Stillings Et Al 1995
Stillings Et Al 1995
Stillings Et Al 1995
iii
Cognitive Science
An Introduction
Second Edition
Neil A. Stillings,
Steven E. Weisler,
Christopher H. Chase,
Mark H. Feinstein,
Jay L. Garfield, and
Edwina L. Rissland
A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England
Page iv
© 1995 Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information
storage and retrieval) without permission in writing from the publisher.
This book was set in Palatino by Asco Trade Typesetting Ltd., Hong Kong and was printed and bound in the United States of America.
Library of Congress CataloginginPublication Data
Cognitive science: an introduction / Neil A. Stillings . . . [et al.].
—2nd ed.
p. cm.
"A Bradford book."
Includes bibliographical references.
ISBN 0262193531
1. Cognition. 2. Cognitive science. I. Stillings, Neil A.
BF311.C5523 1995
153—dc20 9429553
CIP
Second printing, 1998
Page v
To our parents
Page vii
Contents
List of Authors xi
Preface to the Second Edition xiii
Preface to the First Edition xv
Note to the Teacher xvii
Chapter 1 What Is Cognitive Science? 1
1.1 The Cognitive View 1
1.2 Some Fundamental Concepts 2
1.3 Information Processes Can Be Analyzed at Several Levels 7
1.4 Computers in Cognitive Science 11
1.5 Applied Cognitive Science 12
1.6 The Interdisciplinary Nature of Cognitive Science 13
Suggested Readings 14
References 14
Chapter 2 Cognitive Psychology: The Architecture of the Mind 15
2.1 The Nature of Cognitive Psychology 15
2.2 The Notion of Cognitive Architecture 16
2.3 A Global View of the Cognitive Architecture 18
2.4 Propositional Representation 26
2.5 Schematic Representation 32
2.6 Cognitive Processes, Working Memory, and Attention 37
2.7 Mental Images 42
2.8 Automatic and Controlled Processes 55
2.9 The Acquisition of Skill 58
2.10 The Connectionist Approach to Cognitive Architecture 63
Suggested Readings 83
References 83
Page viii
Chapter 3 Cognitive Psychology: Further Explorations 87
3.1 Concepts and Categories 87
3.2 Memory 99
3.3 Reasoning 116
3.4 Problem Solving 129
Suggested Readings 135
References 136
Chapter 4 Artificial Intelligence: Knowledge Representation 139
4.1 The Nature of Artificial Intelligence 139
4.2 Knowledge Representation 151
Suggested Readings 173
References 174
Chapter 5 Artificial Intelligence: Search, Control, and Learning 177
5.1 Search and Control 177
5.2 Learning 192
Suggested Readings 212
References 213
Chapter 6 Linguistics: The Representation of Language 215
6.1 The Study of Linguistic Knowledge 215
6.2 Phonology 220
6.3 Syntax 241
6.4 Universals 260
Suggested Readings 267
References 267
Chapter 7 Neuroscience: Brain and Cognition 269
7.1 Introduction to the Study of the Nervous System 269
7.2 Organization of the Central Nervous System 270
7.3 Neural Representation 291
7.4 Neuropsychology 306
7.5 Computational Neuroscience 323
Suggested Readings 325
References 326
Page ix
Chapter 8 Philosophy: Foundations of Cognitive Science 331
8.1 Philosophy in Cognitive Science 331
8.2 The Enterprise of Cognitive Science 335
8.3 Ontological Issues 345
8.4 Epistemological Issues 367
8.5 The State of Cognitive Science 373
Suggested Readings 374
References 375
Chapter 9 Language Acquisition 379
9.1 Milestones in Acquisition 380
9.2 Theoretical Perspectives 392
Suggested Readings 405
References 406
Chapter 10 Semantics 409
10.1 Semantics and Cognitive Science 409
10.2 Meaning and Entailment 410
10.3 Reference 411
10.4 Sense 418
10.5 Problems in PossibleWorlds Semantics 424
10.6 Cognitive and Computational Models of Semantic Processing 426
Suggested Readings 432
References 433
Chapter 11 Natural Language Processing 435
11.1 Preliminaries 435
11.2 On the Role of Grammar in Language Processing 436
11.3 Connectionist Models 445
11.4 On the Role of Discourse 447
11.5 More on the Role of General Knowledge 452
11.6 Production 454
11.7 Conclusion 458
Suggested Readings 458
References 458
Page x
Chapter 12 Vision 461
12.1 The Problem of Vision 461
12.2 LowLevel Visual Processes 464
12.3 Intermediate Processes and Representations in Vision 479
12.4 HighLevel Visual Processes 490
12.5 The Architecture of Visual Computation 502
Suggested Readings 512
References 512
Index 517
Page xi
List of Authors
Neil A. Stillings
Ph.D. in psychology, Stanford University. Professor of Psychology in the School of Communications and Cognitive Science, Hampshire College. A founding member
of Hampshire's cognitive science program with over twenty years of teaching experience. Organizer of the 1986 national workshop on teaching cognitive science,
funded by the Sloan Foundation, and the 1993 national workshop on undergraduate cognitive science, funded by the National Science Foundation. Coeditor of the
entire textbook and primary author of chapters 1, 2, 3, and 12.
Steven E. Weisler
Ph.D. in linguistics, Stanford University. Associate Professor of Linguistics in the School of Communications and Cognitive Science, Hampshire College. Research and
publications in syntax and semantics. Coorganizer of the 1993 national workshop on undergraduate cognitive science, funded by the National Science Foundation.
Coeditor of the entire textbook, primary author of chapter 9, and coauthor of chapters 6 and 10.
Christopher H. Chase
Ph.D. in neuroscience, University of California at San Diego. Associate Professor of Cognitive Science in the School of Communications and Cognitive Science,
Hampshire College. Research and publications in cognitive neuroscience, particularly reading development and developmental dyslexia. Primary author of chapter 7.
Mark H. Feinstein
Ph.D. in linguistics, City University of New York. Professor of Linguistics in the School of Communications and Cognitive Science, Hampshire College. Member of
the Hampshire faculty since 1976. Research and publications in phonology, mammalian vocalization, animal cognition, and evolution of cognition and behavior. Primary
author of chapter 11 and coauthor of chapter 6.
Jay L. Garfield
Ph.D. in philosophy, University of Pittsburgh. Professor of Philosophy in the School of Communications and Cognitive Science, Hampshire College. Research and
publications in foundations of cognitive science, philosophy of mind, philosophy of language, metaphysics, epistemology, and Buddhist philosophy. Directs
Hampshire's exchange program with the Tibetan universities in exile in India. Author of Belief In Psychology (MIT Press) and author or editor of other books.
Primary author of chapter 8 and coauthor of chapter 10.
Page xii
Edwina L. Rissland
Ph.D. in mathematics, MIT. Professor of Computer Science, University of Massachusetts at Amherst and Lecturer on Law, Harvard University Law School.
Research and publications on knowledge representation, casebased reasoning, artificial intelligence and the law, and machine learning. Fellow of the American
Association for Artificial Intelligence and Vice President of the International Association for Artificial Intelligence and Law. Primary author of chapters 4 and 5.
