Learning the states: A brain inspired neural model

Abstract

AGI relies on Markov Decision Processes, which assume deterministic states. However, such states must be learned. We propose that states are deterministic spatio-temporal chunks of observations and notice that learning of such episodic memory is attributed to the entorhinal hippocampal complex in the brain. EHC receives information from the neocortex and encodes learned episodes into neocortical memory traces thus it changes its input without changing its emerged representations. Motivated by recent results in exact matrix completion we argue that step-wise decomposition of observations into 'typical' (deterministic) and 'atypical' (stochastic) constituents is EHC's trick of learning episodic memory. © 2011 Springer-Verlag Berlin Heidelberg.