The Regularization Effects of Anisotropic Noise in Stochastic Gradient Descent.

Zhanxing Zhu,Jingfeng Wu,Bing Yu,Lei Wu,Jinwen Ma
2018-01-01
Abstract:Understanding the generalization of deep learning has raised lots of concerns recently, where the learning algorithms play an important role in generalization performance, such as stochastic gradient descent (SGD). Along this line, we particularly study the anisotropic noise introduced by SGD, and investigate its importance for the generalization in deep neural networks. Through a thorough empirical analysis, it is shown that the anisotropic diffusion of SGD tends to follow the curvature information of the loss landscape, and thus is beneficial for escaping from sharp and poor minima effectively, towards more stable and flat minima. We verify our understanding through comparing this anisotropic diffusion with full gradient descent plus isotropic diffusion (i.e. Langevin dynamics) and other types of position-dependent noise.
What problem does this paper attempt to address?