随机赋值权重的深度网络具有物体分类能力

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。

Several studies have reported that deep networks still possess certain classification capabilities even when their weight parameters are randomly initialized. Taking AlexNet as the baseline network, we analyzed and discussed whether randomly initialized deep networks have image object classification capabilities from three perspectives. First, we randomly initialized the weights of AlexNet. Next, we performed correlation analysis between the Representational Dissimilarity Matrices (RDMs) of neuronal responses evoked by multiple categories of image object stimuli and those of the original AlexNet, and found that these two types of RDMs exhibited significant correlation. Given that the convolution operation in each layer of a deep convolutional network is a weighted summation, and according to the Central Limit Theorem, the sum of a large number of random variables approximately follows a normal distribution, we further fitted the goodness of fit between the neuronal response distributions of the original AlexNet and those of the randomly initialized AlexNet under the same input image to the Gaussian distribution, respectively, and conducted correlation analysis on these two types of goodness of fit. A large number of simulation experiments showed that for samples from the real world, the corresponding Gaussian goodness of fit values exhibited significant correlation. Finally, we directly used the high-level responses output by the randomly initialized AlexNet for K-nearest neighbor (KNN) classification, and found that its classification accuracy was higher than that of directly performing K-nearest neighbor classification on the original color images. Therefore, consistent with previous literature reports, the experimental results once again demonstrate that randomly initialized deep networks do possess certain object classification capabilities.