Energy-Based Probing for Spiking Neural Networks

Spiking neural networks (SNNs) have been studied not only for their biological plausibility but also for computational efficiency that stems from information processing with binary spikes. In this project, we will explore the adversarial robustness of SNNs, and study various aspects of supervised learning algorithms for SNNs. The first topic we would like to study is the adversarial robustness of SNNs. Due to the event-driven scheme, it is challenging to train SNNs compared to the training procedure of traditional deep neural networks. Thus, we will first investigate how to calculate the gradients for SNNs, then investigate the adversarial attacks for SNNs. Based on these, we will investigate the techniques of defending adversarial examples for spiking neural networks.

Another topic we want to explore is developing efficient and effective supervised algorithms for SNNs. Due to their intricately discontinuous and implicit nonlinear mechanisms, efficient learning algorithms for spiking neural networks are difficult. It has become an essential problem in this field.