This work proposes a task-specific pruning pipeline, named Cut-ViT, which first construct gram anchoring matrices from both spatial and semantic perspectives, and performs the subspace decomposition to extract the corresponding subspace bases.
Abstract
Pruning visual foundation models has attracted considerable attention. However, existing methods focus on rigid point-to-point token alignment on a single dataset for pruning, suffering from two limitations: i) robustness degradation, and ii) task-specificity deficiency. To address these limitations, we propose a task-specific pruning pipeline, named Cut-ViT. Specifically, we first construct gram anchoring matrices from both spatial and semantic perspectives, and perform the subspace decomposition to extract the corresponding subspace bases. Basis-agnostic and residual constraints are then adopted to align the gram subspaces between the native and pruned DINOv3 models along spatial and channel dimensions, enabling subnetworks to inherit robust feature representations of native DINOv3. Furthermore, we design spectral entropy adaptation, which quantifies the information density of feature manifolds along spatial and channel dimensions, thereby adapting the pruning objective to specific downstream tasks. Experiments show that Cut-ViT requires approximately one minute on a single A100 GPU to obtain subnetworks at various sparsity levels, using only 20.9% of the time and 45.5% of the GPU memory compared with previous methods, while achieving SOTA performance on six tasks across nine datasets.
TASSO, a new paradigm that efficiently preserves the latent space geometry while ensuring network plasticity, is introduced with two complementary techniques: subspace learning and geometry-aware knowledge distillation.
Changming Sun, Francesco Barbato, Matteo Caligiuri et al.· 0 citations
Kernel Distance Metric Learning (KDML) is a powerful tool for enhancing clustering performance by adapting the data geometry to correct structural mismatches. However, traditional global KDML methods suffer from severe scalability issues, often incurring O (n3) computational costs and O (n2) memory usage, where n is the sample size. Furthermore, these holistic deformations can lead to overfitting when supervision is sparse. To address these limitations, we propose Constraint-Aligned Subspace Transformation (CAST), a novel framework that restricts metric adaptation to a low dimensional subspace strictly induced by pairwise constraints. By orthonormalizing constraint directions, we reduce the learning problem from estimating a full kernel matrix to a rank-efficient Linear Program (LP). We further derive an implicit update formulation that enables distance evaluation in O (r) time, effectively reducing memory complexity to O (nr). Extensive experiments on diverse datasets demonstrate that CAST matches or exceeds state-of-the-art KDML competitors in downstream clustering tasks while dramatically reducing optimization time, delivering 33×-259× speedups over the fastest baseline in the weak-supervision (1%) regime, while still achieving 3×–20× speedups at 20% supervision.
Fei Wang, Le Li, P. Fränti· Proceedings of the 32nd ACM...· 0 citations
Structural Parameter-free Affinity Regularization (SPARE), a regularizer that matches the pairwise affinities of intermediate tokens to those of the clean latents across images, is proposed, a regularizer that attains the lowest FID among parameter-free regularizers in every tested setting.
Zong-Wei Hong, Jinglun Li, Shen Zhang et al.· 0 citations
SnapViT: single-shot network approximation for pruned Vision Transformers is introduced, a new post-pretraining structured pruning method that enables elastic inference across a continuum of compute budgets, and a self-supervised importance scoring mechanism that maintains strong performance without requiring retraining or labels.
Walter Simoncini, Michael Dorkenwald, Tijmen Blankevoort et al.· 0 citations
Dataset distillation aims to synthesize compact datasets that can approximate the performance of full-data training while significantly reducing computational and storage costs. However, diffusion-based distillation methods often struggle to preserve structural coherence and generalization, especially in visually complex domains. This issue often stems from latent prototypes that are weakly aligned with class-discriminative regions and contaminated by irrelevant background, thereby degrading generation quality and generalization. To address this limitation, we propose a saliency-driven distillation framework that constructs class-discriminative latent prototypes to enhance representativeness and generalization. The framework proceeds in two stages: (1) ensemble Grad-CAM++ saliency is used to construct prototypes emphasizing class-discriminative regions, and (2) hard-prototype refinement is then applied to construct challenging yet class-consistent prototypes, thereby enhancing discriminability and diversity. Importantly, the diffusion backbones (e.g., LDM and DiT) remain frozen; only lightweight classifiers used for saliency extraction are trained. Extensive experiments across multiple benchmarks demonstrate consistent performance improvements over strong baselines. Code will be released.
Yawen Zou, Wenqi Cai, Guang Li et al.· arXiv.org· 0 citations
Prototypical part-based models provide explainable predictions by comparing input regions to learned prototypes. However, current approaches are burdened by complex, multi-stage training pipelines and heavily rely on auxiliary regularization to prevent prototype collapse. To overcome these limitations, we introduce Orthonormal Prototype Alignment Learning (OPAL), a single-stage, end-to-end framework that simplifies interpretable classification. Our approach anchors the latent space using predefined orthonormal bases, embedding each class within a dedicated subspace spanned by fixed part-prototypes. To achieve precise part localization, OPAL enforces spatial competition across feature maps. This mechanism isolates sparse, discriminative regions, directing each prototype to consistently attend to the same semantic concept across different images. By framing classification as a direct representation alignment task, our method eliminates the need for auxiliary losses. Extensive experiments on fine-grained benchmarks demonstrate that OPAL outperforms both its non-interpretable counterparts and state-of-the-art part-prototype methods, delivering granular visual explanations by explicitly revealing the specific image regions driving every prediction. Code is available at https://github.com/ilancarretero/OPAL.
I. Carretero, G. Angulo, R. del Amor et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
