Test-Time Domain-Agnostic Meta-Prompt Learning for Multi-Source Few-Shot Domain Adaptation

Multi-source few-shot domain adaptation (MFDA) is a more common and challenging scenario, as only limited annotated source domain data is provided and a large amount of data is unlabeled. Conventional solutions are difficult to achieve for large-scale vision-language models (VLM) since multiple-source domains need to be aligned and more parameters need to be tuned. To efficiently transfer VLM to the target domain in the MFDA, this study first summarizes the previous prompt tuning for domain adaptation methods as a transductive prompt learning (TPL) paradigm. Then, it introduces a new inductive and transductive prompt learning (I&TPL) paradigm for MFDA. Based on the I&TPL paradigm, a test-time domain-agnostic meta-prompt learning (TDMP) method is further proposed, which is suitable for few-shot annotated multi-source domain data and is compatible with existing prompt tuning methods. As a result, the proposed TDMP does not require multiple complex prompts, constructed source-target pairs, extra auxiliary loss, and pseudo-target labels. Specifically, the proposed TDMP includes domain-agnostic meta-prompt learning and test-time domain-agnostic prompt tuning for target domain adaptation. The first stage is mainly optimized based on the Reptile optimization algorithm. Domain mixup is used to expand the diversity and the number of meta-training tasks. In the second stage, the learned domain-agnostic meta-prompt initializes the test-time prompt to further adapt to the target domain. Extensive experiments are conducted on the OfficeHome, DomainNet, Office, and TerraIncognita datasets of MFDA, achieving better performance with fewer learnable parameters and demonstrating the effectiveness of TDMP.