Research map / Label-efficient and robust learning / OOD and anomaly detection
Mitigating Neural Network Overconfidence with Logit Normalization
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 |
|---|---|---|---|
| Replace cross-entropy training with LogitNorm | WRN-40-2 using softmax confidence · CIFAR-10, six OOD datasets (average) | FPR95: 49.52 -> 15.65 | Helped |
| Replace cross-entropy training with LogitNorm | WRN-40-2 using softmax confidence · CIFAR-10, SVHN | FPR95: 50.33 -> 8.03 | Helped |
| Train with LogitNorm before applying ODIN | WRN-40-2 + ODIN · CIFAR-10, six OOD datasets (average) | FPR95: 40.32 -> 12.95 | Helped |
| Train with LogitNorm before applying Energy scoring | WRN-40-2 + Energy · CIFAR-10, six OOD datasets (average) | FPR95: 26.82 -> 19.14 | Helped |
| Train with LogitNorm before applying GradNorm scoring | WRN-40-2 + GradNorm · CIFAR-10, six OOD datasets (average) | FPR95: 58.98 -> 17.78 | Helped |
| Replace logit normalization with an L2 logit-norm penalty (ablation) | WRN-40-2 trained on CIFAR-10 · CIFAR-10, six OOD datasets (average) | FPR95: 15.65 -> 57.62 | Hurt |
| Use LogitNorm instead of GODIN | WRN-40-2 on CIFAR-10 · CIFAR-10, six OOD datasets (average) | FPR95: 25.24 -> 15.65 | Helped |
| Add post-hoc temperature scaling to LogitNorm | WRN-40-2 trained on CIFAR-10 with LogitNorm · CIFAR-10 | ECE: 66.95 -> 0.41 | Helped |
| Increase the LogitNorm temperature parameter tau (ablation) | LogitNorm · CIFAR-10, six OOD datasets (average) | FPR95 | Hurt |
Closest papers by meaning
- MaNo: Exploiting Matrix Norm for Unsupervised Accuracy Estimation Under Distribution Shifts (NeurIPS 2024 · similarity 0.88)
- Simplification Is All You Need against Out-of-Distribution Overconfidence (CVPR 2025 · similarity 0.86)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.84)
- Boosting Adversarial Training via Fisher-Rao Norm-based Regularization (CVPR 2024 · similarity 0.84)
- Deep Neural Networks Tend To Extrapolate Predictably (ICLR 2024 · similarity 0.83)
- Typicalness-Aware Learning for Failure Detection (NeurIPS 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.