Chapter 1 provides a historical overview tracing the rise of cybernetics, its dissolution into a number of specialized disciplines, and the new rapprochement that makes the Second Edition of Brains, Machines, and Mathematics so timely. Chapter 2 then presents the classic two-state model of the neuron due to McCulloch and Pitts and shows how any finite-state machine can be simulated by a network of such neurons. Chapter 3 introduces the crucial cybernetic concept of feedback and analyzes a problem crucial to the adaptive control of a system, namely, the realization problem of inferring a state-space description of the system from samples of its behavior. Chapter 4 then looks at the role of layered networks of neurons in pattern recognition, introducing two classic schemes for making such networks "learn from experience": Rosenblatt's Perceptron and Hebb's scheme for unsupervised learning. The chapter closes by presenting results due to Minsky-Papert and Winograd Spira, which show how limitations on network topology and neuron complexity limit what a network can do, irrespective of whatever learning rule is employed. It must be stressed that the emphasis here is on automata theory, analyzing the capabilities of networks of simple components similar to McCulloch-Pitts neurons equipped with some "learning rule" to adjust their connections. These are networks "in the style of the brain," rather than models responsible to the current data on cellular neuroanatomy and neurophysiology. This story is continued when we examine a number of studies in the new connectionist or PDP paradigm, which builds AI systems or models cognitive functions in terms of such highly parallel networks. Chapter 5 presents a sampling of current studies of "semi-neural" learning networks, specifically studies of synaptic matrices, Hopfield nets, Boltzmann machines, reinforcement learning, and an interesting back-propagation algorithm. It is in the companion volume The Metaphorical Brain that I present realistic models of the regions of real brains, in addition to related studies in machine vision and robotics. In 1936 (when the word "computer" denoted a person carrying out computations) Alan Turing introduced a model of a computer as a formal machine that could be programmed to carry out any effective procedure, i.e., to do anything that a human could do in the way of symbol manipulation when constrained to follow a well-specified set of instructions. Turing's research was part of an effort by mathematical logicians in the 1930s to delimit what could and could not be achieved by formal procedures. The classic 1943 paper

7 Preface to the Second Edition ix of McCulloch and Pitts showed that networks of their formal neurons could provide the "brain" for any Turing machine. Having devoted several chapters to the sequels to McCulloch and Pitts, we devote the last three chapters to research rooted in the logic of the 1930s, charting the limits of the computable whether or not the computation is implemented in a neural network. Chapter 6 focuses on the capabilities of Turing machines, including the demonstration that there is a universal Turing machine which, when suitably programmed, can simulate any other Turing machine. It also demonstrates the unsolvability of the halting problem, i.e., that there is no program to tell whether an arbitrary machine will eventually halt computation. Chapter 7 studies von Neumann's extension of Turing's universal computer to a universal constructor, and then examines a variety of approaches to self-reproducing machines, all in the framework of cellular or tesellation automata. It is ironic that today many computer scientists use the term "von Neumann computer" to refer to serial computer architectures, for von Neumann's work on cellular automata makes him a pioneer in "non-von Neumann architecture"! Finally, Chapter 8 provides two accessible proofs of G6del's celebrated Incompleteness Theorem, and deepens the reader's understanding by contrasting it with G6del's Completeness Theorem, for which a proof is also given. We include a number of little known results, such as that the incompleteness found by G6del can be removed incrementally in a mechanical fashion, and use this deeper understanding to argue against those philosophers who believe that G6del's Incompleteness Theorem places inherent limitations, which are not shared by humans, on the possible intelligence of machines. Havmg thus outlined the present volume, I now briefly present its relation to five other volumes to be published at much the same time. The Metaphorical Brain (Michael A. Arbib, 2nd edition, Wiley-Interscience, 1988) presents my approach to brain theory in which perceptual and motor schemas provide the functional analysis of behavior, while layered neural networks provide models for neural mechanisms of visuomotor coordination that can be put to experimental test. Neural Models of Visuomotor Coordination (Michael A. Arbib, Rolando Lara, Francisco Cervantes-Perez, and Donald House, in preparation) will provide a much more detailed look at our research on modeling the mechanisms of vi suo motor coordination, especially in frog and toad; on pattern recognition, depth perception, and detour behavior; and reviewing experimental data and offering a careful analysis of both mathematical and computer models. From Schema Theory in Language (Michael A. Arbib, E. Jeffrey Conklin, and Jane C. Hill, Oxford University Press, 1987) develops a schema-theoretic approach to language that is exemplified by three studies: of the effects of brain damage on sentence comprehension, of language acquisition by a two-year-old child, and of the use of salience in generating verbal descriptions. In each case, the result is a computational model linked to cognitive studies of human language performance. The Construction of Reality (Michael A. Arbib and Mary B. Hesse, Cambridge Uni-

