报告人
Yanbiao Ma Xidian University 时间 2025年4月1日 星期二 上午 10:00-11:00 地点
102 会议室
Abstract Building fair and trustworthy artificial intelligence systems is a fundamental requirement for the intelligent transformation of industries, and it holds critical significance for reliable decision-making in high-risk fields such as smart healthcare and autonomous driving. As the core theory of the current AI boom, deep neural networks commonly suffer from representational bias in practical applications, manifesting as significant disparities in recognition capabilities across different categories, attributes, or scenarios. Such biases may lead to misjudgments in high-risk tasks like medical diagnosis and security surveillance, potentially compromising public safety. The causes of representational bias are complex, and existing theories simplistically attribute it to the long-tailed distribution characteristics of real-world data, failing to explain bias phenomena in general scenarios. Consequently, debiasing methods based on these theories have limited applicability. Addressing the lack of a systematic framework for quantitative analysis and optimization in current theories, this report will focus on exploring the mechanisms of representational bias, designing debiasing methods, and quantifying and optimizing the imbalance in model noise generalization capabilities, thereby advancing and refining existing theories.
Biography Yanbiao Ma is a Ph.D. candidate at the School of Artificial Intelligence, Xidian University, under the supervision of Professor Licheng Jiao. His primary research interests include fairness in deep neural networks, long-tailed learning, and federated learning. As the first author, he has published over 10 academic papers in prestigious journals and conferences such as TPAMI, IJCV, CVPR, ICLR, ACM MM, and TMM. He has won five international competition championships organized by CVPR, ICCV, ECCV, and IGARSS, along with one second-place and one third-place finish. He has also served as a program committee member and reviewer for several top-tier international conferences and journals, including CVPR, NeurIPS, ICML, and ICCV.
