The fifth edition of the neural network introductory text An Introduction to Neural Networks Ben Kr\"ose and Patrick van der Smagt Dept. of Computer Systems University of Amsterdam is now available by anonymous ftp. This text is in use at our department for an introductory neural network course, given by the authors. This version differs from the previous (1991) one in several aspects: - many corrected errors & prettified figures - the chapter on Unsupervised Learning is rewritten & expanded - the chapter on Robot Control is adapted - the chapter on Vision is expanded - the chapter on simulators has been removed - the complete list of references (which are also available per chapter) has been removed The book consists of 131 pages. Comments on its context, additions, corrections, and flames are very much appreciated at smagt@fwi.uva.nl. For those people who want to use this manuscript for their courses, or in any other way want to distribute it or multiply it, please get in touch with me. Patrick van der Smagt Department of Computer Systems, University of Amsterdam Kruislaan 403, 1098 SJ Amsterdam, NETHERLANDS Phone +31 20 525-7524, Fax +31 20 525-7490 Email: , ----------------------------------------------------------------------------- TABLE OF CONTENTS Preface 9 I FUNDAMENTALS 11 1 Introduction 13 2 Fundamentals 15 2.1 A framework for distributed representation 15 2.1.1 Processing units 16 2.1.2 Connections between units 16 2.1.3 Activation and output rules 17 2.2 Network topologies 17 2.3 Training of artificial neural networks 18 2.3.1 Paradigms of learning 18 2.3.2 Modifying patterns of connectivity 18 2.4 Notation and terminology 19 2.4.1 Notation 19 2.4.2 Terminology 20 II THEORY 23 3 Adaline and Perceptron 25 3.1 The adaptive linear element (Adaline) 25 3.2 The Perceptron 26 3.3 Exclusive-or problem 27 3.4 Multi-layer perceptrons can do everything 28 3.5 Perceptron learning rule and convergence theorem 30 3.6 The delta rule 31 4 Back-Propagation 33 4.1 Multi-layer feed-forward networks 33 4.2 The generalised delta rule 34 4.3 Working with back-propagation 36 4.4 Other activation functions 37 4.5 Deficiencies of back-propagation 38 4.6 Advanced algorithms 39 4.7 Applications 42 5 Self-Organising Networks 45 5.1 Competitive learning 46 5.1.1 Clustering 46 5.1.2 Vector quantisation 49 5.1.3 Using vector quantisation 49 5.2 Kohonen network 52 5.3 Principal component networks 55 5.3.1 Introduction 55 5.3.2 Normalised Hebbian rule 56 5.3.3 Principal component extractor 56 5.3.4 More eigenvectors 57 5.4 Adaptive resonance theory 58 5.4.1 Background Adaptive resonance theory 58 5.4.2 ART1 The simplified neural network mo del 58 5.4.3 ART1 The original model 61 5.5 Reinforcement learning 63 5.5.1 Associative search 63 5.5.2 Adaptive critic 64 5.5.3 Example The cartpole system 65 6 Recurrent Networks 69 6.1 The Hopfield network 70 6.1.1 Description 70 6.1.2 Hopfield network as associative memory 71 6.1.3 Neurons with graded response 72 6.2 Boltzmann machines 73 III APPLICATIONS 77 7 Robot Control 79 7.1 End-effector positioning 80 7.1.1 Camera-robot coordination is function approximation 81 7.2 Robot arm dynamics 86 7.3 Mobile robots 88 7.3.1 Model based navigation 88 7.3.2 Sensor based control 90 8 Vision 93 8.1 Introduction 93 8.2 Feed-forward types of networks 94 8.3 Self-organising networks for image compression 94 8.3.1 Back-propagation 95 8.3.2 Linear networks 95 8.3.3 Principal components as features 96 8.4 The cognitron and neocognitron 97 8.4.1 Description of the cells 97 8.4.2 Structure of the cognitron 98 8.4.3 Simulation results 99 8.5 Relaxation types of networks 99 8.5.1 Depth from stereo 99 8.5.2 Image restoration and image segmentation 101 8.5.3 Silicon retina 101 IV IMPLEMENTATIONS 105 9 General Purpose Hardware 109 9.1 The Connection Machine 110 9.1.1 Architecture 110 9.1.2 Applicability to neural networks 111 9.2 Systolic arrays 112 10 Dedicated Neuro-Hardware 115 10.1 General issues 115 10.1.1 Connectivity constraints 115 10.1.2 Analogue vs. digital 116 10.1.3 Optics 116 10.1.4 Learning vs. non-learning 117 10.2 Implementation examples 118 10.2.1 Carver Mead's silicon retina 118 10.2.2 LEP's LNeuro chip 120 Author Index 123 Subject Index 126 ----------------------------------------------------------------------------- To retrieve the document by anonymous ftp : Unix> ftp galba.mbfys.kun.nl (or ftp 131.174.82.73) Name (galba.mbfys.kun.nl ) anonymous 331 Guest login ok, send ident as password. Password ftp> bin ftp> cd neuro-intro ftp> get neuro-intro.400.ps.Z 150 Opening ASCII mode data connection for neuro-intro.400.ps.Z (xxxxxx bytes). ftp> bye Unix> uncompress neuro-intro.400.ps.Z Unix> lpr -s neuro-intro.400.ps ;; optionally ----------------------------------------------------------------------------- The file neuro-intro.400.ps.Z is the manuscript for 400dpi printers. If you have a 300dpi printer, get neuro-intro.300.ps.Z instead. The 1991 version is still available as neuro-intro.1991.ps.Z. 1991 Is not the #dots per inch! We don't have such good printers here. Do preview the manuscript before you print it, since otherwise 131 pages of virginal paper are wasted. Some systems cannot handle the large postscript file (around 2M). On Unix systems it helps to give lpr the "-s" flag, such that the postscript file is not spooled but linked (see man lpr). On others, you may have no choice but extract (chunks of) pages manually and print them separately. Unix filters like pstops, psselect, and psxlate (the source code of the latter is available from various ftp sites) can be used to select pages to be printed. Alternatively, print from your previewer. Better still, don't print at all! Enjoy!