Hippocampal formation trains independent components via forcing input reconstruction

Abstract

It is assumed that higher order concept formation utilizes independent components (ICs). It is argued that ICs require dynamic input reconstruction networks (RNs) to form a reliable internal representation. Input reconstruction, however, can be slow and poor with ICs on substrates with lossy dynamics. A model of the hippocampal formation is proposed that develops the ICs on lossy RNs by means of locking inputs to the internal representation and thus forcing fast reconstruction and cancelling losses. It is assumed that upon training ICs can lock themselves, thus hippocampal lesion mostly affects anterograde memories.