Rethinking Noise-Robust Training for Frozen Vision Foundation Models: A Cross-Dataset Benchmark with a Case Study of Small-Loss Failure

• URL: http://arxiv.org/abs/2605.22591v1
• Date: Thu, 21 May 2026 15:02:44 GMT
• Title: Rethinking Noise-Robust Training for Frozen Vision Foundation Models: A Cross-Dataset Benchmark with a Case Study of Small-Loss Failure
• Abstract summary: Frozen Vision Foundation Models (VFMs) with lightweight classification heads are increasingly used in medical imaging.<n>We present a benchmark of eight noisy-label methods across five medical datasets, three backbones, two noise types, and five noise rates.<n>We conclude with evidence-based guidance and a low-regret feature-space selector for practical recommendation.
• Score: 7.512887610816016
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Frozen Vision Foundation Models (VFMs) with lightweight classification heads are increasingly used in medical imaging because they offer efficient and reproducible deployment. Yet noisy-label learning methods for this frozen-feature regime remain poorly understood, and most existing methods still rely on a small-loss assumption inherited from end-to-end training. We present a controlled benchmark of eight noisy-label methods across five medical datasets, three backbones, two noise types, and five noise rates (150 conditions, 6,000 training runs), evaluated with balanced accuracy. The benchmark shows that there is no universal winner: Friedman ranking over the 150 conditions yields $χ^2 = 333.2$ ($p = 4.77 \times 10^{-68}$), ELR wins the most conditions (49/150), while CUFIT attains the best mean rank (2.51). The practical cost of method choice grows sharply with noise severity, from 4.5pp on clean data to 18.8pp at asymmetric 40\% noise. To explain these benchmark-level patterns, we revisit the small-loss assumption in a representative high-risk regime. Under frozen DINOv2 features, clean and noisy loss distributions overlap by 53--61\%, and matched-rate clean-sample detection shows that prediction agreement is markedly more stable than loss ranking under asymmetric noise (3pp vs.\ 13pp precision drop). On ISIC2019 with asymmetric 40\% noise, Co-Teaching reaches 68\% overall accuracy while collapsing to 35.1\% balanced accuracy with zero recall on three minority classes. Together, these results recast noisy-label learning for frozen VFMs as a regime-aware method-selection problem rather than a search for a single dominant algorithm. We conclude with evidence-based guidance and a low-regret feature-space selector for practical recommendation.

