An Event-based Hierarchy Model for Object Recognition

In this work, we propose a new event-based hierarchical architecture based on the spiking neural network (SNN) for object recognition. Unlike the existing hierarchical architecture like HMAX for object recognition, the proposed model makes full use of the precise time information of spatial-temporal spiking signals. In the feature extraction layer in our proposed hierarchical model, we combine the advantages of Hfirst algorithm that extracts features containing the precise time information and HOTS algorithm that considers the relative timings of events. Then, the extracted features are learned by the Tempotron learning rule, which is a biologically plausible supervised learning algorithm. In order to show the the effectiveness of the proposed method, we conducted the comprehensive experiments in the current benchmark datasets (i.e., AER Posture, Poker Card, Letter & Digits). The experimental results show the significant improvements by combining the advantages of event-based feature representation and the precise timing process by SNN over two original state-of-the-art methods. To further demonstrate the applicability and generalization of our proposed method, we extend our experiments to a new and more challenging facial recognition task and our model achieves 92.5% accuracy on face recognition. Furthermore, we also investigate the influence of related parameters on the performance in our model.