Page xiii
Preface to the Second Edition
The second edition of Cognitive Science: An Introduction retains the organizational format and the level of presentation of the first edition. Many chapters of the
book have been substantially revised, however, and several new chapter sections have been added. The changes are detailed in the Note to the Teacher. The team of
authors who wrote the first edition remains largely unchanged. David Rosenbaum and Lynne BakerWard have gone on to other pursuits. We have particularly missed
David's colleagueship and his ability to put his love of the field into words. Fortunately, Christopher Chase joined the Hampshire College faculty at just the right time
and volunteered to take over the chapter on neuroscience.
One or two people were primarily responsible for each chapter: Stillings, chapters 1, 2, 3, 12, and several passages on neural computation in chapter 7; Rissland,
chapters 4 and 5; Feinstein and Weisler, chapter 6; Chase, chapter 7; Garfield, chapter 8; Weisler, chapter 9; Garfield and Weisler, chapter 10; and Feinstein,
chapter 11. Neil Stillings and Steven Weisler edited the entire manuscript.
Many people deserve our thanks for helping with the preparation of the second edition. The authors are part of the cognitive science community in the fivecollege
area centered in Amherst, Massachusetts. It is impossible to mention all of the colleagues and students who contributed to our work, but we thank them for their help.
We appreciate the suggestions and corrections of people who wrote to us about the first edition. Patricia Churchland, Richard Thompson, Lee Spector, Jamie Callan,
Paul Utgoff, and Jody Daniels reviewed parts of the manuscript for the second edition and made many helpful suggestions. Conversations with Gary Marcus, Sean
Stromsten, Sean Hill, Andy Barto, and Richard Yee were also very helpful. Rebecca Neimark prepared the new figures with admirable patience and professionalism.
Michael Zenner, Michael Patterson, and Stacey Guess helped with references and with the computerized scanning of the first edition. Leni Bowen helped with
permissions, typing, and editing at several crucial points, and she kept Hampshire College's cognitive science program on an even keel throughout our work.
We could not have seen this revision through without Betty Stanton's generosity, faith, and friendly persistence and Teri Mendelsohn's professionalism and unfailing
good cheer. The copy editors, Anne Mark and David Anderson, caught our lapses in clarity and our stylistic imperfections with almost frightening accuracy, and they
even corrected our antiquated Beethoven scholarship.
Katherine Pfister and Monica and Tim Stillings deserve special thanks for their support for Neil Stillings during the preparation of this edition.
Page xv
Preface to the First Edition
The seven authors of this book have worked together for many years in the cognitive science community that has grown up in the fivecollege area surrounding
Amherst, Massachusetts. Six of us (Stillings, Feinstein, Garfield, Rosenbaum, Weisler, and BakerWard) teach in Hampshire College's School of Communications and
Cognitive Science (BakerWard recently moved to North Carolina State University). The seventh, Edwina Rissland, is a member of the Computer and Information
Science Department at the University of Massachusetts. We have all also been members of the interdepartmental graduate program in cognitive science at the
University of Massachusetts.
The cognitive science program at Hampshire College is a unique educational experiment. Hampshire was planned during the 1960s as an experimenting college, and
cognitive science was one of the innovative programs it offered when it opened its doors in 1970. In 1972 a single academic department (then called the School of
Language and Communication) was formed to house the cognitive science faculty along with faculty members in computer science and communications. The authors
would like to thank the college and its leadership for allowing us to develop our commitment to the idea that cognitive science could be an exciting area of
undergraduate study. We would also like to thank William Marsh, who was the cofounder and for many years the leader of the program.
Although we worked jointly to lay out the book and rework the manuscript, one or two people were primarily responsible for each chapter: Stillings, chapters 1, 2, 3,
and 12; Rissland, chapters 4 and 5; Feinstein and Weisler, chapter 6; Rosenbaum, chapter 7; Garfield, chapter 8; Weisler, chapter 9; Garfield and Weisler, chapter
10; Feinstein, chapter 11; and BakerWard, the cognitive development section of chapter 3. Neil Stillings edited the entire manuscript.
Because seven people contributed to the book and because it is rooted in fifteen years of curricular experimentation and development at Hampshire College and the
University of Massachusetts, it is impossible to mention everyone who contributed to its development. We thank all of the friends and colleagues who helped bring the
book into existence.
The first draft of the book was written with the support of a faculty development grant to the authors from the Charles A. Dana foundation. The completion of the
book would have been impossible without the enthusiastic support of Betty Stanton of The MIT Press.
The following people reviewed various parts of the manuscript at various points in its preparation. Gary Dell, Stevan Harnad, Keith Holyoak, and Zenon Pylyshyn
deserve particular thanks for reading large chunks of the manuscript. Michael Arbib, Kevin Ashley, Robert Berwick, Charles Briggs, Carol Christensen, Charles
Clifton, Bo Dahlbohm, Willem DeVries, Michael Gazzaniga, Allen Hanson, Norbert Hornstein,
Page xvi
John Haugeland, Haym Hirsch, David Kelley, Judith Kroll, David LeBeaux, Wendy Lehnert, Ken Livingston, Dan Lloyd, Lynn Nadel, Marty Ringle, Penni Sibun,
Catherine Sophian, Devika Subramaniam, Paul Utgoff, Marie Vaughn, Robert Wall, Thomas Wasow, Sandy Waxman, and Bonnie Webber all read one or more
chapters. All of these people exercised their professional expertise with care, and in addition many of them shared their thoughts on the nature of cognitive science and
the challenge of teaching it to undergraduates.
We owe a great debt to Anne Mark, who served as manuscript editor. Her sense of expository and organizational clarity and her eye for both conceptual and
typographical error are staggering. Nelda Jansen turned our ideas for figures into finished art work with efficiency and unfailing good humor. Erik Antelman and John
Gunther also provided invaluable assistance in preparing the figures.
David Rosenbaum's preparation of chapter 7 was partially supported by a Research Career Development Award from the National Institute of Neurological and
Communicative Disorders and Stroke (1 K04 NS0094201) and by a grant from the National Science Foundation (BNS8408634). The preparation of chapter 12
was partially supported by a grant from the National Science Foundation (SER8163019) to Neil Stillings.
Ruth Hammen and Leni Bowen kept the School of Communications and Cognitive Science at Hampshire College running smoothly throughout the preparation of the
book and lent a hand at several crucial points. The family members and friends of the authors helped with the special stresses of textbook writing. Katherine Pfister
deserves special thanks for her steadfast support and good advice while Neil Stillings was editing the manuscript.
Page xvii
Note to the Teacher
We wrote Cognitive Science: An Introduction because we believe that undergraduate cognitive science programs ought to be offered to students and that an
introductory course ought to be part of the cognitive science curriculum. At the undergraduate level, we feel that cognitive science is best conceived of as a broad
interdisciplinary field that draws primarily on psychology, artificial intelligence, linguistics, philosophy, and neuroscience. The disciplines are to some extent distinct in
their methods, theories, and results, yet they are strikingly unified by the convergence of their core questions and by the emergence in each of them of a computational,
or information processing, view. In this text we try to maintain a consistent computational viewpoint, while honoring the distinctive contributions of each of the
disciplines.
