THE HIPPOCAMPUS AS A CASCADED AUTOENCODER NETWORK

Suzanna Becker, Ted Meeds and Adeline Chin

The hippocampus is a key brain structure involved in episodic memory formation. Its unique multi-layered circuitry may provide the key to understanding its hallmark abilities of rapid encoding and high capacity. Unlike many hippocampal models which propose separate learning rules in each layer, we propose a global objective function for learning: optimal input reconstruction. From this goal, approximately optimal, locally computable reconstruction terms are derived for each layer using a novel "Cascaded Auto-encoder" architecture. In an initial series of benchmark studies with our model, designed to shed light on the computational benefits of various known architectural and physiological features of the hippocampus, we compare several versions of our model to multi-layer autoencoder networks trained with back-propagation. These studies account for a wide range of physiological and anatomical data. Additionally, our model provides a powerful tool for testing computational theories of plasticity and generating experimentally testable predictions.