Related papers

• Error-Type-Aware Loss Reweighting for Robust Named Entity Recognition with Noisy LLM Labels [11.267113571494392]
  We propose error-type-aware loss reweighting for named entity recognition.<n>Our approach is simple and efficient, does not require additional training resources, and improves F1 by 0.8 - 2.0 percentage points on dataset-level average for noise levels between 15% and 40%, with a maximum improvement of 4.6 percentage points with 24.1% noise on Wikigold.
  arXiv  Detail & Related papers  (2026-08-31T14:05:08Z)
• Forget or Fine-tune? A Comparative Study of Machine Unlearning Strategies for Noisy Label Correction [0.5647577824219207]
  Noisy labels remain a critical challenge for training deep neural networks.<n>Standard solution is to retrain the model from scratch on the cleaned dataset.<n>Machine Unlearning has emerged as a computationally efficient alternative.
  arXiv  Detail & Related papers  (2026-08-30T21:16:29Z)
• ER-KANs: Efficient and Robust Kolmogorov-Arnold Networks for Data-Scarce Scientific Machine Learning [0.0]
  We show that ChebyKAN's test MSE increases by a factor of 10.6x when training data is corrupted with sigma=0.1 noise.<n> ER-KAN combines three design choices targeting the noisy, data-scarce setting.
  arXiv  Detail & Related papers  (2026-08-14T16:15:37Z)
• Agentic Autoresearch for CT Reconstruction [6.987253902215927]
  We benchmarked 26 methods on Mayo low-dose CT (noise-limited) and a 128-view sparse-view breast task.<n>Every trained breast model was then re-scored on noisy inputs without retraining, and separately retrained on matched noise.<n> Benchmarking gives a tier of statistically indistinguishable top methods, not one winner.
  arXiv  Detail & Related papers  (2026-07-24T18:03:21Z)
• Traffic-Aware Randomized Smoothing for LLM-Based Network Intrusion Detection [1.4915002678801434]
  We present Traffic-Aware Randomized Smoothing (TA-RS) for intrusion detection systems.<n>It is a certified defense that injects Gaussian noise exclusively into the controllable (DC) subspace.<n>It achieves 55-100% certified accuracy on two datasets.
  arXiv  Detail & Related papers  (2026-07-15T13:08:19Z)
• Can LLMs Learn to Reason Robustly under Noisy Supervision? [31.39996460068968]
  We take the first step toward a systematic analysis of noisy label mechanisms in RLVR.<n>We distinguish two types of noise: inactive noisy labels, which reduce data efficiency, and active noisy labels, which are reinforced and risk skewing the model.<n>Motivated by this dynamic, we propose Online Label Refinement (OLR), which progressively corrects potentially noisy labels with majority-voted answers.
  arXiv  Detail & Related papers  (2026-04-05T06:30:50Z)
• Deconstructing the Failure of Ideal Noise Correction: A Three-Pillar Diagnosis [82.06444754727156]
  We show that noise-correction methods still suffer from performance collapse during training.<n>This compellingly demonstrates that the failure is not fundamentally a $T$-estimation problem, but stems from a more deeply rooted flaw.<n>We provide a unified analysis that links three levels: macroscopic convergence states, microscopic optimisation dynamics, and information-theoretic limits on what can be learned from noisy labels.
  arXiv  Detail & Related papers  (2026-03-13T13:53:04Z)
• ReSup: Reliable Label Noise Suppression for Facial Expression Recognition [20.74719263734951]
  We propose a more reliable noise-label suppression method called ReSup. To achieve optimal distribution modeling, ReSup models the similarity distribution of all samples. To further enhance the reliability of our noise decision results, ReSup uses two networks to jointly achieve noise suppression.
  arXiv  Detail & Related papers  (2023-05-29T06:02:06Z)
• Latent Class-Conditional Noise Model [54.56899309997246]
  We introduce a Latent Class-Conditional Noise model (LCCN) to parameterize the noise transition under a Bayesian framework. We then deduce a dynamic label regression method for LCCN, whose Gibbs sampler allows us efficiently infer the latent true labels. Our approach safeguards the stable update of the noise transition, which avoids previous arbitrarily tuning from a mini-batch of samples.
  arXiv  Detail & Related papers  (2023-02-19T15:24:37Z)
• Class Prototype-based Cleaner for Label Noise Learning [73.007001454085]
  Semi-supervised learning methods are current SOTA solutions to the noisy-label learning problem. We propose a simple yet effective solution, named textbfClass textbfPrototype-based label noise textbfCleaner.
  arXiv  Detail & Related papers  (2022-12-21T04:56:41Z)
• Identifying Hard Noise in Long-Tailed Sample Distribution [71.8462682319137]
  We introduce Noisy Long-Tailed Classification (NLT)<n>Most de-noising methods fail to identify the hard noises.<n>We design an iterative noisy learning framework called Hard-to-Easy (H2E)
  arXiv  Detail & Related papers  (2022-07-27T09:03:03Z)
• Uncertainty-Aware Learning Against Label Noise on Imbalanced Datasets [23.4536532321199]
  We propose an Uncertainty-aware Label Correction framework to handle label noise on imbalanced datasets. Inspired by our observations, we propose an Uncertainty-aware Label Correction framework to handle label noise on imbalanced datasets.
  arXiv  Detail & Related papers  (2022-07-12T11:35:55Z)
• (Certified!!) Adversarial Robustness for Free! [116.6052628829344]
  We certify 71% accuracy on ImageNet under adversarial perturbations constrained to be within a 2-norm of 0.5. We obtain these results using only pretrained diffusion models and image classifiers, without requiring any fine tuning or retraining of model parameters.
  arXiv  Detail & Related papers  (2022-06-21T17:27:27Z)
• UNICON: Combating Label Noise Through Uniform Selection and Contrastive Learning [89.56465237941013]
  We propose UNICON, a simple yet effective sample selection method which is robust to high label noise. We obtain an 11.4% improvement over the current state-of-the-art on CIFAR100 dataset with a 90% noise rate.
  arXiv  Detail & Related papers  (2022-03-28T07:36:36Z)

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.