An Event-based Categorization Model Using Spatio-temporal Features in a Spiking Neural Network

Abstract

This paper introduces an event-based categorization model to explore neuromorphic computing for object recognition. The proposed method makes full use of the precise timing information inherently present in the output of a bio-inspired vision sensor and utilizes event-driven processing to keep the form of address event representation (AER). In this model, a hierarchy of event-based time-surface is used to extract the spatio-temporal features of the AER data, and then these extracted features are classified by spiking neural network (SNN) with event-driven Tempotron rule. To demonstrate the effectiveness of the presented method, we conducted a series of experiments on some accessible AER datasets and compared our method to other developed event-based categorization models. Additional pooling operation in feature extraction can reduce the number of parameters and computing time. Experimental results show that our method has better classification performance and real-time computation. That is, the proposed model can extract useful features and consumes much less simulation time while still maintaining competitive accuracy.