Label Shift Aware Adaptation for Online Zero-shot Learning with Contrastive Language-Image Pre-Training (CLIP)
A new research paper introduces Label Shift Aware (LSA), a method designed to improve online zero-shot learning performance for vision-language models like Contrastive Language-Image Pre-Training (CLIP). LSA addresses the challenge of differing label distributions between training and testing data by formulating the problem as a domain adaptation task. The approach adapts CLIP's predictions using unlabeled test data and incorporates label shift correction to mitigate performance degradation. AI
IMPACT This research could enhance the accuracy of vision-language models in real-world scenarios where data distributions shift.