In our view the claim that cognitive science is distinguished by a computational or informationprocessing approach should not be taken too narrowly. The term
computational should not be taken to mean that artificial intelligence ought to be the central or dominant discipline in cognitive science rather than an equal partner.
The term information processing should not be restricted to a particular kind of process modeling that has been popular in cognitive psychology. As explained in
chapters I and 2, we use these terms to refer to any research that concerns cognitive phenomena, such as perception, thought, or language, and that includes abstract
levels of analysis that are designed to explain the functional significance of the inner workings of intelligent systems. In this book linguistic competence theories,
production systems, parallel distributed processing models, the analysis of functional neural pathways in perceptual systems, and a wide range of other research is all
considered computational.
Chapters 2 through 8 have explicit disciplinary origins. The reasons for maintaining some degree of disciplinary identity are both theoretical and practical. On the
theoretical side, there is not yet a unified view of cognitive science that is widely accepted and that erases disciplinary boundaries. We feel that students deserve early
exposure to the actual diversity and theoretical ferment in cognitive science. We have tried to write a book that is not a polemic for a premature and unstable
unification of the field. On the practical side, the institutional structure of cognitive science continues to reflect its origins in the contributing disciplines. Most cognitive
science programs are interdepartmental. Most faculty members hold degrees in one of the contributing disciplines. Many students gravitate toward cognitive science
after an initial interest in one of the disciplines and go on to graduate study in disciplinary departments. For the time being, then, most teaching staffs and students are
comfortable with an introductory course that contains some clear disciplinary signposts. Cognitive science groups should find it easy to build their locally developed
emphases and syntheses into a course by emphasizing certain chapters in the text over others, stressing selected issues in lectures, and possibly including some
supplementary reading.
Page xviii
The twelve chapters of the book reflect two useful themes for course planning. First, emphasizing the disciplines that contribute to cognitive science, the book consists
of an introductory chapter, followed by seven chapters with disciplinary orientations, followed by four chapters on interdisciplinary research topics. Second,
emphasizing the core topics of cognitive science, chapters 1 through 5 present a unified introduction to representation and thought, chapters 6, 9, 10, and 11 treat
language, chapter 12 treats vision, and chapters 7 and 8 treat the neural and philosophical foundations of the field. A course that emphasized topics might cluster the
chapters on language and also shift chapters 7, 8, or 12 out of order. The following remarks about the chapters can be used as an initial guide to course planning. The
varying lengths of the chapters should be taken into account in deciding how much time to devote to them.
Instructors who are familiar with the first edition of the book should pay particular attention to the changes that have been made for the second edition. The new
edition is about twenty percent longer than the previous one. Although the organizational format of the first edition has been retained, significant revisions, noted in the
remarks below, have been made in most chapters.
1. ''What Is Cognitive Science?'' The concepts of representation and formal system are briefly introduced. The notion of levels of analysis is covered more
thoroughly than in the first edition. The role of computers in cognitive science is discussed. The interdisciplinary nature of cognitive science is outlined (although
the example analyzing the word the has been dropped from this edition). The material introduced in this chapter is expanded and reviewed throughout the
book.
2. "Cognitive Psychology: The Architecture of the Mind" This chapter was substantially revised for the second edition. The classical view of the cognitive
architecture, organized around the concept of physical symbol systems, is explicitly compared to the connectionist, or parallel distributed processing, view, of
the cognitive architecture. The foundational material in chapter 1 is deepened considerably in sections 2.3 and 2.10, which are new to this edition. In the
context of the classical view, propositional networks, schemas, working memory, imagery, and skill acquisition are introduced in some depth. The presentation
is similar to the first edition. Section 2.10 introduces both the mechanics of connectionist networks and the controversy over the relation between
connectionism and the classical view. Chapters I and 2 constitute an introduction to the foundations of cognitive science and should be covered in sequence at
the beginning of the course.
3. "Cognitive Psychology: Further Explorations" In this chapter the classical and connectionist notions of cognitive architecture developed in chapter 2 are
applied to the study of concepts, memory, reasoning, and problem solving. The treatment in the first edition has been substantially revised to reflect
connectionist research and other new developments. The chapter can be taught out of order, and the sections within it are largely independent of each other.
Unfortunately, the section on cognitive development had to be dropped from this edition to hold down the length of the book.
4. "Artificial Intelligence: Knowledge Representation" The chapter begins with several case studies of AI programs, which give beginning students a feel for AI.
Several approaches to knowledge representation are then introduced and
Page xix
compared, building on chapter 2. Semantic networks, frames and scripts, production rules, and formal logic are covered. The presentation is largely the
same as that in the first edition.
5. "Artificial Intelligence: Search, Control, and Learning" This chapter continues the exploration of AI, beginning with an introduction to search and control.
Search algorithms are covered in some detail in order to introduce beginning students to algorithmic thinking. Symbolic approaches to learning are treated in
some detail, establishing a useful contrast with the material in section 2.10. The first five chapters can be taught as a unified introduction to central cognitive
representations and processes.
6. "Linguistics: The Representation of Language" Linguistic competence theories are introduced as theories of linguistic representation. The section on
phonology has been substantially revised to reflect recent developments. The material on syntax is largely unchanged, although the presentation has been
strengthened in various ways. The final section on linguistic universals has been expanded and deepened. One or more of chapters 9, 10, and II can be read
immediately following this chapter.
7. "Neuroscience: Brain and Cognition" This chapter has been entirely rewritten. The introductory material on neuroscience has been expanded, reflecting our
view that it is increasingly important for cognitive science students to be familiar with basic concepts in neuroscience. We decided to put more emphasis on
fundamental knowledge than on the very latest developments, although the treatment of neural representation and computation and of neuropsychology has
been deepened considerably as well. The chapter should be assigned after chapters 1 and 2 and before chapter 12.
8. "Philosophy: Foundations of Cognitive Science" The chapter begins with a review and extension of the foundational concepts introduced in chapters 1 and 2
and goes on to cover philosophical issues concerning functionalism, propositional attitudes, qualia, and knowledge representation. The presentation is largely
the same as that in the first edition, although several passages on philosophical issues raised by connectionism have been added. The chapter should be
assigned after chapters I and 2.
9. "Language Acquisition" A description of the stages of language acquisition is followed by an introduction to the theoretical perspectives that have arisen out
of linguistic innateness theories and recent formal work on learnability and parsing. The chapter has been revised and expanded to include more recent
research.
10. "Semantics" This chapter is a relatively informal introduction to formal semantics as it is studied by linguists and philosophers. It begins with a treatment of
quantifiers, names, tense, and scope. Possible worlds and their application to the analysis of necessity and propositional attitudes are then introduced. The
chapter closes with a revised section on the role of formal semantics in psychology and AI. Although the chapter is selfcontained, many of the basic ideas in it
are introduced in chapters 1 through 4 and chapter 6.
11. "Natural Language Processing" The roles of grammar, discourse, and general knowledge in human and machine language understanding are discussed. The
chapter also includes new sections on connectionist models and on language production.
