Fig. 6: Illustration of the surrogate gradient of MSF neurons.
From: A multisynaptic spiking neuron for simultaneously encoding spatiotemporal dynamics
a The forward and backward computation graph of spiking neurons as used for backpropagation through time (BPTT). b Activation functions of ReLU, LIF, and MSF. Surrogate gradients \({f}^{{\prime} }(\cdot )\) of MSF (D = 4) neurons based on rectangular (c) and sigmoid (d) functions σ with different threshold intervals h. e The impact of the threshold interval h on the classification performance on CIFAR-10. Surrogate gradients of MSF (D = 4) neurons based on rectangular (f) and sigmoid (g) functions with different curve steepness α. h The impact of the curve steepness α on the classification performance on CIFAR-10. i Classification performance comparison between different depths of ResNet with LIF and MSF neurons.
