可全域表达的高精度低延时脉冲网络转换方法  

A high-precision,low-latency conversion method with globally representing spiking neural network

作　　者：马钟 徐克欣 李申 王钟犀 MA Zhong;XU Kexin;LI Shen;WANG Zhongxi(Xi'an Microelectronics Technology Institute,Xi'an 710054,China;93170 Unit of People's Liberation Army of China,Xi'an 710000,China;School of Microelectronics,Xi'an Jiaotong University,Xi'an 710049,China)

机构地区：[1]西安微电子技术研究所,西安710054 [2]中国人民解放军93170部队,西安710000 [3]西安交通大学微电子学院,西安710049

出　　处：《集成电路与嵌入式系统》2025年第3期15-23,共9页Integrated Circuits and Embedded Systems

摘　　要：不同于人工神经网络(ANN),脉冲神经网络(SNN)作为第三代神经网络技术的代表,基于生物神经元机制进行计算,使用脉冲信号序列来传递信息,展现出可观的能耗优势和海量数据的高速处理能力。然而,由于脉冲神经元具有复杂的动力学行为和脉冲计算不可微分的特性,现有的SNN直接训练方法效果欠佳,一定程度阻碍了SNN的广泛应用。目前,将高精度ANN转换为SNN被认为是最有前途的生成SNN的方法之一。然而,主流的ANN转换SNN方法存在局限性:首先,不支持负值脉冲,难以表达由动态视觉传感器相机采集的负向脉冲;其次,转换过程中低延时和高精度难以两全。针对以上问题,本文提出了一种可全域表达的新型脉冲神经元,对传统ANN中正负数值和DVS的正负极性均能进行全域表示,并且提出了一种阶梯式Leaky ReLU激活函数和一种区域收敛测试算法,以实现ANN至SNN的零误差转换。通过以上方法,实现可全域表达的高精度、低延迟和高鲁棒的ANN至SNN转换,本文方法在CIFAR10和CIFAR100数据集上表现出卓越性能。Unlike Artificial Neural Networks(ANN),Spiking Neural Networks(SNN),as a representative of the third generation of neural network technologies,perform computations based on biological neuron mechanisms,using sequences of spike signals to transmit information.This exhibits significant energy efficiency advantages and high-speed processing capabilities for massive data.However,due to the complex dynamics of spiking neurons and the non-differentiability of spike computations,the current direct training methods for SNNs are not very effective,hindering their widespread application.At present,converting high-precision ANN to SNN is considered one of the most promising methods for generating SNN.However,mainstream ANN-to-SNN conversion methods have their limitations:first,they do not support negative spikes,making it difficult to represent negative spikes from dynamic vision sensor cameras;second,low latency and high precision cannot be achieved simultaneously during the conversion process.To address these issues,this paper proposes a novel spiking neuron that can represent the entire range of values,supporting both positive and negative values in traditional ANN as well as the positive and negative polarities of DVS(Dynamic Vision Sensor)spikes.Additionally,this paper proposes a step-wise Leaky ReLU activation function and a regional convergence testing algorithm to achieve zero-error conversion from ANN to SNN.With these methods,we realize a high-precision,low-latency,and robust ANN-to-SNN conversion.Our method demonstrates outstanding performance on the CIFAR10 and CIFAR100 datasets.

关 键 词：ANN转换SNN 阶梯式Leaky ReLU激活函数 区域收敛测试算法 全域表达 鲁棒性测试 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]