A Novel Adaptive Gradient Compression Scheme: Reducing the Communication Overhead for Distributed Deep Learning in the Internet of Things
Peng Luo, F. Richard Yu, Jianyong Chen, Jianqiang Li, Victor C. M. Leung
Abstract
Distributed deep learning deployed in an edge computing environment is a promising approach for extracting accurate information from raw sensor data from Internet of Things (IoT). But the distributed training suffers from heavy communication overheads between a master node and multiple compute nodes due to frequent transmission of gradients, which limits the training efficiency of the distributed deep learning. In this article, we propose a novel algorithm named ProbComp-LPAC (ProbComp: probability compression and LPAC: layer parameters adaptive compression), which can reduce the communication overhead and improve the training efficiency of the distributed deep learning. ProbComp-LPAC adopts a probability equation to select the gradients and uses different compression rates in different layers of deep neural networks. Comparing with other methods, such as adaptive compression (AdaComp) and lazily aggregated quantized compression (LAQ), the performance of ProbComp-LPAC is not only faster in the training speed but also higher in the accuracy of the test.