Enabling deep models to generalize in non-stationary environments is vital for real-world machine learning, as data distributions are often found to continually change. Recently, evolving domain generalization (EDG) has emerged to tackle the domain generalization in a time-varying system, where the domain gradually evolves over time in an underlying continuous structure. Nevertheless, it typically assumes multiple source domains simultaneously ready. It still remains an open problem to address EDG in the domain-incremental setting, where source domains are non-static and arrive sequentially to mimic the evolution of training domains. To this end, we propose Weight Diffusion (W-Diff), a novel framework that utilizes the conditional diffusion model in the parameter space to learn the evolving pattern of classifiers during the domain-incremental training process. Specifically, the diffusion model is conditioned on the classifier weights of different historical domain (regarded as a reference point) and the prototypes of current domain, to learn the evolution from the reference point to the classifier weights of current domain (regarded as the anchor point). In addition, a domain-shared feature encoder is learned by enforcing prediction consistency among multiple classifiers, so as to mitigate the overfitting problem and restrict the evolving pattern to be reflected in the classifier as much as possible. During inference, we adopt the ensemble manner based on a great number of target domain-customized classifiers, which are cheaply obtained via the conditional diffusion model, for robust prediction. Comprehensive experiments on both synthetic and real-world datasets show the superior generalization performance of W-Diff on unseen domains in the future.

Recent studies on Neural Collapse (NC) reveal that, under class-balanced conditions, the class feature means and classifier weights spontaneously align into a simplex equiangular tight frame (ETF). In long-tailed regimes, however, severe sample imbalance tends to prevent the emergence of the NC phenomenon, resulting in poor generalization performance. Current efforts predominantly seek to recover the ETF geometry by imposing constraints on features or classifier weights, yet overlook a critical problem: There is a pronounced misalignment between the feature and the classifier weight spaces. In this paper, we theoretically quantify the harm of such misalignment through an optimal error exponent analysis. Built on this insight, we propose three explicit alignment strategies that plug-and-play into existing long-tail methods without architectural change. Extensive experiments on the CIFAR-10-L T, CIFAR-100-L T, and ImageNet-L T datasets consistently boost examined baselines and achieve the state-of-the-art performances.

Large output spaces, also referred to as Extreme multilabel classification (XMC), is a setting that arises, e.g., in large-scale tagging and product-to-product recommendation, and is characterized by the number of labels ranging from hundreds of thousands to millions. This means that the linear classification head, usually only a tiny fraction of the overall model, turns into the main driver for compute and memory demand. Current state-of-the-art XMC methods predominantly rely on FP16-FP32 mixed-precision training, which we show can be unstable, and inefficient in terms of memory usage and computational overhead. Meanwhile, existing low-precision methods typically retain higher precision for the classification layer. In this work, we propose ELMO, a pure low-precision training framework for XMC models using BFloat16 and Float8 data types. By leveraging Kahan summation and stochastic rounding, we demonstrate that XMC models can be effectively trained entirely in Float8, without relying on single-precision master weights or tensor scaling. Low-precision training, combined with our proposed memory optimizations -- gradient fusion and chunking -- enables significant reductions in GPU memory usage. For example, we train a 3-million-label XMC model with only 6.6 GiB of GPU memory, compared to the 39.7 GiB required by the optimized SOTA method, Renee without compromising accuracy.

