THEORETICAL AND EXPERIMENTAL INVESTIGATIONS IN TRAINABLE PATTERN-CLASSIFYING SYSTEMS.
Final rept. Jun 64-65,
STANFORD RESEARCH INST MENLO PARK CA
Pagination or Media Count:
Motivation is given for the use of various trainable pattern-classifying structures called linear and piecewise linear PWL machines. The results of various experiments in training these machines are presented. Two different types of training methods were investigated mode-seeking and error-correction methods. These methods are illustrated by experiments using two-dimensional patterns so that the results of training can be easily visualized. More thorough experiments with 10-dimensional and 100-dimensional patterns are also described. It is concluded that certain of these training methods can be expected to give good performance in complex pattern classifying tasks involving multimodal pattern distributions. The report concludes with a list of recommendations for future research, and contains an Appendix presenting a new theorem on training a linear machine. Author