James (Jay) McClelland, Professor, Stanford University

James (Jay) McClelland, Professor, Stanford University

Title: Integrating New Findings into the Complementary Learning Systems Theory of Memory


Over 15 years ago, Bruce McNaughton, Randy O'Reilly, and I introduced the Complementary Learning Systems (CLS) theory, in which we argued that hippocampus and neocortex play complementary roles in learning a memory:  The hippocampus allows for rapid initial storage of memory traces of specific items and events while the neocortex gradually extracts structure integrating over ensembles of experiences.  The theory was motivated by empirical results and also by the observation that neural networks that learn structure-sensitive distributed representations do so through a gradual interleaved learning process. 

Two challenges to the CLS have emerged in the form of empirical findings (a) that the hippocampus appears to be necessary for some forms of generalization and inference, and (b) that new information can be assimilated rapidly into neocortical networks as long as that information is consistent with previously established schemas.  In this talk I will review these empirical findings and present recent simulations indicating that the findings of type (b) are compatible with the theory as originally formulated, and findings of type (a) can be accommodated in a version of the theory that emphasizes the use of recurrent activation to allow separate hippocampal memory traces to interact.  Questions, however, about how to define 'item', 'event', and 'schema', and about what makes something 'schema consistent' remain.

 Generalization Through the Recurrent Interaction of Episodic Memories:  A Model of the Hippocampal System