Antidromic Spikes Drive Hebbian Learning in an Artificial Dendritic Tree

Hebbian learning using the timing between pre-synaptic and post-synaptic spiking allows a network of silicon neuromorphs to learn and playback complex spatiotemporal input patterns. Learning occurred dynamically and in a stimulus dependent manner by potentiating active synapses that contributed to post-synaptic spike production and depressing active synapses that were anti-causal. Active synapses that were neither causal nor anti-causal remained at their pre-activated efficacy. The network used to evaluate hebbian synaptic plasticity was fully connected with each neuromorph making a prescribed number of connections to the dendrites of all the other neuromorphs. To enable learning of spatiotemporal spiking activity, efferents from each neuromorph had to make connections along the entire length of their target dendrites so as to produce a temporally distributed response. Upon repetitive presentation of an input pattern those synapses that had appropriate causal timing were strengthened while those that were anti-causal were depressed.