Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Trustworthy Predictions from Deep Neural Networks with Fast Adversarial Calibration
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 |
|---|---|---|---|
| combine predictive entropy loss and adversarial calibration loss | FALCON · MNIST and CIFAR10 across perturbation strategies | ECE: MNIST: 0.102 -> 0.082; CIFAR10: 0.137 -> 0.127 | Helped |
| use FALCON instead of Deep Ensembles | ResNet50 · ImageNet, 19 image corruptions | mean ECE: 0.056 vs. 0.036; 35% lower | Helped |
| use FALCON instead of Deep Ensembles | LSTM and GRU sequence models · Sequential MNIST and 20 Newsgroups | average ECE: 0.222 -> 0.118; 0.168 -> 0.108; 0.218 -> 0.158 | Helped |
| add predictive entropy loss L_S (ablation) | FALCON · MNIST, perturbation calibration | ECE | Helped |
| add adversarial calibration loss L_adv (ablation) | FALCON · MNIST, perturbation calibration | ECE | Helped |
| combine both loss terms rather than use either term alone (ablation) | FALCON · MNIST, perturbation calibration | ECE | Helped |
| vary lambda_adv across the tested range (ablation) | FALCON · MNIST y-zoom | ECE: varied by less than 0.04 | No clear effect |
| vary lambda_S across the tested range (ablation) | FALCON · MNIST y-zoom | ECE: varied by less than 0.06 | No clear effect |
| apply temperature scaling after training (ablation) | FALCON · MNIST under large domain shift | calibration | Hurt |
| train an ensemble of FALCON models (ablation) | FALCON · ImageNet, ObjectNet overlapping classes | accuracy: 0.754 -> 0.782 | Helped |
Closest papers by meaning
- Post-Hoc Uncertainty Calibration for Domain Drift Scenarios (CVPR 2021 · similarity 0.89)
- Better Uncertainty Calibration via Proper Scores for Classification and Beyond (NeurIPS 2022 · similarity 0.89)
- Towards Certification of Uncertainty Calibration under Adversarial Attacks (ICLR 2025 · similarity 0.88)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.86)
- Towards Trustworthy Deep Learning (AAAI 2024 · similarity 0.86)
- A Benchmark Study on Calibration (ICLR 2024 · 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.