Vendi Anomaly Score flags outliers via dataset diversity
Original title: Vendi Anomaly Scores for Efficient and Accurate Anomaly Detection
The score needs no hyperparameter tuning, scales linearly, and led 148 embedding benchmarks plus ImageNet tests.
The paper defines the Vendi Anomaly Score, which marks a point as anomalous by how much removing it changes the dataset's Vendi diversity score. VAS is non-parametric, needs no hyperparameter tuning, and scales linearly with dataset size, without density estimation or partitioning. Across 148 embedding datasets from 10 sources and 4 embedding architectures, the authors report state-of-the-art anomaly detection, and ImageNet tests stayed robust as contamination and scale grew.
- VAS measures diversity change when a sample is removed.
- Evaluated on 148 embedding datasets from 10 sources and 4 architectures.
- ImageNet tests remained robust as contamination rate and scale increased.
From the original article
1 min read at arxiv.orgAnomaly detection (AD) is a critical for trustworthy machine learning and scientific discovery. Existing methods typically define anomalies through local density, isolation heuristics, or distance from a fitted distribution, making them sensitive to neighborhood scale, partitioning choices, or restrictive distributional assumptions.
By Amey Pasarkar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.