Data augmentation and model generalization are fundamental topics in machine learning. This talk presents a systematic overview that bridges these areas from both theoretical and practical perspectives. It examines the theoretical foundations and practical effects of data augmentation on model generalization and robustness, covering local and global augmentation techniques, extensions of adversarial training, large language model (LLM)-based augmentation, and physics-informed strategies. By clarifying the theoretical connections between these approaches and model generalization, the talk also demonstrates their practical value in medical analysis, industrial anomaly detection, and point tracking. The presentation draws primarily on the team’s recent research published at leading AI conferences, including AAAI, NeurIPS, CVPR, SIGGRAPH, and ACL.
Kaizhu Huang works in trustworthy AI and its applciation in computer vision, large language model and pattern recognition. He is a Full Professor of Electrical and Computer Engineering and Director of the Digital Innovation Research Center at Duke Kunshan University. He earned his PhD in Computer Science and Engineering from The Chinese University of Hong Kong in 2004 and subsequently held research positions at Fujitsu Research Centre, the University of Bristol, and the Chinese Academy of Sciences, and Xi’an Jiaotong-Liverpool University. His honors include the 2024 IEEE ICDM 10-Year Highest-Impact Paper Award, the 2011 Asia-Pacific Neural Network Society Young Researcher Award, and 2006 Fujitsu President Award. He has published more than 300 peer-reviewed papers including 150+ international journal articles, received over nine best-paper, runner-up, or book awards, serves in editorial roles for leading journals and book series. He has delivered over 60 keynote or tutorial talks at international conferences and workshops.
To be announced soon.