More about HKUST
Resource-efficient Learning for Large Foundation Models
PhD Qualifying Examination Title: "Resource-efficient Learning for Large Foundation Models" by Mr. Sikai BAI Abstract: The success of Large Foundation Models (LFM) is shadowed by their immense computational and memory costs, creating a critical bottleneck for practical deployment. This survey explores resource-efficient learning, a paradigm that shifts the focus from raw model scale to maximizing inference performance relative to its deployment costs in latency, memory, and energy. We systematically review three primary strategies to achieve this: 1) Parameter-Efficient Fine-tuning: Adapting models to new tasks using a minimal fraction of trainable parameters, avoiding costly full fine-tuning. 2) Model Compression: Aggressively reducing the model's deployment footprint and accelerating computation through techniques like quantization, pruning, and knowledge distillation. 3) Data-Efficient Learning: Reducing the reliance on expensive supervised data to build and align high-performance models. Finally, future research aims to apply the philosophy of resource-efficiency into novel strategies tailored for large-scale reasoning models, focusing on optimizing long-context inference and distilling complex reasoning capabilities into smaller, faster architectures. This ensures these powerful reasoning tools can be developed and deployed efficiently and responsibly across a wider range of applications. Date: Thursday, 28 August 2025 Time: 3:00pm - 5:00pm Venue: Room 3494 Lifts 25/26 Committee Members: Prof. Song Guo (Supervisor) Prof. Ke Yi (Chairperson) Prof. Long Quan