Vendi anomaly scores improve detection without tuning parameters
Vendi Anomaly Scores for Efficient and Accurate Anomaly Detection
Machine Learning
Summary
Finding unusual or suspicious data points, called anomalies, is important for trustworthy AI and scientific research. The authors propose a new way to spot anomalies by measuring how removing a data point changes the overall diversity of the dataset, rather than relying on local density or distance measures. This method, called Vendi Anomaly Score (VAS), adapts automatically to the data, needs no tuning, and works well on many types of data. It performs better than existing methods in tests across various datasets and scales well to large datasets like ImageNet.
What this means in practice
- •For machine learning engineers: Detect anomalies in large-scale datasets automatically without tuning hyperparameters, improving robustness of AI models in production.
- •For data science teams: Identify rare or novel patterns in high-dimensional embeddings from images or other data, aiding scientific discovery and data quality assessment.
Authors
Amey Pasarkar, Adji Bousso Dieng
Abstract
Anomaly 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. In this work, we introduce a new paradigm by formulating anomaly detection from a diversity perspective. We propose the $\textit{Vendi Anomaly Score}$ (VAS), which detects anomalies by quantifying how much a sample changes the diversity of the dataset, as measured by the Vendi score, when removed. VAS captures both local redundancy and global data structure without relying on density estimation or partitioning heuristics. Moreover, VAS requires no hyperparameter tuning, instead adapting automatically to the spectral structure of the data. VAS is non-parametric and scales linearly with dataset size. Across 148 benchmark embedding datasets derived from 10 source datasets and 4 embedding architectures, VAS achieves state-of-the-art AD performance. We further validate VAS on large-scale ImageNet experiments, where it remains robust as both contamination rate and dataset scale increase.