Towards Accurate and Robust Domain Adaptation Under Multiple Noisy Environments.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Machine learning algorithms face distribution shifts in non-stationary environments.
• Existing domain adaptation methods lack robustness against various noises (label, feature, open-set) in source data.
• Noise significantly impacts algorithm performance and reliability.

Purpose of the Study:

• To develop a noise-robust domain adaptation technique.
• To theoretically analyze the impact of different noises on target risk.
• To enhance the reliability of machine learning models in noisy, shifting environments.

Main Methods:

• Proposed offline curriculum learning to minimize a new empirical source risk.
• Introduced proxy distribution-based margin discrepancy to reduce noisy distribution distance.
• Developed an energy estimator to identify and mitigate open-set noise.
• Implemented robust parameter learning for domain-invariant features.
• Integrated components into an adversarial network for joint optimization.

Main Results:

• Theoretical analysis revealed distinct impacts of noise types on expected target risk.
• The proposed method significantly reduced the influence of source domain noises.
• Open-set noise was effectively handled using the energy estimator.
• Achieved over 10% accuracy improvement in some transfer tasks on benchmark datasets.
• Demonstrated superior performance on a COVID-19 screening task.

Conclusions:

• The developed noise-robust domain adaptation method effectively handles multiple noise types.
• The approach enhances algorithm robustness and accuracy in challenging environments.
• The method shows significant potential for real-world applications, including medical screening.