8 x Preface to the Second Edition versity Press, 1986) synthesizes my individualist view of schemas as "units of knowledge in the head" with Professor Hesse's view of knowledge as a social construct shared, e.g., by a community of scientists or of religious believers. We thus develop and extend the view of knowledge, including language, as inherently metaphorical and dynamic. Within this shared perspective, we then debate such issues as the freedom of the will and the reality of God. Finally, In Search of the Person (Michael A. Arbib, University of Massachusetts Press, 1985) is a small volume, designed to be read in an evening or two, either for itself alone or as an invitation to the weightier development of its themes in The Construction of Reality. It is subtitled "Philosophical Explorations in Cognitive Science." Much of Chapter 5 was written while I was a Visiting Scholar at the Institute for Cognitive Science at the University of California at San Diego (UCSD) during the Academic Year , on leave from the University of Massachusetts at Amherst. I would like to thank Andy Barto for his invaluable survey of the material for me before I left Amherst and for his vital help with the final draft, as well as to thank David Rumelhart, David Zipser, Gary Cottrell, Ron Williams, and other members of the PDP group for their scientific hospitality and insightful discussions during my stay at UCSD, and Geoffrey Hinton and Terry Sejnowski for their helpful comments on the semifinal draft. Finally, I thank Darlene Freedman for her excellent help in bringing the manuscript to publication. Los Angeles, CA August 1986 Michael A. Arbib

9 From the Preface to the First Edition This book forms an introduction to the common ground of brains, machines, and mathematics, where mathematics is used to exploit analogies between the working of brains and the control- computation - communication aspects of machines. It is designed for a reader who has heard of such currently fashionable topics as cybernetics and G6del's theorem and wants to gain from one source more of an understanding of them than is afforded by popularizations. Here the reader will find not only what certain results are, but also why. A lot of ground is covered, and the reader who wants to go further should find himself reasonably well prepared to go on to study the technical literature. To make full use of this book requires a moderate mathematical background-a year of college mathematics (or the equivalent "mathematical maturity"). However, much of the book should be intelligible to the reader who chooses to skip the mathematical proofs, and no previous study of biology or computers is required at all. The use of microelectrodes, electron microscopes, and radioactive tracers has yielded a huge increase in neurophysiological knowledge in the past few decades. Even a multivolume work such as the Handbook of Neurophysiology cannot fully cover all the facts. Many neurophysiological theories, once widely held, are being questioned as improved techniques reveal finer structures and more sophisticated chemicoelectrical cellular mechanisms. This means that our presentation of mathematical models in this book will have to be based on a grossly simplified view of the brain and the central nervous system. The reader may well begin to wonder what value or interest the study of such systems can have. There is a variety of properties-memory, computation, learning, purposiveness, reliability despite component malfunction-which it might seem

