Neural networks have shown enormous potential for commercial exploitation over the last few years but it is easy to overestimate their capabilities. A few simple algorithms will learn relationships between cause and effect or organise large volumes of data into orderly and informative patterns but they cannot solve every problem and consequently their application must be chosen carefully and appropriately. This book outlines how best to make use of neural networks. It enables newcomers to the technology to construct robust and meaningful non-linear models and classifiers and benefits the more experienced practitioner who, through over familiarity, might otherwise be inclined to jump to unwarranted conclusions. The book is an invaluable resource not only for those in industry who are interested in neural computing solutions, but also for final year undergraduates or graduate students who are working on neural computing projects. It provides advice which will help make the best use of the growing number of commercial and public domain neural network software products, freeing the specialist from dependence upon external consultants.
Author(s): Lionel Tarassenko
Year: 1998
Language: English
Pages: 160
Front Cover......Page 1
A Guide to Neural Computing Applications......Page 4
Copyright Page......Page 5
Contents......Page 6
Foreword......Page 10
1.1 Neural computing–today's perspective......Page 12
1.2 The purpose of this book......Page 13
1.4 Acknowledgements......Page 14
2.2 Why neural networks?......Page 16
2.3 Brief historical background......Page 17
2.5 Pattern classification......Page 19
2.6 The single-layer perceptron......Page 20
2.7 From the 1960s to today: multi-layer networks......Page 23
2.8 Multi-layer perceptrons and the error back-propagation algorithm......Page 25
2.9 Training a multi-layer perceptron......Page 27
2.10 Probabilistic interpretation of network outputs......Page 29
2.11 Unsupervised learning–the motivation......Page 30
2.12 Cluster analysis......Page 31
2.13 Clustering algorithms......Page 32
2.14 Data visualisation-Kohonen's feature map......Page 35
2.15 From the feature map to classification......Page 38
2.16 Radial Basis Function networks......Page 39
2.17 Training an RBF network......Page 41
2.18 Comparison between RBF networks and MLPs......Page 42
2.19 Auto-associative neural networks......Page 43
2.20 Recurrent networks......Page 44
2.21 Conclusion......Page 46
3.2 Neural computing projects are different......Page 48
3.3 The project life cycle......Page 49
3.4 Project planning......Page 50
3.5 Project monitoring and control......Page 53
3.6 Reviewing......Page 54
3.7 Configuration management......Page 55
3.8 Documentation......Page 56
3.9 The deliverable system......Page 57
4.1 Introduction......Page 60
4.2 Identifying neural computing applications......Page 61
4.3 Typical examples of neural computing applications......Page 62
4.5 Technical feasibility......Page 64
4.6 Data availability and cost of collection......Page 65
4.7 The business case......Page 66
4.8 Conclusion......Page 68
5.2 Computational requirements......Page 70
5.3 Platforms for software solutions......Page 72
5.4 Special-purpose hardware......Page 75
5.5 Deliverable system......Page 77
6.2 Glossary......Page 78
6.3 Data requirements......Page 79
6.4 Data collection and data understanding......Page 82
7.2 Overview of design......Page 88
7.3 Pre-processing......Page 90
7.4 Input/output encoding......Page 93
7.5 Selection of neural network type......Page 98
7.6 Selection of neural network architecture......Page 99
7.7 Training and testing the prototype......Page 100
7.8 From prototype to deliverable system......Page 105
7.9 Common problems in training and/or testing the prototype......Page 106
8.1 Overview of the case studies......Page 110
8.2 Benchmark results......Page 113
8.3 Application of data visualisation to the case studies......Page 115
8.4 Application of MLPs to the case studies......Page 120
8.5 Application of RBF networks to the case studies......Page 127
8.6 Conclusions......Page 130
9.2 Data visualisation......Page 132
9.3 Multi-layer perceptrons......Page 134
9.4 On-line learning......Page 136
9.5 Introduction to Netlab......Page 137
Appendix A: The error back-propagation algorithm for weight updates in an MLP......Page 140
Appendix B: Use of Bayes' theorem to compensate for different prior probabilities......Page 142
References......Page 144
Index......Page 148
