Novelty detection in a Kohonen-like network with a long-term depression learning rule

摘要

In the cerebellar cortex, long-term depression (LTD) of synapses between parallel fibers (PF) and Purkinje neurons can spread to neighboring ones, independently of their activation by PF input. This spread of non-specific LTD around the activated synapses resembles how units are affected in the neighborhood of the winner in a Kohonen Network (KN). However in a classic KN the weight vectors become more similar to the input vector with learning, while in the LTD case they should become more dissimilar. We devised a new LTD-KN where units, opposite to the classic KN, decrease their response (LTD-like) when a pattern is learned and we show that this LTD-KN functions as a novelty detector.