Natural Attribute-based Shift Detection

TL;DR

This paper introduces the task of natural attribute-based shift detection to identify samples that differ from training data due to natural attributes, providing new benchmarks and analyzing existing OOD detection methods.

Contribution

It defines NAS detection, creates benchmark datasets across vision, language, and medical domains, and proposes a simple training modification to improve OOD detection across NAS categories.

Findings

Existing OOD methods show inconsistent performance on NAS datasets.

Analysis reveals the relationship between NAS sample locations and detection performance.

A simple training modification enhances detection of all NAS categories.

Abstract

Despite the impressive performance of deep networks in vision, language, and healthcare, unpredictable behaviors on samples from the distribution different than the training distribution cause severe problems in deployment. For better reliability of neural-network-based classifiers, we define a new task, natural attribute-based shift (NAS) detection, to detect the samples shifted from the training distribution by some natural attribute such as age of subjects or brightness of images. Using the natural attributes present in existing datasets, we introduce benchmark datasets in vision, language, and medical for NAS detection. Further, we conduct an extensive evaluation of prior representative out-of-distribution (OOD) detection methods on NAS datasets and observe an inconsistency in their performance. To understand this, we provide an analysis on the relationship between the location of NAS samples in the feature space and the performance of distance- and confidence-based OOD detection methods. Based on the analysis, we split NAS samples into three categories and further suggest a simple modification to the training objective to obtain an improved OOD detection method that is capable of detecting samples from all NAS categories.