基于神经元转换方法的混合卷积脉冲神经网络

Hybrid convolution and spiking neural network based on neuron conversion

脉冲神经网络(Spiking Neural Network,SNN)有较强的生物可解释性,但因其训练算法尚不成熟而导致在复杂图像数据集上存在分类准确率不高、测试耗时过长等问题。针对上述问题,提出一种基于神经元转换方法的混合卷积脉冲神经网络,把卷积神经网络(Convolution Neural Network,CNN)的学习能力和特征提取能力应用于SNN。在CNN上完成训练过程后将权重映射到经过神经元转换的全连接层中,网络前期使用CNN以提取高层次图像特征信息,后期使用浅层的SNN进行分类以减少测试耗时。在3个数据集上和当前识别准确率最高的方法进行对比实验,在MNIST数据集上准确率和耗时基本持平,在cifar10数据集上准确率提升了1.32%,耗时减少了98.87%,在cifar100数据集上准确率提升了3.97%。

Spiking Neural Networks(SNN)is more biologically plausible than traditional neural networks.However,typical SNN has some defects such as low classification accuracy and high time consumption on complex image datasets because of its imperfect training algorithm.To solve these problems,a hybrid convolution and spiking neural network based on neuron conversion is proposed,this network combines the learning and feature extraction ability of Convolution Neural Network(CNN)with SNN.The training process is completed on CNN,then the connection weights are mapped to the full connection layers whose neurons are converted to spiking neurons,the network uses CNN to extract high level features of image in the first stage,and uses shallow SNN to classify in the next stage for reducing time consumption.The performance of this network is evaluated on three datasets,compared with state-of-the-art methods,the classification accuracy and time consumption are both comparable on MNIST,the classification accuracy increase 1.32%and the time consumption decrease 98.87%on cifar10,and the classification accuracy increase 3.97%on cifar100.