Page xx
12. "Vision" The format of this chapter is unchanged. Low, intermediate, and highlevel vision are all covered in some depth. The perspectives of AI,
psychology, and neuroscience are integrated. Although most of the material from the first edition has been retained, several recent lines of research have been
added. The final section of the chapter on the architecture of visual computation has been completely rewritten to reflect new results in neuroscience and
connectionism.
We believe that the introductory course in cognitive science should be offered to students in their first or second year of collegelevel study. The course will typically
contain a mix of beginning students and older students who have some experience with one (or possibly more) of the contributing disciplines. We have tried to make
the book accessible to committed firstyear students. We have also included material that is challenging to these students and at the same time holds the interest of
older, more sophisticated students. Lectures, discussions, and assignments should also be geared toward a range of intellectual sophistication, covering the
fundamentals as well as pushing into more difficult material.
We also believe that an introductory course is a useful component of graduate programs in cognitive science. Because it provides a quick introduction to the multiple
theories, results, and methods of cognitive science, this book should be useful in graduate courses, particularly when it is supplemented by more advanced readings.
Although the initial investment of time can be substantial, planning and teaching a cognitive science course can be an intellectually exciting experience that involves
reflecting on the foundations of one's field, learning some new areas, and working closely with colleagues. The field of cognitive science depends on the strength of its
introductory courses. We urge our colleagues around the world to make the necessary investment in undergraduate instruction.
Writing the second edition of this text has been an even more humbling experience than writing the first. Cognitive science has changed at a dazzling rate in recent
years. Please write to the first author at the address below with your comments and suggestions. You can also request a copy of the National Science Foundation
sponsored report on undergraduate cognitive science.
NEIL SILLINGS
[email protected]
SCHOOL OF COMMUNICATIONS AND COGNITIVE SCIENCE
HAMPSHIRE COLLEGE
AMHERST, MA 01002
Page 1
Chapter 1
What Is Cognitive Science?
One of the most important intellectual developments of the past few decades has been the birth of an exciting new interdisciplinary field called cognitive science.
Researchers in psychology, linguistics, computer science, philosophy, and neuroscience realized that they were asking many of the same questions about the nature of
the human mind and that they had developed complementary and potentially synergistic methods of investigation. The word cognitive refers to perceiving and
knowing. Thus, cognitive science is the science of mind. Cognitive scientists seek to understand perceiving, thinking, remembering, understanding language, learning,
and other mental phenomena. Their research is remarkably diverse. It includes, for example, observing children, programming computers to do complex problem
solving, analyzing the nature of meaning, and studying the principles of neural circuitry in the brain.
1.1 The Cognitive View
Like all intellectual disciplines, cognitive science involves the adoption of a definite perspective. Cognitive scientists view the human mind as a complex system that
receives, stores, retrieves, transforms, and transmits information. These operations on information are called computations or information processes, and the view of
the mind is called the computational or informationprocessing view.
The perspective of cognitive science, although it is necessarily partial, provides a unique and rich set of insights into human nature and the human potential, including
our potential to develop more powerful information technologies. The cognitive view arose from some quite natural puzzles that fascinate all of us at one time or
another. One historical origin of cognitive science was the Greek philosophers' interest in deductive reasoning, the process by which one assumes some information to
be true and derives further information that follows logically from the assumptions. For example, from the premises All dogs have fleas and Fido is a dog, one can
logically derive the conclusion Fido has fleas. In his theory of syllogistic reasoning, Aristotle showed that deductively valid arguments often take one of a small number
of general forms. Learning to reason deductively, then, can be viewed as learning an information process by which valid forms of argument can be recognized and
produced. By the seventeenth century the philosophers Leibniz and Hobbes were arguing that all thought was a kind of calculation with nonnumerical information.
Today, as we will see in the chapters that follow, linguists and philosophers continue to study the logical properties of language, cognitive psychologists compare
people's actual reasoning processes to the idealized systems devised by philosophers and mathematicians, and researchers in artificial intelligence write computer
programs to do logical reasoning.
Page 2
To take another example, consider Louis Armstrong's brilliant improvised solo on his 1927 recording of "Potato Head Blues," which was a landmark in his decisive
contribution to the development of jazz. We could adopt many different perspectives on this event. A physicist and biologist might collaborate to study how Armstrong
shaped his notes in terms of the production of an air stream by certain complex muscle contractions and the physical properties of the vibrating trumpet. A musicologist
might study how the AfricanAmerican community in New Orleans combined properties of African and European music during the late nineteenth and early twentieth
centuries. A music theorist might study how jazz improvisation is governed by the harmonic structure, or chord "changes," of a tune. But the music theorist's analysis
raises fascinating questions about human information processing. How was Armstrong able to learn the rules and creative possibilities of harmony without any formal
training in music theory? In his improvisation he was essentially composing music while he played. How could he possibly create compositions of great structural
elegance so quickly? Part of the answer must be that the human mind has general characteristics that make it possible to learn the rules of improvisation and to apply
them with great facility. Recent research on skill acquisition, described in chapter 2, has uncovered some of these characteristics. The results of the research could be
used to help design new methods to teach musicians to improvise.
Many other examples could be given of the naturalness of the informationprocessing view. Whatever else people are doing, they are always taking in, storing,
retrieving, transforming, transmitting, and acting on the basis of information. Our knowledge of the general characteristics of human information processing always
provides one interesting perspective on human activity. This knowledge is equally relevant as we confront the intellectual potential of computers, wondering what their
special strengths and limitations might be.
1.2 Some Fundamental Concepts
Understanding how the computational, or informationprocessing, view of the mind has led to a significant body of scientific research requires a more detailed
understanding of what it means to say that something is a computational system. To set the stage for the following chapters, we will begin to explain and illustrate some
fundamental concepts here. We will illustrate these concepts with the example of a familiar information process, doing arithmetic. Specifically, we will analyze the
multiplication of nonnegative whole numbers.
Information Processes Are Contentful and Purposeful
The first concept is that an important part of understanding an information process is understanding its significance, or the purpose that it serves. An information
process typically allows an organism or system to make systematic responses to some range of environmental conditions. The responses are typically adaptive or goal
oriented. The information process has a quality of meaningfulness about it. That is, we think of the information in the system as being about the world, as having
content, significance, or meaning. Such contentful qualities are also sometimes called semantic, or intentional, qualities. We find it natural to ascribe understanding,
beliefs, or knowledge to any system that acts adaptively in response to information that it takes in from its environment. An understanding of an information process (or
informationprocessing system)
Page 3
will thus include an account of the content of the information it deals with as well as an account of its competence to employ that information in the service of certain
goals.
Let us consider a simple, but instructive, example, wholenumber multiplication. Without philosophical pause, we will stipulate that there is a world of numbers and
relations among them that exists independently of any organism or device that carries out numerical computations. The product of two numbers is a third number that is
determined by a function, or mapping, from pairs of numbers onto single numbers. This function, call it ∙ , is defined by the following characteristics, which must hold
for all numbers x, y, and z: x ∙ 0 = 0; x ∙ 1 = x; x ∙ y = y ∙ x; x ∙ (y ∙ z) = (x ∙ y) z; and if z > 0 and x > y, then x ∙ z > y ∙ z. Also, the multiplication and addition functions
are related. If we call addition + , then for all x, y, and z, x ∙ (y + z) = x ∙ y + x ∙ z.
