learn class prototypes

This work addresses the degradation of embedding space coherence and model performance in federated learning caused by premature alignment of immature prototypes, particularly under highly non-IID data distributions. To mitigate this issue, the authors propose FedSAP, a novel framework that integrates a delayed-alignment curriculum scheduling mechanism with a geometry-driven proxy separation loss on the unit hypersphere. This approach enhances intra-class compactness and inter-class separability without increasing communication overhead. As the first study to formalize scheduled alignment as a general design principle in federated prototype learning, FedSAP stabilizes representation learning without introducing additional parameters and naturally extends to semi-supervised settings. Extensive experiments demonstrate consistent improvements, with up to a 4-percentage-point gain over state-of-the-art methods across three benchmark datasets, especially excelling in high-heterogeneity scenarios.

This work addresses the high memory overhead and privacy concerns of conventional replay-based methods in online continual learning, which typically require storing large volumes of historical data. To mitigate catastrophic forgetting under stringent memory constraints, the authors propose a prototype-based compressed replay strategy that synthesizes a small set of representative prototype samples per class and augments them via a perturbation mechanism to generate diverse synthetic variants. By integrating prototype synthesis, feature extraction, and perturbation-based augmentation, the method drastically reduces storage requirements while preserving data privacy. Extensive experiments on multiple benchmarks and large-scale multitask settings demonstrate that the approach consistently outperforms existing replay techniques—even when retaining only a minimal number of samples per class—thereby achieving superior performance with significantly lower memory consumption.

Existing prototype alignment methods in heterogeneous federated learning enforce clients with diverse architectures to align within a unified feature subspace, thereby constraining model expressiveness. This work proposes FedSAF, a novel structural alignment paradigm that shifts the alignment objective from coordinate-wise matching to preserving the consistency of inter-class relational structures. By decoupling semantic structure alignment from shared feature bases, FedSAF models class relationships through prototypes and integrates them into a distributed optimization framework. Extensive experiments demonstrate that FedSAF significantly outperforms current heterogeneous federated learning approaches across multiple benchmarks, achieving accuracy improvements of up to 3.52%.

This work addresses the severe domain shift caused by heterogeneous client data distributions in federated learning, which significantly degrades the generalization performance of the global model. To mitigate this issue, the authors propose a domain-aware prototype learning mechanism that explicitly preserves and leverages domain information—a first in federated learning. Specifically, the method constructs a dedicated global prototype for each domain and aggregates local prototypes from clients within the same domain through similarity-based weighting. During local training, features are encouraged to align with prototypes from their own domain while being separated from those of other domains, thereby jointly optimizing intra-domain consistency and inter-domain separability. Extensive experiments on DomainNet, Office-10, and PACS benchmarks demonstrate that the proposed approach substantially outperforms existing methods and effectively alleviates domain shift.

Prototype memory constructs identity prototypes via uniform averaging of embeddings from the same identity, rendering it vulnerable to corruption by low-quality samples—such as blurry, occluded, or large-pose face images—thereby distorting prototypes and biasing training signals. To address this, we propose a quality-aware prototype generation mechanism: embeddings from the same identity are dynamically weighted—based on estimated face quality metrics (e.g., blur, occlusion, and pose confidence)—before averaging. This is the first work to incorporate explicit quality awareness into the prototype memory framework. The method is modular and compatible with mainstream backbone networks and diverse quality estimation modules. Experiments demonstrate substantial improvements in prototype robustness and discriminability. Our approach consistently outperforms the original Prototype Memory on standard benchmarks—including LFW, CFP-FP, and AgeDB-30—with up to a 1.2% absolute accuracy gain, while also markedly enhancing training stability under small-batch settings.

Existing federated prototype learning methods are constrained by a single global prototype, struggling to simultaneously preserve feature fidelity and discriminability under data and model heterogeneity. To address this limitation, this work proposes FedDBP, which employs a dual-branch feature projector on the client side, integrating L2 alignment with contrastive learning to enhance feature quality. On the server side, it leverages Fisher information estimation to dynamically weight feature channels for personalized fusion of global prototypes. This approach overcomes the shortcomings of conventional methods—namely, limited feature expressiveness and rigid prototype representations—and achieves significant performance gains over ten state-of-the-art baselines across multiple benchmarks, demonstrating its effectiveness and robustness in heterogeneous federated learning settings.

This work addresses the issue of semantic drift among clients in federated learning, which often leads to inaccurate global prototypes and degrades model generalization. To mitigate this, the authors propose a hyper-prototype mechanism that aligns local sample features to learnable global class prototypes through gradient matching. The approach further enhances inter-class separability and intra-class consistency by integrating mutual contrastive learning with client-adaptive margins and consistency regularization. Unlike conventional prototype averaging strategies that induce semantic shift, the proposed method preserves semantic coherence in the global representation across diverse heterogeneous settings. Extensive experiments demonstrate state-of-the-art performance on multiple benchmark datasets, validating its effectiveness in achieving robust and semantically consistent federated models.

This work addresses catastrophic forgetting in exemplar-free class-incremental learning, which arises from prototype replay neglecting boundary information between adversarial classes and suffering from class imbalance. To mitigate these issues, the authors propose a manifold-aware boundary sampling strategy coupled with an adaptive class-balancing mechanism. Specifically, synthetic samples are generated near decision boundaries in the feature space through constrained interpolation, while a novel loss function dynamically adjusts the gradient weights of historical classes based on training progression. This approach enhances the discriminability and robustness of prototypes without storing original data, effectively alleviating representation drift and class imbalance. The method achieves state-of-the-art performance across multiple standard benchmarks, even outperforming existing drift-compensation techniques.

This work addresses class-incremental learning without access to any original samples from previous classes by proposing a memory-free framework. It leverages a frozen ImageNet-pretrained encoder to construct a stable latent space, models the latent distributions of old classes using prototype-centered representations, and jointly optimizes the geometric structure of both old and new classes through lightweight adapters and supervised contrastive learning. To the best of our knowledge, this is the first approach to achieve effective incremental learning without storing any raw images. The method significantly outperforms existing techniques on Split CIFAR-100, achieving LastAcc of 31.64%, 37.06%, and 43.10% and AvgAcc of 45.86%, 52.19%, and 56.18% under the Inc5, Inc10, and Inc20 settings, respectively.

This work proposes a generative federated prototype learning framework to address two key challenges in federated learning: model bias toward majority classes caused by data imbalance and high communication overhead due to the transmission of high-dimensional parameters. The approach constructs class-level feature prototypes using Gaussian mixture models and aggregates semantically similar knowledge across clients via Bhattacharyya distance. To mitigate inter-client distributional heterogeneity, it generates synthetic features that augment local representations. A dual-classifier architecture combined with a hybrid loss function—integrating Dot regression and cross-entropy—is employed to optimize local training. Experimental results demonstrate that the proposed method improves accuracy by 3.6% under imbalanced settings while significantly reducing communication costs.