Why Normalizing Flows Fail to Detect Out-of-Distribution Data

TL;DR

Normalizing flows often fail at out-of-distribution detection because they learn generic features rather than dataset-specific semantics, but architectural modifications can improve their performance.

Contribution

This paper identifies why normalizing flows struggle with OOD detection and proposes architectural changes to enhance their ability to learn dataset-specific semantic features.

Findings

Flows learn local pixel correlations and generic transformations.

Architectural modifications improve OOD detection performance.

Properties enabling high-fidelity image generation can hinder OOD detection.

Abstract

Detecting out-of-distribution (OOD) data is crucial for robust machine learning systems. Normalizing flows are flexible deep generative models that often surprisingly fail to distinguish between in- and out-of-distribution data: a flow trained on pictures of clothing assigns higher likelihood to handwritten digits. We investigate why normalizing flows perform poorly for OOD detection. We demonstrate that flows learn local pixel correlations and generic image-to-latent-space transformations which are not specific to the target image dataset. We show that by modifying the architecture of flow coupling layers we can bias the flow towards learning the semantic structure of the target data, improving OOD detection. Our investigation reveals that properties that enable flows to generate high-fidelity images can have a detrimental effect on OOD detection.