PhD Dissertation Defense by MA Xiao | Practical On-Device Deep Intelligence for Resource-Constrained Embedded and Wearable Devices

ABOUT THE TALK

Embedded and wearable devices increasingly rely on deep neural networks to support intelligent sensing directly on the device. However, practical on-device intelligence remains difficult across the deployment lifecycle. Before deployment, models must operate accurately under severe memory, computation, and energy constraints, while often supporting multiple sensing applications on a single device. After deployment, they must remain reliable as real-world data distributions change over time.

This dissertation studies four connected challenges. First, highly constrained devices may not have sufficient capacity to run a single compressed model accurately. We therefore explore how multiple complementary tiny models can be coordinated and selectively executed to improve accuracy without proportionally increasing per-input computation or peak memory. Second, wearable devices require reusable intelligence across multiple applications. We investigate how a wearable foundation model can be compressed and specialized for initial tasks while preserving transferable representations for future tasks. Third, continuous adaptation to every incoming batch can waste resources when the environment is stable. We develop an on-demand strategy that detects distribution shifts and triggers adaptation only when necessary. Fourth, conventional adaptation often depends on backpropagation and architecture-specific components, leading to high memory and latency overhead. We instead study lightweight embedding alignment that corrects distribution-induced feature distortions using forward-only computation and applies across convolutional and transformer-based architectures.

Together, these contributions present a unified lifecycle perspective on practical on-device deep intelligence. They show how model capacity can be allocated more effectively before deployment, how reusable representations can be maintained as applications evolve, and how post-deployment adaptation can be triggered and executed efficiently.

ABOUT THE SPEAKER

Xiao MA is a PhD candidate in Computer Science at Singapore Management University (SMU), supervised by Professor Dong Ma and Professor Baihua Zheng. He received both his Bachelor's degree and his Master's degree from Beijing Institute of Technology. He was also a visiting PhD student at the University of Cambridge during Jan to May 2026. His research focuses on on-device deep learning, spanning embedded machine learning (TinyML), test-time adaptation, and mobile sensing. His work has been published at top-tier venues including ICLR, IEEE PerCom, IEEE TMC, and ACM IMWUT. Notable achievements include the Mark Weiser Best Paper Award at PerCom 2024 for DiTMoS, a Distinguished Paper Award at IMWUT 2024 for LR-Auth, the SMU Presidential Doctoral Fellowship (2024, 2025), and the SMU Dean's List (2025).