Encoding and Decoding of Patterns which are Correlated in Space and Time

Abstract

In the late forties, Hebb postulated his mechanism for learning: Information presented to a neural network during a learning session is stored in the synaptic efficacies in such a way that at a synapse only concurrent events are correlated and encoded. In this paper, three things are done. We study the performance of Hebbian learning in networks with a broad distribution of transmission delays and show that it can handle both stationary and dynamic objects such as single patterns and cycles by the very same principle, the delays taking care of temporal changes, if any. Second, we discuss unlearning, an unsupervised anti-Hebbian procedure that can deal with unbiased and correlated patterns alike and greatly improves the efficiency of a neural network. Finally, we indicate how to code and retrieve correlated (slow) motion. Here the key idea is to code the changes (per unit of time) and use an asymmetric learning rule.