베이지안 테스트 시간 적응을 위한 인과적 특징 분해 = Causal Feature Decomposition for Bayesian Test-Time Adaptation

다국어 초록 (Multilingual Abstract)

Test-time adaptation(TTA) aims to stably adapt a trained model without labels to maintain generalization performance on unseen data distributions. However, existing TTA methods are designed to trust predictions that are not guaranteed to be ground-truth labels, accumulating errors that use the model's own misjudgments as adaptation signals. We attribute this instability to the failure to distinguish causal features from non-causal nuisance factors during adaptation. The proposed method decomposes features into content and style, adapts only content features under a causal model, and guarantees identifiability through label sufficiency, domain invariance, content-style de-correlation, and information compression. During testing, the model selectively performs Bayesian updates only to content features, while style features remain fixed. We further employ a calibration-aware control mechanism that regulates update strength to prevent model drift. Experimental results on classification and segmentation tasks show that the proposed method provides consistent improvements over existing TTA methods in terms of both predictive performance and calibration robustness. This study demonstrates that selective Bayesian adaptation via causal feature decomposition at test time enables stable and reliable model adaptation under distribution shifts.

목차 (Table of Contents)

• 제1장 Introduction 5
• 제1절 Domain Shift 5
• 제2절 Approaches to Address Domain Shift 6
• 1 Domain Adaptation 6
• 2 Domain Generalization 7
• 3 Test-Time Adaptation 8
• 제3절 Test-Time Adaptation 9
• 1 Entropy Minimization 9
• 2 Normalization Adaptation 10
• 3 Bayesian TTA 12
• 4 Limitations 12
• 제4절 Bayesian Modeling 13
• 1 Uncertainty Modeling 13
• 2 Bayesian Probability 14
• 3 Bayesian Theorem 15
• 제5절 Objectives 15
• 1 Identifiable Causal Feature Decomposition 16
• 2 Selective Bayesian Adaptation 17
• 3 Contribution 18
• 제6절 Overview 18
• 제2장 Problem Setup and Assumptions 19
• 제1절 Problem Statement 19
• 제2절 Structural Causal Model 21
• 제3절 Proposed Alternatives 23
• 제3장 Methodology 25
• 제1절 Representation Decomposition 25
• 제2절 Selective Bayesian Adaptation 28
• 제3절 Test-Time Updates 30
• 제4절 Calibration-Aware Control 33
• 제4장 Theoretical Analysis 35
• 제1절 Risk Decomposition under the SCM 35
• 제2절 Bayesian Optimality of NIW Updates 36
• 제3절 Stability of Online Updates 36
• 제4절 Calibration Control 37
• 제5장 Experiments 38
• 제1절 Experimental Setting 38
• 1 Datasets 38
• 2 Implementation Details 40
• 3 Comparison Methods 40
• 4 Evaluation Metrics 41
• 제2절 Main Results 42
• 1 Classification Results 42
• 2 Segmentation Results 44
• 제6장 Conclusion 46