A brief discussion on moderatism based local gradient learning rules

Abstract

Moderatism [Y. Okabe et al., 1988], which is a learning rule for ANNs, is based on the principle that individual neurons and neural nets as a whole try to sustain a "moderate" level in their input and output signals. In this way, a close mutual relationship with the outside environment is maintained. In this paper, two potential moderatism-based local, gradient learning rules are proposed. Then, a pattern learning experiment is performed to compare the learning performances of these two learning rules, the error based weight update (EBWU) rule [Tanvir Islam, M et al., December 2001][Tanvir Islam, M et al., September 2001], and error backpropagation [Bishop, CM et al., 1995].