Models and mechanisms for integrating cortical feature spaces

Abstract

The medial temporal lobe memory system is the seat of explicit memory in mammals, including both recognition and episodic memory. Recognition is the sense that an object, a scene, or some other context has been experienced previously. Episodic memory is the automatic and long-term storage of life's experiences, each separate and distinct. These very different forms of memory are served by different subsystems in the medial temporal lobe and hippocampal formation. The difference in computational function may result from differences in the structure of the inputs that drive their respective networks. We examine models and mechanisms participating in both of these memory subsystems. Particularly, we consider how the high-order feature spaces that they integrate participate in the underlying computation. The primate perirhinal cortex computes a familiarity signal for high-order objective features, an input space with structured correlations. We present a minimal model of this familiarity signal and show that word-frequency information is encoded in the space of semantic similarity. The hippocampus in rodents is critical to their ability to navigate through the world and learn their spatial surroundings, but it also provides a model for episodic memory. We present a mechanism that provides rapid sparsification of the spatial inputs to hippocampus and produces informative spatial representations. This mechanism may enable fast contextual learning in unfamiliar environments and, simultaneously, provide the distinct basis of activity necessary for encoding episodic memory.