Self-Supervised Spiking Neural Networks applied to Digit Classification

The self-supervised learning (SSL) paradigm is a rapidly growing research area in recent years with promising results, especially in the field of image processing. In order for these models to converge towards the creation of discriminative representations, a data augmentation is applied to the input data that feeds two-branch networks. On the other hand, Spiking Neural Networks (SNNs) are attracting a growing community due to their ability to process temporal information, their low-energy consumption and their high biological plausibility. Thanks to the use of Poisson process stochasticity to encode the same data into different temporal representations, and the success of using surrogate gradient on learning, we propose a self-supervised learning method applied to an SNN network, and we make a preliminary study on the generated representations. We have shown its feasibility by training our architecture on a dataset of images of digits (MNIST), then we have evaluated the representations with two classification methods.