Deletion of Trained Patterns by Incremental Learning in Artificial Neural Network using Fahlman-Lebiere Learning Algorithm

Abstract

In this paper, we describe a method for the deletion of some of the trained patterns which are previously learned in the already-trained network in order to be adapted in a changeable environment. The deletion is performed by inhibiting all the actual outputs for the trained patterns to be deleted. Our approach is to realize the deletion by incremental learning without destroying the previously trained network but by incorporating some new hidden units so as to inhibit all the actual outputs for the trained patterns to be deleted. Fahlman-Lebiere (FL) learning algorithm is particularly suitable for this purpose since this algorithm can gradually add the required number of new hidden units. Previous studies show that the addition of categories and patterns can be performed by incremental learning [9], [10]. In this paper, we apply FL algorithm to the deletion of some of the trained patterns by incremental learning. Investigation shows that FL network realizing this deletion task has better generalization ability than Backpropagation (BP) network due to the attainment of well-saturated hidden outputs in FL network. By performing the deletion, the generalization ability of the resultant network does not degrade in comparison with the network before deletion.