Scalable Second Order Optimization for Deep Learning

TL;DR

This paper introduces a scalable second-order optimization method tailored for deep learning, achieving faster convergence and better performance on large-scale tasks by leveraging heterogeneous hardware architectures.

Contribution

The paper presents a practical, scalable second-order optimization algorithm that significantly improves training efficiency and effectiveness for deep neural networks.

Findings

Outperforms first-order methods in convergence speed

Achieves better wall-clock time on large models

Demonstrates superior results on diverse large-scale tasks

Abstract

Optimization in machine learning, both theoretical and applied, is presently dominated by first-order gradient methods such as stochastic gradient descent. Second-order optimization methods, that involve second derivatives and/or second order statistics of the data, are far less prevalent despite strong theoretical properties, due to their prohibitive computation, memory and communication costs. In an attempt to bridge this gap between theoretical and practical optimization, we present a scalable implementation of a second-order preconditioned method (concretely, a variant of full-matrix Adagrad), that along with several critical algorithmic and numerical improvements, provides significant convergence and wall-clock time improvements compared to conventional first-order methods on state-of-the-art deep models. Our novel design effectively utilizes the prevalent heterogeneous hardware architecture for training deep models, consisting of a multicore CPU coupled with multiple accelerator units. We demonstrate superior performance compared to state-of-the-art on very large learning tasks such as machine translation with Transformers, language modeling with BERT, click-through rate prediction on Criteo, and image classification on ImageNet with ResNet-50.