검색 상세

From Cross-Domain to Cross-Model: Invariant Anchors for Reliable Adaptation under Distribution Shift

  • 김태훈 (Taehoon Kim, 아주대학교 일반대학원)

초록/요약

Modern machine learning rests on the assumption that training and deployment data share a distribution, yet in practice this premise fails: a detector trained on clear daytime scenes meets fog and night, and a classifier trusted in one setting meets corruption in another. Under such shift a model’s confident predictions become an unreliable guide to its own adaptation, because the largest learning weight falls on the most confident errors, and self-training on those errors only reinforces them. This dissertation develops a single principle: recover the target’s invariant identity and discount the nuisance, rather than imitate the model’s nuisance-corrupted output. It shows that this principle organizes into two axes of invariance, instantiated by three independently developed methods. On the first axis the nuisance is the domain and the invariant anchor is grounded in a labeled source; on the second the nuisance is a single model’s bias and the anchor is recovered from a complementary model, with two studies developing the first axis and one the second. The first study locates the cross-domain invariant in the object category: instead of trusting the detector’s own confidence on an unlabeled target region, it grounds each target region’s label in the most similar labeled source object, yielding more reliable supervision than confidence-based self-training across four different adaptation scenarios. The second study keeps the cross-domain axis but moves the invariant to a structural regularity—the proportion and arrangement of foreground and background a detector produces stays nearly constant across domains—and exploits it to supervise the unlabeled target, achieving detection competitive with the state of the art across seven adaptation benchmarks. The third study turns to the cross-model axis, where the source data and its labels are gone and a vision–language model must adapt online to corrupted images. Because no source-grounded anchor can be built, it is recovered instead from a frozen self-supervised model whose bias differs from the adapting one and stays reliable precisely where it fails, and fused with it to disrupt the confirmation loop that defeats adaptation based on entropy minimization alone. This yields consistent gains over CLIP-based online test-time adaptation across standard corruption benchmarks, remaining competitive in all settings; furthermore, under the continual and mixed-domain protocols, it remains competitive with existing methods. Taken together, the three studies trace one idea through a progression of decreasing external support—from an explicitly labeled category, to a hand-identified structural statistic, to a learned representation in a second model—while the role of the anchor never changes: it is the quantity the nuisance cannot corrupt, against which the model’s raw output is reconciled, as the ablations and analyses in each study indicate.

more

목차

1 Introduction 1
1.1 The Gap Between Where a Model Is Trained and Where It Is Used 1
1.2 Why the Gap Cannot Be Closed by Trusting the Model's Outputs 2
1.3 Thesis Statement: Recover the Invariant Identity, Discount the Nuisance 3
1.4 Two Axes: Where the Invariant Anchor Is Sourced 5
1.5 Outline of Dissertation 6
2 Preliminaries 8
2.1 Overview 8
2.2 The Training–Inference Gap 8
2.3 Representation Versus Decision 9
2.4 Two Adaptation Regimes 10
2.5 The Invariant Anchor and the Two Axes 10
3 Categorical Invariance as an Anchor: Source-grounded Object Correspondence for Cross-Domain Detection 13
3.1 Overview 13
3.2 Background 15
3.3 Methodology 17
3.3.1 Overview 17
3.3.2 Target Surrogate Selection Module 17
3.3.3 PutMix 20
3.4 Experiments 21
3.4.1 Experimental Setups 21
3.4.2 Comparison Results 22
3.4.3 Ablation Studies 25
3.5 Discussion 27
4 Structural Invariance as an Anchor: Dual-Prototype Objectness Decoupling for Cross-Domain Detection 29
4.1 Overview 29
4.2 Background 31
4.3 Methodology 32
4.3.1 Overview 32
4.3.2 Categorical Moving Alignment 33
4.3.3 Regional Moving Alignment 34
4.3.4 Objectness Decoupling 36
4.3.5 Objectness Decoupled Loss 37
4.4 Experiments 38
4.4.1 Experimental Setups 39
4.4.2 Comparison Results 40
4.4.3 Ablation Studies 42
4.5 Discussion 45
5 Semantic Invariance as a Cross-Model Anchor: Bridging CLIP and DINO for Test-Time Adaptation 47
5.1 Overview 47
5.2 Background 49
5.3 Methodology 51
5.3.1 Overview 51
5.3.2 Representation level: Cross-Modal Logit Ensemble 53
5.3.3 Optimization level: Gated Test-time Information Maximization 54
5.3.4 Prediction level: Online Logit Adjustment 56
5.3.5 Inference pipeline 58
5.4 Experiments 58
5.4.1 Experimental Setups 58
5.4.2 Results on Test-Time Adaptation 59
5.4.3 Empirical Studies 61
5.5 Discussion 66
6 Conclusion 68
6.1 Summary 68
6.2 Objectives Revisited 69
6.3 Limitations and Scope 70
6.4 Future Directions 71
Bibliography 73

more
반출 Meta View 목록