Enhancing Information Maximization With Distance-Aware Contrastive Learning for Source-Free Cross-Domain Few-Shot...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Existing Cross-Domain Few-Shot Learning (CDFSL) methods necessitate access to source domain data for model pre-training.
• Growing concerns regarding data privacy, transmission costs, and training expenses necessitate CDFSL solutions that do not require source data access.

Purpose of the Study:

• To address the novel Source-Free CDFSL (SF-CDFSL) problem, enabling CDFSL using pre-trained models without accessing source data.
• To overcome challenges in SF-CDFSL, including limited labeled target samples and the inability to align source and target domain distributions.

Main Methods:

• Proposes Enhanced Information Maximization with Distance-Aware Contrastive Learning (IM-DCL) for SF-CDFSL.
• Introduces a transductive mechanism for query set learning and Information Maximization (IM) for fitting target data distributions.
• Develops Distance-Aware Contrastive Learning (DCL) with weighted distance calculations for soft classification of positive and negative feature sets, addressing IM's limitations.

Main Results:

• IM-DCL demonstrates superior performance over existing methods on the BSCD-FSL benchmark, particularly for distant domain tasks.
• Evaluations on four datasets confirm IM-DCL's effectiveness in handling SF-CDFSL without source data access.
• Ablation studies validate the contributions of IM and DCL components to the overall performance.

Conclusions:

• The proposed IM-DCL method effectively solves the SF-CDFSL problem by leveraging pre-trained models and novel learning strategies.
• SF-CDFSL is achievable without source data, offering a privacy-preserving and cost-efficient alternative to traditional CDFSL.
• The method shows significant potential for real-world applications where source data is inaccessible or sensitive.