Title: A Theory of Learning and Generalization
Subtitle: With Applications to Neural Networks and Control Systems
Author: M. Vidyasagar
Publisher: Springer-Verlag London, 1997.
Details: 382 pages, 36 figures, 150 references
This book contains a comprehensive treatment of so-called statistical
learning theory and its applications both to neural networks and control
systems. Other names for the subject are PAC learning theory and
computational learning theory. In addition to a thorough treatment of known
results, the book also presents several new results for the first time, as
well as a list of sixteen open problems.
The Table of Contents of the book is shown below:
Preface xiii
1. Introduction 3
2. Preliminaries 15
2.1 Pseudometric Spaces, Packing and Covering Numbers 15
2.2 Probability Measures 19
2.3 Large Deviation Type Inequalities 21
2.4 Some Advanced Ideas 25
3. Problem Formulations 29
3.1 Uniform Convergence of Empirical Means 29
3.2 Learning Concepts and Functions 39
3.3 Model-Free Learning 59
4. Vapnik-Chervonenkis and Pollard (Pseudo-) Dimensions 69
4.1 Definitions 69
4.2 Bounds on Growth Functions 75
4.3 Growth Functions of Iterated Families 87
5. Uniform Convergence of Empirical Means 95
5.1 Restatement of the Problems Under Study 95
5.2 Equivalence of the UCEM and ASCEM Properties 99
5.3 Main Theorems 101
5.4 Preliminary Lemmas 108
5.5 Theorem 5.1: Proof of Necessity 119
5.6 Theorem 5.1: Proof of Sufficiency 124
5.7 Proofs of the Remaining Theorems 136
5.8 Uniform Convergence Properties of Iterated Families 140
6. Learning Under a Fixed Probability Measure 153
6.1 UCEM Property Implies ASEC Learnability 155
6.2 Shrinking Width is Equivalent to Consistent Learnability 161
6.3 Finite Metric Entropy Implies Learnability 167
6.4 Examples 176
6.5 Learnable Concept Classes Have Finite Metric Entropy 179
6.6 Model-Free Learning 185
7. Distribution-Free Learning 195
7.1 Uniform Convergence of Empirical Means 195
7.2 Function Learning 202
7.3 Concept Learning 207
7.4 Learnability of Functions with a Finite Range 219
8. Learning Under an Intermediate Family of Probabilities 223
8.1 General Families of Probabilities 225
8.2 Totally Bounded Families of Probabilities 235
8.3 Families of Probabilities with a Nonempty Interior 246
9. Alternate Models of Learning 249
9.1 Efficient Learning 250
9.2 Active Learning 264
9.3 Learning with Prior Information 273
10. Applications to Neural Networks 291
10.1 What is a Neural Network? 292
10.2 Learning in Neural Networks 295
10.2.1 Problem Formulation 295
10.2.2 Reprise of Sample Complexity Estimates 298
10.2.3 Complexity-Theoretic Limits to Learnability 303
10.3 Estimates of VC-Dimensions of Families of Networks 307
10.3.1 Multi-Layer Perceptron Networks 308
10.3.2 A Network with Infinite VC-Dimension 314
10.3.3 Neural Networks as Verifiers of Formulas 316
10.3.4 Neural Networks with Piecewise-Polynomial Activation Functions
323
10.3.5 A General Approach 328
10.3.6 Networks with Pfaffian Activation Functions 332
10.3.7 Results Based on Order-Minimality 335
10.4 Structural Risk Minimization 337
11. Applications to Control Systems 343
11.1 Identification 344
11.2 Robust Control 352
12. Some Open Problems 361