centroid computation

Conventional compositional zero-shot learning (CZSL) methods model primitive concepts (attributes/objects) in a monolithic manner, neglecting their semantic diversity across different compositions. Method: We propose ClusPro, a parameter-free framework that dynamically discovers diverse prototypes via adaptive clustering in the embedding space, thereby constructing a disentangled and structured representation space for primitives. It introduces an updateable prototype set to characterize concept boundaries and jointly optimizes prototype-driven contrastive learning and decorrelation learning to enforce intra-class compactness and inter-class disentanglement—without introducing any additional parameters. Contribution/Results: ClusPro achieves significant improvements over state-of-the-art methods on three standard benchmarks under both closed-world and open-world evaluation protocols, demonstrating superior generalization capability and principled modeling of semantic compositionality.

This work addresses the low efficiency and poor structural preservation inherent in prototype selection for large-scale datasets. We propose TPS, a topology-aware prototype selection framework grounded in topological data analysis (TDA). TPS leverages persistent homology to characterize the intrinsic geometry and connectivity structure of data, enabling adaptive identification of topologically salient samples as prototypes; it further supports parallel implementation. Compared with conventional methods, TPS achieves substantial data compression—retaining only 5–15% of samples on multiple synthetic and real-world benchmarks—while maintaining or improving classification accuracy by 1.2–3.8 percentage points. The approach also exhibits strong interpretability and robustness. Its core innovation lies in the first systematic integration of TDA’s structural awareness into prototype selection, thereby unifying computational efficiency, structural fidelity, and interpretability.

To address classification of unordered sets (e.g., image or text collections), this paper proposes a Grassmann manifold-based prototypical learning framework: inputs are modeled as linear subspaces, with learnable subspace prototypes and differentiable, adaptive dimension-selection factors. It introduces the first subspace prototype mechanism—preserving model transparency while explicitly quantifying each input vector’s contribution to the decision, thereby significantly enhancing interpretability. End-to-end training employs Riemannian optimization. On multiple image and text set classification benchmarks, the method matches Transformer performance while reducing parameter count by 42%–68% and accelerating inference by 2.1–3.4×. Key innovations include geometrically aware prototype representation, adaptive dimensionality selection, and unified modeling of interpretable decision policies.

To address the high computational cost of k-nearest neighbors (kNN) in multi-label classification—stemming from large-scale training sets—this paper pioneers the extension of prototype learning to the multi-label setting. We propose a label-aware prototype generation method that jointly optimizes label structure consistency and instance similarity. Our approach integrates a multi-label distance metric, greedy initialization, iterative optimization guided by label coverage, and an adaptive kNN reweighting mechanism. Experiments across multiple benchmark datasets demonstrate that our method compresses the training set by over 80%, while maintaining or improving macro-F1 score and classification accuracy. Crucially, it significantly reduces inference cost without sacrificing performance. The core contribution is the first interpretable and efficient prototype learning framework specifically designed for multi-label classification, bridging scalability and fidelity in label-space modeling.

Generalized Category Discovery (GCD) aims to jointly cluster unlabeled data—containing both known and novel classes—by leveraging labeled data from known classes; its core challenge lies in accuracy imbalance caused by distributional ambiguity between known and novel classes. This paper proposes a Joint Prototype Learning (JPL) framework: (1) unifying prototype modeling for both known and novel classes to eliminate classifier bias; (2) introducing a two-level adaptive pseudo-labeling mechanism to mitigate confirmation bias; and (3) integrating contrastive regularization and clustering consistency constraints to align feature representations with clustering objectives, further enhanced by novel-class cardinality estimation and outlier detection for task-level co-optimization. Evaluated on both generic and fine-grained benchmarks, JPL achieves state-of-the-art performance, significantly improving balanced accuracy across known and novel classes while enhancing representation discriminability.

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 the critical challenge in few-shot knowledge distillation of efficiently selecting an extremely small yet representative subset of samples to enhance student model performance. The authors propose a minimalist coreset selection method that computes the medoid for each class and selects the sample closest to this medoid as the distilled data. By eschewing complex sampling strategies and relying solely on intra-class geometric structure, the approach achieves remarkable simplicity while significantly outperforming random sampling and existing coreset methods. Extensive experiments demonstrate consistent improvements across four image classification benchmarks and three teacher–student architectures—including both CNNs and Transformers—establishing the method as a strong standard baseline for few-shot knowledge distillation.

This work addresses the mismatch between the geometric structure of pretrained embeddings and downstream tasks in practical systems such as digital governance, where label scarcity, domain shift, and the infeasibility of retraining large models degrade nearest-neighbor retrieval performance. To bridge the gap between unsupervised post-processing and fully supervised projection, the authors propose a label-efficient method that leverages limited supervision to softly align embeddings with class prototypes while preserving embedding dimensionality. This approach optimizes local neighborhood structure and significantly enhances similarity-based retrieval and lightweight classification under extremely low-label regimes. Experimental results demonstrate a 25.7% improvement in local neighborhood quality over raw embeddings and outperform strong unsupervised post-processing baselines by more than 21.1%.

This work addresses the limitations of prototype-based models in large-scale image classification—namely, their weak generalization, reliance on costly fine-tuning, and susceptibility to prototype drift—by introducing vector quantization into the latent space. The proposed approach employs a discrete, learnable codebook to constrain prototype representations, enabling stable, data-anchored, and interpretable prototype modeling without requiring fine-tuning of the backbone network. Evaluated on benchmark datasets including ImageNet, CUB-200, and Cars-196, the method achieves competitive classification accuracy while significantly enhancing model interpretability and prototype consistency.

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.