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

Forte : Finding Outliers with Representation Typicality Estimation

ICLR 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Robust fine-tuning for OOD generalization · 0 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.

Forte estimates typical sets by combining self-supervised representations (CLIP, ViT-MSN, DINOv2) with per-point PRDC summary statistics and one-class SVM/KDE/GMM density models to detect OOD and Stable-Diffusion synthetic images without labels or OOD exposure.

Problem

Detecting out-of-distribution inputs and synthetic images generated by foundation/generative models, where likelihood-based generative OOD detectors fail due to likelihood misestimation and typicality issues and focus on pixels rather than semantics.

Approach

Adapt per-point precision, recall, density and coverage (manifold-estimation statistics from generative evaluation) as summary statistics over self-supervised representation spaces, then model their ID typical set with non-parametric density estimators (OCSVM, KDE, GMM), extending DoSE.

Key results

• On CIFAR-10 vs SVHN, Forte+SVM reports AUROC 99.84 and FPR95 0.00 versus DoSE's 97.30 AUROC and 13.16 FPR95.
• On ImageNet near-OOD SSB-Hard, Forte+GMM reaches AUROC 94.95 and FPR95 22.30, versus 77.28 AUROC and 72.90 FPR95 for DINOv2+MLS.
• On Golden Retriever Img2Img S=1.0 synthetic detection, Forte+GMM reaches AUROC 98.11 versus 60.15 for the CLIP baseline.

Measured effects reported in the paper

Surprising findings

• Standard statistical tests (Z-test, KS, Mann-Whitney) and distribution divergences fail to give consistent signals for Stable Diffusion-generated images, even when generated images are highly realistic.
• OpenOOD's linear-probe DINOv2 with MLS performs poorly on OOD detection despite strong representations.

Limitations

• Theoretical justification relies on Gaussian assumptions with equal covariance and large-mean-shift approximations.
• Precision and density alone make it a single-sample test; recall and coverage require the full reference set, making it a two-sample test.
• Some reported results have high variance (e.g., FastMRI/OAI KDE, Golden Retriever low-strength settings).

Experimental setup

Datasets: CIFAR-10 CIFAR-100 CelebA SVHN ImageNet-1k iNaturalist Texture OpenImage-O NINCO SSB-Hard ImageNet-C ImageNet-R

Benchmarks: CIFAR-10 vs CIFAR-100 CIFAR-10 vs CelebA CIFAR-10 vs SVHN ImageNet-1k OpenOOD v1.5 suite Synthetic ImageNet Golden Retriever and Volleyball FastMRI vs OAI medical datasets

Metrics: AUROC FPR95 FD FD-infinity CMMD

Baselines: NNGuide ViM ViT-B+CE+RMDS DINOv2+MLS DoSE WAIC Typicality test LLR CLIP-based density baseline FD FD-infinity CMMD

Builds on: DoSE Precision and Recall metrics Density and Coverage metrics Typicality test Glow

Closest papers by meaning

• A Geometric Explanation of the Likelihood OOD Detection Paradox (ICML 2024 · similarity 0.89)
• Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.89)
• ODIM: Outlier Detection via Likelihood of Under-Fitted Generative Models (ICML 2024 · similarity 0.88)
• Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.86)
• Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.86)
• ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.86)