Vendi Novelty Scores for Out-of-Distribution Detection

TL;DR

The paper introduces Vendi Novelty Scores (VNS), a new OOD detection method based on diversity metrics that outperforms existing techniques and works efficiently with limited training data.

Contribution

It proposes VNS, a novel diversity-based OOD detection approach that is linear-time, non-parametric, and effective with minimal training data, advancing the state-of-the-art.

Findings

VNS achieves state-of-the-art OOD detection performance across benchmarks.

VNS maintains high performance using only 1% of training data.

VNS combines local and global novelty signals effectively.

Abstract

Out-of-distribution (OOD) detection is critical for the safe deployment of machine learning systems. Existing post-hoc detectors typically rely on model confidence scores or likelihood estimates in feature space, often under restrictive distributional assumptions. In this work, we introduce a third paradigm and formulate OOD detection from a diversity perspective. We propose the Vendi Novelty Score (VNS), an OOD detector based on the Vendi Scores (VS), a family of similarity-based diversity metrics. VNS quantifies how much a test sample increases the VS of the in-distribution feature set, providing a principled notion of novelty that does not require density modeling. VNS is linear-time, non-parametric, and naturally combines class-conditional (local) and dataset-level (global) novelty signals. Across multiple image classification benchmarks and network architectures, VNS achieves state-of-the-art OOD detection performance. Remarkably, VNS retains this performance when computed using only 1% of the training data, enabling deployment in memory- or access-constrained settings.