Research map / Label-efficient and robust learning / Recognition and long-tail
Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty
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.
Proposes Hyper-Evidential Neural Network (HENN), which places a grouped Dirichlet distribution over class probabilities to learn from composite labels and quantify a new "vagueness" uncertainty, trained with a UPCE loss plus KL regularizer.
Problem
Deep classifiers trained on composite (set-valued) labels from ambiguous images cannot quantify uncertainty arising from composite labels in the training data; existing partial-label, conformal, and evidential methods do not address this.
Approach
Model predictions as hyper-opinions in Subjective Logic via a grouped Dirichlet distribution, so the network outputs evidence for singleton classes and composite sets; define a vagueness measure (total belief mass on composite sets) and train with an uncertainty partial cross-entropy loss plus a KL regularizer.
Key results
- On tinyImageNet with M=10 composite labels, HENN reaches OverJS 84.4 and CompJS 93.4 versus 83.4 and 66.9 for DNN.
- CompJS improves by over 15-20% across the three datasets relative to the best baseline reported in the main table.
- Singleton accuracy improves by roughly 2-5% over baselines on the three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the proposed KL regularization to UPCE (ablation) | HENN · Nonliving26, M=15, 7×7 blur | OverJS: 78.25 -> 86.93 | Helped |
| Add the proposed KL regularization to UPCE (ablation) | HENN · Nonliving26, M=15, 7×7 blur | CompJS: 62.89 -> 94.78 | Helped |
| Add the proposed KL regularization to UPCE (ablation) | HENN · Nonliving26, M=15, 7×7 blur | Acc: 84.96 -> 85.19 | Mixed |
| Restrict KL regularization to singleton classes (ablation) | HENN · Nonliving26, M=15, 7×7 blur | CompJS: 94.78 -> 94.15 | Hurt |
| Replace KL regularization with GDD entropy regularization (ablation) | HENN · Nonliving26, M=15, 7×7 blur | OverJS: 86.93 vs 86.68 | No clear effect |
| Use HENN instead of the strongest reported baseline (DNN) | composite prediction · tinyImageNet, M=10, 3×3 blur | CompJS: 66.9 -> 93.4 | Helped |
| Use HENN instead of the strongest reported baseline (DNN) | composite prediction · Living17, M=10, 3×3 blur | CompJS: 81.0 -> 96.5 | Helped |
| Use HENN instead of the strongest reported baseline (DNN) | composite prediction · Nonliving26, M=10, 3×3 blur | CompJS: 62.0 -> 96.8 | Helped |
Surprising findings
- RAPS and E-CNN underperform even plain DNN on composite-set quality because they over-predict sets for hard non-vague images.
Limitations
- The universal approximation assumption behind the propositions may not hold in practice, as the authors acknowledge.
- Combinatorial growth of composite sets (2^K) is avoided by considering only a limited number of composite sets.
- Vagueness is tied to annotator labeling bias, which the synthetic blur-based construction only approximates.
Experimental setup
Datasets: TinyImageNet Living17 Nonliving26 CIFAR100
Benchmarks: TinyImageNet Living17 Nonliving26 CIFAR100
Metrics: OverJS CompJS Accuracy AUROC
Baselines: DNN ENN E-CNN RAPS PiCO
Builds on: evidential deep learning subjective logic uncertainty cross-entropy loss partial cross-entropy loss
Closest papers by meaning
- Hyper-opinion Evidential Deep Learning for Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.87)
- Uncertainty Regularized Evidential Regression (AAAI 2024 · similarity 0.86)
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.86)
- Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.86)
- Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? (NeurIPS 2024 · similarity 0.85)
- Random-Set Neural Networks (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.