Regularized Top-$k$: A Bayesian Framework for Gradient Sparsification