報告題目	Understanding Machine Learning with Noisy Labels: Challenges, Techniques, and Solutions
報告人	S M Hasan Mahmud, Associate Professor, Department of Software Engineering，Daffodil International University，Bangladesh
報告时间	2025年4月12日（周六）下午15:30-16:00
報告所在	PP电子學院科技綜合樓711
報告內容簡介	報告內容簡介： In recent years, the rise of deep learning has significantly advanced the performance of models across various domains. However, the effectiveness of these models heavily depends on the quality of labeled data. Noisy data labels—incorrect or imprecise annotations—pose a serious challenge, particularly when data is sourced from multiple or unreliable origins. This study explores the different data sources that contribute to label noise and highlights the diverse challenges associated with them, including class imbalance, mislabeled instances, and inconsistencies across datasets. We categorize label noise into symmetric and asymmetric types and analyze how these affect model training and generalization. Furthermore, this lecture focuses on key techniques developed to address these challenges, including robust loss functions, noise-tolerant algorithms, label correction methods, and semi-supervised learning approaches. Special emphasis is placed on the role of deep learning in mitigating the impact of noisy labels through architectures that inherently model uncertainty or adapt during training. By synthesizing current research and identifying practical solutions, this lecture aims to provide a comprehensive overview of the state-of-the-art in handling noisy labeled data. 報告人簡介： Dr. S M Hasan Mahmud is an Associate Professor of Department of Software Engineering，Daffodil International University in Bangladesh. His research interests includes Bio-informatics，Data Science，Machine Learning，AI Drug Discovery and Health Informatics. He has over 40 publications as journal and conference papers. His google scholar citations are 1335 and total impact factor over 170.
承辦學院	電子與信息工程學院
發布日期	2025-4-8
歡迎廣大教師、學生參加！