Learning to Diversify for Single Domain Generalization

Introduction

• Domain shifts such as changes in illumination, appearance, or background can degrade source-trained models; domain generalization targets unseen domains, unlike adaptation to unlabeled targets.DG methods broadly use alignment across source domains or augmentation through transformations and pseudo-novel samples.Source: §1 · p. 1
• Single-DG trains with only one source domain, limiting access to domain identifiers; prior adversarial-noise methods [37] [50] produce images insufficiently diverse to cover real target distributions.Without multiple sources or domain information, alignment- and augmentation-based DG methods cannot readily identify robust features or transformations.Source: §1 · p. 1
• L2D’s style-complement module synthesizes unseen-style samples in latent feature space and alternates their generation with discriminative, style-invariant representation learning.The generated samples challenge the task model and are intended to improve its resilience to target shifts.Source: §1 · p. 2
• L2D minimizes a tractable mutual-information upper bound between generated and source samples, then maximizes mutual information between samples in the same semantic category.This iterative min-max strategy diversifies generated samples while bringing same-category samples close in latent space.Source: §1 · p. 2
• Experiments on digits recognition, corrupted CIFAR-10, and PACS report that L2D surpasses state-of-the-art DG and Single-DG methods on all datasets.The paper also evaluates standard DG under a leave-one-domain-out protocol.Source: §1 · p. 2

Domain Shift

• Domain-adaptation methods reduce domain shift by matching marginal [26] [13] or conditional distributions [27] [29], often using partially labelled or unlabelled target data [38] [26] [28].Few-shot domain adaptation uses a small number of labelled target samples alongside source samples [32].Source: §2 · p. 2

Domain Generalization

• Domain generalization does not use target-domain data during training and commonly learns domain-invariant representations or augments data.Invariant-representation approaches include kernel-based features [34], semantic alignment [33], and meta-learning methods [10] [24] [11].Source: §2 · p. 2

Domain Generalization: Data Augmentation

• Domain-generalization augmentation generates out-of-domain samples for training, including adversarial hard examples [42], domain-change perturbations [39], pseudo-novel-domain synthesis [51], and jigsaw self-supervision [5].Source: §2 · pp. 2–3

Single-Domain Generalization

• Single-DG trains on one source domain and evaluates on unseen domains; this paper aims to generate samples with novel styles, textures, and appearances that complement the source distribution [37] [50].Prior Single-DG methods use gradient-based augmentation to encourage semantic consistency [37] or generate challenging perturbations through entropy maximization [50]; their visual changes mostly concern color and texture.Source: §2 · p. 3

Style-Complement Module

• The generator uses K convolution–style-learning–transposed-convolution transformations with learnable mean and variance shifts, then combines their reconstructions using normally sampled weights.The shifts are applied at the pixel level, and the module produces images that retain the source image’s semantic information while changing its style.Source: §3 · pp. 3–4

Objective Function

• Training alternates between minimizing source/generated mutual information to diversify styles and freezing the generator while training the task network to maximize same-class mutual information.Source: §3 · pp. 3, 5

Synthesizing Novel Styles

• To decorrelate source and generated latent features, the method minimizes a variational approximation to the MI upper bound, using qθ(z+|z) because p(z+|z) is intractable [7].The approximation is considered reliable when the variational and conditional distributions are close; optimization uses negative log-likelihood rather than directly minimizing the bound gap.Source: §3 · p. 4

Semantic Consistency

• Class-conditional maximum mean discrepancy constrains latent distribution shifts among samples of the same class to mitigate semantic distortion in generated images.Source: §3 · p. 4

MI Maximization

• The task network increases same-class mutual information with supervised contrastive loss and applies cross-entropy to predictions on both source and generated images.Source: §3 · pp. 4–5

4.2. Comparisons on Corrupted CIFAR-10

• Approximately 2.6% above the best baseline, L2D achieves the highest average accuracy across level-5 CIFAR-10-C corruption categories.The evaluation averages accuracy across Weather, Blur, Noise, and Digital corruptions.Source: Table 2 · p. 6, §4.2 · p. 6

4.1. Comparisons on Digits

• 14.9% improvement in overall accuracy accompanies improvements of 20.3% on SVHN, 29.36% on MNIST-M, and 13.33% on SYN.L2D achieves the second-best performance on USPS; the paper suggests the small style shift from MNIST may limit the benefit of diverse generated images.Source: §4.1 · p. 6

4.3. Comparisons on PACS

• L2D achieves the best average accuracy on PACS when trained on photo and evaluated on art painting, cartoon, and sketch.It has a relatively large margin on sketch, which has the largest domain shift from photo in this evaluation.Source: Table 3 · p. 6, §4.3 · p. 6

4.2. Comparisons on Corrupted CIFAR-10

• 13.2% improvement on noise corruptions accompanies a performance margin that grows with corruption severity, supporting L2D's resilience to severe corruptions.The margin over other methods is relatively small at severity level 1 and gradually enlarges at higher levels.Source: §4.2 · p. 6, Figure 3 · p. 7

4.3. Comparisons on PACS

• L2D achieves state-of-the-art leave-one-domain-out generalization on PACS without using domain identifiers during training.The image augmentation module gradually enlarges the domain shift between generated and source images.Source: Table 4 · p. 7, §4.3 · p. 7

4.4. Empirical Analysis

• 1.35%, 1.26%, and 3.76% lower accuracy follows removing style modification, MI minimization, and MI maximization, respectively.Removing the style-complement module also produces a large performance margin below the full model; MI minimization constrains generated images to diversify from the source, while MI maximization brings same-class diverse images closer in the embedding space.Source: §4.4 · pp. 7–8
• On unseen Digits target domains, L2D produces tighter, better class-wise-separated feature clusters than the baselines.The visualized features are from 1,000 randomly selected samples per target domain; ERM, ADA, and ME-ADA show multiple sub-clusters within classes.Source: Figure 4 · p. 8, §4.4 · p. 8
• Average performance rises as transformations increase and stabilizes after k = 5; excessive transformations may add noise, while α1 = α2 = 1 yields the best accuracy.The method surpasses state-of-the-art performance across a wide range of α settings; the sensitivity analysis is on PACS under Single-DG.Source: Figure 2 · p. 7, §4.4 · p. 8

• Across three benchmark datasets, L2D outperforms state-of-the-art methods on both single-domain and standard leave-one-domain-out domain generalization.Source: §5 · p. 8
• L2D uses a style-complement module to generate diverse out-of-domain images from a single source domain.Source: §5 · p. 8
• Its iterative min-max mutual-information strategy minimizes a tractable upper bound for generated-image diversity and maximizes MI among same-category samples for style-invariant representations.Source: §5 · p. 8