A neural-network model of the cortico-hippocampal interplay: Contexts and generalization

Abstract This paper presents computer simulations of a neural network comprising two sensory pathways, each built of preprocessing and associative memory modules perhaps corresponding to a primary and higher sensory area, and a hippocampal area that serves as an integration of fusion zone during learning and retrieval of polymodal information. The network is able to store unimodal details about a complex environment in local assemblies restricted to the corresponding associative memory, whereas a representation of the simultaneous occurrences of several stimuli is constituted and stored in a self-organizing manner in the hippocampal area. This can be viewed as storage of a “particular context”. If many stimulus constellations are presented to the network during learning, it may over-learn, that is, the hippocampal area can no longer distinguish particular situations, but instead represents more general contexts or categories that a given environmental situation may belong to. Feedback from the hippocampal region to association areas can restore particular memories; it can still act as a threshold control gate raising sensitivity in the appropriate cortex regions when it is overloaded.