Paper Overview
Field: Machine Learning Authors: Amey Pasarkar, Adji Bousso Dieng Published: 2026-11-09 arXiv: 2611.12871
Introduction
Anomaly detection (AD) is critical for trustworthy machine learning and scientific discovery. Existing methods typically define anomalies through local density, isolation heuristics, or distances to fitted distributions, making them sensitive to neighborhood scales, partitioning choices, or restrictive distributional assumptions.
Key Contribution
This work proposes a diversity-based paradigm for anomaly detection: the Vendi Anomaly Score (VAS). VAS detects anomalies by quantifying how much a sample changes the diversity of the dataset—measured by the Vendi score—when that sample is removed.
Key Properties
- Captures both local redundancy and global data structure without density estimation or partitioning heuristics
- Requires no hyperparameter tuning; automatically adapts to the spectral structure of the data
- Non-parametric and scales linearly with dataset size
- Benchmark evaluation: VAS achieves state-of-the-art AD performance across 148 benchmark embedding datasets, drawn from 10 source datasets and 4 embedding architectures.
- Large-scale validation: Experiments on ImageNet confirm that VAS remains robust as contamination rates and dataset sizes grow.
Results
Original Abstract
> We propose the Vendi Anomaly Score (VAS), which detects anomalies by quantifying how much a sample changes the diversity of the dataset when removed. VAS captures both local redundancy and global data structure without density estimation, requires no hyperparameter tuning, and achieves state-of-the-art AD performance across 148 benchmark datasets.
---
Paper link: https://arxiv.org/abs/2611.12871