Learning machine with a hierarchial structure without mutual connections

Abstract

A learning machine has plural multiple-input single-output signal processing circuits connected in a hierarchical structure. The learning machine sets a threshold value, which is a evaluation standard for change in weight coefficients, high during the early part of the learning process and enables rough learning to progress without changing the weight coefficients for those multiple-input single-output signal processing circuits for which errors are sufficiently small. On the other hand, the learning machine gradually reduces the threshold value as learning progresses and advances learning by a non-linear optimization method (including a conjugate gradient method, a linear search method, or a combination of conjugate gradient and linear search methods) during the later part of the learning process, and thereby improves the learning speed.