EnergySense: A Fine-Grained Energy Analysis Framework for DNN Processing with Low-Power Ubiquitous Sensors

Energy efficiency is the key requirement to maximize the lifetime of ubiquitous sensors, such as mobile phones, wearable devices and resource-constrained sensors. Such sensors offer new possibilities for providing AI-based applications and services in low-power, low-cost local processing with real-time feedback. However, they have difficulty handling computationally intensive feature extraction for deep learning because they are typically powered by batteries with a finite lifetime. To address this challenge, we propose an energy-efficient analysis framework. 1) A response-adaptive energy model is proposed to evaluate the global consumption of a ubiquitous computing system in low-power operation mode. 2) We also report the design and implementation of EnergySense, which is an energy-efficient scheduling framework, and further optimize the energy efficiency of the execution of DNN-based tasks. Extensive experiments verify the framework's effectiveness in reducing energy consumption compared to the baseline method. With up to 10 scheduled sensors, the energy consumption for DNN computing is reduced by 67.7% compared to the case of a single device. As a result, Energy-Sense provides a longer life cycle of the feature map extraction and improves the computing power of low-power ubiquitous sensors.