跨脉冲传播的深度脉冲神经网络训练方法

基于反向传播的脉冲神经网络（SNNs）的训练方法仍面临着诸多问题与挑战，包括脉冲发放过程不可微分、脉冲神经元具有复杂的时空动力过程等。此外，SNNs反向传播训练方法往往没有考虑误差信号在相邻脉冲间的关系，大大降低了网络模型的准确性。为此，提出一种跨脉冲误差传播的深度脉冲神经网络训练方法（cross-spike error backpropagation，CSBP），将神经元的误差反向传播分成脉冲发放时间随突触后膜电位变化关系和相邻脉冲发放时刻点间的依赖关系两种依赖关系。其中，通过前者解决了脉冲不可微分的问题，通过后者明确了脉冲间的依赖关系，使得误差信号能跨脉冲传播，提升了生物合理性。此外，并对早期脉冲残差网络架构存在的模型表示能力不足问题进行研究，通过修改脉冲残余块的结构顺序，进一步提高了网络性能。实验结果表明，所提方法比基于脉冲时间的最优训练算法有着明显的提升，相同架构下，在CIFAR10数据集上提升2.98%，在DVS-CIFAR10数据集上提升2.26%。

Deep spiking neural network training method across spiking propagation

Backpropagation-based training methods for SNNs still face many problems and challenges, including that the spike firing process is non-differentiable and spike neurons have complex spatiotemporal dynamics processes. In addition, SNNs backpropagation training methods often did not consider the relationship of the error signal between adjacent spikes, greatly reducing the accuracy of the model. To this end, this paper proposed a cross-spike error backpropagation training method for deep spiking neural networks(CSBP), which divided the error backpropagation of neurons into two dependencies: the dependency of spike firing time with the postsynaptic membrane potential(DSFT) and the dependency between spike firing time(DBSFT). Among them, DSFT solved the problem of spike non-differentiability and DBSFT clarified the dependence between spikes, allowing error signals to propagate across spikes, improving biological rationality. In addition, this paper solved the problem of insufficient expressive ability in early spiking ResNet network architecture by modifying the structural order of the spike residual block. Experimental results show that the proposed method is significantly improved compared to the SOTA(state-of-the-art) training algorithms based on spike time. Under the same architecture, the improvement is 2.98% on the CIFAR10 dataset, and 2.26% on the DVS-CIFAR10 dataset.