Paper introduces a new adaptive gradient method with gradient compression for distributed training.
problem Communication overhead in distributed machine learning systems.
method Adaptive gradient method with gradient compression, scalable system BytePS-Compress.
result Convergence rate of for non-convex problems.