For a device to qualify as a multiplier it must possess representations or symbols that stand for numbers. It must be able to accept and produce these symbols, and it
must be able to transform them in a way that faithfully represents the product function as it was just defined. That is, when the device is fed two symbols, representing
two numbers, it must produce the symbol that represents the product of the two numbers. Notice that by focusing on the purpose of a process, on the competence
that it has to display, we have achieved a certain kind of purity. We have specified what a multiplier must do (its response function or inputoutput behavior) without
specifying how it might do it, beyond requiring that it possess some kind of representation of numbers.
An analysis concerned with the structure of a system's environment, with the information that the system has at its disposal, and with the goals it can satisfy through its
deployment of that information, can be called a semantic analysis, a competence theory, a knowledgelevel analysis, or an ecological theory (if there is a special
emphasis on an analysis of the environment and adaptiveness to it). We all use this kind of explanation in our everyday lives. We make assumptions about what people
know and what their goals are, and we assume that they can, for the most part, use their knowledge rationally to meet their goals. This folk psychology is quite
successful. It is a crucial part of our lives that we are able to predict and explain our own and others' behavior at the knowledge level much of the time.
Given our conscious awareness of our own beliefs and goals, and our everyday methods of discovering other people's, it might be imagined that analyzing the human
mind at the knowledge level is not a significant scientific problem. However, cognitive scientists have found that in several important areas our awareness of the
information we possess is limited and even misleading. An example of the limitations of awareness is our judgment of the distance from us of an object in our visual
field. We know that we possess depth perception. In scientific research on vision, however, it becomes important to know just how accurate depth judgments are,
whether they are affected by the familiarity of the object, whether they depend on the presence of other objects in the scene or on the visibility of the ground, and so
on. We cannot answer such questions just by consulting our visual awareness. Careful laboratory studies as well as careful studies of the visual environment are
required. By studying the environment, we might find that the presence of other objects in a scene can provide information about depth. For example, a nearer object
can partly occlude the visibility of a more distant object because it blocks the light reflected from part of that object. Occlusion is thus a possible cue to depth. We
could then study in the laboratory whether our depth judgments make use of occlusion information, and, if so, how that information interacts with other information
about depth.
Page 4
An example of how awareness can be misleading is our feeling that we have a uniformly vivid impression of the visual world that is nearly 180 degrees wide. It is rather
easy to demonstrate that this impression is an illusion and that we are not able, for example, to perceive much detail in the periphery of the visual field without moving
our eyes. Given that our awareness is a poor guide to the information that is immediately available in different parts of the visual field, careful laboratory research is
again the only route to discovering our visual competence. In line with these examples from vision, it will be shown in chapter 6 that we are largely unaware of our
competence with the sound and grammatical structures of our native languages.
If, rather than trying to understand an existing system, we are trying to construct a system to carry out a process, it is obviously also very helpful to begin with a clear
understanding of what we are trying to accomplish. It would be unfortunate, for example, if the designers of an electronic calculator did not fully understand the correct
behavior of the arithmetic functions that their machine was supposed to calculate. For this reason, a surprising amount of the research in artificial intelligence involves
trying to figure out just what knowledge is involved in certain types of intelligent behavior.
Information Processes are Representational
To characterize an analysis at the knowledge level, we have already invoked our second key concept, representation. The information that figures in a computation
must be represented in some way. Further understanding of a computation requires an understanding of how the information is represented. Returning to the example
of multiplication, we have so far said that a multiplier must represent numbers, but we have not said how. To develop more detail, let us suppose that the
representation is the familiar decimal placevalue notation. In this notation there are ten digit symbols, "0," "1," "2,'' . . . ''9," which represent the numbers 0 through 9.
Notice that we distinguish here between symbols, such as "5," which are part of the representation, and numbers, such as 5, which are being represented, by placing
quotation marks around the symbols. The distinction can be brought out more clearly by recalling the Roman notation system, in which "V" is the symbol for 5, there is
no single digit symbol for the number 2, there is a single digit symbol for the number 50, namely, "L," and so on. In decimal notation numbers larger than 9 are
represented by concatenating the digit symbols into strings, such as "65," which is a string of length 2 with "5" in the first position and "6" in the second position. The
assignment of a number to a symbolic string is determined systematically by the decimal placevalue function, in which the contribution of each digit symbol to the
assignment is determined jointly by its basic assignment and its position in the string. For example, knowing that the basic assignments of "6" and "5" are 6 and 5 and
knowing the placevalue function, one can determine that the assignment of "65" is 6 ∙ 101 + 5 ∙ 100, which equals 60 + 5, or 65. The decimal placevalue function
establishes a onetoone correspondence, or mapping, between digit strings and numbers. Note that although the Roman system also uses strings of digit symbols to
represent numbers, the mapping between strings and numbers cannot be defined as a placevalue function (consider, for example, the significance of the digit symbol
"I" in the left position of the strings "II" and "IV").
Placevalue representation illustrates some properties that can give representations generality and power. First, a potentially infinite set of symbols can be constructed
by beginning with a relatively small stock of basic symbols and using one or more rules of construction, or syntax, to build new, complex symbols. In the case of
whole num
Page 5
bers the ten digit symbols are the basic symbols, and concatenation is the sole syntactic rule. Allowing strings of digits to be of unlimited length produces a potentially
infinite stock of symbols. Representational schemes that include rules for building complex symbolic structures out of simpler ones are called combinatorial,
generative, or productive. Second, the meaning, or semantic interpretation, of a complex symbol is built up from the meanings of the syntactic parts of the symbol.
Typically, each basic symbol has a fixed meaning, and each syntactic rule is associated with a semantic rule that contributes to the interpretation of the complex
symbol. For whole numbers, the numbers 0 through 9 are the fixed interpretations of the ten basic digit symbols "0" through "9." The syntactic rule of concatenation is
associated with the placevalue rule of interpretation. For example, if we already have a digit string "xyz," and we concatenate ''5" onto it to make "5xyz,'' then the
placevalue rule says that the numerical interpretation of "5xyz" is 5 ∙ 103 (or 5,000) plus the numerical interpretation of "xyz." Representational schemes in which the
interpretation of complex symbolic structures is determined by the interpretations of their syntactic parts are said to have a compositional semantics. Third,
information processes that transform symbolic input structures into symbolic outputs can be defined in terms of syntactic structures of the inputs and outputs. Such
information processes analyze the syntactic structures of inputs and build syntactically structured outputs. Such information processes are also called algorithms.
Because the meanings of symbolic structures are a function of their syntax, information processes that operate on syntax can produce meaningful results and represent
meaningful operations in the domain that is represented.
These properties of algorithms can be illustrated by the example of multiplication. We have to define a process that operates syntactically on decimal placevalue
representation in such a way that the process succeeds in representing the multiplication function, ∙. The wellknown process for multiplying numbers with paper and
pencil, call it , will serve our purposes. It is illustrated in the following example:
65
32
130
195
2080
The decimal multiplication algorithm is a fairly complicated information process. It requires, among other things, a memory for the products of all of the pairs of digits
(e.g., "2" "3" yields "6"), an ability to process columns in concatenation order starting with the "ones" column, an ability to "carry," an ability to save partial results
until all columns are processed, and an ability to carry out paperandpencil decimal addition.
