Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DICS: Find Domain-Invariant and Class-Specific Features for Out-of-Distribution Generalization

Qiaowei Miao, Yawei Luo, Yi Yang

TL;DR

This paper proposes a DICS model to extract Domain-Invariant and Class-Specific features, including Domain Invariance Testing (DIT) and Class Specificity Testing (CST), which mitigate the effects of spurious correlations introduced by confounders.

Abstract

While deep neural networks have made remarkable progress in various vision tasks, their performance typically deteriorates when tested in out-of-distribution (OOD) scenarios. Many OOD methods focus on extracting domain-invariant features but neglect whether these features are unique to each class. Even if some features are domain-invariant, they cannot serve as key classification criteria if shared across different classes. In OOD tasks, both domain-related and class-shared features act as confounders that hinder generalization. In this paper, we propose a DICS model to extract Domain-Invariant and Class-Specific features, including Domain Invariance Testing (DIT) and Class Specificity Testing (CST), which mitigate the effects of spurious correlations introduced by confounders. DIT learns domain-related features of each source domain and removes them from inputs to isolate domain-invariant class-related features. DIT ensures domain invariance by aligning same-class features across different domains. Then, CST calculates soft labels for those features by comparing them with features learned in previous steps. We optimize the cross-entropy between the soft labels and their true labels, which enhances same-class similarity and different-class distinctiveness, thereby reinforcing class specificity. Extensive experiments on widely-used benchmarks demonstrate the effectiveness of our proposed algorithm. Additional visualizations further demonstrate that DICS effectively identifies the key features of each class in target domains.

DICS: Find Domain-Invariant and Class-Specific Features for Out-of-Distribution Generalization

TL;DR

This paper proposes a DICS model to extract Domain-Invariant and Class-Specific features, including Domain Invariance Testing (DIT) and Class Specificity Testing (CST), which mitigate the effects of spurious correlations introduced by confounders.

Abstract

While deep neural networks have made remarkable progress in various vision tasks, their performance typically deteriorates when tested in out-of-distribution (OOD) scenarios. Many OOD methods focus on extracting domain-invariant features but neglect whether these features are unique to each class. Even if some features are domain-invariant, they cannot serve as key classification criteria if shared across different classes. In OOD tasks, both domain-related and class-shared features act as confounders that hinder generalization. In this paper, we propose a DICS model to extract Domain-Invariant and Class-Specific features, including Domain Invariance Testing (DIT) and Class Specificity Testing (CST), which mitigate the effects of spurious correlations introduced by confounders. DIT learns domain-related features of each source domain and removes them from inputs to isolate domain-invariant class-related features. DIT ensures domain invariance by aligning same-class features across different domains. Then, CST calculates soft labels for those features by comparing them with features learned in previous steps. We optimize the cross-entropy between the soft labels and their true labels, which enhances same-class similarity and different-class distinctiveness, thereby reinforcing class specificity. Extensive experiments on widely-used benchmarks demonstrate the effectiveness of our proposed algorithm. Additional visualizations further demonstrate that DICS effectively identifies the key features of each class in target domains.
Paper Structure (9 sections, 7 equations, 4 figures, 3 tables)

This paper contains 9 sections, 7 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Method
  4. Preliminary
  5. Domain Invariance Testing
  6. Class Specificity Testing
  7. Optimization
  8. Experiments
  9. Conclusion

Figures (4)

  • Figure 1: Domain-related and class-shared features are confounders that undermine models' out-of-distribution generalization. A model might mistakenly classify an image from the target domain as a dog due to similar textures or because its sketch style more closely resembles art than photographic styles. Our goal is to eliminate the influence of these confounders and identify domain-invariant, class-specific features that genuinely define each class, such as the giraffe's long neck.
  • Figure 2: SCM of OOD task.
  • Figure 3: DICS model includes domain invariance testing and class specificity testing. The former maximizes the similarity of class features across different source domains to ensure domain invariance. The latter forces the current input's class features to be closer to those of the same class and farther from those of other classes to maintain class specificity.
  • Figure 4: Visual results of RDM and DICS on PACS. (a) The background serves as a domain-specific confounder. (b) The fence in front of the horse is a feature shared with the "guitar" class, which confuses RDM. (c) DICS focuses more on the elephant's trunk, the key to identifying elephants. (d) DICS recognizes the giraffe's long neck as the basis for prediction.