Neuromorphic Temporal Pattern Detection with Spiking Neural Networks using Synaptic Delays

Hugo Bulzomi, Yuta Nakano, Rémy Bendahan, Jean Martinet

Abstract: Whereas conventional artificial neural networks rely on weights and biases to parametrize connections between neurons, spiking neural networks aim to emulate biological systems more closely by modeling synaptic delays as well. Delays, and training method to adjust them, have been studied for their potential to increase spiking neurons’ expressiveness in tasks involving processing of temporal information. Although there exists methods for training synaptic delays in order to recognize temporal patterns, those are still not widely used in the literature due to the difficulty of efficiently applying these methods to concrete tasks using real-life data. In this paper, we present a new method for identifying temporal patterns of interest from an event flow. Once these patterns are identified, specific detectors can be defined in the form of polychronous groups of spiking neurons able to extract these movements from event data. We present experimental results on both synthetic and real event data from a neuromorphic sensor, and show how our method could be used to create efficient classification systems.