PhD Dissertation Proposal by MA Xiao | Towards Efficient Deep Learning on Resource-Constrained Embedded Devices

Towards Efficient Deep Learning on Resource-Constrained Embedded Devices

ABOUT THE TALK

The widespread adoption of deep learning has enabled breakthroughs in perception and decision-making tasks across domains such as vision, speech, and human activity recognition. However, deploying such capabilities on resource-constrained embedded devices—like microcontrollers and edge sensors—remains a major challenge due to the tight limits on memory, compute, and energy. These challenges are particularly severe in on-device settings, where models must operate autonomously, adapt in real time, and process data locally without access to powerful cloud infrastructure. This thesis investigates two complementary research directions to address these challenges. First, it proposes a bottom-up design paradigm that builds accurate, ultra-compact models from scratch using diversity-driven ensemble techniques and lightweight fusion mechanisms. Second, it explores robust, unsupervised adaptation strategies to maintain model accuracy in the presence of domain shifts. Together, these contributions aim to establish a practical foundation for deploying deep learning in real-world embedded systems, where reliability, adaptability, and efficiency are critical.

SPEAKER BIOGRAPHY

MA Xiao is a third-year Ph.D. candidate in Computer Science at the School of Computing and Information Systems under the supervision of Associate Professor Dong MA. His research focuses on developing efficient machine learning algorithms for resource-constrained devices.