摘要
为了提升模型性能的同时不引入额外的计算量与能量消耗,提出了一种脉冲非对称卷积算法。利用卷积核交叉部分的权重大的特点,采用多个尺寸的卷积核替换普通卷积的单个卷积核进行卷积运算与叠加,提高中心卷积核的决策作用,在推理阶段将脉冲非对称卷积层和批量归一化层进行合并,实现简化运算。结果表明,基于脉冲非对称卷积算法的图像与事件分类模型在DVS Gesture数据集上分类精度可达98.1%,同时不引入额外的计算量和能耗。
This paper proposes a spiking asymmetric convolution algorithm to enhance model performance without introducing additional computational complexity and energy consumption.The study is accomplished by using multiple-sized convolution kernels to replace a single convolution kernel of ordinary convolution for convolution operating and superposing based on the greater weight of the cross-part of convolution kernels aiming to improve the decision-making role of the central convolution kernel;and merging the spiking asymmetric convolutional layers with batch normalization layers to simplify computations during inference.The results demonstrate that the image and event classification models based on the spiking asymmetric convolution algorithm achieve a classification accuracy of 98.1%on DVS Gesture dataset,without introducing additional computational complexity and energy consumption.
作者
桑林
Sang Lin(College of Innovation&Entrepreneurship,Heilongjiang University of Science&Technology,Harbin 150022,China)
出处
《黑龙江科技大学学报》
CAS
2024年第2期323-328,共6页
Journal of Heilongjiang University of Science And Technology
关键词
脉冲神经网络
类脑计算
残差学习
非对称卷积
spiking neural networks
neuromorphic computing
residual learning
asymmetric convolution
