·Intelligence

The Eye, ·. the Brain,

and theComputer

Intelligence

The Eye,the Brain,and the

ComputerMartin A. Fischler

SRI International

Oscar FirscheinSRI International

A....ADDISON-WESLEY PUBLISHING COMPANY

Reading, Massachusetts • Menlo Park, California • Don Mills. OntarioWokingham, England • Amsterdam> Sydney • Singapore • TokyoMadrid • Bogota • Santiago • San Juan :, ,

Sponsoring Editor • Peter S. GordonProduction Supervisor • BetteJ.. AaronsonCopy Editor • G. W. HelfrichText Designer and Art Coordinator • Geri Davis, Quadrata Inc..Cover Designer • Marshall HenrichsManufacturing Supervisor • Hugh Crawford

Library of Congress Cataloging-in-Publication Data

Fischler, Martin A.Intelligence : the eye, the brain, and the computer..

Bibliography: p.Includes index..1. Artificial intelligence. 2. Machine learning.

3. Cognition. 4. Perception. I. Firschein, Oscar.II. Title.Q335.F57 1987 006.3 86-3557ISBN 0-201-12001-1

Reprinted with corrections April, 1987

Copyright © 1987byAddison-Wesley Publishing Company, Inc. All rights reserved. Nopart ofthis publication may be reproduced, stored in a retrieval system, or transmitted inanyform or byan means, electronic, mechanical, photocopying, recording, or otherwise,without the prior written permission of the publisher. Printed in the United States ofAmerica. Published simultaneously in Canada.

BCDEFGHU-HA-8987

Preface

This book is intended to be an intellectual journey into the domain of human andmachine intelligence. The subjectmatter has been approached from a conceptualand sometimes even philosophical point of view, one biasedby our experience in thatbranch of computer and cognitive sciencecalled artificial intelligence (AI). On thisjourney we will oftenbe dealing with topics,such as the operation of the brain, whereknowledge is lacking, and where there are only vague conjectures as to possible mecha­nisms. In our review of machine intelligence, wewill discuss both the present-day andultimate limits of machine performance.

We intend the book to provide the Scientific Americanlevel reader with an under­standingof the conceptof intelligence, the nature of the cognitive and perceptual capa­bilities of peopleand machines, and the representations and algorithms used to attainintelligent behavior. While we have attempted to makethe material understandable byan educated layman, weare equally concernedwith having somethingimportant to sayto our professional colleagues and peers.

"Boxes" have been used to augment the text. These boxes,and longer appendices,present material of interestthat expand on the topics beingdiscussed, and sometimesthey may contain technical material of more interestto the specialist. It is intended thatthe text shouldstand on its own; the reader can usually omit the boxes until a laterreading.

v

vi

PREFACE

A unifying themein our exposition is the critically important conceptof represen­tation. Issues relatedto this and other importantconcepts may be raised in an earlierchapter, our position stated, and then elaborated and supported in later chapters. Ifourinitial discussion on someimportantpoint appears to be briefor unsupported, havefaith, sincewewill probably return to this topic many times beforeyour journeythrough the book is completed.

Our primary purpose in the firstpart of this book is to provide a basis for under­standingthe nature of intelligence and intelligent behavior: (a) as it exists in man andhigher animals, (b) as it couldexist in a machine, and (c) as a scientific discipline con­cerned with the mechanisms and limits involved in acquiring, representing, and apply­ing knowledge.

The lastpart of the book deals with perceptionand primarily vision, the means bywhich knowledge about our physical environment is acquired. Wewill see that percep­tual behavior, far from beingpassive or mechanical, requires a reasoningability at leastequal to that needed for the most difficult problem-solving tasks.

The flavor ofthis book can be gleanedfrom the following brief description ofchapter topics.

Intelligence

This chapter discusses the nature of intelligence, indicating its characteristics or com­ponents. Weexamine the issue of whetherintelligence is primarily associated with func­tional behavior (performance) or with the structuresand machinery that give risetobehavior (competence). The role of language in intelligence is explored and wespecu­late about the role of emotion (pain.pleasure), aesthetic appreciation, and physical inter­action with the external world, in achieving intelligent behavior. The subjectof artificial(machine) intelligence is introduced and something of its history, goals, and approach isindicated. Finally, the problem of measuring or evaluating intelligence is treated.

