Catastrophic Interference in Learning Processes by Neural Networks

Starting from the Eighties, there has been a fluorishing of models of memorization and learning processes, based on neural networks, which characterized by a multilayer feedforward architecture and the supervised back-propagation learning rule are plagued by the so-called catastrophic interference problem.

Abstract

Starting from the Eighties, there has been a fluorishing of models of memorization and learning processes, based on neural networks. As shown by McCloskey e Cohen (1989), and Ratcliff (1990), the ones characterized by a multilayer feedforward architecture and the supervised back-propagation learning rule are plagued by the so-called catastrophic interference problem. This latter arises when, after a network learned a certain number of items belonging to a suitable learning set, the same network is submitted to a second learning process with a new learning set.