[논문 리뷰] Exploiting Domain-Specific Features to Enhance Domain Generalization

Manh-Ha Bui, Toan Tran외 2명|arXiv (Cornell University)|2021. 10. 18.

Domain Adaptation and Few-Shot Learning인용 수 69

한 줄 요약

The paper introduces meta-Domain Specific-Domain Invariant (mDSDI), a framework that disentangles domain-invariant and domain-specific features and uses meta-learning to adapt domain-specific information to unseen domains, achieving competitive DG results across benchmarks.

ABSTRACT

Domain Generalization (DG) aims to train a model, from multiple observed source domains, in order to perform well on unseen target domains. To obtain the generalization capability, prior DG approaches have focused on extracting domain-invariant information across sources to generalize on target domains, while useful domain-specific information which strongly correlates with labels in individual domains and the generalization to target domains is usually ignored. In this paper, we propose meta-Domain Specific-Domain Invariant (mDSDI) - a novel theoretically sound framework that extends beyond the invariance view to further capture the usefulness of domain-specific information. Our key insight is to disentangle features in the latent space while jointly learning both domain-invariant and domain-specific features in a unified framework. The domain-specific representation is optimized through the meta-learning framework to adapt from source domains, targeting a robust generalization on unseen domains. We empirically show that mDSDI provides competitive results with state-of-the-art techniques in DG. A further ablation study with our generated dataset, Background-Colored-MNIST, confirms the hypothesis that domain-specific is essential, leading to better results when compared with only using domain-invariant.

연구 동기 및 목표

• Motivate the limitations of solely learning domain-invariant representations for domain generalization.
• Propose a theoretically grounded framework that jointly learns domain-invariant and domain-specific features.
• Use meta-learning to adapt domain-specific information from source domains to unseen target domains.
• Disentangle representations to preserve useful information for robust generalization.
• Evaluate mDSDI on standard DG benchmarks and a crafted ablation dataset to verify the usefulness of domain-specific features.

제안 방법

• Define domain-invariant and domain-specific latent representations in a probabilistic framework.
• Use adversarial training to learn domain-invariant features across sources.
• Use a domain classifier to learn domain-specific features from each source domain.
• Enforce disentanglement of Z_I and Z_S by minimizing their covariance.
• Employ a meta-learning procedure to adapt domain-specific features from multiple source domains to unseen targets.
• Train end-to-end with a combined objective that integrates invariance, specificity, and meta-learning terms.

실험 결과

연구 질문

• RQ1 Can domain-specific information improve generalization beyond domain-invariant representations in DG?
• RQ2 Does disentangling domain-invariant and domain-specific features lead to better robustness on unseen domains?
• RQ3 Can meta-learning effectively adapt domain-specific information to unseen target domains without target data?
• RQ4 How does the proposed framework perform on standard DG benchmarks compared to state-of-the-art methods?

주요 결과

방법	CMNIST	RMNIST	VLCS	PACS	OfficeHome	TerraInc.	DomainNet	평균
ERM (25)	51.5 ± 0.1	98.0 ± 0.0	77.5 ± 0.4	85.5 ± 0.2	66.5 ± 0.3	46.1 ± 1.8	40.9 ± 0.1	66.6
IRM (18)	52.0 ± 0.1	97.7 ± 0.1	78.5 ± 0.5	83.5 ± 0.8	64.3 ± 2.2	47.6 ± 0.8	33.9 ± 2.8	65.4
GroupDRO (26)	52.1 ± 0.0	98.0 ± 0.0	76.7 ± 0.6	84.4 ± 0.8	66.0 ± 0.7	43.2 ± 1.1	33.3 ± 0.2	64.8
Mixup (34;35;36)	52.1 ± 0.2	98.0 ± 0.1	77.4 ± 0.6	84.6 ± 0.6	68.1 ± 0.3	47.9 ± 0.8	39.2 ± 0.1	66.7
MLDG (11)	51.5 ± 0.1	97.9 ± 0.0	77.2 ± 0.4	84.9 ± 1.0	66.8 ± 0.6	47.7 ± 0.9	41.2 ± 0.1	66.7
CORAL (29)	51.5 ± 0.1	98.0 ± 0.1	78.8 ± 0.6	86.2 ± 0.3	68.7 ± 0.3	47.6 ± 1.0	41.5 ± 0.1	67.5
MMD (30)	51.5 ± 0.2	97.9 ± 0.0	77.5 ± 0.9	84.6 ± 0.5	66.3 ± 0.1	42.2 ± 1.6	23.4 ± 9.5	63.3
DANN (31)	51.5 ± 0.3	97.8 ± 0.1	78.6 ± 0.4	83.6 ± 0.4	65.9 ± 0.6	46.7 ± 0.5	38.3 ± 0.1	66.1
CDANN (32)	51.7 ± 0.1	97.9 ± 0.1	77.5 ± 0.1	82.6 ± 0.9	65.8 ± 1.3	45.8 ± 1.6	38.3 ± 0.3	65.6
MTL (1;27)	51.4 ± 0.1	97.9 ± 0.0	77.2 ± 0.4	84.6 ± 0.5	66.4 ± 0.5	45.6 ± 1.2	40.6 ± 0.1	66.2
SagNets (37)	51.7 ± 0.0	98.0 ± 0.0	77.8 ± 0.5	86.3 ± 0.2	68.1 ± 0.1	48.6 ± 1.0	40.3 ± 0.1	67.2
ARM (28)	56.2 ± 0.2	98.2 ± 0.1	77.6 ± 0.3	85.1 ± 0.4	64.8 ± 0.3	45.5 ± 0.3	35.5 ± 0.2	66.1
VREx (33)	51.8 ± 0.1	97.9 ± 0.1	78.3 ± 0.2	84.9 ± 0.6	66.4 ± 0.6	46.4 ± 0.6	33.6 ± 2.9	65.6
RSC (38)	51.7 ± 0.2	97.6 ± 0.1	77.1 ± 0.5	85.2 ± 0.9	65.5 ± 0.9	46.6 ± 1.0	38.9 ± 0.5	66.1
mDSDI (Ours)	52.2 ± 0.2	98.0 ± 0.1	79.0 ± 0.3	86.2 ± 0.2	69.2 ± 0.4	48.1 ± 1.4	42.8 ± 0.1	67.9

• mDSDI achieves the highest average accuracy across 7 DG benchmarks compared to 14 baselines.
• mDSDI preserves domain-invariant information while also leveraging domain-specific features to improve performance in background-related scenarios.
• On VLCS and PACS, mDSDI shows notable gains due to effective use of domain-specific information.
• mDSDI attains strong results on large-scale DomainNet (42.8% in DomainNet) and Office-Home (69.2%).
• Ablation on Background-Colored-MNIST confirms domain-specific features contribute to improved classification when backgrounds provide label-related cues.