Properties of a trainable linear threshold logic unit.

Abstract

Many tasks can be reduced to the problem of pattern recognition and the vast majority of applications of learning machines is concerned with such problems. The examples of pattern recognition are speech recognition, handprinted characters recognition, weather forecasting, automatic control of technological processes, etc. The subject matter of this work is the detailed analysis of the basic element of the neuron-net-like learning systems - the Linear Threshold Logic Unit. As a mathematical model a many-dimensional vector space is used. This approach gives clear insight into the properties of the element and is particularly fruitful in the analysis of process of training. In the first part of the work, the general problem of the pattern recognition is presented and some properties of the basic element of the learning machine are discussed. The second part is concerned with the training procedures for the LTLU. Both, geometrical and analytical treatments of the Error Correction Procedures are discussed in details. Other learning procedures are also surveyed.