Professor Cho Sung-in's Research Team Has a Paper Accepted at NeurIPS 2025, a Top International Conference in Artificial Intelligence

▶ (From left) Professor Cho Sung-in; Lee Jae-yoon, master's student

The paper 'DUET: Dual-Facet Pseudo Labeling and Uncertainty-aware Exploration & Exploitation Training for Source-Free Domain Adaptation', written by master's student Lee Jae-yoon (first author) of Professor Cho Sung-in's research team (Smart Vision & Media (SVM) Lab) and Professor Cho Sung-in (corresponding author), has been accepted for presentation at NeurIPS 2025, a top international conference in artificial intelligence.

▶ The Neural Information Processing Systems (NeurIPS) 2025 logo (above) and the link to the poster paper description (below)

The paper proposes a source-free domain adaptation technique that adapts a pre-trained source model to unlabelled target domain data. Existing techniques that use the multimodal model CLIP for domain adaptation assign pseudo labels to all data without any correction despite the target model's low accuracy early in training, and thus suffer performance degradation caused by low-quality pseudo labels. In addition, the existing loss functions used to train CLIP on target domain data for knowledge transfer to the target model do not take account of how far the model has already adapted to the target domain, and so fail to provide a proper training signal.

The proposed technique, DUET, assigns pseudo labels only to samples where the prediction of CLIP — with its excellent domain generalization — agrees with that of the task-specific pre-trained source model, and applies consistency regularization to the unassigned samples for rich feature learning, greatly raising the quality of the pseudo labels. The newly proposed loss function also uses an entropy computation algorithm that takes domain adaptation into account, so that it grasps the model's training state more sensitively than existing loss functions and provides the correct training signal. Implementation and evaluation demonstrated an average 2% performance improvement across three benchmark datasets, together with an improvement in the model's overconfidence problem.

▶ The corrected pseudo label generation process (CPG), dynamic entropy-based visual optimization (DVO) and the pseudo labelling matching framework (PLMatch)

NeurIPS (Conference on Neural Information Processing Systems) is the world's most prestigious conference in artificial intelligence and machine learning. It focuses on evaluating originality, technical contribution, reproducibility and academic and industrial impact, and comprises a main research track together with Datasets & Benchmarks, workshops and tutorials, competitions and an industry expo. NeurIPS 2025 runs as parallel in-person events in San Diego, United States (2–7 December) and Mexico City (30 November – 5 December).

References:

NeurIPS 2025 homepage: https://neurips.cc/

Paper link: https://neurips.cc/virtual/2025/poster/120326

Homepage of Professor Cho Sung-in's research team: https://sites.google.com/view/csi2267svm/