Research map / Label-efficient and robust learning / OOD and anomaly detection
Vendi Anomaly Scores for Efficient and Accurate Anomaly Detection
Amey Pasarkar, Adji Bousso Dieng
Read the paperCheck your idea against 137,217 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 Vendi anomaly scores, a new family of embedding-based anomaly scores built on the Vendi diversity measure, for anomaly detection in learned embedding spaces.
Problem
Existing embedding-based anomaly detectors define anomalies via local density, isolation heuristics, or distance from a fitted distribution, which makes them sensitive to neighborhood scale, partitioning choices, or restrictive distributional assumptions.
Approach
Formulates anomaly scoring as a diversity-based quantity (the Vendi score) rather than density, isolation, or distributional fit, offering a new paradigm for anomaly detection on learned embeddings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| score samples by Vendi Anomaly Score instead of baseline detectors | embedding anomaly detection · ADBench image and text embeddings (148 instances) | AUROC average rank: 2.35 vs 3.80 (GMD), 4.05 (DN2) | Improved |
| use VAS instead of baselines on small contaminating subpopulations | ImageNet-1K DINOv3 embeddings · ImageNet contamination sweeps (up to 100k samples) | AUROC | Improved |
| replace exact leave-one-out Vendi change with first-order approximation (ablation) | VAS · ADBench 148 instances | AUROC / Spearman rho: AUROC change at most 1.16%; Spearman over 0.98 | No clear effect |
| use adaptive q from principal eigenvalue instead of alternative q rules (ablation) | VAS · ADBench 148 tasks | AUROC | No clear effect |
| grid search baseline hyperparameters (tuned on evaluation sets) (ablation) | baseline detectors · ADBench | AUROC | No clear effect |
| use VAS rather than KPCA/DN2/LOF on some datasets | ADBench · CIFAR10 ResNet18, SVHN, Agnews BERT, Imdb BERT | AUROC: Agnews BERT 0.676 vs LOF 0.725; CIFAR10 ResNet18 0.745 vs KPCA 0.751 | Mixed |
Experimental setup
Builds on: Vendi score
Closest papers by meaning
- Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.87)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.87)
- We Need to Rethink Benchmarking in Anomaly Detection (NeurIPS 2026 · similarity 0.86)
- Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.86)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.86)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 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.
