An Event-Driven Computational System with Spiking Neurons for Object Recognition

We propose a biologically plausible computational system using spiking neural networks for object recognition, which processes the data from a temporal contrast address event representation (AER) sensor. The spike-based features are obtained through event-driven Gabor function and LIF neurons. And a time-to-first spike operation (also as a temporal Winner-Take-All (WTA) operation) with lateral reset in the same pooling area is implemented for reducing memory and computational costs. An address lookup table (LUT) is also applied to adjust the feature maps via address mapping and reordering. Then, the extracted spike feature patterns are classified by tempotron neurons. Our system can not only capture temporal visual information, but also learn features entirely based on the timing spikes information. Experiments conducted on two AER datasets have proved its efficiency for object recognition.