Research map / Label-efficient and robust learning / Recognition and long-tail
Weakly-Supervised Residual Evidential Learning for Multi-Instance Uncertainty Estimation
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 MIREL, a weakly-supervised residual evidential learning scheme that gives Dirichlet-based bag- and instance-level uncertainty estimates for multiple instance learning (MIUE).
Problem
Uncertainty estimation for multiple instance learning, where only bag labels are available and instance-level uncertainty must be estimated without instance labels.
Approach
Derives an instance estimator g∘f from a permutation-invariant bag classifier (via the Fundamental Theorem of Symmetric Functions), adds an instance-specific residual head trained with a weakly-supervised evidential loss (weighted evidence aggregation over positive bags), and models bag- and instance-level predictions as Dirichlet distributions with a single forward pass.
Key results
- On MNIST-bags with ABMIL, MIREL raises average instance-level UE AUROC from 59.20 to 78.63 and bag-level UE from 81.41 to 86.47.
- On CAMELYON16, MIREL achieves average bag-level UE 72.06 and instance-level UE 72.85, versus 64.95 and 68.33 for the best competing UE method (Bayes-MIL at instance level).
- MIREL improves instance-level confidence AUROC by about 16.49% on average when added to existing MIL networks on MNIST-bags.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MIREL to ABMIL | ABMIL · MNIST-bags, instance level | average UE AUROC: 59.20 -> 78.63 | Helped |
| Add MIREL to ABMIL | ABMIL · MNIST-bags, bag level | average UE AUROC: 81.41 -> 86.47 | Helped |
| Add MIREL to ABMIL | ABMIL · CAMELYON16, bag level | average UE AUROC: 56.98 -> 72.06 | Helped |
| Add MIREL to ABMIL | ABMIL · CAMELYON16, instance level | average UE AUROC: 40.60 -> 72.85 | Helped |
| Replace attention score a_k with derived estimator T (ablation) | ABMIL instance estimator · MNIST-bags, instance level | average UE AUROC: 59.20 -> 67.20 | Helped |
| Add I-EDL loss while using derived estimator T (ablation) | ABMIL instance estimator T · MNIST-bags, instance level | average UE AUROC: 67.20 -> 72.60 | Helped |
| Add I-EDL loss while using attention score a_k (ablation) | ABMIL instance estimator · MNIST-bags, instance level | average UE AUROC: 59.20 -> 59.59 | No clear effect |
| Add residual estimator R trained with MIREL loss (ablation) | I-EDL with ABMIL estimator T · MNIST-bags, instance level | average UE AUROC: 72.60 -> 78.63 | Helped |
| Add residual estimator R trained with MIREL loss (ablation) | I-EDL with ABMIL estimator T · MNIST-bags, bag level | average UE AUROC: 81.82 -> 86.47 | Helped |
Surprising findings
- The attention-based instance scores of ABMIL are a poor basis for instance uncertainty; the instance estimator derived from the bag classifier (g∘f) already improves instance UE substantially over attention scores.
- General UE methods not designed for MIL often yield instance- and near-OOD AUROC below 0.5 on CAMELYON16.
Limitations
- Datasets are limited in scale (bags are expensive); larger datasets with more than 10,000 bags are needed for validation.
- Only AUROC is used; calibration metrics such as ECE or Brier score are not reported.
- Scope is binary MIL; multi-label MIL is left for future work.
Experimental setup
Datasets: MNIST-bags CIFAR10-bags CAMELYON16 FMNIST-bags KMNIST-bags SVHN-bags Texture-bags TCGA-PRAD
Benchmarks: MNIST-bags CIFAR10-bags CAMELYON16
Metrics: AUROC accuracy (reference only)
Baselines: Mean pooling MIL Max pooling MIL DSMIL Deep Ensemble MC Dropout I-EDL Bayes-MIL
Builds on: evidential deep learning attention-based MIL Fisher information-based EDL (I-EDL) RED regularization Fundamental Theorem of Symmetric Functions (deep sets)
Closest papers by meaning
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.87)
- Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? (NeurIPS 2024 · similarity 0.84)
- Second-Order Uncertainty Quantification: A Distance-Based Approach (ICML 2024 · similarity 0.84)
- Benchmarking Uncertainty Disentanglement: Specialized Uncertainties for Specialized Tasks (NeurIPS 2024 · similarity 0.84)
- Uncertainty Regularized Evidential Regression (AAAI 2024 · similarity 0.83)
- The Evidence Contraction Issue in Deep Evidential Regression: Discussion and Solution (AAAI 2024 · similarity 0.83)
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.