Dynamic instance-aware layer-bit-select network on human activity recognition using wearable sensors

During recent years, deep convolutional neural networks have achieved remarkable success in a wide range of sensor-based human activity recognition (HAR) applications, which often require high computational cost and memory footprint, hence hindering practical HAR deployment on resource-limited mobile and wearable devices. Quantization has provided an effective solution to compress models and accelerate activity inference in real-world situations. However, most previous quantization schemes are static, which always utilize the same bit-width for all activity samples in a given layer. Intuitively, since activity samples are highly diverse according to their difficulty level, it is rather unrealistic to maintain a consistent bit-width quantization configuration for different activity samples. Based on dynamic quantization strategy, this paper introduces a novel Layer-Bit-Select Network named LBSNet to adaptively determine the optimal bit-widths of each convolutional layer according to the difficulty level of recognized activities. To achieve this goal, we design a lightweight Bit-selector, which is jointly optimized with a given main network. In such a way, easy activities such as sitting may be allocated to lower bit-widths, while high bit-widths may handle more complicated or hard activities like falls. Extensive experiments are conducted on several mainstream HAR benchmarks including WISDM, UCI-HAR, UniMiB-SHAR, and PAMAP2 to validate the effectiveness of our proposed approach. For instance, it can achieve round 5.3$\times$ speedup and 6.2$\times$ model size compression, with merely 0.6% accuracy drop on WISDM dataset, compared to full-precision model. This approach has great potential to yield more computation-efficient and faster activity inference on mobile embedded platforms.