Research map / Label-efficient and robust learning / OOD and anomaly detection
LINe: Out-of-Distribution Detection by Leveraging Important Neurons
Yong Hyun Ahn, Gyeong-Moon Park, Seong Tae Kim
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 LINe instead of DICE + ReAct | DenseNet · CIFAR-10 (six OOD datasets, average) | FPR95: 16.48 -> 14.71 | Improved |
| Use LINe instead of DICE + ReAct | DenseNet · CIFAR-100 (six OOD datasets, average) | FPR95: 49.57 -> 35.67 | Improved |
| Use LINe instead of DICE + ReAct | ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 27.25 -> 20.70 | Improved |
| Add activation clipping (ablation) | Energy OOD scoring on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 58.41 -> 35.40 | Improved |
| Add activation pruning to activation clipping (ablation) | LINe on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 26.88 -> 20.70 | Improved |
| Add weight pruning to activation clipping (ablation) | LINe on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 23.19 -> 20.70 | Improved |
| Set activation clipping threshold δ to 0.8 rather than no clipping (ablation) | LINe on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 44.88 -> 20.70 | Improved |
| Lower clipping threshold from δ = 0.8 to δ = 0.1 (ablation) | LINe on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 20.70 -> 41.18 | Hurt |
| Raise clipping threshold from δ = 0.8 to δ = 2.0 (ablation) | LINe on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 20.70 -> 31.88 | Hurt |
| Set pruning percentiles p_w = 10 and p_a = 10 (ablation) | LINe on ResNet-50 · ImageNet (four OOD datasets, average) | FPR95: 27.64 -> 20.70 | Improved |
| Replace Taylor with IntGrad for Shapley approximation (ablation) | LINe on DenseNet and ResNet-50 · CIFAR-10, CIFAR-100, and ImageNet | FPR95: 14.71 -> 14.71; 35.67 -> 35.67; 20.70 -> 20.70 | No clear effect |
Later papers that compared with this method
4 later accepted papers report Line as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- Improving Out-of-Distribution Detection via Dynamic Covariance Calibration (ICML 2025 · better on 1 of 1 shared result)
- Discriminability-Driven Channel Selection for Out-of-Distribution Detection (CVPR 2024 · better on 5 of 5 shared results)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · better on 4 of 4 shared results)
- Gradient-based Out-of-Distribution Detection (ECCV 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Optimal Parameter and Neuron Pruning for Out-of-Distribution Detection (NeurIPS 2023 · similarity 0.91)
- Distribution-Aware Feature Selection for Post-hoc Out-of-Distribution Detection (ECCV 2026 · similarity 0.89)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.89)
- NECO: NEural Collapse Based Out-of-distribution detection (ICLR 2024 · similarity 0.88)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.88)
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · 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.
