Machine Learning with Noisy Labels
演讲者
时间
2025年11月19日 17:05 至 18:30
地点
Shuangqing-C654
线上
Zoom 787 662 9899
(BIMSA)
摘要
With the rise of large AI models, data assets have gained increasing importance. Understanding how to identify and correct label errors in our datasets is crucial. This is because label errors are pervasive in the era of big data and rectifying them can significantly enhance our knowledge. Moreover, large AI models are prone to overfitting label errors, which hinders their ability to generalize. In this talk, we will present typical approaches to handle label noise, such as extracting confident examples (indicating likely correct/incorrect labels) using deep network properties. Additionally, we will explore methods that focus on directly modelling the label noise, providing theoretical guarantees for designing statistically consistent algorithms. By illustrating the intuitions behind state-of-the-art techniques, we would equip researchers and practitioners with valuable insights into effectively managing label noise.
演讲者介绍
Tongliang Liu is the Director of Sydney AI Centre at the University of Sydney, Australia. He received his PhD from UTS and his Bachelor from USTC. He is broadly interested in the fields of trustworthy machine learning and its interdisciplinary applications, with a particular emphasis on learning with noisy labels, adversarial learning, causal representation learning, 3D computer vision, and AI for science. He is/was a senior meta reviewer for many conferences, such as ICML, ICLR, NeurIPS, AAAI, and IJCAI. He is a co-Editor-in-Chief for Neural Networks, an Associate Editor of IEEE TPAMI, IEEE TIP, JAIR, MLJ, TMLR, and ACM Computing Surveys.