Research map / Label-efficient and robust learning / Recognition and long-tail
Discriminative Complementary-Label Learning with Weighted Loss
Yi Gao, Min-Ling Zhang
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 |
|---|---|---|---|
| Add weighted loss (L-W vs L-UW) to discriminative CLL model (ablation) | discriminative CLL model · unbiased complementary labels, 6 dataset/model combos | test accuracy: +0.24 MNIST linear, -0.37 MNIST MLP, +0.25 Fashion linear, +0.25 Fashion MLP, +0.42 Kuzushiji linear, +1.81 Kuzushiji ML | Mixed |
| Add weighted loss to discriminative CLL model (ablation) | discriminative CLL model · biased complementary labels (Table 4, sets 1-3) | test accuracy: L-W below L-UW on MNIST linear set 3 (79.72 vs 81.16) and MNIST MLP set 3 (85.91 vs 84.98 is above; set 3 Kuzushiji MLP | Mixed |
| Discriminative complementary-loss model (L-UW) without weighting vs PC, NN, GA baselines | discriminative CLL model · unbiased complementary labels, MNIST/Fashion/Kuzushiji | test accuracy: L-UW best or near-best in most cases, e.g. Kuzushiji linear 60.87 vs Forward 60.57 | Improved |
| Full L-W vs strongest baseline (Forward) on linear models | discriminative CLL model · MNIST linear unbiased | test accuracy: L-W 90.22 vs Forward 90.42 (below Forward) | Mixed |
| Use MLP instead of linear model with GA baseline | GA baseline · Kuzushiji unbiased | test accuracy: GA 69.56 MLP vs 52.52 linear, yet L-W leads on linear | Mixed |
| Discriminative model (L-W) vs Forward on biased labels with linear model | discriminative CLL model · biased complementary labels, Kuzushiji linear set 3 | test accuracy: +3.44 (57.85 vs 54.41) | Improved |
| Empirical risk of URE methods (PC, NN) with MLP goes to zero or negative, while GA stays stable | URE-based CLL losses · MLP training, unbiased | test accuracy | Hurt |
Closest papers by meaning
- Complementary Label Learning with Positive Label Guessing and Negative Label Enhancement (ICLR 2025 · similarity 0.93)
- Multi-Label Learning With Multiple Complementary Labels (TPAMI 2025 · similarity 0.91)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.90)
- Partial Label Learning with a Partner (AAAI 2024 · similarity 0.87)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.87)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.86)
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.