As we said in the previous section, the algorithm can be said to represent the product function if it operates in such a way that the representational mapping between
symbols and numbers is maintained. Let us explore this in more detail by looking at the example above. The representation assigns the string "65" to the number 65,
"32" to 32, and "2,080" to 2,080. The product of 65 and 32 is 2080. Therefore, when the longhand decimal multiplication algorithm is applied to "65" and "32," the
output must be "2,080." More generally, if we map any two input numbers onto digit strings, apply the algorithm, and map the resulting string back to a number, that
number must be the product of the input numbers. If the algorithm always mimics
Page 6
the operation in the world, we say that they have the same structure, or are isomorphic. Notice, again, that it is critical to distinguish between the operation on
symbols and the operation in the domain of numbers that it represents. Many different algorithms could be used to represent multiplication. For example, repeated
addition is a possible algorithm for multiplication. To multiply "65" by "32" using this approach, we would apply an algorithm for addition to 32 copies of "65."
Information Processes Can be Described Formally
The third basic concept results from further attention to the notion of algorithm that was just developed. An algorithm is defined completely in terms of processes that
operate on a representation. The processes do not operate on the domain being represented. They are not even defined in terms of the meaning of the representation,
which is carried separately by the semantic mapping from the representation to the domain. An algorithm is a formal procedure or system, because it is defined in
terms of the form of the representation rather than its meaning. It is purely a matter of manipulating patterns in the representation.
The algorithm for paperandpencil decimal multiplication, for example, is entirely a matter of manipulating strings of digit symbols according to certain rules. A person
with no knowledge that the strings stand for numbers could be taught to carry out the rules correctly. In fact, most of us probably carry out these rules without thinking
about why they lead to correct results. Why, for example, does carrying the 1 when multiplying 65 by 2 lead to a correct result? Exactly why is the 195 placed one
space to the left before adding it to the 130? Answering these questions requires referring to the placevalue mapping that determines the meaning of decimal notation.
But one need not know about the placevalue mapping to carry out the process. The process is meaningful, but it acquires its meaning indirectly.
An appreciation of the formal nature of information processes brings out two further points. First, since algorithms can be carried out without any higher knowledge
about their meanings, they can be carried out by physical systems, which can be biological or engineered. This insight is one of the intellectual foundations for computer
science and for the belief that a scientific understanding of mind can be achieved. Second, a formal analysis of an information process provides a particularly
convincing demonstration that we understand its inner workings. Because it makes no reference to the meaning of a process, an algorithmic analysis of a process
shows exactly how it is done without leaning on our understanding of its meaning.
In summary, the cognitive scientist approaches information processes by distinguishing between formal operations on symbols and the representational relations
between symbols and what they stand for. The realization that an organism or machine can produce meaningful behavior by performing formal operations on symbolic
structures that bear a representational relationship to the world is a key insight of cognitive science.
Our choice of example should not be taken to mean that cognitive science is easy. Cases like the multiplication algorithm or typical computer programs are good
illustrations, because they show that the strategy of cognitive science makes sense. A computer program that multiplies numbers or does database management is
obviously performing formal operations on information that has a welldefined representational relationship to a particular domain. Just as obviously, the success of the
programs is
Page 7
due entirely to the fact that the formal operations preserve the representational relationship. No inaccessible form of intelligence or understanding is present. Things are
not so simple with biological organisms, however. The human mind was not built by a team of engineers and programmers who can just hand over the blueprints and
programs. The information processes and representational relationships required for human intelligence are extremely complex, and they can only be discovered by
careful and creative research.
Cognitive Science Is a Basic Science
A fourth point about the informationprocessing view is that the concepts just introduced are employed by cognitive scientists in the pursuit of basic scientific
knowledge. Cognitive scientists seek to discover highly general and explanatory fundamental principles of information processing. This goal collides with the
tremendous variability that we see in human thought and behavior. We might doubt that any principles of human information processing can be found that hold across
all cultures and historical epochs. Certainly we will have to dig deeper than our simple example of the decimal multiplication algorithm, which is a relatively recent
human invention that is spottily distributed among the world's people and may slowly disappear from human cognition as electronic calculators become more common.
To take another example that is explored in detail in chapters 6 and 9, although people seem to communicate by making noises with their vocal tracts in all of the
world's societies, languages show tremendous variation. Do the formal structures and processes of all of the world's languages have anything in common? Research
suggests that important linguistic universals do indeed exist and that they play an important role in children's acquisition of their native languages. Of course, the
hypothesized universals must also explain how the great surface variety in the world's languages is possible.
1.3 Information Processes Can Be Analyzed At Several Levels
The distinction between studying the competence or knowledge of a system and studying its formal information processes can be thought of as a distinction between
levels of analysis. The formal analysis is at a lower level, providing an account of the information processes that underlie the competence that is visible at a higher,
behavioral level. The analysis of the semantic mapping from the formal representations to the domain can be thought of as a bridge between the formal and knowledge
levels. It explains why a formal system is a successful implementation of a particular competence. Although the formal analysis can be thought of as deeper than a
knowledgelevel analysis, it does not replace it. Each level of analysis contributes its own insights to the overall picture. Without the knowledgelevel analysis, including
the understanding of the representational mapping, we wouldn't have an understanding of what the algorithm accomplishes, and we wouldn't be able to capture the fact
that two different algorithms with different representations both compute the same function. Without the formal analysis we would know what a system does but not
how it does it. Reviewing the example of decimal multiplication, the abstract competence analysis (x ∙ 0 = 0, etc.) tells us what the product function is, the
representational analysis shows that decimal notation systematically represents numbers, and the formal analysis fully specifies the mechanics of the algorithm. If we
ignore a level of analysis, we miss an important part of the picture.
Page 8
The use of multiple levels of analysis to study intelligent systems is one of the hallmarks of cognitive science. Research on a topic can be focused on one or another
level of analysis at a given time, depending on the most promising avenues of study, and findings at one level can influence understanding at other levels. The goal is to
develop an ever more insightful and complete picture of cognition at all levels of analysis.
Information Processes Must Be Physically Implemented
The analysis of systems at the knowledge level and the formal level, and the use of the concept of representation to tie these levels together, distinguish cognitive
science (including computer science) from other sciences. To a surprising extent information processes can be studied at the knowledge and formal levels
independently of their physical implementations. This is somewhat remarkable because an information process cannot actually occur unless it is implemented in some
physical medium. For any actual informationprocessing system there obviously are physical levels of analysis aimed at uncovering its biology or physics. Further, there
must be some relationship between the physical and formal levels of analysis, because the ability of a physical system to carry out a formal information process
depends completely on its physical construction. Just as biologists concluded in the nineteenth century that life arises from particular organizations of matter and energy
and not from a special life force, so cognitive scientists proceed from the assumption that cognition arises from material structure and processes and not from any
mysterious extraphysical powers. These considerations raise the question of what role physical levels of analysis should play in cognitive science.
