ADP-MTL: An Adaptive Multi-Task Learning for Efficient and Simultaneous Task Execution in Resource-Constrained Mobile Human Activity Recognition

Jiashuo Wang

 Co-Presenters: Individual Presentation
College: Dorothy and George Hennings College of Science, Mathematics and Technology
Major: Mathematical Sciences
Faculty Research Mentor: Bin Hu
 

Abstract:

This paper presents ADP-MTL, an adaptive multi-task learning (MTL) framework designed for efficient and simultaneous task execution in human activity recognition (HAR) on resource-constrained mobile devices. By leveraging dynamic masking and progressive weight pruning, the framework isolates task-specific parameters while reallocating redundant ones, ensuring efficient parameter utilization and scalability. We utilize a Transformer-based model architecture optimized for capturing temporal dependencies in wearable sensor data. The proposed approach achieves higher accuracy and significantly reduces computational, memory, and time overheads compared to deploying multiple independent deep learning models. Our evaluation across multiple HAR datasets and deep learning models demonstrates that ADP-MTL improves accuracy by up to 3.5% while reducing the memory footprint and FLOPs by 30–50% compared to conventional approaches. This framework provides a robust, scalable, and resource-efficient solution for HAR applications in mobile and edge computing environments.