Distributed learning with compressed gradients
This work addresses scalability issues in large-scale machine learning for researchers and practitioners, but it is incremental as it provides a unified analysis framework rather than introducing new methods.
The paper tackled the problem of scalability in distributed optimization by analyzing the convergence rates and communication complexity of gradient methods using stale and compressed gradients, deriving explicit step-size expressions and showing that fast convergence with limited information exchange is possible.
Asynchronous computation and gradient compression have emerged as two key techniques for achieving scalability in distributed optimization for large-scale machine learning. This paper presents a unified analysis framework for distributed gradient methods operating with staled and compressed gradients. Non-asymptotic bounds on convergence rates and information exchange are derived for several optimization algorithms. These bounds give explicit expressions for step-sizes and characterize how the amount of asynchrony and the compression accuracy affect iteration and communication complexity guarantees. Numerical results highlight convergence properties of different gradient compression algorithms and confirm that fast convergence under limited information exchange is indeed possible.