Non-Linear Outlier Synthesis for Out-of-Distribution Detection

TL;DR

This paper introduces NCIS, a novel method that improves out-of-distribution detection by generating high-quality synthetic outliers in the diffusion model's embedding space, achieving state-of-the-art results on standard benchmarks.

Contribution

NCIS enhances synthetic outlier quality by operating in the diffusion embedding space and modeling class-conditional manifolds with a volume-preserving network, advancing OOD detection performance.

Findings

Achieves new state-of-the-art OOD detection results on ImageNet100 and CIFAR100.

Highlights the importance of data pre-processing and key design choices.

Demonstrates the effectiveness of embedding space operations for outlier synthesis.

Abstract

The reliability of supervised classifiers is severely hampered by their limitations in dealing with unexpected inputs, leading to great interest in out-of-distribution (OOD) detection. Recently, OOD detectors trained on synthetic outliers, especially those generated by large diffusion models, have shown promising results in defining robust OOD decision boundaries. Building on this progress, we present NCIS, which enhances the quality of synthetic outliers by operating directly in the diffusion's model embedding space rather than combining disjoint models as in previous work and by modeling class-conditional manifolds with a conditional volume-preserving network for more expressive characterization of the training distribution. We demonstrate that these improvements yield new state-of-the-art OOD detection results on standard ImageNet100 and CIFAR100 benchmarks and provide insights into the importance of data pre-processing and other key design choices. We make our code available at \url{https://github.com/LarsDoorenbos/NCIS}.