Computational Model of Episodic Memory: Encoding, Recalling, Forgetting

This paper introduces a computational model of episodic memory that learns episodes in response to continuous sensory input. The proposed model utilizes a Hebbian based associative learning network to dynamically encode events by associating sensory inputs from different fields as activities within an event and applies a mapping technique to store episodes. The proposed mechanism makes use of ART that can continuously learn new activities with and without forgetting past ones, making them suitable for real-time incremental and lifelong learning. The ART makes event activities generalized as it continuously adds new information on match with the existing activities, and forgets past ones. In proposed model event is divided into a set of simultaneously occurred activities which represents a binding of current sensory values with its corresponding category variables. The model further incorporates both episode and event forgetting mechanism and combined with a mechanism of episode updating. It also incorporates a consolidation mechanism that prevents false information (events) from being added to the episode and ensures that no events are forgotten after a certain number of recalls over time. An empirical investigation was done in order to validate the suggested episodic memory model. The recall accuracy of the model was assessed utilizing incomplete and erroneous cues. The results of the investigation demonstrate that even with imperfect and noisy cues, the suggested model is highly connected with encoding and recalling events and episodes.