Determination and Detection of Features in Patterns

Abstract

In this paper feature determination as a method of training the first layer of weights in a two layer learning machine (Perceptron) is investigated. The problem is viewed as one of examining a set of patterns and determining a set of simpler patterns, or features, so that each of the original patterns can be formed by superposing the features. While the general problem of finding a minimal set of features was not solved, two algorithms were given that solve the problem for restricted pattern sets.