Communication and Cybernetics 17

Transcription

1 Communication and Cybernetics 17 Editors: K. S. Fu W. D. Keidel W. J. M. Levelt H. Wolter

2 Communication and Cybernetics Editors: K.S.Fu, W.D.Keidel, W.1.M.Levelt, H.Wolter Vol. Vol. 2 Vol. 3 Vol. 4 Vol. 5 Vol. 6 Vol. 7 Vol. 8 Vol. 9 Vol. 10 Vol. 11 Vol. 12 Vol. 13 Vol. 14 Vol. 15 Vol. 16 Vol. 17 W.Meyer-Eppler Grundlagen und Anwendungen der Informationstheorie B.Malmberg Structural Linguistics and Human Communication 1. L. Flanagan Speech Analysis/Synthesis and Perception G.Herdan The Advanced Theory of Language as Choice and Chance G. Hammarstrom Linguistische Einheiten im Rahmen der modernen Sprachwissenschaft 1. Peters Einfiihrung in die allgemeine Informationstheorie K.Weltner The Measurement of Verbal Information in Psychology and Education Facts and Models in Hearing Edited by E.Zwicker, E. Terhardt G. Hammarstrom Linguistic Units and Items Digital Pattern Recognition Edited by K. S. Fu Structure and Process in Speech Perception Edited by A. Cohen, S.G.Nooteboom 1. D. Markel, A. H.Gray,lr. Linear Prediction of Speech R. G. Busnel, A. Classe Whistled Languages Applications of Syntactic Pattern Recognition, Applications Edited by K.S.Fu P.Kummel Formalization of Natural Languages K. Steinbuch Kommunikationstechnik T.Kohonen Associative Memory A System-Theoretical Approach

3 Teuvo Kohonen Associative Memory A System-Theoretical Approach Corrected Printing of the First Edition With 54 Figures Springer-Verlag Berlin Heidelberg New York 1978

4 Professor Dr. TEUVO KOHONEN Helsinki University of Technology, Dept. of Technical Physics SF OtaniemijFinland Professor KING SUN Fu, PhD Purdue University, School of Electrical Engineering West Lafayette, IN 47907, USA Professor Dr. WOLF DIETER KEIDEL 1. Physiologisches Institut der UniversiUit Erlangen-Niirnberg D-8520 Erlangen, Fed. Rep. of Germany Professor Dr. WILLEM J. M. LEVELT Katholieke Universiteit, Faculteit der Sociale Wetenschappen Nijmegen, Holland Professor Dr. HANS WOLTER Institut fur Angewandte Physik der Universitat D-3550 Marburg/Lahn, Fed. Rep. of Germany ISBN-13: DOl: / e-isbn-13: Library of Congress Cataloging in Publication Data. Kohonen, Teuvo. Associative memory. (Communication and cybernetics; v. 17). Bibliography: p. Includes index.!. Memory. 2. Association storage. 1. Title. Q360.K ' This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically those of translation, reprinting, re-use of illustrations, broadcasting, reproduction by photocopying machine or similar means, and storage in data banks. Under 54 of the German Copyright Law where copies are made for other than private use, a fee is payable to the publisher, the amount of the fee to be determined by agreement with the publisher. by Springer-Verlag Berlin Heidelberg 1977 Softcover reprint of the hardcover I st edition 1977 The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Offset printing, and book binding: Briihlsche Universitatsdruckerei Giessen

5 Preface About the Scope of This Text This book contains two types of material ~ first, the many divergent and often diffuse meanings given to the concepts of association, associative memory, and associative recazz are expounded. A review of this kind was felt necessary because there apparently does not exist any single monograph which could serve as a reference to these topics. But the presentation of the main body of this text is motivated by quite other reasons: in recent years, plenty of interesting mathematical and system-theoretical material has been published which makes it possible to gain a view of associative memory which is different from the conventional abstract and computationally oriented approaches. It seems that the basic operation of associative memory, the storage of information together with the relations or links between the data items, and the selective recall of stored information relative to a piece of key or cue information presented, is not restricted to certain computer-technological implementations but can also be reflected in more general mathematically describable processes in certain physical or other systems, especially in their adaptive state changes. It further seems that some generally known forms of associative memory, namely, certain computer technological artifacts, or abstract systems of concepts or data, are in fact special representations of a class of processes characterized as associative memory. The text in presentation is thus a new system-theoretically motivated approach to the phenomenon of associative memory, and it is hoped that the formalisms and some applications presented here might interest scientists working in the fields of cybernetics, computer science, mathematical psychology, physiology, and physics. Various Aspects of Memory. The subject of memory has been approached by many branches of science. The rather incoherent views thereby gained may be consequent of emphasis laid on different aspects. For instance, in a similar way as the concept of computing involves the machine, the program code, and the information process that is realized when a program is run, so the concept of memory in a cybernetic system may mean 1) the material facilities, for instance, a computer memory, a physical system, or a set of organic elements which serve as a substrate for the representation of information, 2) the entirety of stored representations per se (in life sciences, sometimes named "memories"), 3) the abstract structure of knowledge that is implicit in

