Priming an artificial neural classifier

Abstract

Repetition priming capacity enables biological systems to manage easily with recently met situations. Priming an artificial neural network is of great interest in some modeling tasks. The network is an incremental neural classifier. This system creates units when it is not able to recognize a pattern correctly. Repetition priming is introduce through a priming function, by reinforcing the recognition of recently seen categories. Characteristics of this function are discused in order to find the more suitable shape. Experiments are performed on handwritten recognition application. Methods described enable to detect easily priming with low computation (computation of a simple linear regression). More computation enables to measure the phenomenon (difference between the slope of the regression lines, with and without priming).