郑重声明:原文参见标题,如有侵权,请联系作者,将会撤销发布!
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO. 1, JANUARY 2023
Abstract
脉冲神经网络(SNN)代表了神经形态计算(NC)架构中最突出的生物启发计算模型。然而,由于尖峰神经元函数的不可微分性质,标准误差反向传播算法不直接适用于SNN。在这项工作中,我们提出了一个串联学习框架,该框架由SNN和通过权重共享耦合的人工神经网络(ANN)组成。ANN是一种辅助结构,有助于在尖峰训练水平上训练SNN的误差反向传播。为此,我们将尖峰计数作为SNN中的离散神经表示,并设计了一个ANN神经元激活函数,该函数可以有效地近似耦合SNN的尖峰计数。所提出的串联学习规则在传统的基于帧和基于事件的视觉数据集上展示了竞争性的模式识别和回归能力,与其他最先进的SNN实现相比,推理时间和总突触操作至少减少了一个数量级。因此,所提出的串联学习规则为在低计算资源下训练高效、低延迟和高精度的深度SNN提供了一种新的解决方案。
I. INTRODUCTION
II. LEARNING THROUGH A TANDEM NETWORK
A. Neuron Model
B. Encoding and Decoding Schemes
C. Spike Count as a Discrete Neural Representation
D. Credit Assignment in the Tandem Network
III. EXPERIMENTAL EVALUATION AND DISCUSSION
A. Experimental Setups
B. Frame-Based Object Recognition Results
C. Event-Based Object Recognition Results
D. Superior Regression Capability
E. Activation Direction Preservation and Weight-Activation Dot Product Proportionality Within the Interlaced Layers
F. Efficient Learning Through Spike-Train Level Surrogate Gradient
G. Rapid Inference With Reduced Synaptic Operations
IV. DISCUSSION AND CONCLUSION
标签:Training,Rapid,Inference,Neural,SNN,ANN,尖峰,Learning From: https://www.cnblogs.com/lucifer1997/p/17148891.html