自校准首脉冲时间编码神经元模型  

Self-calibrating First Spike Temporal Encoding Neuron Model

作　　者：冯忍 陈云华[1] 熊志民 陈平华[1] FENG Ren;CHEN Yunhua;XIONG Zhimin;CHEN Pinghua(School of Computer Science,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广州510006

出　　处：《计算机科学》2024年第3期244-250,共7页Computer Science

基　　金：广东省自然科学基金(2021A1515012233)。

摘　　要：由于脉冲神经元具有复杂的时空动力过程且脉冲信息不可导,脉冲神经网络(SNN)的训练一直是一个难题。基于人工神经网络(ANN)转SNN间接训练深度SNN的方法,避免了直接训练深度SNN的难题,但该方法所获得的SNN的性能在很大程度上会受到脉冲信息编码机制的影响。在众多编码机制中,首脉冲时间编码(TTFS)具有良好的生物学基础和更高的能效,但现有TTFS编码采用单脉冲形式,信息表征能力较弱,编码所需时间窗较大。为此,在TTFS的单脉冲编码基础上,增加一个校准脉冲,形成一种自校准首脉冲时间(SC-TTFS)编码机制,并构建相应的SC-TTFS神经元模型。在SC-TTFS中,首脉冲为必定发放的脉冲,而校准脉冲根据首脉冲发放后剩余的膜电位来确定是否发放,用于对编码脉冲所引起的转换量化误差和截断误差进行补偿,同时缩小编码所需的时间窗。通过对多种编码对应的转换误差进行对比分析,以及在多种网络结构上进行ANN-SNN转换实验,验证了所提方法的优越性。采用CIFAR10和CIFAR100数据集,基于VGG和ResNet两种网络结构进行了实验验证。结果表明,所提方法在两类网络结构和两种数据集上均实现了精度无损的ANN-SNN转换,且相较于最先进的同类方法,所提方法所构建的SNN具有最短的网络推理延迟。另外,在VGG结构上,所提方法相比TTFS编码能源效率提升了约80%。Because of the complex spatio-temporal dynamic process of spike neurons and the non-differentiable spike information,the training of spike neural network(SNN)has always been very difficult.The ANN-to-SNN method for indirect training of deep SNN avoids the difficulties of direct training of deep SNN.However,the performance of the SNN obtained in this approach is greatly affected by the spike information encoding mechanism.Among many coding mechanisms,TTFS has a good biological basis and is energy efficient,but existing TTFS codes use a single-spike formalism,which has weak information representation capability and large time windows for encoding.Therefore,based on the single spike coding of TTFS,a calibration spike is added to form a self-calibrating first spike time to first spike coding mechanism,and the corresponding SC-TTFS neuron model is constructed.In SC-TTFS,the first spike is the spike that must be emitted,while the calibration spike determines whether it is emitted according to the residual membrane potential after the first spike is emitted,which is used to compensate the quantification error and truncation error caused by the coding spike and to reduce the time window required for coding.The advantages of this approach are verified by comparing and analyzing the corresponding conversion errors of various codes and ANN-SNN conversion experiments on various network architectures.On CIFAR10 and CIFAR100 datasets,the proposed algorithm is verified by experiments based on VGG and ResNet network structures,and it achieves ANN-SNN transformation with non-destructive accuracy on both network structures and two data sets.Compared to state-of-the-art similar methods,the SNN constructed by the proposed method has the smallest network inference latency.In addition,on the VGG structure,the proposed method improves the energy efficiency by about 80%compared with TTFS coding.

关 键 词：脉冲神经网络 脉冲编码机制 ANN-SNN转化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]