Research map / Label-efficient and robust learning / Recognition and long-tail
Learning with Biased Complementary Labels
Xiyu Yu, Tongliang Liu, Mingming Gong, Dacheng Tao
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 the estimated transition matrix instead of the true one (ablation) | LM · CIFAR10 (uniform) | accuracy (%): 73.38 -> 42.96 | Hurt |
| use the estimated transition matrix instead of the true one (ablation) | LM · CIFAR10 (with0) | accuracy (%): 85.32 -> 84.60 | No clear effect |
| use the estimated transition matrix instead of the true one (ablation) | LM · MNIST (three settings) | accuracy (%): 98.63 -> 98.61 | No clear effect |
| model biased (non-uniform) complementary labels with a transition matrix, versus the uniform assumption PC/S | LM · CIFAR10 (with0) | accuracy (%): 18.12 -> 85.32 | Improved |
| replace PC/S with LM under the uniform setting | LM/T · MNIST (uniform) | accuracy (%): 86.59 -> 97.18 | Improved |
| complementary labels drawn from a small subset of classes (with0) rather than uniformly | LM/T · CIFAR10 | accuracy (%): 73.38 -> 85.32 | Improved |
| complementary labels drawn from a small subset of classes (with0) rather than uniformly | LM/T · MNIST | accuracy (%): 97.18 -> 98.63 | Improved |
| use a non-invertible transition matrix instead of an invertible one (ablation) | LM/T · MNIST (with0) | accuracy (%): 98.63 -> 98.65 | No clear effect |
| use a manually designed Q instead of a randomly generated one (ablation) | LM/T · MNIST (without0) | accuracy (%): 97.65 -> 97.15 | No clear effect |
| use a manually designed Q instead of a randomly generated one (ablation) | LM/T · MNIST (with0) | accuracy (%): 98.63 -> 98.56 | No clear effect |
| replace PC/S with LM on CIFAR100 | LM/T · CIFAR100 (with0) | accuracy (%): 8.95 -> 62.84 | Improved |
| replace PC/S with LM on UCI PENDIGITS, all 10 classes | LM · PENDIGITS (classes 1 ~ 10) | accuracy (%): 38.0 -> 62.8 | Improved |
Closest papers by meaning
- Learning from Complementary Labels (NeurIPS 2017 · similarity 0.90)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.90)
- Complementary-Label Learning for Arbitrary Losses and Models (ICML 2019 · similarity 0.89)
- Complementary Label Learning with Positive Label Guessing and Negative Label Enhancement (ICLR 2025 · similarity 0.89)
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.86)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 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.
