Classification capacity of spiking neural networks

We investigate a minimalistic abstraction for a multi-layer spiking neural network (SNN), in order to gain fundamental insight into classifying spatiotemporal patterns using temporal coding. The firing model is based on coincidence detection with temporal binning: a neuron X fires at time t if at least η > 1 upstream neurons fire prior to t such that the spikes arrive at X in an interval [t−δ/2, t+δ/2] (we set η = 2 for our results here). The input layer contains K neurons, each of which emits a single spike in one of T time bins, so that there are T possible spatiotemporal patterns. There are M output neurons, connected to the input layer through a layered reservoir of N neurons. Different input patterns lead to different firing patterns because the synapses connecting neurons are of different, random, lengths. The parameters K and T are measures of the spatial and temporal resolution, respectively, of the patterns to be classified.