Neuromorphic Data Augmentation for Training Spiking Neural Networks

Developing neuromorphic intelligence on event-based datasets with SpikingNeural Networks (SNNs) has recently attracted much research attention. However,the limited size of event-based datasets makes SNNs prone to overfitting andunstable convergence. This issue remains unexplored by previous academic works.In an effort to minimize this generalization gap, we propose Neuromorphic DataAugmentation (NDA), a family of geometric augmentations specifically designedfor event-based datasets with the goal of significantly stabilizing the SNNtraining and reducing the generalization gap between training and testperformance. The proposed method is simple and compatible with existing SNNtraining pipelines. Using the proposed augmentation, for the first time, wedemonstrate the feasibility of unsupervised contrastive learning for SNNs. Weconduct comprehensive experiments on prevailing neuromorphic vision benchmarksand show that NDA yields substantial improvements over previousstate-of-the-art results. For example, the NDA-based SNN achieves accuracy gainon CIFAR10-DVS and N-Caltech 101 by 10.1% and 13.7%, respectively. Code isavailable on GitHub https://github.com/Intelligent-Computing-Lab-Yale/NDA_SNN