Unsupervised Learning of Spatio-temporal Patterns Using Spike Timing Dependent Plasticity

This paper presents an unsupervised approach for learning of patterns with spatial and temporal information from a very small number of training samples. The method employs a spiking network with axonal conductance delays that learns the encoding of individual patterns as sets of polychronous neural groups, which emerge as a result of training. A similarity metric between sets, based on a modified version of the Jaccard index, is used for pattern classification. Two different neural connectivity models are evaluated on a data set consisting of hand-drawn digits that encode temporal information (i.e., from the starting to the end point of the digit). The results demonstrate that the approach can successfully generalize these patterns from a significantly small number of training samples.