Interference and discrimination in trained networks by the backpropagation algorithm

Summary form only given. A number of recent simulation studies have shown that when a connectionist net is trained, using backpropagation, to memorize sets of items in sequence and without negative exemplars, newly learned information seriously interferes with old. Three converging methods have been employed to show why and under what circumstances such retroactive interference arises. A geometrical analysis technique has shown that the elimination of interference always results in a breakdown of old-new discrimination. A formally guaranteed solution to the problems of interference and discrimination, presented as the HARM model. has been used to assess the relative merits of other proposed solutions. Two simulation studies have assessed the effects of providing nets with experience of the experimental task. The results indicate that interference and discrimination problems are closely related