Understanding Sensor Data Using Deep Learning Methods on Resource-Constrained Edge Devices.

With the development of the Internet of Things, more and more edge devices (such as smart-phones, tablets, wearable devices, embedded devices, gateway equipment and etc.) generate huge amounts of rich sensor data every day. With them, some deep learning based recognition applications provide users with various recognition services on edge devices. However, a fundamental problem these applications meet is how to perform deep learning algorithms effectively and promptly on a resource-constrained platform. Some researchers have proposed completing all computation tasks on the cloud side then returning the results back to edge devices, but such procedure is always time-consuming because of data transmission. In this case, training deep learning models on cloud side and executing the trained model directly on edge devices for inference is a better choice. Meanwhile, the deep learning based mobile applications also need to satisfy the requirements of low latency, low storage and low consumption. To fulfill above objectives, we aim to propose a new deep learning compression algorithm. We conduct comprehensive experiments to compare the proposed light-weight model with other standard state-of-the-art compression algorithms in terms of inference accuracy, process delay, CPU load, energy cost and storage coverage, based on an audio recognition system.