검색 상세

Optimizing Domain-Wise Thresholds for Multi-Domain Semi-Supervised Learning

  • 곽태우 (아주대학교 일반대학원)

초록/요약

Semi-Supervised Domain Generalization (SSDG) aims to improve generalization when the source domains contain few labeled samples and abundant unlabeled data. Most SSDG methods leverage unlabeled data via pseudo-labeling, a technique commonly used in semi- supervised learning (SSL). This work focuses on the role and limitations of pseudo-labeling in SSDG. As in SSL, SSDG exhibits an inherent trade-off between pseudo-label accuracy and coverage. Raising the confidence threshold increases the accuracy of pseudo-labels but reduces the coverage of unlabeled data; lowering the threshold incorporates more unlabeled samples at the cost of increased noise. To quantify and optimize this trade-off, we propose a function G that combines the unsupervised loss on selected samples with a penalty for unse- lected samples, enabling data-dependent threshold selection. Building on G, we theoretically show that applying domain-specific thresholds is more effective than using a single global threshold in SSDG. By accounting for domain-specific distributions, this approach achieves a more favorable balance between pseudo-label quality and coverage. We validate the effec- tiveness of our method on standard SSDG benchmarks, including PACS and Office-Home. Keywords: Deep learning, Semi-supervised learning, Domain Generalization, Domain-Wise, Pseudo-labeling.

more

목차

1 Introduction 1
1.1 Introduction 1
2 Related Work 4
2.1 Semi-Supervised Learning 4
2.2 Domain Generalization 4
2.3 Semi-Supervised Domain Generalization 5
3 Proposed Method 7
3.1 Preliminary 7
3.2 Theoretical Motivation 8
3.3 Domain-Wise Thresholds 12
4 Experiments 15
4.1 Experimental settings 15
4.1.1 Datasets 15
4.1.2 Protocol and Metrics 15
4.1.3 Implementation detail 16
4.2 Main results 16
4.3 Ablation study 16
4.3.1 Domain-wise variation in threshold values 16
4.3.2 Accuracy-Coverage Analysis 17
5 Conclusion 20
Bibliography 21

more
반출 Meta View 목록