References & Citations
Computer Science > Machine Learning
Title: CD-SGD: Distributed Stochastic Gradient Descent with Compression and Delay Compensation
(Submitted on 21 Jun 2021 (v1), last revised 7 Sep 2021 (this version, v2))
Abstract: Communication overhead is the key challenge for distributed training. Gradient compression is a widely used approach to reduce communication traffic. When combining with parallel communication mechanism method like pipeline, gradient compression technique can greatly alleviate the impact of communication overhead. However, there exists two problems of gradient compression technique to be solved. Firstly, gradient compression brings in extra computation cost, which will delay the next training iteration. Secondly, gradient compression usually leads to the decrease of convergence accuracy.
Submission history
From: Enda Yu [view email][v1] Mon, 21 Jun 2021 01:15:12 GMT (5633kb,D)
[v2] Tue, 7 Sep 2021 01:44:00 GMT (5634kb,D)
Link back to: arXiv, form interface, contact.