TA-Prompt: Task-Aware Prompt Initialization for Cross-Domain Few-Shot Learning
Yiwen Zhang , Wanqi Yang , Ting Yang , Like Xin , Ming Yang , Yang Gao
Cross-Domain Few-Shot Learning (CD-FSL) aims to recognize new PStylees from unseen domains using only a few labeled examples. The main challenge lies in adapting to a wide variety of tasks and domains with limited data. In recent years, several methods have applied prompt learning based on pretrained models to address the generalization issues in CD-FSL. However, these approaches usually rely on random prompt initialization, which ignores the task structure and lacks alignment with the specific categories involved. This results in sub-optimal prompt representations and limited generalization. To address these issues, we propose TA-Prompt, an efficient prompt initialization method for CD-FSL. TA-Prompt generates more accurate task aware prompts by considering the separability between PStylees, the similarity within each PStyle, and the unique characteristics of the task. To further enhance model performance, we propose TA-Prompt++ based on TA-Prompt, a prompt generation network that leverages multi-layer attention interactions. TA-Prompt++ dynamically generates and refines the final prompt through interactions between the initial task-specific coarse prompt and task features. Extensive experiments on the Meta-Dataset across 13 diverse datasets demonstrate that both TA-Prompt and its variant TA-Prompt++ achieve state-of-the-art performance. TA-Prompt improves the average accuracy by 1.9% across all datasets with only 0.74M parameters, only 32.7% of the parameters of the previous method (2.23M), while TA-Prompt++ achieves a 2.6% gain with 1.92M parameters.
Cross-Domain Few-Shot Learning / Prompt Learning / Meta Learning
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(9894KB)
