Research map / Label-efficient and robust learning / Noisy labels
Dimensionality-Driven Learning with Noisy Labels
Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah Erfani, Shutao Xia, Sudanthi Wijewickrema, James Bailey
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use D2L's LID-triggered adaptive interpolation of raw and predicted labels instead of the strongest reported baseline, Forward | LeNet-5 · MNIST, 20%–60% symmetric label noise | test accuracy (%): 88.02±0.1 / 68.46±0.1 / 45.51±0.2 -> 98.84±0.1 / 98.49±0.1 / 94.73±0.2 (cross-entropy); Forward: 96.45±0.1 / 94.90±0.1 / | Improved |
| Use D2L's LID-triggered adaptive interpolation of raw and predicted labels instead of Forward | 6-layer CNN · SVHN, 20%–60% symmetric label noise | test accuracy (%): 87.63±0.1 / 82.68±0.1 / 80.92±0.2 vs 85.51±0.1 / 79.09±0.2 / 62.57±0.2 | Improved |
| Use D2L's LID-triggered adaptive interpolation of raw and predicted labels instead of Forward | 12-layer CNN · CIFAR-10, 20%–60% symmetric label noise | test accuracy (%): 85.13±0.2 / 83.36±0.3 / 72.84±0.3 vs 84.61±0.3 / 82.84±0.2 / 72.41±0.4 | Improved |
| Use D2L's LID-triggered adaptive interpolation of raw and predicted labels instead of Forward | ResNet-44 · CIFAR-100, 20%–60% symmetric label noise | test accuracy (%): 62.20±0.4 / 52.01±0.3 / 42.27±0.2 vs 60.25±0.2 / 51.27±0.3 / 41.22±0.3 | Improved |
| Use D2L rather than Forward | LeNet-5 · MNIST, clean labels | test accuracy (%): 99.30±0.0 -> 99.28±0.0 | Hurt |
| Use D2L rather than Forward | 12-layer CNN · CIFAR-10, clean labels | test accuracy (%): 90.27±0.1 -> 89.41±0.2 | Hurt |
| Monitor LID during training and adjust the loss after dimensionality expansion begins | D2L on CIFAR-10 · CIFAR-10 with 40% and 60% noisy labels | test accuracy | Improved |
| Use LID-based loss adaptation to regulate training | D2L on CIFAR-10 · CIFAR-10 with 40% and 60% noisy labels | representation subspace dimensionality (LID) | Improved |
| Vary neighborhood size k and batch count m for LID estimation (ablation) | D2L on CIFAR-10 · CIFAR-10, varying noise rates | test accuracy | No clear effect |
Closest papers by meaning
- Understanding and Utilizing Deep Neural Networks Trained with Noisy Labels (ICML 2019 · similarity 0.88)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.87)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.85)
- Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection (CVPR 2024 · similarity 0.85)
- On the Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics (ICML 2025 · similarity 0.85)
- How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
