Selective Input Sparsity in Spiking Neural Networks for Pattern Classification

The concept of input sparsity in Spiking Neural Networks for pattern recognition is introduced and explored with the goals of reductions in network inference time and size, leading to lower resource requirements in hardware implementations. A method is proposed by which selective input sparsity can be inferred from the training set to reduce the size of the network before training and decrease the network inference time. This method also requires no additional pre-processing steps during the testing phase, making it an excellent candidate for edge applications. For a basic fully connected spiking neural network trained to solve the MNIST handwritten digits, selective input sparsity is applied and the network size is reduced by 58.16% and a 41.07% decrease in the network's inference time is observed without notable accuracy hinderance. In the case of the Fashion MNIST dataset, selective input sparsity reduced the network size by 55.99% and reduced the network's inference time by 59.05%.