The Convergence of Sparsified Gradient Methods

Abstract

Distributed training of massive machine learning models, in particular deep neural networks, via Stochastic Gradient Descent (SGD) is becoming commonplace. Several families of communication-reduction methods, such as quantization, large-batch methods, and gradient sparsification, have been proposed. To date, gradient sparsification methods - where each node sorts gradients by magnitude, and only communicates a subset of the components, accumulating the rest locally - are known to yield some of the largest practical gains. Such methods can reduce the amount of communication per step by up to three orders of magnitude, while preserving model accuracy. Yet, this family of methods currently has no theoretical justification. This is the question we address in this paper. We prove that, under analytic assumptions, sparsifying gradients by magnitude with local error correction provides convergence guarantees, for both convex and non-convex smooth objectives, for data-parallel SGD. The main insight is that sparsification methods implicitly maintain bounds on the maximum impact of stale updates, thanks to selection by magnitude. Our analysis and empirical validation also reveal that these methods do require analytical conditions to converge well, justifying existing heuristics.

arXiv commentNIPS 2018 - Advances in Neural Information Processing Systems; Authors in alphabetic order

References in corpus (13)

• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
• MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distributed Systems
• HOGWILD!: A Lock-Free Approach to Parallelizing Stochastic Gradient Descent
• Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training
• TernGrad: Ternary Gradients to Reduce Communication in Distributed Deep Learning
• Large Batch Training of Convolutional Networks
• Asynchronous Parallel Stochastic Gradient for Nonconvex Optimization
• Gradient Sparsification for Communication-Efficient Distributed Optimization
• meProp: Sparsified Back Propagation for Accelerated Deep Learning with Reduced Overfitting
• YellowFin and the Art of Momentum Tuning
• Asynchronous stochastic convex optimization
• SparCML: High-Performance Sparse Communication for Machine Learning
• The Convergence of Stochastic Gradient Descent in Asynchronous Shared Memory

Cited by in corpus (5)

• Machine Learning at the Wireless Edge: Distributed Stochastic Gradient Descent Over-the-Air
• Federated Learning With Quantized Global Model Updates
• ScaDLES: Scalable Deep Learning over Streaming data at the Edge
• EventGraD: Event-Triggered Communication in Parallel Machine Learning
• Communication-Efficient Distributed Optimization with Quantized Preconditioners