Selection, Ensemble, and Adaptation: Advancing Multi-Source-Free Domain Adaptation via Architecture Zoo

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional Multi-Source Free Domain Adaptation (MSFDA) relies on uniform source models.
• This limits adaptability when source domains offer diverse model architectures.

Purpose of the Study:

• To introduce Zoo-MSFDA, a generalized setting accommodating multiple diverse source models per domain.
• To address the challenge of suboptimal models in Zoo-MSFDA.
• To propose a novel method for transferability estimation.

Main Methods:

• Theoretical analysis of model selection using transferability and diversity principles.
• Development of Source-Free Unsupervised Transferability Estimation (SUTE) for cross-architecture transferability assessment.
• Introduction of a Selection, Ensemble, and Adaptation (SEA) framework.

Main Results:

• The SEA framework significantly enhances adaptation performance in 2D image classification.
• SUTE demonstrates state-of-the-art performance in transferability estimation.
• Zoo-MSFDA setting proves effective with diverse source models.

Conclusions:

• Zoo-MSFDA offers a more general and effective approach to multi-source domain adaptation.
• SUTE provides a robust solution for evaluating model transferability without source data or target labels.
• The SEA framework successfully leverages diverse source models for improved adaptation.