10 xii From the Preface to the First Edition difficult to attribute to "mere mechanisms." However, herein lies one important reason for our study: By making mathematical models, we have proved that there do exist purely electrochemical mechanisms that have the above properties. In other words, we have helped to "banish the ghost from the machine." We may not yet have modeled the mechanisms that the brain employs, but we have at least modeled possible mechanisms, and that in itself is a great stride forward. There is another reason for such a study, and it goes much deeper. Many of the most spectacular advances in physical science have come from the wedding of the mathematicodeductive method and the experimental method. The mathematics of the last 300 years has grown largely out of the needs of physics-applied mathematics directly, and pure mathematics indirectly by a process of abstraction from applied mathematics (often for purely esthetic reasons far removed from any practical considerations). In these pages we coerce what is essentially still the mathematics of the physicist to help our slowly dawning comprehension of the brain and its electromechanical analogs. It is probable that the dim beginnings of biological mathematics here discernible will one day happily bloom into new and exciting systems of pure mathematics. Here, however, we apply mathematics to derive far-reaching conclusions from clearly stated premises. We can test the adequacy of a model by expressing it in mathematical form and using our mathematical tools to prove general theorems. In the light of any discrepancies we find between these theorems and experiments, we may return to our premises and reformulate them, thus gaining a deeper understanding of the workings of the brain. Furthermore, such theories can guide us in building more useful and sophisticated machines. The beauty of this mathematicodeductive method is that it allows us to prove general properties of our models and thus affords a powerful adjunct to model making in the wire and test-tube sense. Biological systems are so much more complicated than the usual systems of physics that we cannot expect to achieve a fully satisfactory biological mathematics for many years to come. However, the quest is a very real and important one. This book strives to introduce the reader to its early stages. He or she will, I hope, find that the results so far obtained are of interest. Certainly they represent only a very minute fraction of what remains to be found-but the start of a quest is nonetheless exciting for being the start. I do not believe that the application of mathematics will solve all our physiological and psychological problems. What I do believe, though, is that the mathematicodeductive method must take an important place beside the experiments and clinical studies of the neurophysiologist and the psychologist in our drive to understand brains, just as it has already helped the electrical engineer to build the electronic computers that, although many, many degrees of magnitude less sophisticated than biological organisms, still represent our closest man-made analog to brains.

11 From the Preface to the First Edition xiii This book is a revision of the lecture notes of a course delivered in June-August of 1962 at the University of New South Wales in Sydney, Australia. I want to thank John Blatt for inviting me to the Visiting Lectureship; Derek Broadbent for inviting me to broadcast the lectures; and Joyce Kean for her superb job of typing up the original lecture notes. I have spent the last two years with the Research Laboratory of Electronics and the Department of Mathematics at the Massachusetts Institute of Technology on a research assistantship (supported by the U.S. Armed Forces and National Institute of Health). [I was a graduate student at MIT from January of 1961 till September of 1963.] lowe so many debts to the people there that 1 cannot fully do justice to them. 1 do particularly want to thank Warren McCulloch for his continual help and encouragement. It was George W. Zopf who first urged publication of the lectures. Bill Kilmer gave the original lecture notes a helpful and critical reading. For years 1 have nurtured the desire to claim at a point such as this, "Any mistakes which remain are thus solely his responsibility." However, this would be a sorry expression of a very genuine gratitude, and so 1 follow convention and admit that any errors that remain are my responsibility. Finally 1 should like to thank all those authors whose work I have quoted and their publishers for so graciously granting me permission to use their material. M.A.A.

12 Contents Preface to the Second Edition From the Preface to the First Edition... vii xi CHAPTER 1 A Historical Perspective The Road to Cybernetics Defined and Dissolved The New Rapprochement CHAPTER 2 Neural Nets and Finite Automata Logical Models of Neural Networks States, Automata, and Neural Nets CHAPTER 3 Feedback and Realization The Cybernetics of Feedback From External to Internal Descriptions CHAPTER 4 Pattern Recognition Networks Universals and Feature Detectors The Perceptron Learning without a Teacher Network Complexity

13 xvi Contents CHAPTERS Learning Networks Connectionism Synaptic Matrices Hopfield Nets and Boltzmann Machines Reinforcement Learning and Back-Propagation CHAPTER 6 Turing Machines and Effective Computations Manipulating Strings of Symbols Turing Machines Introduced Recursive and Recursively Enumerable Sets CHAPTER? Automata that Construct as well as Compute Self-Reproducing Automata Tesselations and the Garden of Eden Toward Biological Models CHAPTER 8 GOdel's Incompleteness Theorem The Foundations of Mathematics Incompleteness and Its Incremental Removal Predicate Logic and Godel's Completeness Theorem Speed-Up and Incompleteness The Brain - Machine Controversy APPENDIX Basic Notions of Set Theory Index