Improving Human Activity Recognition With Wearable Sensors Through BEE: Leveraging Early Exit and Gradient Boosting

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Early-exiting in deep neural networks accelerates activity inference by using internal classifiers at shallower layers.
• Existing methods often optimize all classifiers equally, leading to a training-test data distribution mismatch and suboptimal performance.

Purpose of the Study:

• To address the limitations of uniform classifier optimization in early-exiting networks for activity recognition.
• To propose a novel dynamic-depth network architecture for improved accuracy-speed trade-offs in human activity recognition.

Main Methods:

• Organizing classifiers into a dynamic-depth network and jointly optimizing them using a gradient-boosting approach.
• Implementing gradient-rescaling for stable training of classifiers at different depths.
• Employing prediction reweighting to prioritize deeper classifiers and mitigate data scarcity.

Main Results:

• The proposed dynamic-depth network achieves state-of-the-art accuracy and speed on multiple human activity recognition benchmarks (UCI-HAR, PAMAP2, UniMiB-SHAR, USC-HAD).
• Demonstrated effectiveness through comprehensive experiments and a real-world implementation on an ARM-based mobile device.

Conclusions:

• The dynamic-depth network effectively resolves the training-test data distribution mismatch in early-exiting models.
• This approach offers significant improvements in both accuracy and computational efficiency for human activity recognition tasks.