Research map / Label-efficient and robust learning / OOD and anomaly detection
Exploring Channel-Aware Typical Features for Out-of-Distribution Detection
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 LAPS, a test-time OOD detector that replaces BATS's fixed per-channel typical-set proportion with channel-aware typical sets calibrated by channel mean and std, then computes an energy score.
Problem
Test-time out-of-distribution detection in classifiers is hurt by abnormal hidden activations, and existing typical-set methods assume a fixed proportion of typical features for every channel even though channels contribute differently.
Approach
Calibrate each channel's typical-set bounds (lambda) using the deviation of its mean and standard deviation from global statistics, rectify activations into channel-aware typical sets, and compute the energy score on the rectified features; claimed to give a better bias-variance trade-off than BATS.
Key results
- On ImageNet-1K with ResNet50, LAPS reduces average FPR95 to 23.68% from BATS's 30.16% and raises average AUROC to 94.78% from 93.59%.
- On CIFAR-100, LAPS achieves average FPR95 60.78% and AUROC 85.66%, versus BATS at 65.85% and 85.01%.
- With MobileNet-V2 on ImageNet-1K, LAPS lowers average FPR95 by 18.12 points relative to BATS (34.99% vs 53.11%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace BATS fixed typical sets with channel-aware sets calibrated from mean and standard deviation | Energy OOD score with ResNet50 · ImageNet-1K ID, OOD average | FPR95: 30.16 -> 23.68 | Helped |
| replace BATS fixed typical sets with channel-aware sets calibrated from mean and standard deviation | Energy OOD score with ResNet50 · ImageNet-1K ID, OOD average | AUROC: 93.59 -> 94.78 | Helped |
| use channel-aware typical features instead of BATS fixed typical features | Energy OOD score with MobileNet-V2 · ImageNet-1K ID, OOD average | FPR95: 53.11 -> 34.99 | Helped |
| use channel-aware typical features instead of BATS fixed typical features | Energy OOD score with ResNet18 · ImageNet-1K ID, OOD average | FPR95: 41.90 -> 26.11 | Helped |
| use channel-aware typical features instead of BATS fixed typical features | Energy OOD score · CIFAR-100 ID, OOD average | FPR95: 65.85 -> 60.78 | Helped |
| add calibration from mean to BATS (ablation) | BATS with Energy score and ResNet50 · ImageNet-1K ID, OOD average | FPR95: 30.16 -> 26.73 | Helped |
| add calibration from standard deviation to BATS (ablation) | BATS with Energy score and ResNet50 · ImageNet-1K ID, OOD average | FPR95: 30.16 -> 29.33 | Helped |
| add calibration from mean to BATS + variance (ablation) | BATS + variance with Energy score and ResNet50 · ImageNet-1K ID, OOD average | FPR95: 29.33 -> 23.68 | Helped |
| add calibration from standard deviation to BATS + mean (ablation) | BATS + mean with Energy score and ResNet50 · ImageNet-1K ID, OOD average | FPR95: 26.73 -> 23.68 | Helped |
| vary n across 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0 (ablation) | LAPS · OOD detection benchmark in Table 5 | FPR95: 29.62, 29.49, 29.33, 29.43, 29.43, 29.46 | No clear effect |
| replace BATS with LAPS for the MSP score | MSP on ImageNet-1K with ResNet50 · ImageNet-1K ID, OOD average | FPR95: 53.89 -> 44.73 | Helped |
| replace BATS with LAPS for the ODIN score | ODIN on ImageNet-1K with ResNet50 · ImageNet-1K ID, OOD average | FPR95: 38.54 -> 34.40 | Helped |
Surprising findings
- Fixing the typical-set proportion per channel is shown to hurt OOD detection, since channels contribute differently to OOD detection.
Limitations
- Relies on three hyperparameters (lambda, m, n) whose selection affects the bias-variance trade-off.
- Theoretical analysis assumes Gaussian-distributed deep features, following BATS.
Experimental setup
Datasets: CIFAR-100 ImageNet-1K Textures SVHN LSUN iSUN iNaturalist SUN Places
Benchmarks: CIFAR-100 OOD benchmark (Textures, SVHN, LSUN, iSUN) ImageNet-1K OOD benchmark (iNaturalist, SUN, Places, Textures)
Metrics: FPR95 AUROC
Baselines: MSP ODIN Energy ReAct BFAct BATS
Builds on: BATS Energy score ReAct BFAct
Closest papers by meaning
- Discriminability-Driven Channel Selection for Out-of-Distribution Detection (CVPR 2024 · similarity 0.90)
- Test-Time Linear Out-of-Distribution Detection (CVPR 2024 · similarity 0.89)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.88)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.88)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.88)
- Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection (AAAI 2024 · similarity 0.87)
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.