The Brainand the Computer

We examine the ultimate capacity of the computeras an intelligence engine andwhetherman can createa machine moreintelligent than himself. Are there componentsof man's intelligence which cannot be found in anyanimal or duplicatedin a machine?Are there essential differences between the architecture ofthe human brain and thedigital computer which imply a difference in capacity or competence? Are there prob­lemsthat cannot in theory be solved by a machine?

The Representation of Knowledge

We explore the conceptof knowledge, and note that the nature of the physical encod­ing of knowledge is an importantconsideration in achieving intelligent behavior. Wedescribehow knowledge can be represented in the memory of a computerand raise thequestionas to whether there are elements of knowledge that cannot be described ordiscussed.

vii

PREFACE

Reasoning

This chapterdiscusses the role of reasoning in intelligent behavior. Wedescribe how areasoning system can usea formal language to representthings in the world and theirrelationships, and how it can solve problems using such a representation. The limits offormal representations in their ability to solve problems and to dealwith vaguely formu­lated problems are indicated. The crucial difficulty arises of how a reasoningsystemmight selectthe best representation for a given problem, since how can a reasoningsystem know which facts in its database are relevant to the problem at hand?

Learning

While a primary attribute of learning is the formulation of concepts and representationsto dealeffectively with new situations, an important aspect ofthe learning process is theidentification ofa particular situation as beingan instance ofan already-learned con­cept. We will find that the noting ofsuch similarities is far from trivial, and that analogylies at the heart ofmuch ofthe learning process. This chapterexamines the differentmodes of learning, and indicates the extentofpresent-day machine learning.

Language and Communication

This chapter examines the purpose of language and methods ofcommunicating, andexplores the roleof language in intelligent behavior. The question arisesas to whethermanis the only organism with a true capacity for language. Finally, we discuss the ap­proaches usedfor machine "understanding" ofnatural language.

ExpertSystems

We describe expert systems, programs that duplicate human expertise in a specializedfield such as medicine or engineering. These systems are of interestbecause they canact at a high level of competence by relying on a detailed knowledge baseof informa­tion, despite a rather limited reasoning ability.

Biological Vision

Visual perception is one of the mostdifficult tasks yetfaced in attempting to designmachines that can duplicate human behavior. The relationship between the evolution­ary development ofvision in organisms and theirneedsand limitations is explored. Weexamine the universal mechanisms that nature has devised and which offer a solutiontothe problem of visual understanding ofthe world.

Computational Vision

This chapterdescribes the representations and algorithms that are used in computeranalysis ofimages. We indicate how sensor information is converted to an array ofnumbers and the problems involved in deducing a model ofthe three-dimensional world

viii

PREFACE

by means ofoperations carried out on such image arrays. The major problems andparadigms underlying currentattempts to achieve machine vision are discussed.

Epilogue

In this concludingchapterwerestateand summarize the mostimportant views andarguments relevant to the modeling of intelligence as an information processing activitythat can be carried out bya machine. We discuss whether it is possible to construct anintelligent machine that can function in the world.

Acknowledgments

We owe a maior debt to three peoplefor their extensive helpwith respect to both tech­nical and organizational aspects of this book: Professor Walter Gong, SanJoseStateUniversity; Dr. Richard Duda, Syntelligence; and Professor Nils Nilsson, StanfordUniversity.

The Addison-Wesley reviewers, Anthony Paseraand Werner Feibel, were masterfulin their criticisms, not only pointing out deficiencies, but also providingthe referencesthat broadened our thinking on a number of topics. Chris Varley, our original editoratAddison-Wesley, was very enthusiastic about the project, and wegreatly appreciate thecontinuing supportof Peter Gordon, our current editor.

