Step-Wise Distribution-Aligned Style Prompt Tuning for Source-Free Cross-Domain Few-Shot Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Cross-domain few-shot learning (CDFSL) methods struggle with large pre-trained models (LMs) due to inaccessible source data and training strategies.
• Fine-tuning LMs for CDFSL is computationally expensive, limiting practical applications.

Purpose of the Study:

• Investigate the source-free CDFSL (SF-CDFSL) problem, enabling few-shot learning (FSL) in target domains using only pre-trained models and limited target samples.
• Address the challenge of implicitly narrowing domain gaps without access to source data.

Main Methods:

• Propose Step-wise Distribution-aligned Style Prompt Tuning (StepSPT), a novel approach for SF-CDFSL.
• Utilize a style prompt to adjust target samples towards an expected distribution.
• Employ a dual-phase optimization process: an external process for step-wise distribution alignment of the style prompt and an internal process for classifier updates.

Main Results:

• StepSPT demonstrates superiority over existing prompt tuning methods and state-of-the-art approaches on five datasets.
• Ablation studies confirm the efficacy of the proposed StepSPT method.
• Performance analyses highlight the effectiveness of the distribution optimization strategy.

Conclusions:

• StepSPT offers a practical and effective solution for SF-CDFSL, particularly for large pre-trained models.
• The method implicitly reduces domain gaps by optimizing prediction distributions, overcoming limitations of source-free scenarios.
• StepSPT advances the field of few-shot learning by enabling efficient adaptation of LMs to new domains.