Research map / Label-efficient and robust learning / Semi-supervised learning
ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ANEDL, which adds an evidential deep learning (EDL) outlier detector with adaptive negative optimization to a FixMatch-style open-set semi-supervised learner, using different uncertainty metrics for inlier selection and inference.
Problem
Open-set semi-supervised learning, where unlabeled and test data contain outliers from unseen classes that can corrupt self-training; existing softmax/binary one-vs-all outlier detectors scale poorly with many classes and cannot distinguish epistemic from aleatoric uncertainty.
Approach
Replace binary/softmax outlier detectors with a multinomial EDL head trained with an adaptive negative loss: a Fisher-information-weighted KL term pushes evidence down on unlabeled samples for uncertain classes while sparing confident (inlier) classes, and uses a separate uncertainty metric for self-training (evidence of predicted class) and inference (sum of top-M evidence).
Key results
- On CIFAR-100 (55/45 split, 50 labels), ANEDL reaches 90.3% AUROC versus 87.0% for OpenMatch, and 26.7% error versus 27.7%.
- ANEDL reports best error rate in 7 of 8 settings in the main comparison, per the authors.
- Ablation on CIFAR-100: removing both negative optimization and adaptive weighting drops AUROC to 88.4%, adding NO gives 89.0%, and full ANO gives 90.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ANEDL instead of OpenMatch | Open-set SSL · CIFAR-100, 55/45 classes, 50 labels | AUROC (%): 87.0±1.1 -> 90.3±0.1 | Helped |
| use ANEDL instead of OpenMatch | Open-set SSL · CIFAR-100, 55/45 classes, 50 labels | Error rate (%): 27.7±0.4 -> 26.7±0.2 | Helped |
| remove adaptive loss weights while retaining negative optimization (ablation) | ANEDL on CIFAR-100, 55 inlier classes and 50 labels per class · CIFAR-100 | AUROC: 90.3±0.1 -> 89.0±0.3 | Hurt |
| remove adaptive loss weights while retaining negative optimization (ablation) | ANEDL on CIFAR-100, 55 inlier classes and 50 labels per class · CIFAR-100 | Error rate: 26.7±0.2 -> 26.9±0.3 | Hurt |
| add negative optimization and adaptive weighting instead of classical EDL loss (ablation) | ANEDL on CIFAR-100, 55 inlier classes and 50 labels per class · CIFAR-100 | AUROC: 88.4±0.4 -> 90.3±0.1 | Helped |
| use M_Inference for self-training and inference instead of M_Self-training for both (ablation) | ANEDL on CIFAR-100 · CIFAR-100 | AUROC: 88.4±0.3 -> 89.8±0.2 | Helped |
| use M_Inference for self-training and inference instead of M_Self-training for both (ablation) | ANEDL on CIFAR-100 · CIFAR-100 | Error rate: 26.7±0.2 -> 27.8±0.2 | Hurt |
| use M_Self-training for self-training and M_Inference for inference (ablation) | ANEDL on CIFAR-100 · CIFAR-100 | AUROC: 88.4±0.3 -> 90.3±0.1 | Helped |
| use top-M=25 evidence sum instead of all K=55 evidence values for inference (ablation) | ANEDL on CIFAR-100 · CIFAR-100 | AUROC | Helped |
| set target evidence P=200 instead of P=100 in strengthened KL divergence (ablation) | ANEDL on CIFAR-100 · CIFAR-100 | AUROC: 90.3±0.1 -> 89.9±0.9 | Hurt |
| remove DebiasPL (ablation) | ANEDL on CIFAR-100 · CIFAR-100 | AUROC: 90.3±0.1 -> 90.0±0.5 | Hurt |
Surprising findings
- Using the sum of top-M evidence values (M=25) rather than total evidence for inference improves outlier detection over the sum of all K=55 evidence values.
Limitations
- Mini-ImageNet baseline numbers for OpenMatch were produced by the authors' own reimplementation, while other baselines are quoted from OpenMatch.
- Several hyperparameters (M, P, lambda terms) are tuned and sensitive; too large P destabilizes training.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-30 Mini-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-30 Mini-ImageNet
Metrics: AUROC error rate
Baselines: FixMatch MTC T2T OpenMatch
Builds on: OpenMatch evidential deep learning I-EDL FixMatch DebiasPL
Closest papers by meaning
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.88)
- Hyper-opinion Evidential Deep Learning for Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.87)
- Uncertainty Estimation by Density Aware Evidential Deep Learning (ICML 2024 · similarity 0.87)
- Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? (NeurIPS 2024 · similarity 0.87)
- Calibrating LLMs with Information-Theoretic Evidential Deep Learning (ICLR 2025 · similarity 0.86)
- Uncertainty Regularized Evidential Regression (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.