Incremental learning of sequence patterns with a modular network model

Abstract

The relationship between generalization and differentiation fluctuates depending on the ongoing context, which is extracted by the current adaptive capability of the learner. In the present report, we numerically examine the relationship between generalization and differentiation using a novel connectionist model. The simulation results of incremental learning indicate that the newly added sequence modifies the previously learned internal representations in a different manner, depending on the inconsistency with the preceding task. This observation supports our assertion that it is fundamentally important to investigate how the transition dynamics of learning toward a goal affects the finally acquired structure of the learner.