Explaining the Computational Mind PDF
Preview Explaining the Computational Mind
Explaining the Computational Mind Explaining the Computational Mind Marcin Milkowski The MIT Press Cambridge, Massachusetts London, England © 2 013 M assachusetts 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. MIT Press books may be purchased at special quantity discounts for business or sales promotional use. For information, please email [email protected] or write to Special Sales Department, The MIT Press, 55 Hayward Street, Cambridge, MA 02142. This book was set in Stone Sans by Toppan Best-set Premedia Limited. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Milkowski, Marcin. Explaining the computational mind / Marcin Milkowski. pages cm Includes bibliographical references and index. ISBN 978-0-262-01886-9 (hardcover : alk. paper) 1. Cognitive neuroscience—Data processing. 2. Computational neuroscience. 3. Computational complexity. I. Title. QP360.5.M535 2013 612.8'233—dc23 2012036420 10 9 8 7 6 5 4 3 2 1 Contents Preface vii Acknowledgments xi 1 Computation in Cognitive Science: Four Case Studies and a Funeral 1 2 Computational Processes 25 3 Computational Explanation 85 4 Computation and Representation 137 5 Limits of Computational Explanation 177 Notes 203 References 209 Index 233 Preface This book is about explaining cognitive processes by appeal to computa- tion. The mind can be explained computationally because it is computa- tional; this is true whether it is engaging in mental arithmetic, parsing natural language, or processing the auditory signals we attend to in order to experience music. All these capacities arise from complex information- processing operations of the mind. My central claim reflects my adherence to realism: a computational account of the mind can constitute a genuine explanation only insofar as the mind is itself computational . This stands in stark contrast to computational models of, say, the weather, which, rather than construing clouds and rainfall as the processing of information, describe the purely physical processes underlying the meteorological phenomena. Why write yet another book about computation and cognition? So much ink has been spilled; positions have been refined and hardened; it ’ s difficult to envisage anything but a scholastic exercise in adding even more distinctions to the discussion. Well, in fact, the reason for writing this book was my growing impatience with extant accounts of computational expla- nation of cognition, which proved either too limited in scope or too sketchy. There simply wasn’ t a book that I could refer to in discussions with cognitive scientists and philosophers alike, even if parts of the puzzle were easily available; so, I had to write that book myself. The term “ computation ” in cognitive science had been misappropri- ated to refer to a class of processes that were supposed to occur only during rule-governed discrete symbol manipulation. But this way of framing com- putation is highly idiosyncratic and leads to numerous misunderstandings — especially because it ’ s hard to say what the term “ symbol ” stands for — in viii Preface an interdisciplinary debate taking place between people of varied back- grounds. It is unclear whether, and if so how, these rule-governed pro- cesses should differ from rule-following processes or, for that matter, rule-compatible processes. Symbolic computations in cognitive science are supposed to be radically different from operations of classical connection- ist artificial neural networks (ANNs). To a software engineer, however, ANNs are just another kind of machine learning. Why would they be noncomputational if we build dedicated computers composed of ANNs? Even worse, why exactly should we refrain from calling neuro computational models computational? Is this a kind of institutionalized terminological schizophrenia? Superficial distinctions obscure a common pattern present in the differ- ent branches of cognitive modeling. This, in turn, leads to debates over the alleged supremacy of high-level theories over neurocomputational modeling, or vice versa, and to dismissing whole bodies of research as inessential in cases where it would be more useful to strive for integration rather than elimination. I suggest that we will be happier if, instead of adding more distinctions, we have fewer of them. Computation— understood generally as informa- tion processing — is a basic ingredient in the majority of cognitive theories and models. Granted, the notion is broad, but the alternative ways of carving it are not viable when it comes to saying when a computation is being implemented. The symbolic, semantic, syntactic, and modeling accounts of implementation are plagued by simple yet devastating objec- tions; for instance, one can easily gerrymander the physical processes or entities to which the computations are said to correspond. Real progress in determining when a computational process is being implemented occurs once we have been able to replace the wildcard term “ symbol ” with some- thing more physically tangible, and the structure of the process is no longer open to gerrymandering. Given that we already have fairly robust theories of causation in the philosophy of science, I suggest that the best way to accomplish these aims is by relying on the causal structure of the world. Of course, there is another way. Adding a number of provisos to the traditional, semantic, syntactic, or modeling account could block gerry- mandering. But that would lead to an explosion of epicycles, and, at any rate, the provisos would boil down to a redescription, in semantic or syn- Preface ix tactic terms, of the conditions that differentiate genuine causal structures capable of information processing from mere hodgepodge mixtures of physical states. There is nothing to be gained by such redescription except more verbosity and unnecessary complexity. Although the vision I offer in this book is much more ecumenical than the received view, this does not mean that I cannot distinguish between the different approaches in cognitive science (or anywhere else the computational idiom is used). Those differences are not really essential, however, when it comes to general methodological principles. If you want to confirm that your computational model is empirically sound, you are not compelled to rely o nly on the specific methodology of Bayesian, bioro- botic, cybernetic, connectionist, dynamic, neurocomputational, or sym- bolic modeling — whatever the case may be. The basic principles of explanation and confirmation are the same for all the methodologies, and this makes comparisons possible even if the mechanisms posited by differ- ent methodologies are dissimilar in important ways. A word about mechanisms is in order. I started to think of physical computation about a dozen years ago. In my doctoral dissertation on Daniel Dennett’ s philosophy of mind (unfortunately for the English- speaking readers, I wrote it in Polish), I discussed his claim that evolution is algorithmic. The notions I used to describe algorithms and computation coalesced around causation, structures, and processes rather than merely around function terms. Only later did I find that the terminology of orga- nized entities that I had used was in fact equivalent to the vocabulary of the neomechanistic philosophy of science. Having discovered, like Mr. Jourdain, that I had been talking prose, I decided to get rid of my own terminology and settled for mechanistic talk. This is beneficial because the mechanistic account of explanation fits the practice of cognitive science much better than do the competing accounts. As I will show, it fares better than classical functionalism in particular. Computationalism is here to stay — but it ’ s not what most people had taken it to be. In particular, it does not rely on a Cartesian gulf between either software and hardware or mind and brain. The computational method of describing the ways information is processed is usually abstract — but cognition is possible only when computation is realized physically, and the physical realization is not the same thing as its description. The mechanistic construal of computation allows me to show that no purely
The list of books you might like
