Learning under Less and Noisy Labels:
The world’s data are overwhelmingly unlabelled, and semi-supervised learning (SSL) has long provided an effective pathway for exploiting such data with minimal annotation. SSL has achieved substantial progress across classification, segmentation, generation, and language understanding, enabling more data-efficient modelling. This study aims to advance core Semi/weakly-supervised learning principles, covering efficient augmentation, stable and uncertainty-aware pseudo-labeling, sample-adaptive supervision, and multi-teacher learning, etc., to improve segmentation, classification, and medical imaging. Robust methods are also developped for learning from noisy labels in medical imaging, VQA, and captioning.
Z. Zhao, L. Zhou*, Y. Duan, L. Wang, L. Qi, and Y. Shi, "DC-SSL: Addressing Mismatched Class Distribution in Semi-supervised Learning", CVPR, 2022
ZC. Wang, Z. Zhao, X. Xing, D. Xu, X. Kong, and L. Zhou*, "Conflict-Based Cross-View Consistency for Semi-Supervised Semantic Segmentation", CVPR, 2023.
Z. Zhao, S. Long, J. Pi, J. Wang, and L. Zhou*, "Instance-specific and Model-adaptive Supervision for Semi-supervised Semantic Segmentation", CVPR, 2023.
Z. Zhao, L. Yang, S. Long, J. Pi, L. Zhou*, and J. Wang*, "Augmentation Matters: A Simple-yet-Effective Approach to Semi-supervised Semantic Segmentation", CVPR, 2023