Various chaptersof the book were reviewed bythe following people, but we takefull blame for any remaining errors in the text: James Dolby, Martin Billik, and JonPearce,San JoseState University; Gordon Uber, Lockheed Palo Alto Research Labora­tory; Peter Cheeseman, NASA Ames; Hugh Crane, Don Kelly, Michael Georgeff RayPerrault, Jerry Hobbs, Phil Cohen, and Enrique Ruspini, all ofSRI. SRIcolleagues whohelped in other ~ays include Ken Laws, Grahame Smith, Lynn Quam, Helen Wolf,Steve Barnard, BobBolles, Alex (Sandy) Pentland, and Andrew Hanson.

Weare grateful for the environment of the SRIArtificial Intelligence Center madepossible bythe past leadership of Peter Hartand Nils Nilsson, and now byStan Rosen­schein. We acknowledge the earlierinfluence on our thinking by Russell Kirsch andDavid Willis. . .

Wewish explicitly to acknowledge the contribution made by DARPA, whose en­lightened supportat the national level was critical to the development ofartificial intel­ligence as a scientific discipline. In particular, we cite thosetechnical monitors withwhom wehad directcontact with in our work in machine vision: Dave Carlstrom, LarryDruffel, RonOhlander, BobSimpson, and (at a higher 'management level) Bob Kahnand SaulAmarel.

Our families served as intelligent layreviewers: Beverly, Sharon, and Ron Fischler;and Ben,Joseph, and ThedaFirschein. Sharon Fischler also contributed some oftheideas for the final artwork.

Menlo Park, CA M.A.F.'O.F.

1

Contents

Part One

Foundations

Intelligence

WhatIs Intelligence? "3• Theories of Intelligence 4• Theories ofMind 8How CanIntelligence Be Measuredor Evaluated? 10• Assessing Human Intelligence 10~ Assessing Machine Intelligence 12IsMan The Only Intelligent Animal? 12The Machinery of Intelligence• Reliance on Paradigms 13• Two Basic Paradigms 13Artificial Intelligence (AI) 15• The Mechanization of Thought 15• The Computer and the Tho Paradigms 18• How CanWe Distinguish Between Mechanical and Intelligent Behavior? 18• The Role of Representation in Intelligent Behavior 20

SUMMARY AND DISCUSSION 20

ix

1

3

\

x

CONTENTS

2 The Brain and The ComputerThe Human Brain 24• Evolution of the Brain 24• Architecture of the Brain 30The Computer 39• The Nature ofComputer Programs and Algorithms 40• The Universal 'luring Machine 43Limitations on the Computational Ability of a Logical Device 43• The G5del Incompleteness Theorem 43• Unsolvability byMachine 45• Implications of G6del's Theorem 46• Computational Complexity-the Existence ofSolvable but Intrinsically

Difficult Problems 47Limitations on the Computational Ability ofa Physical Device 49• Reliable Computation WithUnreliable Components 51

DISCUSSION 55

23

Appendixes 58

2-1 The NerveCell and Nervous System Organization2-2 The Digital Computer

3 The Representation of KnowledgeRepresentation: Concepts 64• Formvs. Contentof Knowledge 64• Representing Knowledge 65• The Relation Between a Representation and ThingsRepresented 66Role of Representation 67Representations Employed in Human Thinking 67• The Use ofModeIs and Representations 68• The Use of "Visual" Representations 69Effectiveness of a Representation 69Representations Employed in Artificial Intelligence 71• FeatureSpace (orDecision Space) 74• Decision Tree/Game Tree 75• Isomorphic/Iconic/Analogical Representations 77

DISCUSSION 80

5861

63

•4

xi

CONTENTS

Part Two

Cognition

Reasoning and Problem Solving

Human Reasoning 84• Human Logical Reasoning 85• Human Probabilistic Reasoning 86Formal Reasoning and Problem Solving 87• Requirements fora Problem Solver 87• Categories of Reasoning 88The Deductive Logic Formalism 90• Propositional Calculus 90• Propositional Resolution 91• Predicates 93• Quantifiers 93• Semantics 93• Computationallssues 94• Nonstandard Logics 95Inductive Reasoning 96• Measures of Belief 97• Bayesian Reasoning 98• Belief Functions 100• Representing a Problem in a Probabilistic Formalism 103• Comments Concerning the Probabilistic Formalism 103Additional Formalisms for Reasoning 106• Algebraic/Mathematical Systems 106• Heuristic Search 106• Programming Systems that Facilitate Reasoning and Problem Solving 108• Common-Sense Reasoning 109Problem Solving andTheorem Proving 110• Representing the Problem III• The Predicate Calculus Representation for the Monkey/Banana

