Research map / Label-efficient and robust learning / OOD and anomaly detection
On the Importance of Gradients for Detecting Distributional Shifts in the Wild
Rui Huang, Andrew Geng, Yixuan Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use gradient norm of KL(uniform||softmax) instead of Energy score | Energy · ImageNet-1k OOD (ResNetv2-101) | FPR95: 71.03 -> 54.70 | Improved |
| use GradNorm instead of Energy score on DenseNet-121 | Energy · ImageNet-1k OOD (DenseNet-121) | FPR95: 51.29 -> 41.00 | Improved |
| use gradients from last FC layer only instead of all parameters (ablation) | GradNorm · ImageNet-1k OOD avg | FPR95: 69.35 -> 54.70 | Improved |
| use gradients from last layer instead of Block 1 (ablation) | GradNorm · ImageNet-1k OOD avg | FPR95: 73.52 -> 54.70 | Improved |
| use uniform target (all labels) instead of one-hot target for backpropagation (ablation) | GradNorm · SUN | FPR95: reduces FPR95 by 48.35% | Improved |
| use L1 norm instead of higher-order Lp norms (ablation) | GradNorm · ImageNet-1k OOD | FPR95: up to 22.31% FPR95 over L2 | Improved |
| change temperature from T=1 (ablation) | GradNorm · ImageNet-1k OOD | FPR95 | Hurt |
| use gradient norm instead of KL divergence directly from output space (ablation) | KL divergence score · ImageNet-1k OOD | FPR95 | Improved |
| combine feature term U and output term V (U*V) instead of either alone (ablation) | GradNorm · ImageNet-1k OOD avg | FPR95: 71.00 -> 54.70 (U alone), 73.18 -> 54.70 (V alone) | Improved |
| increase backbone capacity beyond ResNetv2-101x1 (ablation) | GradNorm · ImageNet-1k OOD avg | FPR95: 55.84 -> 66.20 (101x3) | Hurt |
| use GradNorm instead of Mahalanobis | Mahalanobis · ImageNet-1k OOD (ResNetv2-101) | FPR95: 81.69 -> 54.70 | Improved |
Closest papers by meaning
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.90)
- Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity (NeurIPS 2025 · similarity 0.90)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.89)
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · similarity 0.88)
- ConjNorm: Tractable Density Estimation for Out-of-Distribution Detection (ICLR 2024 · similarity 0.88)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.88)
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.
