This study analyzes signSGD and medianSGD for heterogeneous data and proposes a noise correction mechanism.
problem The convergence of signSGD and medianSGD is non-convergent in distributed settings with heterogeneous data.
method The study analyzes signSGD and medianSGD for heterogeneous data and proposes a noise correction mechanism to overcome the convergence gap.
result The proposed noise correction mechanism provably closes the gap between mean and median of the gradients, leading to global convergence to stationary solutions.