Taming the Wild: A Unified Analysis of Hogwild!-Style Algorithms
It provides a unified theoretical foundation for optimizing SGD in distributed settings, addressing noise from asynchrony and low-precision arithmetic, which is incremental but important for scalable machine learning.
The paper tackles the analysis of asynchronous and reduced-precision SGD algorithms by introducing a martingale-based framework, enabling convergence proofs for convex and non-convex problems like matrix completion, and demonstrating efficient experimental performance on modern hardware.
Stochastic gradient descent (SGD) is a ubiquitous algorithm for a variety of machine learning problems. Researchers and industry have developed several techniques to optimize SGD's runtime performance, including asynchronous execution and reduced precision. Our main result is a martingale-based analysis that enables us to capture the rich noise models that may arise from such techniques. Specifically, we use our new analysis in three ways: (1) we derive convergence rates for the convex case (Hogwild!) with relaxed assumptions on the sparsity of the problem; (2) we analyze asynchronous SGD algorithms for non-convex matrix problems including matrix completion; and (3) we design and analyze an asynchronous SGD algorithm, called Buckwild!, that uses lower-precision arithmetic. We show experimentally that our algorithms run efficiently for a variety of problems on modern hardware.