Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reliable Active Learning from Unreliable Labels via Neural Collapse Geometry

Atharv Goel, Sharat Agarwal, Saket Anand, Chetan Arora

TL;DR

Real-world active learning suffers when labels are noisy or distributions shift, leading standard heuristics to misselect samples. The paper introduces NCAL-R, a neural collapse–guided active learning framework that uses Class-Mean Alignment Perturbation (CMAP) and Feature Fluctuation (FF) to acquire samples that preserve class separation and reveal ambiguities. NCAL-R achieves higher accuracy with fewer labels and demonstrates robustness to synthetic label noise and stronger out-of-distribution generalization on ImageNet-100 and CIFAR-100. The results suggest that incorporating geometric reliability criteria into acquisition decisions improves the trustworthiness and practical deployability of active learning in real labeling pipelines.

Abstract

Active Learning (AL) promises to reduce annotation cost by prioritizing informative samples, yet its reliability is undermined when labels are noisy or when the data distribution shifts. In practice, annotators make mistakes, rare categories are ambiguous, and conventional AL heuristics (uncertainty, diversity) often amplify such errors by repeatedly selecting mislabeled or redundant samples. We propose Reliable Active Learning via Neural Collapse Geometry (NCAL-R), a framework that leverages the emergent geometric regularities of deep networks to counteract unreliable supervision. Our method introduces two complementary signals: (i) a Class-Mean Alignment Perturbation score, which quantifies how candidate samples structurally stabilize or distort inter-class geometry, and (ii) a Feature Fluctuation score, which captures temporal instability of representations across training checkpoints. By combining these signals, NCAL-R prioritizes samples that both preserve class separation and highlight ambiguous regions, mitigating the effect of noisy or redundant labels. Experiments on ImageNet-100 and CIFAR100 show that NCAL-R consistently outperforms standard AL baselines, achieving higher accuracy with fewer labels, improved robustness under synthetic label noise, and stronger generalization to out-of-distribution data. These results suggest that incorporating geometric reliability criteria into acquisition decisions can make Active Learning less brittle to annotation errors and distribution shifts, a key step toward trustworthy deployment in real-world labeling pipelines. Our code is available at https://github.com/Vision-IIITD/NCAL.

Reliable Active Learning from Unreliable Labels via Neural Collapse Geometry

TL;DR

Real-world active learning suffers when labels are noisy or distributions shift, leading standard heuristics to misselect samples. The paper introduces NCAL-R, a neural collapse–guided active learning framework that uses Class-Mean Alignment Perturbation (CMAP) and Feature Fluctuation (FF) to acquire samples that preserve class separation and reveal ambiguities. NCAL-R achieves higher accuracy with fewer labels and demonstrates robustness to synthetic label noise and stronger out-of-distribution generalization on ImageNet-100 and CIFAR-100. The results suggest that incorporating geometric reliability criteria into acquisition decisions improves the trustworthiness and practical deployability of active learning in real labeling pipelines.

Abstract

Active Learning (AL) promises to reduce annotation cost by prioritizing informative samples, yet its reliability is undermined when labels are noisy or when the data distribution shifts. In practice, annotators make mistakes, rare categories are ambiguous, and conventional AL heuristics (uncertainty, diversity) often amplify such errors by repeatedly selecting mislabeled or redundant samples. We propose Reliable Active Learning via Neural Collapse Geometry (NCAL-R), a framework that leverages the emergent geometric regularities of deep networks to counteract unreliable supervision. Our method introduces two complementary signals: (i) a Class-Mean Alignment Perturbation score, which quantifies how candidate samples structurally stabilize or distort inter-class geometry, and (ii) a Feature Fluctuation score, which captures temporal instability of representations across training checkpoints. By combining these signals, NCAL-R prioritizes samples that both preserve class separation and highlight ambiguous regions, mitigating the effect of noisy or redundant labels. Experiments on ImageNet-100 and CIFAR100 show that NCAL-R consistently outperforms standard AL baselines, achieving higher accuracy with fewer labels, improved robustness under synthetic label noise, and stronger generalization to out-of-distribution data. These results suggest that incorporating geometric reliability criteria into acquisition decisions can make Active Learning less brittle to annotation errors and distribution shifts, a key step toward trustworthy deployment in real-world labeling pipelines. Our code is available at https://github.com/Vision-IIITD/NCAL.

Paper Structure

This paper contains 17 sections, 5 equations, 2 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Experiments
  4. Experimental Setup.
  5. Evaluation Metrics.
  6. Covariate Shift Results.
  7. Label Drift and GCD.
  8. Related Work
  9. Conclusion
  10. Deriving the CMAP
  11. Implementation note:
  12. Training Protocol
  13. Experimental settings.
  14. Compute.
  15. Algorithm Pseudo Code
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Comparison of test accuracy across varying label budgets on three benchmark datasets—ImageNet100, CIFAR100, and CIFAR10. NCAL's good performance even at lower annotation budgets suggests that its Neural Collapse-guided selection promotes more structured and representative feature learning. (Note: accuracy for 100% data of ImageNet100, CIFAR100 and CIFAR10 are: 79.16%, 70.75% and 90% respectively. Reported results are average of 3 independent runs.)
  • Figure 2: Ablation