UC Santa Barbara

Spiking neural networks (SNNs) are known as the third generation of neural networks. For an SNN, the bio-inspired neural dynamics endow the great potential to simulate the neural behaviors of the brain; the additional temporal information propagation provides a larger space to make a comprehensive decision; the binary format and the sparse activities of spikes make SNNs quite energy efficient when considering the real deployment. High accuracy, high efficiency, and high robustness are several attractive features of the brain.

In the early stage, the bio-plausible unsupervised training methods are the mainstream but restrict the learning accuracy of SNNs. Recently, the emerging supervised training algorithms inspired by backpropagation through time (BPTT) have successfully boosted the accuracy. However, the implementation complexity of these BPTT-based algorithms is explosively growing, which raises a much higher demand for hardware resources. To improve the training efficiency, this dissertation proposes two solutions to optimize the BPTT-based training. The first solution is to directly design an ASIC accelerator for SNNs while the other is to optimize the dataflows on GPU.

On the other side, how to improve the robustness of SNNs is critical for building a reliable neuromorphic system. This dissertation first discusses how to disturb an SNN model through adversarial examples, and then conducts an in-depth analysis of the SNN robustness. With the observations, a robust training method for SNNs is inspired by the robustness certification in neural networks.