Automatic Open-World Reliability Assessment

TL;DR

This paper introduces automated methods to assess the reliability of open-world image classifiers by analyzing score distributions, enabling better detection of when system performance may degrade due to out-of-distribution inputs.

Contribution

It formalizes the open-world recognition reliability problem and proposes multiple distribution-based algorithms for automatic reliability assessment applicable to various classifiers.

Findings

Algorithms outperform SoftMax mean detection

Effective in identifying reliability drops

Applicable to both classic and open-world classifiers

Abstract

Image classification in the open-world must handle out-of-distribution (OOD) images. Systems should ideally reject OOD images, or they will map atop of known classes and reduce reliability. Using open-set classifiers that can reject OOD inputs can help. However, optimal accuracy of open-set classifiers depend on the frequency of OOD data. Thus, for either standard or open-set classifiers, it is important to be able to determine when the world changes and increasing OOD inputs will result in reduced system reliability. However, during operations, we cannot directly assess accuracy as there are no labels. Thus, the reliability assessment of these classifiers must be done by human operators, made more complex because networks are not 100% accurate, so some failures are to be expected. To automate this process, herein, we formalize the open-world recognition reliability problem and propose multiple automatic reliability assessment policies to address this new problem using only the distribution of reported scores/probability data. The distributional algorithms can be applied to both classic classifiers with SoftMax as well as the open-world Extreme Value Machine (EVM) to provide automated reliability assessment. We show that all of the new algorithms significantly outperform detection using the mean of SoftMax.