适用于SCNN的多维度注意力方法

Multi-dimensional attention method for SCNN

脉冲卷积神经网络(Spiking Convolutional Neural Network, SCNN)具有强大的局部特征提取能力,但维度分布复杂,对脉冲事件易作出错误判断,影响网络的识别精度与收敛速度。受卷积神经网络的多维度注意力方法(Convolutional Block Attention Module, CBAM)启发,采用双路压缩-提取技术来获取各维度的注意力,提出一种适用于SCNN的多维度注意力方法,提升了网络对脉冲事件的感知能力,并优化了网络整体性能。实验结果表明,相比于基准的空域反向传播(Spatio-Temporal Backpropagation, STBP)算法,提出方法的识别精度提高了4.31%。

Spiking Convolutional Neural Network(SCNN)has become a research hotspot in neuromorphic vision tasks because of its biological rationality and strong local feature extraction ability.However,due to the complex dimensional distribution of SCNN,neurons are easy to make wrong judgment on the importance of spikes,which will affect the recognition accuracy and convergence speed of the network.To solve the above problems,inspired by the multi-dimensional attention method CBAM of Convolutional Neural Network,this paper proposes a multi-dimensional attention method called CTSA(Channel-Temporal-Spatio Attention),which is suitable for the unique spatio-temporal dynamics of SCNN to help the network make a comprehensive and accurate importance judgment of spikes,and improve the overall performance of the network.Extensive experiments verify the effectiveness and efficiency of CTSA method.On CIFAR10-DVS dataset,the recognition accuracy of CTSA is improved by 4.31%compared with the benchmark method.