(M/B) Problem 112• PROLOG Representation of the MIB Problem 113• Production Rule (OPS-5) Representation for the MIB Problem 113• General Problem Solver Representation for the MIB Problem 114• Formalisms or Reasoning Systems? 115• Relating Reasoning Formalisms to the RealWorld 115

DISCUSSION 116

81

83

xii

CONTENTS

Appendixes

4-1 AI Programming Languages4-2 The Monkey/Bananas Problem

5 Learning

Human and Animal Learning 130• Types ofAnimal Learning 131• Piaget's Theory of Human Intellectual Development 132Similarity 135• Similarity Based on ExactMatch 136• Similarity Based on Approximate Match 137Learning 137• Model Instantiation: ParameterLearning 138• Model Construction: Description Models 143• Concept Learning 148

DISCUSSION 151

117

117122

129

Appendix 152

5-1 Parameter Learning for an Implicit Model

6 Language and Communication

Language in Animals and Man 158• Brain Structures Associated with Language Production

and Understanding 159• Human Acquisition of Language 161• Animal Acquisition of Language 164Language and Thought 165Communication 167• The Mechanics ofCommunication 167• Vocabulary ofCommunication 168• Understanding Language 169Machine Understanding of Language 171• Faking Understanding 171• WhatDoes it Mean for a Computer to Understand? 171• The Study ofLanguage 173

DISCUSSION 185

152

157

xiii

CONTENTS

Appendix

6-1 Representing Passing Algorithms

7 Expert/Knowledge-Based Systems

Human Experts 190Production Systems 191• Control StructuresUsed in Production Systems 192Production Systems in Psychological Modeling 195Production Rule-Type ExpertSystems 197• Plausible Reasoning in ExpertSystems 198• Basic AIIssues 200

DISCUSSION 202

186

186

189

Appendix 202

7-1 PROSPECTOR Procedure for Hypothesis Updating 202

•8

PartThree

Perception (Vision)

Vision

The Nature of Organic Vision 207The Evolution and Physiology of Organic Vision 209• Seeing and the Evolution ofIntelligence 209• Evolution and Physiology ofthe Organic Eye 211• Eye and Brain 213The Psychology of Vision 220• Perceiving the Visual World: Recognizing Patterns 220• Perceptual Organization 224• Visual illusions 226• Visual Thinking, Visual Memory, and Cultural Factors 229

DISCUSSION 232

205

207

xiv

CONTENTS

Appendixes

8-1 Color Vision and Light8-2 Stereo Depth Perception and the Structure of the

Human Visual Cortex

9 Computational Vision

Signals-to-Symbols Paradigm 241Low Level SceneAnalysis (LLSA) 242• Image Acquisition (Scanning and Quantizing) 243• Image Preprocessing (Thresholding and Smoothing) 245• Detection of Local Discontinuities and Homogeneities

(Edges, Texture, Color) 248• Local SceneGeometry froma Single Image

(Shape from Shadingand Texture) 256• Local SceneGeometry from Multiple Images (Stereo and Optic Flow) 259Intermediate Level SceneAnalysis (lLSA) 262• Image/Scene Partitioning 264-• Edge Linking and Deriving a LineSketch 269• Recovering Three-Dimensional SceneGeometry from a LineDrawing 272• Image Matching 276• Object Labeling 278• Model Selection and Instantiation 279High Level SceneAnalysis (HLSA) 281• Image/Scene Description 281• Knowledge Representation 283• The Problem of High-Level SceneAnalysis 285• Reasoning Abouta Simple Scene 285

DISCUSSION 286A Basic Concern AboutSignals-to-Symbols 287

• Necessary Attributes of a Machine Vision System 288• Summary 289

233

233

236

239

Append~es 289

9-1 Mathematical Techniques for Information Integration 2899-2 A Path-FindingAlgorithm 2979-3 Relational (Rubber Sheet) Image Matching 299

Epilogue 301

Bibliography 311

Index 325