Spiking neural networks (SNNs) are a new generation of deep learning models inspired by biology, which belong to a subset of deep learning and have a strong biological basis to support them. It has received more and more attention from researchers in recent years due to its advantages of high efficiency, energy saving, and high interpretability. However, compared with traditional ANNs, SNNs are still in the early stage of research and still face many problems.

In this thesis, we first analyze the reasons of the poor performance of SNNs in image classification and propose a new interpretable spiking neuron to improve the learning ability of the network for big datasets. In addition, we propose a new method of adversarial defense to enhance the robustness of SNNs against tiny noise. Besides, we also propose a new training algorithm for optimizing the speed of the SNN in the training and inference process.

In addition to the study of SNN algorithms, we also apply SNN to specific application problems. To address the problem of redundancy in event camera datasets, we propose a SNN-based mask network that selectively deletes redundant pulse signals, thus reducing the space occupied by the dataset and facilitating transmission. Finally, we combine SNNs with engineering problems, and since SNNs are good at handling timing signals, we use SNNs to achieve high-precision classification of tactile signals collected by tactile sensors.

In summary, SNNs are positioned as ANNs with biological plausibility, i.e., they have the interpretability of biological networks and some characteristics of ANNs. In this thesis, we work on developing different algorithms for the speed, accuracy, and robustness of SNNs and design SNNs for problems under a variety of different domains.