Transferability-Guided Cross-Domain Cross-Task Transfer Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Transfer learning aims to leverage knowledge from source tasks to improve performance on target tasks.
• Evaluating the effectiveness of transfer learning, especially across different domains and tasks, remains a challenge.
• Existing methods often require computationally expensive auxiliary tasks for transferability assessment.

Purpose of the Study:

• To propose novel, auxiliary-free transferability metrics for cross-domain and cross-task transfer learning.
• To develop metrics that efficiently estimate the benefit of a source model for a target task.
• To learn more generalizable representations for improved transfer learning performance.

Main Methods:

• Introduced fast optimal transport-based conditional entropy (F-OTCE) using optimal transport (OT) between source and target distributions.
• Developed joint correspondence OTCE (JC-OTCE) by incorporating label distances into the OT problem for enhanced accuracy.
• Utilized these metrics as objective functions for model finetuning and domain generalization (DG).

Main Results:

• F-OTCE and JC-OTCE significantly outperform state-of-the-art auxiliary-free metrics in correlation with transfer accuracy (21.1% and 25.8% higher, respectively).
• The proposed metrics drastically reduce computation time, from 43 minutes to under 11 seconds per task pair.
• F-OTCE demonstrated notable accuracy gains in few-shot classification for finetuning (up to 4.41%) and DG (up to 2.34%).

Conclusions:

• F-OTCE and JC-OTCE offer efficient and effective solutions for evaluating and enhancing transfer learning.
• The auxiliary-free nature of these metrics makes them highly practical for real-world applications.
• These metrics pave the way for more generalizable representations and improved performance in cross-domain and cross-task learning scenarios.