Disambiguation of Pattern Sequences with Recurrent Networks

Recently, there has been great interest in using neural networks to learn sequences of patterns. This is obviously very important from a cognitive point of view. In this paper, we show how a simple network with a hippocampal-like structure can be used to learn complex stimulus sequences. Specifically, we train the system, which has a one-step recurrent dynamics, to disambiguate sequences with temporal overlaps of more than one step. Usually, this is done either through delay lines or by means of capacitive effects. We show that a significant population of unforced, recurrently activated neurons in the system can enable the system to disambiguate quite well over sev eral time steps.