随机赋值权重的深度网络具有物体分类能力

DNNs with Randomly Assigned Parameters Could Classify Object Images

一些文献中报道,即使深度网络的权重参数随机赋值,对应的深度网络仍有一定的分类能力.以AlexNet为模式网络,从3个侧面分析探讨了随机深度网络是否具有图像物体分类能力.首先,将AlexNet的权重随机赋值;然后,对多类不同的图像物体刺激下的神经元响应的表达不相似矩阵(RDM)与原AlexNet的RDM进行了相关性分析,发现这2类RDM具有显著相关性;鉴于深度卷积网络每层的卷积操作为加权求和,且根据中心极限定理,大量随机变量的和近似服从正态分布,进一步分别拟合了在同一输入图像下原AlexNet神经元响应的分布和随机AlexNet神经元响应的分布与高斯分布的拟合优度,并对上述2种优度进行了相关性分析.大量模拟实验表明,对于来自真实世界的样本,对应高斯拟合优度呈现显著相关性.最后,直接利用赋以随机权重的AlexNet输出的高层响应进行K近邻分类,发现其分类精度高于直接对原始彩色图像进行K近邻分类的精度.因此,与文献报道相似,实验结果再次表明随机深度网络的确具备一定的物体分类的能力.

It was reported in some existing works that even if a random set of parameters are assigned to a given deep neural network,this DNN would still exhibit a classification ability to some extent.Could DNNs with randomly assigned parameters indeed classify object images?AlexNet is used as the model network to investigate this problem from three aspects.Firstly,a random set of parameters to AlexNet is assigned,and a correlation analysis is performed between the RDM(representational dissimilarity matrix)of its neuron re-sponses to multi-class image stimuli from ImageNet and that of the original AlexNet.It is found that the corre-lation between these two RDMs is significant.Secondly,considering that the convolution operation at each DNN layer can be regarded as a process of weighting and then summing,as well as the central limit theorem“the sum of a large number of random variables is approximately Gaussian”,the fitness between the distribution of the neuron responses and the real Gauss distribution under the same inputs is further calculated,and the correlation of the fitness between the original and random AlexNet is analyzed.Extensive experimental re-sults show that only the correlation of the fitness using real samples expresses significance.Finally,the high-layer responses from the random AlexNet is used to perform KNN(K-nearest neighbor)classification,and it is found that its classification accuracy is higher than that by KNN with the original images.Hence,similar to the reports in literature,the results re-demonstrate that a random DNN indeed has a classification ability to some extent.