Research map / Label-efficient and robust learning / OOD and anomaly detection
CORES: Convolutional Response-based Score 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 CORES, a training-free, ID-free post-hoc OOD score built from extreme (max positive / min negative) convolutional responses of sample-relevant kernels traced backward from the top and bottom predicted classes.
Problem
Deep classifiers are overconfident on out-of-distribution inputs; post-hoc OOD detection is needed that does not rely on training/ID data or auxiliary OOD samples.
Approach
Measures the magnitude and frequency of extreme convolutional responses (highest positive and lowest negative per kernel), restricted to kernels traced backward through the network from the highest- and lowest-confidence classes, since ID inputs elicit stronger kernel responses than OOD inputs.
Key results
- On CIFAR-10 with ResNet-18, CORES reduces average FPR95 by 19.97% relative to the next best method; for CIFAR-100 reductions are 27.99%, 18.67%, and 13.91% across the three CIFAR architectures.
- On near-OOD CIFAR-10 vs CIFAR-100 with WRN-28, CORES reaches 43.98 FPR95 and 91.64 AUROC, the best among compared methods.
- On ImageNet-1k with VGG-16 (common OOD), CORES averages 31.89 FPR95 and 92.53 AUROC, the best average FPR95 in that table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CORES instead of the strongest reported baseline | OOD detection with ResNet-18 · CIFAR-10 common OOD (average) | FPR95: 25.06 -> 5.09 | Helped |
| Use CORES instead of the strongest reported baseline | OOD detection with WRN-28 · CIFAR-10 common OOD (average) | FPR95: 8.62 -> 3.25 | Helped |
| Use CORES instead of the strongest reported baseline | OOD detection with ResNet-18 · CIFAR-100 common OOD (average) | FPR95: 70.63 -> 25.40 | Helped |
| Use CORES instead of the strongest reported baseline | OOD detection with VGG-16 · ImageNet-1k common OOD (average) | FPR95: 37.63 -> 31.89 | Helped |
| Use CORES instead of the strongest reported baseline | Near-OOD detection with VGG-16 · ImageNet-1k ID / ImageNet-O OOD | AUROC: 74.81 -> 79.10 | Helped |
| Remove response-frequency metrics (RF+ and RF−) (ablation) | CORES with WRN-28 · CIFAR-10 common OOD | FPR95: 3.81 -> 6.33 | Hurt |
| Remove response-magnitude metrics (RM+ and RM−) (ablation) | CORES with WRN-28 · CIFAR-10 common OOD | FPR95: 3.81 -> 5.04 | Hurt |
| Ignore negative convolutional responses (ablation) | CORES with WRN-28 · CIFAR-10 common OOD | FPR95: 3.81 -> 5.46 | Hurt |
| Ignore positive convolutional responses (ablation) | CORES with WRN-28 · CIFAR-10 common OOD | FPR95: 3.81 -> 6.33 | Hurt |
| Add sample-relevant kernel selection (ablation) | CORES · OOD detection across evaluated datasets | FPR95: approximately 55% | Helped |
| Use more convolutional layers in CORES (ablation) | CORES with DenseNet-101 · CIFAR-10 and CIFAR-100 common OOD | FPR95 | Helped |
| Include positive response metrics (ablation) | CORES with WRN-28 · CIFAR-10 common OOD | FPR95: approximately 3% | Helped |
Surprising findings
- Removing the positive-response metrics (RM+ and RF+) hurts more than removing negative-response metrics, with about 3% FPR95 increase in the ablation.
Limitations
- Tuned thresholds tau+ and tau- using Gaussian and uniform noise, which may not generalize to all OOD types.
- Performance on ImageNet-O near-OOD is only moderate (FPR95 72.20 for VGG-16), and gains are not uniform across all settings (e.g., some OOD datasets show weaker results).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-1k SVHN Textures LSUN-R iSUN TinyImageNet-R OpenImage-O SUN Places ImageNet-O
Benchmarks: CIFAR-10 OOD suite (SVHN, Textures, LSUN-R, iSUN, TinyImageNet-R) CIFAR-100 OOD suite ImageNet-1k OOD (OpenImage-O, SUN, Places, Textures) near-OOD CIFAR-10 vs CIFAR-100 near-OOD ImageNet-1k vs ImageNet-O
Metrics: AUROC FPR95
Baselines: MSP ODIN Energy ReAct DICE TFEM FeatureNorm
Builds on: Maximum softmax probability ReAct DICE FeatureNorm Energy score
Closest papers by meaning
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.88)
- Test-Time Linear Out-of-Distribution Detection (CVPR 2024 · similarity 0.86)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.86)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.86)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.85)
- OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary (CVPR 2025 · similarity 0.85)
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.