E²MPL: An Enduring and Efficient Meta Prompt Learning Framework for Few-Shot Unsupervised Domain Adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot unsupervised domain adaptation (FS-UDA) aims to improve model performance on unlabeled target domains using limited source domain data.
• Existing FS-UDA methods struggle with model instability and high time complexity during adaptation to new tasks.

Purpose of the Study:

• To propose a novel framework, Enduring and Efficient Meta-Prompt Learning (E2MPL), to address the instability and time-consuming nature of current FS-UDA approaches.
• To enhance the generalization capabilities and efficiency of models in few-shot unsupervised domain adaptation scenarios.

Main Methods:

• Utilizes the pre-trained CLIP model as a backbone for feature learning.
• Designs domain-shared prompts with virtual tokens to capture meta-knowledge and mitigate domain gaps.
• Employs a task prompt learning network for adaptive, task-specific prompt generation, enabling fast and stable generalization.
• Formulates meta-prompt learning as a bilevel optimization problem with a closed-form solution for efficient, single-step adaptation.

Main Results:

• Achieved significant improvements on the DomainNet benchmark dataset for FS-UDA.
• In 5-way 1-shot tasks, E2MPL improved average accuracy by at least 15 percentage points and reduced adaptation time by 64.67%.
• In 5-way 5-shot tasks, E2MPL improved average accuracy by at least 9 percentage points and reduced adaptation time by 63.18%.
• Demonstrated enhanced stability, reducing average IQR values by over 40.80% (1-shot) and 25.35% (5-shot).

Conclusions:

• E2MPL offers a promising solution for stable and efficient few-shot unsupervised domain adaptation.
• The framework effectively mitigates domain gaps and achieves superior performance compared to state-of-the-art methods.
• E2MPL's bilevel optimization and efficient prompt learning enable rapid and robust adaptation to new tasks.