FedDQ: Communication-Efficient Federated Learning with Descending Quantization

Federated learning (FL) is an emerging privacy-preserving distributed learning scheme. The large model size and frequent model aggregation cause serious communication issue for FL. Many techniques, such as model compression and quantization, have been proposed to reduce the communication volume. Existing adaptive quantization schemes use ascending-trend quantization where the quantization level increases with the training stages. In this paper, we investigate the adaptive quantization scheme by maximizing the convergence rate. It is shown that the optimal quantization level is directly related to the range of the model updates. Given the model is supposed to converge with training, the range of model updates will get smaller, indicating that the quantization level should decrease with the training stages. Based on the theoretical analysis, a descending quantization scheme is thus proposed. Experimental results show that the proposed descending quantization scheme not only reduces more communication volume but also helps FL converge faster, when compared with current ascending quantization scheme.