Abstract
As deep neural networks (DNNs) achieve tremendous success across many applica2 tion domains, researchers tried to explore in many aspects on why they generalize 3 well. In this paper, we provide a novel perspective on these issues using the gradient 4 signal to noise ratio (GSNR) of parameters during training process of DNNs. The 5 GSNR of a parameter is defined as the ratio between its gradient’s squared mean and 6 variance, over the data distribution. Based on several approximations, we establish 7 a quantitative relationship between model parameters’ GSNR and the generaliza8 tion gap. This relationship indicates that larger GSNR during training process leads 9 to better generalization performance. Moreover, we show that, different from that10 of shallow models (e.g. logistic regression, support vector machines), the gradient11 descent optimization dynamics of DNNs naturally produces large GSNR during12 training, which is probably the key to DNNs’ remarkable generalization ability.