Evidential Deep Learning for Open-Set Active Domain Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Open-set domain adaptation (OSDA) addresses knowledge transfer challenges with novel classes in target domains.
• Existing OSDA methods often neglect prediction uncertainty and incur significant training costs.
• Evidential Deep Learning (EDL) models uncertainty using Dirichlet distributions, moving beyond standard softmax outputs.

Purpose of the Study:

• To propose an efficient OSDA method that accounts for prediction uncertainty and minimizes annotation overhead.
• Introduce Evidential Deep Learning for Open-Set Active Domain Adaptation (EOSADA).
• Enhance model performance in OSDA by effectively selecting informative target domain samples.

Main Methods:

• Implemented EDL to create an open-set classifier by replacing softmax with Dirichlet distributions.
• Developed a two-round sample selection strategy based on target domain data uncertainty and semantic similarity.
• Focused on balancing the selection of known and novel classes for informative sample acquisition.

Main Results:

• The proposed EOSADA method demonstrated superior performance in OSDA scenarios.
• The strategy effectively identified informative samples from both known and novel classes.
• Achieved significant performance gains without altering the underlying model architecture.

Conclusions:

• EOSADA offers a robust and efficient solution for open-set domain adaptation.
• Leveraging EDL and strategic sample selection addresses key limitations of traditional OSDA methods.
• The approach is effective under limited annotation budgets, maximizing performance in complex domain adaptation tasks.