2008 Reports

Improving the Dependability of Machine Learning Applications

Murphy, Christian; Kaiser, Gail E.

As machine learning (ML) applications become prevalent in various aspects of everyday life, their dependability takes on increasing importance. It is challenging to test such applications, however, because they are intended to learn properties of data sets where the correct answers are not already known. Our work is not concerned with testing how well an ML algorithm learns, but rather seeks to ensure that an application using the algorithm implements the specification correctly and fulfills the users' expectations. These are critical to ensuring the application's dependability. This paper presents three approaches to testing these types of applications. In the first, we create a set of limited test cases for which it is, in fact, possible to predict what the correct output should be. In the second approach, we use random testing to generate large data sets according to parameterization based on the application's equivalence classes. Our third approach is based on metamorphic testing, in which properties of the application are exploited to define transformation functions on the input, such that the new output can easily be predicted based on the original output. Here we discuss these approaches, and our findings from testing the dependability of three real-world ML applications.