Core Concepts
Efficiently quantifying uncertainty for wearable event detection on microcontrollers is crucial for reliable predictions and system efficiency.
Abstract
Traditional machine learning techniques struggle with shifts in data distribution, especially in mobile healthcare applications. UR2M introduces a novel framework for uncertainty-aware event detection on microcontrollers. It achieves faster inference speed, energy-saving, memory efficiency, and improved uncertainty quantification. The proposed approach utilizes evidential deep learning and cascade learning to optimize model deployment and efficiency. By sharing shallower layers among different event models, UR2M enables efficient multi-event detection while minimizing memory constraints. Extensive experiments demonstrate the effectiveness of UR2M across various wearable datasets.
Stats
Our results demonstrate that UR2M achieves up to 864% faster inference speed.
UR2M achieves 857% energy-saving for uncertainty estimation.
The approach saves 55% of memory compared with existing uncertainty estimation baselines.
A 22% improvement in uncertainty quantification performance was observed.