Proceedings of the Annual Meeting of the Cognitive Science Society

Once the human brain learns a concept, it can easily transfer the learned knowledge across diverse environments without referring back to the original learning materials. Inspired by this cognitive process, Black-Box Domain Adaptation (BBDA) has been purposed to transfer the knowledge learned in a black-box source model to the target domain without any premise for source data or model parameters. Existing BBDA methods mainly rely on knowledge distillation or sample selection with pseudo labels, overlooking the different learning difficulties of classes. This results in easy-learning classes dominating the adaptation process and thus degrades adaptation performance. Motivated by the significant success of Vision-Language models (ViL model), we propose a novel method that integrates the knowledge of ViL model to achieve adaptation while mitigating learning imbalance. Experiments on various datasets demonstrate the effectiveness of the proposed method.