- Abstract: Deep neural networks exhibit complex learning dynamics due to the highly non-convex loss landscape, which causes slow convergence and vanishing gradient problems. Second order approaches, such as natural gradient descent, mitigate such problems by neutralizing the effect of potentially ill-conditioned curvature on the gradient-based updates, yet precise theoretical understanding on how such curvature correction affects the learning dynamics of deep networks has been lack- ing. Here, we analyze the dynamics of training deep neural networks under a generalized family of natural gradient methods that applies curvature corrections, and derive precise analytical solutions. Our analysis reveals that curvature corrected update rules preserve many features of gradient descent, such that the learning trajectory of each singular mode in natural gradient descent follows precisely the same path as gradient descent, while only accelerating the temporal dynamics along the path. We also show that layer-restricted approximations of natural gradient, which are widely used in most second order methods (e.g. K-FAC), can significantly distort the learning trajectory into highly diverging dynamics that significantly differs from true natural gradient, which may lead to undesirable net- work properties. We also introduce fractional natural gradient that applies partial curvature correction, and show that it provides most of the benefit of full curvature correction in terms of convergence speed, with additional benefit of superior numerical stability and neutralizing vanishing/exploding gradient problems, which holds true also in layer-restricted approximations.
- Original Pdf: pdf