A Brain-Inspired Self-Organizing Episodic Memory Model for a Memory Assistance Robot

Abstract

This article discusses the implementation of a brain-inspired episodic memory model, which provides memory assistance and tackles the modern public issue of memory impairment embedded as an end-to-end system on the robot companion, Pepper. Based on the fusion adaptive resonance theory, the proposed model can observe and memorize the content of daily events in five aspects: 1) people; 2) activities; 3) times; 4) places; and 5) objects. The model is based on the human memory pipeline, containing a working memory and a two-layer long-term memory model, which can effectively merge, cluster, and summarize past memories based on their context and relevance in a self-organizing manner. When providing memory assistance, the robot can analyze a user query and find the best matching memory cluster to generate verbal cues to stimulate recalling of the target event. Moreover, using reinforcement learning, the robot eventually learns the most effective mapping of cue types to event type through social interaction. Experiments show the feasibility of the proposed model, which can handle episodic events with elasticity and stability. Moreover, there is evidence that the robot is able to provide robust memory assistance from knowledge obtained through previous observations, with 99% confidence, intervals in the participants’ mean recall percentage of the events increases 19.63% after receiving memory assistance from the robot.