6 VI the stored representations and their semantic relations, or 4) the recollections which, especially in the mental sense, may be sequences of events that are reconstructed in a process of reminiscence. AcknowZedgements Many core ideas presented here have evolved in numerous discussions with my many collaborators. Especially during the writing of this book, I have got much valuable help from Messrs. Pekka Lehtio and Erkki Oja. Many members of the staff of our laboratory have contributed to the computer simulations and other material contained in this book. I am indebted to the following persons who have been of great help in many ways: K.Bry, L. Hippelainen, J. Hyvarinen, L. Hall, T. Junkkari, J. Karhunen, H. Laine, H. Pohjanpalo, J. Rekula, E. Riihimaki, J. Rovamo, P. Teittinen, J. Tuominen, and L. Vainio. Some remarks made on the contents by Academician Erkki Laurila, Professor James A. Anderson and Professor Michael A. Arbib, have been very useful. I am also grateful to my wife for her support. This work has been made under the auspices of the Academy of Finland. Otaniemi, Finland January, 1977 T. Kohonen

7 Contents Chapter 1. Introduction 1.1 On the Physical Embodiment of Associative Information Structures A System-Model Approach to the Associative Memory Relational Structures of Knowledge Acquisition and Retrieval of Knowledge as Structured Sequences Implementations of Associative RecalL Basic Characteristics of Associative Memories The Content-Addressable Memory (CAM) Hash-Codi ng Holographic Associative Memories Nonholographtc, Distributed Associative Memories Mathematical Notations and Methods Vector Space Concepts Matrix Notations Further Properties of Matrices Matrix Equations Projection Operators Matrix Differential Caleulus Chapter 2. Associative Search Methods 2.1 Addressing by the Contents Hash-Coding Principles An Example of Hash-Coding: Multiple-Keyword Search A Processing Language for Associative Data Structures Content-Addressable Memories Associative Recall by the Partial Match Operation Hardware Implementation of the CAM Structure Parallel Comparison of Magnitudes... 67

8 VIII 2.3 Optimal Associative Mappings System Model for an Analog Associative Memory Autoassociative Recall as an Orthogonal Projection The Novelty Filter Autoassociative Encoding Optimal Linear Associative Mappings Optimal Nonlinear Associative, Mappings The Problem of Invariant Recognition Relationship Between Associative Mapping, Linear Regression, and Linear Estimation Relationship of Associative Mappings to Pattern Classification Discriminant Functions Comparison Methods Statistical Formulation of Pattern Classification Chapter> 3. Adaptive Formation of Optimal Associative Mappings 3.1 On the Implementation of Conditioned Responses in Simple Physical Systems A Simple Adaptive Linear System..., On the Physical Realizability of Adaptive Elements Adaptive Filters Which Compute Orthogonal Projections The Novelty Detector Paradigm Analysis of the Adaptive Linear Unit by Means of Matrix Products An Extremely Fast Adaptive Process Which Generates the Novelty Filter Adaptation With Forgetting Recursive Generation of the Optimal Associative Mapping Linear Corrective Algorithms The General Setting for the Computation of M = YX Recursive Evaluation of the Best Exact,Solution (Gradient Projection Method) Recursive Evaluation of the Best Approximate Solution (Regression Solution) Recursive Solution in the General Case

9 Chapter 4. On Biological Associative Memory 4.1 Physiological Foundations of Memory On the Mechanisms of Memory in Biological Systems Structural Features of Some Neural Networks Functional Features of Neurons Modelling of the Synaptic Plasticity Can the Memory Capacity Ensue from Synaptic Changes? Computerized Models of Associative Memory in Neural Networks The Associative Network Paradigm The Novelty Filter Paradigm Review of Related Approaches References Author Index Subj ect Index IX