Reliability Evaluation of ML systems, the oracle problem

Abstract

The growing adoption of machine learning (ML) in safety-critical contexts makes reliability evaluation of ML systems a crucial task. Although testing represents one of the most used practices to evaluate the reliability of “traditional” systems, just few techniques can be used to evaluate ML-systems’ reliability due to the oracle problem. In this paper, I present a test oracle surrogate able to automatically classify tests’ outcome to obtain feedback about tests whose expected output is unknown. For this purpose, various sources of knowledge are considered to evaluate the outcome of each test. The aim is to exploit this test oracle surrogate to apply classical testing strategies to perform reliability assessment of ML systems. Some preliminary experiments have been performed considering a Convolutional Neural Network (CNN) and exploiting the well known MNIST dataset. These results promise that the presented technique can be effectively used to evaluate the reliability of ML systems.