Domain Generalization

 EFFICIENTLY ASSEMBLE NORMALIZATION LAYERS AND REGULARIZATION FOR FEDERATED DOMAIN GENERALIZATION - CVPR 2024

BENCHMARKING ALGORITHMS FOR FEDERATED DOMAIN GENERALIZATION - ICLR 2024

• Setting: Domain generalization where domain data are not shared between the domains due to privacy concerns.
• Challenge: The scale of the no of clients in the federated setting may be too high in addition to high data diversity between the clients or domains.
• Proposal: Introduce a novel data partition scheme that allows us to distribute domain datasets among a few or many clients while controlling client heterogeneity.

 IMPROVING DOMAIN GENERALIZATION WITH DOMAIN RELATIONS - ICLR 2024

• Setting: Domain generalization - multiple source domains, the model should do well on unseen data
• Proposal: Utilizes domain similarities based on domain metadata to learn domain-specific models. A pairwise similarity matrix is used 

MIXUP-INDUCED DOMAIN EXTRAPOLATION FOR DOMAIN GENERALIZATION - AAAI 2024 

• Setting: Domain generalization with the assumption that domains are sampled from a meta-task
• Research Gap: The number of domains and the radius of the Wasserstein ball centered on the target domain were not considered in deriving the generalization bound earlier. This realization prompts the collection of adequate domains through a reverse Mixup scheme to generate extrapolated domains.
• Solution: 
• Mixing the statistics from multiple source domains (mean, std deviation)
• Expand the interpolation space to increase the intersections with the domains sampling from the target environment
• Historical statistical information of domains used through a moving average weight over the iterations.
• An extrapolated domain ($\mathcal{D}^E_{n}$) can contribute to one of the interpolated domains as follows:
• 
• The distribution of this extrapolated domain can be inferred through Gaussian distribution
• AdaIN is used to generate samples of new domains from the distribution statistics

TOWARDS DOMAIN-AWARE KNOWLEDGE DISTILLATION FOR CONTINUAL MODEL GENERALIZATION - WACV 2024

• Setting: Addressing catastrophic forgetting in domain generalization models

 AN ITERATIVE SELF-LEARNING FRAMEWORK FOR MEDICAL DOMAIN GENERALIZATION - NeurIPS 2023

• Setting: Domain generalization where domain categorizations are unknown and exhibit distinct characteristics.
• Challenge: 
• Difficult to obtain domain categorization and even if there are such categorizations, they may be too broad to be useful.
• Training a single model that generalizes across all domains can lead to suboptimal results when they posses different distinct characteristics.
• Solution: A self-learning framework that discovers decoupled domains and trains personalized classifiers for each decoupled domain.

PROMPTSTYLER: PROMPT-DRIVEN STYLE GENERATION FOR SOURCE-FREE DOMAIN GENERALIZATION - ICCV 2023

• Setting: