Neural modeling of sequential inferences and learning over episodic memory

Abstract

a b s t r a c tEpisodic memory is a significant part of cognition for reasoning and decision making. Retrieval in episodic memory depends on the order relationships of memory items which provides flexibility in reasoning and inferences regarding sequential relations for spatio-temporal domain. However, it is still unclear how they are encoded and how they differ from representations in other types of memory like semantic or procedural memory. This paper presents a neural model of sequential representation and inferences on episodic memory. It contrasts with the common views on sequential representation in neural networks that instead of maintaining transitions between events to represent sequences, they are represented as patterns of activation profiles wherein similarity matching operations support inferences and reasoning. Using an extension of multi-channel multi-layered adaptive resonance theory (ART) network, it is shown how episodic memory can be formed and learnt so that the memory performance becomes dependent on the order and the interchange of memory cues. We present experiments as a proof of concepts to show that the model contrasts sequential representations in semantic memory with those in episodic memory and the model can exhibit transitive inferences consistent with human and animals data.

Episodic Memory Model for Learning Robotic Manipulation Tasks

“…Hierarchical emotional episodic memory, using deep adaptive resonance theory network [40] has also been proposed to learn emotions correlated with past experiences. Generalized fusion adaptive resonance theory-based EM-ART has been introduced to make the episodic memory suitable to handle intricate relations of past events [41], [42].…”

Hierarchical emotional episodic memory for social human robot collaboration

“…For these reasons, researchers presented computational models of the episodic memory for the systems to memorize experience sequences (Rinkus 2004;Howard et al 2005;Mueller and Shiffrin 2006;Norman et al 2008;Jockel et al 2008;Starzyk and He 2009;Nuxoll and Laird 2012;Stachowicz and Kruijff 2012). Moreover, to make the episodic memory suitable to handle intricate relations of events, EM-ART was developed (Wang et al 2012;Subagdja and Tan 2015) which is based on a generalization of fusion adaptive resonance theory (Fusion ART model) (Tan et al 2007). The model allows temporal event sequences to be efficiently learned, recognized, and recalled from stored episodes in response to a partial event sequence cue.…”

“…The model allows temporal event sequences to be efficiently learned, recognized, and recalled from stored episodes in response to a partial event sequence cue. Recently, Deep ART was designed (Park and Kim 2016;Park et al 2017) to overcome the episode retrieval error caused by different episode lengths (Park et al 2015) and the same event repetition (Gao and Tan 2014;Subagdja and Tan 2015). In addition, the ART network uses the method of resonance in the learning structure each time a new input is received so that the network structure can be flexibly developmental and is able to learn in real time.…”

“…To encode temporal information to the ART memory, EM-ART (Wang et al 2012;Subagdja and Tan 2015) was developed. There are two concatenated ART networks in EM-ART as depicted in Fig.…”

“…There is a limitation in EM-ART which is episode retrieval error due to different episode lengths (Park et al 2015) and same event repetition (Gao and Tan 2014;Subagdja and Tan 2015). To overcome the the limitation, Deep ART was designed (Park and Kim 2016;Park et al 2017).…”