Research map / Label-efficient and robust learning / OOD and anomaly detection

Neuron Activation Coverage: Rethinking Out-of-distribution Detection and Generalization

ICLR 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Out-of-distribution generalization theory · 4 citations (OpenAlex)

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 neuron activation coverage (NAC), a per-neuron coverage measure over KL-gradient-weighted neuron states computed on InD training data, and applies it to OOD detection (NAC-UE) and model selection for OOD generalization (NAC-ME).

Problem

Out-of-distribution (OOD) inputs cause well-trained classifiers to fail; existing neuron-based OOD methods either modify networks or use binary neuron activation thresholds that discard distributional detail, and InD validation accuracy is a weak criterion for selecting models that generalize OOD.

Approach

Define a neuron activation state as the neuron output multiplied by the gradient of a KL divergence between the softmax output and a uniform vector, passed through a sigmoid, then measure per-neuron coverage from the histogram density of InD training states. Averaging coverage gives an OOD uncertainty score (NAC-UE), and integrating coverage gives a model quality score (NAC-ME).

Key results

• NAC-UE reduces FPR95 by 10.60% over the strongest rival (ViM) on CIFAR-100 and achieves 86.98% average AUROC there, the best average across the three benchmarks.
• NAC-ME improves rank correlation with OOD test accuracy over the validation criterion by 11.61% on average for ViT-b16 on DomainBed (e.g., +19.30 on TerraInc).
• NAC-ME improves average OOD accuracy over validation selection for all four backbones, e.g., 61.77 to 62.47 on ResNet-18 and 73.03 to 74.10 on ViT-b16.

Measured effects reported in the paper

Surprising findings

• Using only one layer (layer4) already gives NAC-UE 23.50% FPR95 on CIFAR-10, beating the previous best KNN.
• Rank correlation of NAC-ME on VLCS with ResNet-18 only reaches about 50%, which the authors attribute to too few OOD test domains.

Limitations

• NAC-UE is sensitive to hyperparameters r (or O*), steepness alpha, and bin count M, which must be tuned on validation sets including OOD validation data.
• Rank correlation for NAC-ME is limited when only a few OOD test domains are available, as on DomainBed.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 ImageNet-1K MNIST SVHN Textures Places365 iNaturalist OpenImage-O PACS VLCS OfficeHome

Benchmarks: CIFAR-10 OpenOOD far-OOD CIFAR-100 OpenOOD far-OOD ImageNet-1K OpenOOD far-OOD DomainBed iWildCam

Metrics: FPR95 AUROC Spearman rank correlation OOD test accuracy

Baselines: ViM KNN ReAct ASH SHE GEN RMDS MDS OpenMax

Builds on: neuron coverage (system testing) ReAct ViM OpenOOD DomainBed

Closest papers by meaning

• Benchmarking Out-of-Distribution Generalization Capabilities of DNN-based Encoding Models for the Ventral Visual Cortex. (NeurIPS 2024 · similarity 0.88)
• Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.88)
• Advancing Out-of-Distribution Detection via Local Neuroplasticity (ICLR 2025 · similarity 0.87)
• Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.86)
• Measure gradients, not activations! Enhancing neuronal activity in deep reinforcement learning (NeurIPS 2025 · similarity 0.86)
• Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations (NeurIPS 2025 · similarity 0.86)