A New Approach to Modeling Cognitive Information Learning Process using Neural Networks

摘要

Recurrent Neural Networks have shown potential for cognitive modeling by showing good results with real-world temporal contextual data. Long Short Term Memory (LSTM) neural network model was designed to address the issue of large time lags in input data successfully. However, for the lower level cognitive processing, specifically, information processing, storage & recall, its performance had room for improvement mainly because LSTM cannot learn adoptively. Sustained Temporal Order Recurrent (STORE) networks are designed to encode the order of temporal data, and then could recall the encoded data in veridical as well as non-veridical order employing unsupervised learning. By the fusion of LSTM and STORE, a new neural network model LSTM-STORE is proposed that can alternate between supervised and unsupervised learning modes. The key purpose of this fusion is to mimic the working of brain for information processing during active as well as sleep time. To mediate the shift in the two learning modes we proposed Consolidation Control System CCS, which through gating mechanism controls the flow of information though the proposed network. The proposed model LSTM-STORE not only aims to investigate the working of LSTM in an unsupervised learning based environment; but also to perform lower level cognitive reasoning tasks. Thus LSTM-STORE basis it's plausibility on a comprehensive cognitive foundation. The proposed model is justified by providing comparative experimental proof.