Enabling Lightweight Device-Free Wireless Sensing With Network Pruning and Quantization

Deep learning based device-free wireless sensing systems have achieved satisfactory performance in sensing human gesture, identity, location, etc.. However, subject to the limited computation and storage resources of wireless devices, complex deep learning algorithms could not run on these devices, which limits the practical implementation. In this paper, motivated by the emerging edge intelligence technique, we try to explore and exploit realizing lightweight device-free wireless sensing with network pruning and network quantization methods, which could reduce the complexity of a sensing system while keeping its sensing performance almost unchanged. Specifically, we propose a Taylor criterion ranking based network pruning strategy to remove the nonessential neurons so as to reduce the computational complexity and storage requirement, and design a network quantization strategy to quantize network parameters so as to further reduce the storage requirement. We design a mmWave-based device-free gesture recognition testbed to evaluate the proposed strategies. Extensive experimental results show that the developed strategies reduce the computational complexity and storage requirement to 27% and 14%, respectively, with the recognition accuracy reduced by only 2%.