Suppressive specialization: implications for generalization and interference in connectionist networks

Abstract

Three experiments were conducted to investigate the relationship between generalization and interference in connectionist networks. Specific modifications to the traditional error-back propagation learning algorithm are proposed to mitigate the interference that results from introducing new knowledge into previously trained networks. Data collected from 29 human subjects was compared to networks using the traditional algorithms and networks using the modified algorithm. The results indicate substantial improvements in connectionist network performance, in both tolerance to interference and improved generalization, can result from simple modifications to the learning algorithm that acknowledge previously acquired knowledge. Furthermore, these improvements result in connectionist network performances that more closely resemble the behaviors of human subjects when faced with similar tasks.