Hard Labels In! Rethinking the Role of Hard Labels in Mitigating Local Semantic Drift

• URL: http://arxiv.org/abs/2512.15647v2
• Date: Mon, 22 Dec 2025 07:47:31 GMT
• Title: Hard Labels In! Rethinking the Role of Hard Labels in Mitigating Local Semantic Drift
• Abstract summary: We show that when only a limited number of crops per image are used, soft labels are prone to local semantic drift.<n>We propose a new training paradigm, Hard Label for Alleviating Local Semantic Drift (HALD), which leverages hard labels as intermediate corrective signals.<n>Our findings re-establish the importance of hard labels as a complementary tool, and call for a rethinking of their role in soft-label-dominated training.
• Score: 34.88652783901701
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Soft labels generated by teacher models have become a dominant paradigm for knowledge transfer and recent large-scale dataset distillation such as SRe2L, RDED, LPLD, offering richer supervision than conventional hard labels. However, we observe that when only a limited number of crops per image are used, soft labels are prone to local semantic drift: a crop may visually resemble another class, causing its soft embedding to deviate from the ground-truth semantics of the original image. This mismatch between local visual content and global semantic meaning introduces systematic errors and distribution misalignment between training and testing. In this work, we revisit the overlooked role of hard labels and show that, when appropriately integrated, they provide a powerful content-agnostic anchor to calibrate semantic drift. We theoretically characterize the emergence of drift under few soft-label supervision and demonstrate that hybridizing soft and hard labels restores alignment between visual content and semantic supervision. Building on this insight, we propose a new training paradigm, Hard Label for Alleviating Local Semantic Drift (HALD), which leverages hard labels as intermediate corrective signals while retaining the fine-grained advantages of soft labels. Extensive experiments on dataset distillation and large-scale conventional classification benchmarks validate our approach, showing consistent improvements in generalization. On ImageNet-1K, we achieve 42.7% with only 285M storage for soft labels, outperforming prior state-of-the-art LPLD by 9.0%. Our findings re-establish the importance of hard labels as a complementary tool, and call for a rethinking of their role in soft-label-dominated training.

Related papers

• Towards Better IncomLDL: We Are Unaware of Hidden Labels in Advance [68.65348495515258]
  We presentHidLDL, which aims to recover a complete label distribution from a real-world incomplete label distribution.<n>We simultaneously use local feature similarity and the global low-rank structure to reveal the mysterious veil of hidden labels.
  arXiv  Detail & Related papers  (2025-11-16T08:10:26Z)
• Label Distribution Learning with Biased Annotations by Learning Multi-Label Representation [120.97262070068224]
  Multi-label learning (MLL) has gained attention for its ability to represent real-world data.<n>Label Distribution Learning (LDL) faces challenges in collecting accurate label distributions.
  arXiv  Detail & Related papers  (2025-02-03T09:04:03Z)
• Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations [91.67511167969934]
  imprecise label learning (ILL) is a framework for the unification of learning with various imprecise label configurations. We demonstrate that ILL can seamlessly adapt to partial label learning, semi-supervised learning, noisy label learning, and, more importantly, a mixture of these settings.
  arXiv  Detail & Related papers  (2023-05-22T04:50:28Z)
• Exploiting Completeness and Uncertainty of Pseudo Labels for Weakly Supervised Video Anomaly Detection [149.23913018423022]
  Weakly supervised video anomaly detection aims to identify abnormal events in videos using only video-level labels. Two-stage self-training methods have achieved significant improvements by self-generating pseudo labels. We propose an enhancement framework by exploiting completeness and uncertainty properties for effective self-training.
  arXiv  Detail & Related papers  (2022-12-08T05:53:53Z)
• Beyond Hard Labels: Investigating data label distributions [0.9668407688201357]
  We compare the disparity of learning with hard and soft labels for a synthetic and a real-world dataset. The application of soft labels leads to improved performance and yields a more regular structure of the internal feature space.
  arXiv  Detail & Related papers  (2022-07-13T14:25:30Z)
• Label Noise-Resistant Mean Teaching for Weakly Supervised Fake News Detection [93.6222609806278]
  We propose a novel label noise-resistant mean teaching approach (LNMT) for weakly supervised fake news detection. LNMT leverages unlabeled news and feedback comments of users to enlarge the amount of training data. LNMT establishes a mean teacher framework equipped with label propagation and label reliability estimation.
  arXiv  Detail & Related papers  (2022-06-10T16:01:58Z)
• PseudoSeg: Designing Pseudo Labels for Semantic Segmentation [78.35515004654553]
  We present a re-design of pseudo-labeling to generate structured pseudo labels for training with unlabeled or weakly-labeled data. We demonstrate the effectiveness of the proposed pseudo-labeling strategy in both low-data and high-data regimes.
  arXiv  Detail & Related papers  (2020-10-19T17:59:30Z)
• Learning Soft Labels via Meta Learning [3.4852307714135375]
  One-hot labels do not represent soft decision boundaries among concepts, and hence, models trained on them are prone to overfitting. We propose a framework, where we treat the labels as learnable parameters, and optimize them along with model parameters. We show that learned labels capture semantic relationship between classes, and thereby improve teacher models for the downstream task of distillation.
  arXiv  Detail & Related papers  (2020-09-20T18:42:13Z)
• SSKD: Self-Supervised Knowledge Distillation for Cross Domain Adaptive Person Re-Identification [25.96221714337815]
  Domain adaptive person re-identification (re-ID) is a challenging task due to the large discrepancy between the source domain and the target domain. Existing methods mainly attempt to generate pseudo labels for unlabeled target images by clustering algorithms. We propose a Self-Supervised Knowledge Distillation (SSKD) technique containing two modules, the identity learning and the soft label learning.
  arXiv  Detail & Related papers  (2020-09-13T10:12:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.