Oscillatory Network and Deep Value Network Based Memory Replay Model of Hippocampus

Abstract

Memory replay is crucial for learning and consolidation. Hippocampal place cells demonstrate neuronal replay of behavioral sequences at a faster timescale in forward and reverse directions during resting states (awake and sleep). We propose a model of the hippocampus to demonstrate replay characteristics. The model comprises two parts - a Neural Oscillator Network to simulate replay and a Deep Value network to learn value function. The Neural Oscillator Network learns the input signal and allows modulation of the speed and direction of replay of the learned signal by modifying a single parameter. Combining reward information with the input signal and when trained with the Deep Value Network, reverse replay achieves faster learning of associations than forward replay in case of a rewarding sequence. The proposed model also explains the changes observed in the replay rate in an experimental study in which a rodent explores a linear track with changing reward conditions.