Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training

Abstract

We study the effect of mini-batching on the loss landscape of deep neural networks using spiked, field-dependent random matrix theory. We demonstrate that the magnitude of the extremal values of the batch Hessian are larger than those of the empirical Hessian. We also derive similar results for the Generalised Gauss-Newton matrix approximation of the Hessian. As a consequence of our theorems we derive an analytical expressions for the maximal learning rates as a function of batch size, informing practical training regimens for both stochastic gradient descent (linear scaling) and adaptive algorithms, such as Adam (square root scaling), for smooth, non-convex deep neural networks. Whilst the linear scaling for stochastic gradient descent has been derived under more restrictive conditions, which we generalise, the square root scaling rule for adaptive optimisers is, to our knowledge, completely novel. %For stochastic second-order methods and adaptive methods, we derive that the minimal damping coefficient is proportional to the ratio of the learning rate to batch size. We validate our claims on the VGG/WideResNet architectures on the CIFAR-$100$ and ImageNet datasets. Based on our investigations of the sub-sampled Hessian we develop a stochastic Lanczos quadrature based on the fly learning rate and momentum learner, which avoids the need for expensive multiple evaluations for these key hyper-parameters and shows good preliminary results on the Pre-Residual Architecure for CIFAR-$100$.

References in corpus (14)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• One weird trick for parallelizing convolutional neural networks
• The Loss Surfaces of Multilayer Networks
• On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima
• Identifying and attacking the saddle point problem in high-dimensional non-convex optimization
• Stochastic Gradient Descent for Non-smooth Optimization: Convergence Results and Optimal Averaging Schemes
• A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasets
• On Empirical Comparisons of Optimizers for Deep Learning
• Towards Understanding Generalization of Deep Learning: Perspective of Loss Landscapes
• Gradient Descent Happens in a Tiny Subspace
• The large learning rate phase of deep learning: the catapult mechanism
• Semicircle Law for a Class of Random Matrices with Dependent Entries
• An information and field theoretic approach to the grand canonical ensemble
• Deep Curvature Suite

Cited by in corpus (3)

• Deformed semicircle law and concentration of nonlinear random matrices for ultra-wide neural networks
• Dynamics of Stochastic Momentum Methods on Large-scale, Quadratic Models
• Efficient fine-tuning of 37-level GraphCast with the Canadian global deterministic analysis