Uncertainty-Calibrated Test-Time Model Adaptation Without Forgetting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Test-time adaptation (TTA) is crucial for models facing changing data distributions.
• Existing TTA methods incur high computational costs and suffer from catastrophic forgetting, degrading performance on original data.
• Current TTA approaches often lead to overconfident predictions, underestimating data uncertainty.

Purpose of the Study:

• To develop efficient and effective TTA methods that minimize computational cost and mitigate catastrophic forgetting.
• To introduce calibrated TTA that accurately reflects model and data uncertainty.
• To improve model robustness and performance in dynamic testing environments.

Main Methods:

• Proposed Efficient Anti-Forgetting Test-Time Adaptation (EATA) with active sample selection and Fisher regularization.
• Introduced EATA with Calibration (EATA-C) to address overconfident predictions by separating model and data uncertainty.
• EATA-C utilizes network divergence for model uncertainty and prediction disagreement for data uncertainty, employing divergence minimization and min-max entropy regularization.

Main Results:

• EATA significantly reduces optimization costs compared to prior methods.
• EATA-C effectively calibrates predictions by addressing both model and data uncertainty.
• Both methods demonstrate strong performance in image classification and semantic segmentation tasks, outperforming existing TTA solutions.

Conclusions:

• EATA and EATA-C offer efficient and robust solutions for test-time adaptation.
• Calibrated TTA (EATA-C) improves prediction reliability by accurately quantifying uncertainty.
• The proposed methods enhance model performance and generalization in the presence of distribution shifts.