The answer to the question must proceed from an understanding of the relationship between the formal and physical levels. Just as a representational mapping ties the
knowledge and formal levels of analysis together, so the formal and physical levels are tied together by an implementational mapping. The formally defined
representations and processes must be mapped onto arrangements of some physical medium and transformations of that medium. The mapping must establish an
isomorphism between the formal and physical levels, so that the states and changes in the physical medium faithfully preserve all of the informational relationships in the
formal system. To return to our multiplication example, any physical system that implements the standard decimal multiplication algorithm must bear a precise
correspondence to the algorithm. Consider the state of having a multiplicand of "65." A human being with pencil and paper physically implements this state by
inscribing certain twodimensional shapes in a particular lefttoright order in a particular position on the paper. In an electronic calculator a multiplicand of "65" might
be implemented by the distribution of electrical charge in certain circuit components. In both cases there is a precise correspondence between physical states and
transformations in the computing system and the formal structure of the algorithm.
The Necessity of Higher Levels of Analysis
We might imagine that our ultimate goal is to understand the physical workings of an informationprocessing system and that having a physical understanding would
render the understanding at the formal and knowledge levels unnecessary. Consider an electronic calculator that can multiply as an example. There is a sense in which
its behavior can be completely understood in terms of the movement of electrical current through
Page 9
circuits made up of conductors and semiconductors. The physical analysis would make use of the circuit diagram, laws governing electrical circuits (e.g., Ohm's and
Kirchhoff's laws), and perhaps some more detailed electron physics needed to understand the behavior of the semiconductors.
The most immediate problem with sticking with the physical analysis is that, by dispensing with the knowledge level, we lose any account of the significance of the
calculator's behavior. The physical description does not capture the fact that the calculator is a multiplier. This loss of meaning has many consequences. A mistake in
the circuit design, for example, would not be reflected in the physical description, since the circuit laws give equally nice descriptions of useful and useless circuits.
Further, at the physical level we could not express the fact that two calculators of different design both multiply, because their physical descriptions could be arbitrarily
different (e.g., an electronic calculator and an oldfashioned mechanical adding machine). This kind of inadequacy is generally referred to as the failure of a level of
analysis to capture a generalization. The knowledge level is needed to capture generalizations that involve the significance or purpose of an information process or the
content of its representations. The physical level of analysis obviously captures other generalizations that cannot be expressed at the knowledge level, such as Ohm's
law.
Acknowledging the need for the knowledge level, we might still hope to dispense with the formal level of description. Given that we have a representational mapping
from the world to the formal system and an implementational mapping from the formal system to a physical medium, we could imagine bypassing the formal level and
working with a single direct map from the world to the physical medium. For example, we could map the world of numbers and the product function directly onto
charge distributions and their fluctuations in the circuits of an electronic calculator without bothering with the formal analysis of placevalue notation and the
multiplication algorithm defined over it. Once again, however, the formal level captures generalizations that cannot be expressed on either the knowledge level or the
physical level. In this case the generalizations have to do with the algorithms a system is using. Physical descriptions are full of physical detail that is irrelevant to the
algorithm in question, and, worse, they cannot express the fact that two systems of different physical construction compute the same algorithm (e.g., a person and an
electronic calculator both doing decimal multiplication). Algorithmic concerns also cannot be expressed at the knowledge level, because the knowledge level abstracts
away from the formal detail of computation in order to lump together processes that have the same semantics in spite of having different algorithms. The generalizations
that can be captured at the formal level are profound, forming a good part of the discipline of computer science. We will see in section 5.1, for good example, that it is
possible to define and compare different algorithms for searching through large sets of alternatives without any consideration of either the semantics or the physical
implementation of the alternatives.
Each level of analysis, then, has its own focus and descriptive vocabulary, allowing a conceptually clear explanation of some aspect of computational systems that is
obscured or even lost at other levels of description. The concepts at each level express different generalizations and impose different categorizations on information
processing systems.
In addition to this conceptual reason for maintaining higher levels of analysis, there are methodological reasons as well. It is obvious that the design of engineered
computational systems can be profitably guided by highlevel decisions about what sorts of
Page 10
algorithms they must compute and what parts of the world they will have to deal with. It is perhaps not so obvious, but equally important, that our attempts to
understand biological informationprocessing systems, such as the human mind/brain, at the physical level can be guided by understanding at the knowledge and formal
levels. Physically, the brain is enormously complex. We can get a much better idea of what to look at in the brain and how to look at it if we have a good idea of what
the brain is computing and what algorithms it is using to do the computations. In particular, a theory of some mental process at the formal level is a source of
hypotheses about the implementational mapping from formal representations and processes to nervous tissue. We will see in chapter 7, for example, that analysis of
auditory localization in the barn owl at the behavioral and formal levels guided a systematic and successful search for the neural circuits that compute the direction of an
environmental sound from the acoustic input to the ears.
The Importance of the Physical Level of Analysis
The fact that the knowledge and formal levels capture unique generalizations and that a particular formal system can be physically implemented in any number of ways
suggests the possibility that cognitive science could be pursued without regard to the physical level of analysis, which could be left to biologists and engineers. Although
it is true that the higher levels of analysis are the distinctive province of cognitive science and that much research in the field takes place exclusively at these levels, the
physical level cannot be ignored, largely for practical reasons.
Understanding the physical implementation of information processes is obviously a crucial piece of an overall understanding of cognition, since all cognition occurs in
some physical medium. Researching mappings between the higher levels of analysis and the physical level is greatly facilitated by an understanding of both levels. It is
therefore not a good idea to enforce a complete division of scientific labor between people who work at the knowledge and formal levels and people who work at the
physical level.
In general, results at one level of analysis can always potentially affect research at other levels, because the levels are mutually constraining. We have already pointed
out that a formal theory of an information process constrains its physical implementation, because the physical implementation has to be a faithful instantiation of the
formal process. The constraints run in the other direction as well. Discoveries about the physical structure of the brain can strongly suggest that it is computing certain
kinds of algorithms and not others or that it represents knowledge in certain ways and not others. We will see in chapter 2, for example, that in recent years relatively
simple facts about the brain, such as the speed with which neural cells can generate signals and the way in which the cells are interconnected, have been used by some
researchers to draw farreaching conclusions about the representations and algorithms that characterize cognition in biological organisms. In the realm of technology
the availability of a particular kind of hardware can spur research into new representations and algorithms that are particularly suited to that hardware. For example, in
recent years it has become feasible to build computers that have thousands of interconnected, active processors rather than just one. This possibility triggered a wave
of research on parallel algorithms that could take advantage of many simultaneously active processors.
Ideally, research on cognition should occur at all levels of analysis, and all fruitful interactions among levels should be encouraged. Building bridges between the higher
Page 11
levels of analysis and the physical level has been, and continues to be, the most difficult problem in the cognitive and neural sciences. In recent years, however, it has
also been one of the most active areas of new research.
1.4 Computers In Cognitive Science
It should be clear from the discussion so far that a number of concepts and distinctions that are important in computer science are part of the underlying assumptions of
cognitive science. More generally, the growth of computer science has greatly accelerated the development of cognitive science. Computers thus have a tendency to
become a metaphor for all information processing. The metaphor is often useful, but it is potentially misleading as well.
