Abstract
Spiking Neural Networks (SNNs) has recently emerged as a prominent neuralcomputing paradigm. However, the typical shallow spiking network architectureshave limited capacity for expressing complex representations, while training avery deep spiking network have not been successful so far. Diverse methods havebeen proposed to get around this issue such as converting off-line trained deepArtificial Neural Networks (ANNs) to SNNs. However, ANN-to-SNN conversionscheme fails to capture the temporal dynamics of a spiking system. On the otherhand, it is still a difficult problem to directly train deep SNNs using inputspike events due to the discontinuous and non-differentiable nature of thespike signals. To overcome this problem, we propose using differentiable (butapproximate) activation for Leaky Integrate-and-Fire (LIF) spiking neurons totrain deep convolutional SNNs with input spike events using spike-basedbackpropagation algorithm. Our experiments show the effectiveness of theproposed spike-based learning strategy on state-of-the-art deep networks (VGGand Residual architectures) by achieving the best classification accuracies inMNIST, SVHN and CIFAR-10 datasets compared to other SNNs trained withspike-based learning. Moreover, we analyze sparse event-driven computations todemonstrate the efficacy of proposed SNN training method for inferenceoperation in the spiking domain.