High Accurate, Low Latency Conversion of Spiking Neural Networks with BLIF Neurons

Spiking Neural Networks (SNNs), acknowledged as the third generation of neural network technology, stand out for their efficiency and closer resemblance to biological brain functions compared to traditional Artificial Neural Networks (ANNs). However, training SNNs can be challenging because it's difficult to differentiate between spike trains. Converting ANNs to SNNs has emerged as a promising solution to this issue. Yet, ANN-SNN conversion has its limitations: it cannot represent negative spikes as captured by Dynamic Vision Sensors (DVS) cameras and suffers from severe accuracy loss with short simulation time-steps. In this paper, by developed a bi-threshold dynamic mechanism, we proposed the Bi-threshold Linear Leaky Integrate-and-Fire (BLIF) neu-ron, relieving constraints on the negative spike expression of SNNs. Secondly, we analyzed the residual error between BLIF neuron and Leaky ReL U activation function, and proposed a Stepwise Leaky ReLU activation function more suitable for BLIF neuron under membrane potential encoding, which ensures zero residual error in the ANN-SNN conversion, even with very short simulation time-steps. Last, we developed a zonal convergence testing method to eliminate the distribution error in the ANN-SNN conversion. These three approaches help us convert the ANNs to SNNs with no distribution error and residual error, which have high accuracy, low latency and good robustness, and have achieved the state-of-the-art performance on various datasets such as CIFAR10, CIFAR100 and ImageNet.