Despite the advantages of discriminative prototype-based methods, their role in adversarial robustness remains underexplored. Meanwhile, current adversarial training methods predominantly focus on robustness against adversarial attacks without explicitly leveraging geometric structures in the latent space, usually resulting in reduced accuracy on the original clean data. We propose a novel framework named Adversarially trained Deep Positive-Negative Prototypes (Adv-DPNP), which integrates discriminative prototype-based learning with adversarial training. Adv-DPNP uses unified class prototypes that serve as both classifier weights and robust anchors in the latent space. Moreover, a novel dual-branch training mechanism maintains stable prototypes by updating them exclusively with clean data, while the feature extractor is trained on both clean and adversarial inputs to increase invariance to adversarial perturbations. In addition, we use a composite loss that combines positive-prototype alignment, negative-prototype repulsion, and consistency regularization to further enhance discrimination, adversarial robustness, and clean accuracy. Extensive experiments on standard benchmarks (CIFAR-10/100 and SVHN) confirm that Adv-DPNP improves clean accuracy over state-of-the-art defenses and baseline methods, while maintaining competitive or superior robustness under a suite of widely used attacks, including FGSM, PGD, C\&W, and AutoAttack. We also evaluate robustness to common corruptions on CIFAR-10-C, where Adv-DPNP achieves the highest average accuracy across severities and corruption types. Additionally, we provide an in-depth analysis of the discriminative quality of the learned feature representations, highlighting the effectiveness of Adv-DPNP in maintaining compactness and clear separation in the latent space.

Enabling deep models to generalize in non-stationary environments is vital for real-world machine learning, as data distributions are often found to continually change. Recently, evolving domain generalization (EDG) has emerged to tackle the domain generalization in a time-varying system, where the domain gradually evolves over time in an underlying continuous structure. Nevertheless, it typically assumes multiple source domains simultaneously ready. It still remains an open problem to address EDG in the domain-incremental setting, where source domains are non-static and arrive sequentially to mimic the evolution of training domains. To this end, we propose Weight Diffusion (W-Diff), a novel framework that utilizes the conditional diffusion model in the parameter space to learn the evolving pattern of classifiers during the domain-incremental training process.

Human Activity Recognition (HAR) aims to recognize activities by training models on massive sensor data. In real-world deployment, a crucial aspect of HAR that has been largely overlooked is that the test sets may have different distributions from training sets due to inter-subject variability including age, gender, behavioral habits, etc., which leads to poor generalization performance. One promising solution is to learn domain-invariant representations to enable a model to generalize on an unseen distribution. However, most existing methods only consider the feature-invariance of the penultimate layer for domain-invariant learning, which leads to suboptimal results. In this paper, we propose a Categorical Concept Invariant Learning (CCIL) framework for generalizable activity recognition, which introduces a concept matrix to regularize the model in the training stage by simultaneously concentrating on feature-invariance and logit-invariance. Our key idea is that the concept matrix for samples belonging to the same activity category should be similar. Extensive experiments on four public HAR benchmarks demonstrate that our CCIL substantially outperforms the state-of-the-art approaches under cross-person, cross-dataset, cross-position, and one-person-to-another settings.

"Extreme Classification" (or XC) is the task of annotating data points (queries) with relevant labels (documents), from an extremely large set of L possible labels, arising in search and recommendations. The most successful deep learning paradigm that has emerged over the last decade or so for XC is to embed the queries (and labels) using a deep encoder (e.g. DistilBERT), and use linear classifiers on top of the query embeddings. This architecture is of appeal because it enables millisecond-time inference using approximate nearest neighbor search (ANNS). The key question is how do we design training algorithms that are accurate as well as scale to O(100M) labels on a limited number of GPUs. State-of-the-art XC techniques that demonstrate high accuracies (e.g., DEXML, Renée, DEXA) on standard datasets have per-epoch training time that scales as O(L) or employ expensive negative sampling strategies, which are prohibitive in XC scenarios. In this work, we develop an accurate and scalable XC algorithm ASTRA with two key observations: (a) building ANNS index on the classifier vectors and retrieving hard negatives using the classifiers aligns the negative sampling strategy to the loss function optimized; (b) keeping the ANNS indices current as the classifiers change through the epochs is prohibitively expensive while using stale negatives (refreshed periodically) results in poor accuracy; to remedy this, we propose a negative sampling strategy that uses a mixture of importance sampling and uniform sampling. By extensive evaluation on standard XC as well as proprietary datasets with 120M labels, we demonstrate that ASTRA achieves SOTA precision, while reducing training time by 4x-15x relative to the second best.