An Investigation into Neural Net Optimization via Hessian Eigenvalue Density

Abstract

To understand the dynamics of optimization in deep neural networks, we develop a tool to study the evolution of the entire Hessian spectrum throughout the optimization process. Using this, we study a number of hypotheses concerning smoothness, curvature, and sharpness in the deep learning literature. We then thoroughly analyze a crucial structural feature of the spectra: in non-batch normalized networks, we observe the rapid appearance of large isolated eigenvalues in the spectrum, along with a surprising concentration of the gradient in the corresponding eigenspaces. In batch normalized networks, these two effects are almost absent. We characterize these effects, and explain how they affect optimization speed through both theory and experiments. As part of this work, we adapt advanced tools from numerical linear algebra that allow scalable and accurate estimation of the entire Hessian spectrum of ImageNet-scale neural networks; this technique may be of independent interest in other applications.

21 pages, 19 figures

Cited by in corpus (7)

• Scaling Laws for Neural Language Models
• The Early Phase of Neural Network Training
• The Break-Even Point on Optimization Trajectories of Deep Neural Networks
• Emergent properties of the local geometry of neural loss landscapes
• How noise affects the Hessian spectrum in overparameterized neural networks
• The asymptotic spectrum of the Hessian of DNN throughout training
• A Dynamic Sampling Adaptive-SGD Method for Machine Learning