Spatially Invariant Convolutional Spiking Neural Network For Resource-Constrained IoT Devices

The image classification accuracy of convolutional spiking neural network (CSNN) decreases substantially for a distorted image dataset (images with affine transformation). To improve the classification accuracy of CSNN for distorted image dataset we need to add a spatial invariance feature to it. We propose to use a spatial transformer network (STN) as a pre-processing unit of CSNN, which improves the spatial invariance of CSNN, hence improving the classification accuracy of CSNN. This is the first attempt where we have implemented a spatial transformer network for improving the classification accuracy of convolutional spiking neural network on distorted MNIST and distorted fashion MNIST datasets using PyTorch. This combined network consisting of STN and CSNN (spatially invariant CSNN) was end-to-end trained. Our results demonstrate an improvement of 5.9% and 5.0% in the image classification accuracy of the CSNN due to the implementation of the spatial transformer network for the distorted MNIST dataset and distorted Fashion-MNIST dataset respectively.