Research map / Label-efficient and robust learning / Recognition and long-tail
Learning from Similarity-Confidence Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Sconf risk correction (ABS or NN) instead of the unbiased risk estimator (ablation) | Sconf classification · MNIST, Kuzushiji-MNIST, Fashion-MNIST, EMNIST-Digits, EMNIST-Letters, EMNIST-Balanced, CIFAR-10, and SVHN | accuracy: MNIST: 87.22± 2.11 -> 96.12± 2.31 (ABS), 96.04± 1.23 (NN); Kuzushiji-MNIST: 78.12± 3.08 -> 89.25± 1.58 (ABS), 90.00± 0.5 | Helped |
| Use Sconf risk correction (ABS or NN) instead of the unbiased risk estimator (ablation) | Sconf classification · EMNIST-Digits, EMNIST-Letters, EMNIST-Balanced, CIFAR-10, and SVHN | accuracy: EMNIST-Digits: 87.96±1.67 -> 96.62± 0.06 (ABS), 96.21± 0.11 (NN); EMNIST-Letters: 77.14±3.71 -> 86.32± 1.20 (ABS), 86.72 | Helped |
| Add Gaussian noise to similarity confidence (ablation) | Sconf linear classifier · Synthetic setups A–D | accuracy: Setup A: 89.91±0.18 -> 89.84±0.03 (std. = 0.1), 89.55±0.64 (std. = 0.2), 89.64±0.33 (std. = 0.3) | No clear effect |
| Add Gaussian noise to similarity confidence (ablation) | Sconf linear classifier · Synthetic setups A–D | accuracy: Setup B: 90.62±0.28 -> 90.34±0.40 (std. = 0.1), 90.03±0.33 (std. = 0.2), 89.49±0.92 (std. = 0.3) | No clear effect |
| Add Gaussian noise to similarity confidence (ablation) | Sconf linear classifier · Synthetic setups A–D | accuracy: Setup C: 88.05±0.30 -> 88.14±0.14 (std. = 0.1), 87.92±0.57 (std. = 0.2), 87.91±0.38 (std. = 0.3) | No clear effect |
| Add Gaussian noise to similarity confidence (ablation) | Sconf linear classifier · Synthetic setups A–D | accuracy: Setup D: 90.43±0.14 -> 90.29±0.30 (std. = 0.1), 90.40±0.15 (std. = 0.2), 90.20±0.14 (std. = 0.3) | No clear effect |
| Train from similarity-confidence data rather than hard similarity labels | Sconf methods versus SD · All eight benchmark datasets | accuracy: MNIST: Sconf-ABS 96.12± 2.31; SD 86.57± 0.78; Kuzushiji-MNIST: 89.25± 1.58; 76.42± 4.09; Fashion-MNIST: 91.44± 0.39; 83. | Helped |
| Use ABS rather than NN correction (ablation) | Sconf risk correction · Benchmark datasets | accuracy: ABS/NN: MNIST 96.12± 2.31/96.04± 1.23; Kuzushiji-MNIST 89.25± 1.58/90.00± 0.55; Fashion-MNIST 91.44± 0.39/91.37± 0.30; E | Mixed |
Closest papers by meaning
- Binary Classification with Confidence Difference (NeurIPS 2023 · similarity 0.89)
- Classification from Pairwise Similarity and Unlabeled Data (ICML 2018 · similarity 0.89)
- Confidence Difference Reflects Various Supervised Signals in Confidence-Difference Classification (ICML 2025 · similarity 0.88)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.85)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.84)
- Stabilizing Sample Similarity in Representation via Mitigating Random Consistency (ICML 2025 · similarity 0.84)
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.