Optimized Potential Initialization for Low-Latency Spiking Neural Networks

Spiking Neural Networks (SNNs) have been attached great importance due to the distinctive properties of low power consumption, biological plausibility, and adversarial robustness. The most effective way train deep SNNs is through ANN-to-SNN conversion, which yielded best performance in network structure large-scale datasets. However, there a trade-off between accuracy latency. In order achieve high precision as original ANNs, long simulation time needed match firing rate spiking neuron with activation value an analog neuron, impedes practical application SNN. this paper, we aim high-performance converted extremely latency (fewer than 32 time-steps). We start by theoretically analyzing conversion show that scaling thresholds does play similar role weight normalization. Instead introducing constraints facilitate at cost model capacity, applied more direct optimizing initial membrane potential reduce loss each layer. Besides, demonstrate optimal initialization potentials can implement expected error-free conversion. evaluate our algorithm on CIFAR-10 dataset CIFAR-100 state-of-the-art accuracy, using fewer time-steps. For example, reach top-1 93.38% 16 Moreover, method be other ANN-SNN methodologies remarkably promote when time-steps small.