Research map / Label-efficient and robust learning / OOD and anomaly detection
Discriminability-Driven Channel Selection 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 DDCS, a test-time OOD detection method that selects channels of the penultimate layer by an ID-prototype discriminability score (inter-class similarity and variance), after mildly pre-rectifying anomalous activations, then scores with Energy.
Problem
Activation-based test-time OOD detection methods assume every penultimate-layer channel is useful and rectify anomalous activations in all of them, even though channels contribute unequally to OOD detection.
Approach
Select channels with high ID class discriminability (low inter-class prototype similarity and high inter-class variance) without needing OOD data, and pre-rectify activations locally (BATS-style) and globally (ReAct-style) before estimating the score; the selected channels are then used for the Energy score.
Key results
- On CIFAR-10 with DenseNet-101, DDCS reaches average FPR95 13.57% and AUROC 97.22%, versus 14.75% / 96.99% for LINe.
- On ImageNet-1K with ResNet50, DDCS achieves average FPR95 19.36% and AUROC 95.55%, outperforming LINe (20.69% / 95.03%).
- Plugged into MSP and Energy, DDCS lowers average FPR95 further than ReAct, BATS and VRA (e.g., Energy + DDCS 19.36% vs Energy + VRA 25.49%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DDCS channel selection with Energy | DenseNet-101 · CIFAR-10, six OOD datasets (average) | FPR95: 14.75 -> 13.57 | Helped |
| Use DDCS channel selection with Energy | DenseNet-101 · CIFAR-100, six OOD datasets (average) | FPR95: 35.67 -> 32.83 | Helped |
| Use DDCS channel selection with Energy | ResNet50 · ImageNet-1K, four OOD datasets (average) | FPR95: 20.69 -> 19.36 | Helped |
| Add CAP pre-rectification (ablation) | DDCS · CIFAR-10, six OOD datasets (average) | FPR95: 23.65 -> 13.57 | Helped |
| Add CAP pre-rectification (ablation) | DDCS · CIFAR-10, six OOD datasets (average) | AUROC: 95.32 -> 97.22 | Helped |
| Add CS channel selection (ablation) | DDCS · CIFAR-10, six OOD datasets (average) | FPR95: 14.70 -> 13.57 | Helped |
| Add CS channel selection (ablation) | DDCS · CIFAR-10, six OOD datasets (average) | AUROC: 97.01 -> 97.22 | Helped |
| Add CAP and CS (ablation) | Energy OOD detection · CIFAR-10, six OOD datasets (average) | FPR95: 26.55 -> 13.57 | Helped |
| Add CAP and CS (ablation) | Energy OOD detection · CIFAR-10, six OOD datasets (average) | AUROC: 94.57 -> 97.22 | Helped |
| Use DDCS channel selection | MobileNetV2 with LHAct CAP · ImageNet-1K, four OOD datasets (average) | FPR95: 40.91 -> 22.61 | Helped |
| Use DDCS channel selection | MobileNetV2 with LHAct CAP · ImageNet-1K, four OOD datasets (average) | AUROC: 90.61 -> 94.84 | Helped |
Surprising findings
- Discarding some penultimate-layer channels can substantially improve OOD detection; using all channels is worse than selecting a subset.
- On CIFAR-100, DDCS does not beat LINe on average FPR95 (32.83% vs 35.67% is better, but LINe is better on some sets such as SVHN and LSUN-resize in FPR95).
Limitations
- Relies on a fixed number of selected channels K and balance factor a, which are tuned via sensitivity analysis.
- Gains on CIFAR-100 and some OOD sets (e.g., Places365) are mixed, with DDCS not best on every dataset.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-1K
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-1K SVHN Textures iSUN LSUN-resize LSUN-crop Places365 iNaturalist SUN
Metrics: FPR95 AUROC
Baselines: MSP ODIN Mahalanobis Energy ReAct BATS DICE SHE VRA LINe LHAct
Builds on: ReAct BATS Energy score LINe LHAct
Closest papers by meaning
- Exploring Channel-Aware Typical Features for Out-of-Distribution Detection (AAAI 2024 · similarity 0.90)
- Learning to Shape In-distribution Feature Space for Out-of-distribution Detection (NeurIPS 2024 · similarity 0.89)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.89)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.89)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 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.