DeepEmote: Towards multi-layer neural networks in a low power wearable multi-sensors bracelet

Abstract

Wearable smart sensing is a promising technology to enhance user experience that has already been exploited in sport/fitness, as well as health and human monitoring. Wearable sensing systems not only provide continuous data monitoring and acquisition, but are also expected to process, and make sense of the acquired data by classification in similar ways as human experts do. Supporting continuous operation on ultra-small batteries poses unique challenges in energy efficiency. In this paper, we present an ultra-low-power bracelet with several sensors that is able to run multi-layer neural networks learning algorithms to process data efficiently. The design combines low-power design, energy efficient algorithms and makes this bracelet suitable for long-term uninterrupted usage with small coin batteries. We demonstrate in-field measurement results that prove that neural networks applications can fit within the mW power and memory envelope of a commercial ARM Cortex M4F microcontroller. We show that a fully connected network of 26 neurons achieve an accuracy of 100% on emotion detection, using only 2% of memory available. Field trials demonstrate that the wearable device can achieve a 2-month lifetime while performing one emotion detection classification every 10 minutes.