Dynamic Computation for Efficient AI: From Foundations to Applications
COM1 Level 3
MR1, COM1-03-19
Abstract:
As AI models grow in scale and complexity, achieving efficiency and adaptability becomes increasingly critical. My research focuses on dynamic neural networks—models that can adjust their behavior and structure based on varying computational demands. This work aims to enable more efficient and scalable AI models by addressing key challenges in training and inference
1. Training: Traditional methods for transforming static models into dynamic ones often rely on full-parameter fine-tuning, which is prohibitively expensive for large-scale models. I propose leveraging parameter-efficient fine-tuning (PEFT) to construct dynamic models, significantly reducing training costs while maintaining performance.
2. Inference: While most research on dynamic networks emphasizes accelerating visual perception tasks, critical areas like visual generation and vision–language understanding remain underexplored. To address this, I extended dynamic modeling into these domains, introducing innovations such as DyDiT and DyDiT++ for visual generation, as well as SGL for vision–language understanding. These advancements highlight the versatility of dynamic networks.
Looking forward, I aim to build a comprehensive and dynamic computational ecosystem by integrating model design with machine learning systems through joint optimization. This vision aims to foster the development of efficient, adaptable, and versatile AI models,
driving meaningful progress toward real-world advancements in Artificial General Intelligence (AGI).