Dataset distillation (DD) has emerged as a widely adopted technique for crafting a synthetic dataset that captures the essential information of a training dataset, facilitating the training of accurate neural models. Its applications span various domains, including transfer learning, federated learning, and neural architecture search. The most popular methods for constructing the synthetic data rely on matching the convergence properties of training the model with the synthetic dataset and the training dataset. However, using the empirical loss as the criterion must be thought of as auxiliary in the same sense that the training set is an approximate substitute for the population distribution, and the latter is the data of interest. Yet despite its popularity, an aspect that remains unexplored is the relationship of DD to its generalization, particularly across uncommon subgroups. That is, how can we ensure that a model trained on the synthetic dataset performs well when faced with samples from regions with low population density? Here, the representativeness and coverage of the dataset become salient over the guaranteed training error at inference. Drawing inspiration from distributionally robust optimization, we introduce an algorithm that combines clustering with the minimization of a risk measure on the loss to conduct DD. We provide a theoretical rationale for our approach and demonstrate its effective generalization and robustness across subgroups through numerical experiments.

Visible–infrared person re-identification (VI-ReID) focuses on accurately matching individuals across different imaging modalities. Existing studies focus on generating modality-consistent images at the pixel level through the use of generative adversarial networks (GANs) to mitigate the impact of modality discrepancies. However, these methods face significant challenges in overcoming the limitation that synthesized samples from different modalities may suffer from semantic distortion. In this work, we propose an online one-stage style fuzzy normalization (SFN) method to generate modality-fuzzy features in the latent space while regularizing the model’s predictions. Specifically, SFN adaptively mixes the feature statistics of two random modality instances of the same identity in a single forward pass during training. In this process, to enhance the richness of modality interaction information, we design a novel causality balance loss, which enforces the generated fuzzy features to be independent of their initial modality while simultaneously encouraging them to align more closely with the other modality. Furthermore, we introduce an identity-aware consistency loss to regularize the predictions between the original and SFN-generated features to ensure semantic consistency. In contrast to prior work, SFN is a plug-and-play module that does not rely on any generative-based models, making it highly adaptable to various network architectures. Extensive experiments were performed on three public cross-modality datasets to ensure fair and reliable comparisons. The empirical results demonstrate the clear superiority of our method over previous state-of-the-art methods.