Research map / Label-efficient and robust learning / OOD and anomaly detection
Advancing Out-of-Distribution Detection via Local Neuroplasticity
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 a post-hoc OOD detector that compares activations of a trained Kolmogorov-Arnold Network (KAN) against an untrained copy, exploiting KANs' local spline plasticity, and reports state-of-the-art average AUROC on seven image and tabular medical benchmarks.
Problem
Out-of-distribution detection is hard because no single post-hoc detector performs consistently across near and far OOD datasets, and many methods are sensitive to training set size.
Approach
Uses the local plasticity of spline-based KAN activations: a KAN trained on in-distribution data changes only the spline coefficients near InD inputs, so the difference between trained and untrained activations (median-scored, max-aggregated over k-means partitions of the InD data) separates InD from OOD samples.
Key results
- KAN detector reaches 94.12 average overall AUROC on CIFAR-10, versus 93.37 for NAC (previous best).
- On ImageNet-1K FS the KAN detector achieves 78.52 average overall AUROC, above NAC at 76.28.
- On CIFAR-100 the KAN detector reaches 83.44 overall AUROC, tied with NAC at 83.36 and ahead of the other baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KAN detector instead of NAC | OOD detection · CIFAR-10 benchmark | Avg Overall AUROC: 93.37 -> 94.12 | Helped |
| use KAN detector instead of NAC | OOD detection · CIFAR-100 benchmark | Avg Overall AUROC: 83.36 -> 83.44 | Helped |
| use KAN detector instead of ASH | OOD detection · ImageNet-200 FS | Avg Overall AUROC: 67.18 -> 71.46 | Helped |
| use KAN detector instead of NAC | OOD detection · ImageNet-1K FS | Avg Overall AUROC: 76.28 -> 78.52 | Helped |
| use KAN detector instead of KLM | OOD detection · eICU Age | AUROC: 51.0 -> 50.5 | Hurt |
| reduce partitions from 10 to 1 (ablation) | KAN detector · CIFAR-10 benchmark | AUROC: 94.12 -> 46.08 | Hurt |
| replace KAN splines with binary histograms (ablation) | KAN detector · CIFAR-10 benchmark | AUROC: 94.12 -> 85.29% | Hurt |
| use maximum instead of median scoring (ablation) | KAN detector · CIFAR-10 benchmark | AUROC: 94.12 -> 54.99 | Hurt |
| replace k-means with DBSCAN partitioning (ablation) | KAN detector · CIFAR-10 benchmark | AUROC: 94.12 -> 89.82 | Hurt |
| train the KAN for regression to a constant instead of classification (ablation) | KAN detector · Image benchmarks | AUROC: +0.2% | Helped |
| train the KAN for regression to a constant instead of classification (ablation) | KAN detector · Tabular medical benchmarks | AUROC: approximately 3% decrease | Hurt |
| increase the number of partitions from 1 to 4 (ablation) | KAN detector · Friedman | AUROC: 52.11 -> 69.42 | Helped |
Surprising findings
- Replacing the KAN training task with regression to a constant changed image-benchmark AUROC by only about 0.2% but reduced tabular performance by about 3%, suggesting the training task need not match the OOD problem.
- Max aggregation with max scoring performs very poorly on CIFAR-10 (54.99 AUROC for max scoring), showing the scoring choice is critical.
Limitations
- Inference is the slowest among compared methods (2.216 ms vs 0.175 to 0.681 ms for KNN, VIM, NAC).
- Partitioning is needed to capture joint feature distributions; on regression datasets one partition often suffices or the approach underperforms, and the paper notes this is unresolved.
- Hyperparameters such as grid size and partition count are benchmark-dependent and tuned on validation data.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200 ImageNet-1K eICU CIFAR-10 near OOD TIN MNIST SVHN Textures Places365 SSB-hard
Benchmarks: OpenOOD CIFAR-10 OpenOOD CIFAR-100 OpenOOD ImageNet-200 FS OpenOOD ImageNet-1K FS TabMed eICU Ethnicity TabMed eICU Age TabMed eICU Synthetic OOD
Metrics: AUROC FPR@95
Baselines: OpenMax ODIN MDS RMDS Gram ReAct VIM KNN ASH SHE GEN NAC
Builds on: Kolmogorov-Arnold Networks OpenOOD NAC KNN
Closest papers by meaning
- 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 0.87)
- Neuron Activation Coverage: Rethinking Out-of-distribution Detection and Generalization (ICLR 2024 · similarity 0.87)
- KAN: Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.87)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.86)
- Strengthen Out-of-Distribution Detection Capability with Progressive Self-Knowledge Distillation (ICML 2025 · similarity 0.86)
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.