FriSem

Andrew Saxe, PI, Theory of Learning Lab, Professorial Research Fellow at the Gatsby Computational Neuroscience Unit & Sainsbury Wellcome Centre

Title: Generalization-optimized episodic and semantic memory interactions

Abstract: The relationship between episodic and semantic memory presents an enduring puzzle for theories of the brain and mind: why do some memories stay permanently dependent on the hippocampus, while other memories seem to transfer partially to neocortex? Why do memories have different characteristics, and why even have multiple memory systems in the first place? We propose that hippocampal-neocortical interactions are orchestrated to optimize generalization performance in novel settings. Using mathematical analysis of a simple formal setting, we show that a dual-system memory achieves better generalization than the best possible single-system memory, providing a normative framework for hippocampal-neocortical replay. Within this framework, the optimal amount of replay depends on the predictability of experience: predictable information can be replayed extensively, while unpredictable information must only be replayed briefly to avoid overfitting. We therefore show that predictable information exhibits retrograde amnesia, while unpredictable information exhibits flat amnesia curves. Turning to semantic memory, we extend a mathematical account of human semantic development to incorporate control. Our findings offer new analytical tools linking nonlinear task demands to behavioral improvements and neural representations over learning in deep network models of controlled semantic cognition. Overall these results contribute to a possible computational basis for the distinction between semantic and episodic memory.