Spike time displacement-based error backpropagation in convolutional spiking neural networks

In this paper, we introduce a supervised learning algorithm, which avoids backward recursive gradient computation, for training deep convolutional spiking neural networks (SNNs) with single-spike-based temporal coding. The algorithm employs linear approximation to compute the derivative of spike latency respect membrane potential, and it uses neurons piecewise postsynaptic potential reduce computational cost complexity processing. To evaluate performance proposed in architectures, employ SNNs image classification task. For two popular benchmarks MNIST Fashion-MNIST datasets, network reaches accuracies of, respectively, 99.2 $$92.8\%$$ . trade-off between memory storage capacity accuracy is analyzed by applying sets weights: real-valued weights that are updated pass their signs, binary weights, employed feedforward process. We CSNN on datasets obtain acceptable negligible drop (about 0.6 $$0.8\%$$ drops, respectively).