Goal conditional episodic replay

Abstract

The human brain employs sophisticated schemes to store and reuse accumulated information. Episodic memory is one such strategy: temporally dependent event sequences are encoded and stored, and then replayed for beneficial effects for the collected information. In this research an algorithm is employed where episodes are created by a series of stacked Recurrent Neural Networks (RNNs). Information at different points of the processing is stored as the hidden states of the corresponding RNN. Another RNN is subsequently used to generate episodes given partial information of the encoded episodes, and a pre-set intended answer. The results show that the use of novel episodic replay allows the network to increase the speed of convergence in some of the 20 QA bAbI tasks.