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

ITP: Instance-Aware Test Pruning for Out-of-Distribution Detection

AAAI 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Provable and feature-based OOD detection · 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 ITP, a post-hoc OOD detection method that combines class-specific coarse-grained redundancy pruning of last-layer weights with instance-level right-tailed Z-score test pruning of overconfident parameters.

Problem

Weight-sparsification OOD detectors select parameters important for ID data, but these parameters can behave overconfidently on OOD inputs and hurt detection.

Approach

Estimate class-specific Gaussian contribution distributions of last-layer weights on ID training data; prune low-average-contribution weights globally (CRP) and, per test sample, prune weights whose contribution has a right-tailed Z-score above a threshold (FTP), then compute energy scores from the pruned model.

Key results

• On CIFAR-100 (DenseNet-101), ITP reduces average FPR95 to 35.03% from Energy's 68.45%, and ITP + ReAct reaches 30.13% FPR95 / 91.91% AUROC.
• On ImageNet-1k (ResNet-50), ITP achieves 22.51% FPR95 and 94.55% AUROC versus Energy's 58.41% / 86.17%; ITP + ReAct reaches 20.39% FPR95.
• Using only 30 samples per class for distribution estimation achieves performance nearly equivalent to the full training set, with preprocessing time of 0.0073 hours vs 0.2411 hours for full ITP.

Measured effects reported in the paper

Surprising findings

• Parameters selected as important for ID data can behave overconfidently on OOD inputs, hurting detection.
• Using only two samples per class already yields a 29.48% FPR95 reduction versus the energy baseline on ImageNet.

Limitations

• FTP gains on ResNet-50 are smaller, attributed to the larger 2048-dimensional feature space allowing noise to dominate.
• Hyperparameters p and lambda are selected via a Gaussian-noise validation set, and the paper reports results on CIFAR-100 and ImageNet with differing per-dataset settings.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 ImageNet-1k

Benchmarks: CIFAR-10 CIFAR-100 ImageNet-1k OpenOOD v1.5

Metrics: FPR95 AUROC

Baselines: MSP Energy ODIN ReAct DICE OPNP LAPS LINe KNN ASH GradNorm MLS

Builds on: DICE LINe Energy score ReAct

Closest papers by meaning

• Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.88)
• OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary (CVPR 2025 · similarity 0.88)
• Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.87)
• Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.87)
• Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.87)
• Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.87)