The rough outline of the analogy between computers and the human mind should already be apparent. Our stress on the independence of an information process from
its physical implementation is akin to the common distinction between software and hardware in the computer world. For example, when we see the same word
processing program running on two different brands of computers, it is obvious that an information process can be rigorously defined as a formal process without
reference to any particular physical device. Current computer software also makes it clear that formal information processes can be powerful and flexible. Highly
complex information, ranging from documents to personnel records to satellite photographs, can be represented and stored in computers. Programming languages
allow finite sets of simple instructions to be flexibly combined to define complex operations on the stored information. Today's computers and computer software are
concrete examples of information processors and processes that make it easier to understand what cognitive science is about.
On the other hand, computers are very far from exhibiting some of the most impressive, yet mundane, aspects of human cognition, such as learning a language,
identifying the objects in a naturally occurring visual scene, or solving simple problems via analogies with other situations. Therefore, it is misleading to take the
computer metaphor to mean that current computer programs, programming languages, and hardware are good models for human cognition. The usefulness for
cognitive science of any current piece of computer science or technology is a matter of research, not prior stipulation. Research is now under way on new
"generations" of computer software and hardware that many workers in the field think will be needed to advance computer intelligence.
This research includes the part of computer science called artificial intelligence, or AI, which overlaps considerably with cognitive science. Many researchers in AI
try to model their computer programs after human intelligence, and they derive inspiration from insights into human information processing that come from other
disciplines in cognitive science, such as psychology and linguistics. The insights also flow in the other direction. Attempts by AI researchers to program systems that
can understand language, see, or solve problems have led to new, testable hypotheses about human cognition. In its interplay with the other cognitive science
disciplines AI provides a powerful alternative methodology for exploring and testing theories of cognition that supplements the empirical methods of psychology and
linguistics. Chapters 4 and 5 of this book are about AI.
The technique of expressing a cognitive theory as a computer program and then running the program to explore the ramifications of the theory is now an important
Page 12
tool throughout cognitive science. Many theories are complicated enough that it is impossible to figure out what their predictions are without simulating them on
computers. There are differences of emphasis between AI and the computer simulation of theories of human cognition. Much AI research is oriented toward practical
results. A program that is too slow, for example, might be of no use. On the other hand, a useful AI program might not use the same algorithms as a human being
solving the same problem. Researchers oriented toward simulation have complementary goals. Their measure of success is whether their programs work the same way
as human (or animal) cognitive processes. It doesn't matter to them if the computer simulation is impractically slow, or slower than the biological process it simulates,
as long as it makes the right predictions. In spite of these differences there is no absolute line between AI and simulation research. Any AI program might turn out to
be a good theory of human cognition, and any simulation program might prove to have practical applications.
1.5 Applied Cognitive Science
Cognitive scientific theories are leading increasingly to practical applications. Many of the applications are in the domain of education and learning. As cognitive
psychologists have come to understand the information processes involved in reading, for example, they have begun to develop new ways of diagnosing and treating
children's reading difficulties. Linguistic theory has led to much more precise knowledge of the speech impairments that follow strokes in the left hemisphere of the
brain. This new knowledge promises to be useful in designing appropriate speech therapy and computerdriven linguistic prostheses for stroke victims.
Other applications of cognitive science are more surprising. Contemporary theories of human memory have been applied to the question of the reliability of legal
witnesses. This research has already led to important changes in the role of eyewitness testimony in the legal process. Another example is the application of new
psychological theories of skill acquisition and visualmotor imagination to the design of training programs for athletes.
Artificial intelligence research is leading to applications in expert systems and robotics. Expert systems have been developed to aid in configuring computer systems,
exploring for oil, establishing the structure of complex organic compounds, and diagnosing diseases.
These developing applications are exciting, particularly to those within the field or thinking of going into it. The excitement should be tempered, however, by an
awareness of the moral and political questions raised by applied cognitive science. Just as physicists must confront the implications of nuclear technology, and
biologists the implications of genetic engineering, so cognitive scientists must confront the implications of knowledge technology. The potentially controversial
applications of cognitive science research range from the possible development of a new generation of intelligence tests, which might be misused, to the largescale
introduction of intelligent robots in manufacturing industries, which might cause a massive loss or displacement of jobs. As in other sciences, the less and the more
controversial applications often flow from the same underlying theoretical research. For example, results in computer vision might be used to design either a visual
prosthesis for the blind or the control system of a cruise missile carrying a nuclear warhead. We hope that this book will provide the basic understanding of cognitive
scientific theory that is needed to think about the
Page 13
policy issues posed by new informationprocessing technologies. We also strongly recommend the study of relevant aspects of history, social science, and the
humanities, because the perspective of cognitive science, although crucial, must be supplemented by other perspectives.
1.6 The Interdisciplinary Nature of Cognitive Science
The Five Disciplines
As noted earlier, cognitive science is an interdisciplinary field that has arisen from the convergence on a common set of questions by psychology, linguistics, computer
science, philosophy, and neuroscience. The five contributing disciplines will undoubtedly retain their separate identities, because each of them involves a much larger
set of concerns than the focus on a basic science of cognition. A more interesting question is whether cognitive science will become a distinct academic discipline in its
own right, within which the contributions of the five converging disciplines become so thoroughly intermingled and transformed that they are no longer identifiable. This
book presents cognitive science in its current form. The distinctive contributions of each of the five disciplines are highlighted, particularly in chapters 2 through 8. The
topics selected for discussion, however, are those where the disciplines have shared the most common ground. Chapters 9 through 12 present several examples of
research areas in which the interdisciplinary collaboration has been particularly close and has begun to obscure some of the differences among disciplines.
Obviously, no one of the five contributing disciplines encompasses the entire subject matter of cognitive science, and each discipline brings to the field a focus on
particular areas. Neuroscientists are primarily concerned with the organization of the nervous system. Linguists are concerned with the structure of human language and
the nature of language acquisition. Philosophers are concerned with logic and meaning, and with clarifying the fundamental concepts of cognitive science, such as
information and knowledge. Psychologists are concerned with general human mental capacities, such as attention and memory. Computer scientists are concerned with
the possibilities for AI. Cognitive science encompasses all of these concerns. Cognitive scientists, although they usually specialize in one or two of the contributing
disciplines, benefit greatly from the crossfertilization of all of them.
The most important differences among the five disciplines are in the research methods that they use to address the nature of mind. Psychologists emphasize controlled
laboratory experiments and detailed, systematic observations of naturally occurring behaviors. Linguists test hypotheses about grammatical structure by analyzing
speakers' intuitions about grammatical and ungrammatical sentences or by observing children's errors in speech. Researchers in AI test their theories by writing
programs that exhibit intelligent behavior and observing where they break down. Philosophers probe the conceptual coherence of cognitive scientific theories and
formulate general constraints that good theories must satisfy. Neuroscientists study the physiological basis of information processing in the brain.
The Study and Practice of Cognitive Science
From the student's point of view one of the main attractions of cognitive science is the diversity of its methods. The field accommodates a wide variety of personal
intellectual