FedDQ: Communication-Efficient Federated Learning with Descending Quantization
Federated learning (FL) is an emerging privacy-preserving distributed learning scheme. Due to the large model size and frequent model aggregation, FL suffers from critical communication bottleneck. Many techniques have been proposed to reduce the communication volume, including model compression and quantization, where quantization with increasing number of levels has been proposed. This paper proposes an opposite approach to do adaptive quantization. First, we present the drawback of ascending-trend quantization based on the characteristics of training. Second, we formulate the quantization optimization problem and theoretical analysis shows that quantization with decreasing number of levels is preferred. Then we propose two strategies to guide the adaptive quantization process by using the change in training loss and the range of model update. Experimental results on three sets of benchmarks show that descending-trend quantization not only saves more communication bits but also helps FL converge faster, when compares with current ascending-trend quantization.
READ FULL TEXT