Adaptive Ratio-Based-Threshold Gradient Sparsification Scheme for Federated Learning

Federated learning (FL) is a distributed learning paradigm that has received great attention over the past several years due to its privacy-preserving property. As the models involved in FL are usually dense and overparameterized however, various studies are being conducted in gradient sparsification to reduce the high communication overhead. While many of the recently-presented schemes that are variations of top-k have shown competitive inference accuracy convergence to the baseline “vanilla” FL, they have a fixed sparsity rate throughout all of the communication rounds, which leads to an unnecessary excessive transmission of gradients as the global model converges. Furthermore, the constant-threshold gradient sparsification method called Threshold-z-, that is well-known for its dynamic rate, does not account for the ratio between the gradient and the pre-update parameter value, causing some gradients that are orders of magnitude larger than the pre-update parameter values to be neglected in the following aggregation process. In this paper, we introduce a new algorithm that addresses both of these issues, called adaptive ratio-based-threshold gradient sparsification method. Our main idea is introducing a custom gradient sparsity threshold for each local parameter based on their pre-update value and a hyperparameter denoted as Ψ. We demonstrate through image classification experiments on MNIST and CIFAR-10 datasets in both independent-and-identically-distributed (IID) and non-IID settings that under optimal Ψ, the gradient sparsity rates adapt & increase as the global model converges, while simultaneously producing inference accuracies that are competitive to vanilla FL.