New algorithm reduces communication costs in distributed deep learning.
problem High communication costs in distributed deep learning.
method Sparse-SignSGD with Majority Vote (S3GD-MV).
result Significantly reduces communication costs while maintaining accuracy.