Machine Learning Testing

Machine Learning (ML) is becoming more and more important with its increasing integration in various domains such as self-driving cars, machine translation, medical systems, many of these being safety-critical applications. Naturally, trustworthiness concerns arise, regarding the decisions or predictions obtained, and also the need for systematic testing of ML software. Compared with testing traditional software, machine learning testing comes with additional challenges, regarding input generation, overcoming the ‘test Oracle problem’ (the challenge of distinguishing the corresponding desired, correct behaviour from potentially incorrect behaviour), test assessment criteria, bug detection in training data or learning program.

The current practice includes adversarial ML testing (finding inputs that, when changed minimally from their original versions, get classified differently), fuzz and search-based test input generation. For solving the ‘Oracle problem’ metamorphic relations have been studied – these are transformations of the test data that are expected to yield the same output, or to generate expected changes in the predictive result.

This project will focus on automating ML testing for autonomous driving, but there are other applications areas, including medical image analysis and machine translation for which collaboration opportunities are possible.

Eligibility:

Candidates are expected to hold (or be about to obtain) a minimum 2:1 honours degree (or equivalent) in a related area / subject, e.g. Computer Science, Software Engineering, Mathematics, Machine Learning, etc. MSc, MA or relevant experience in a related discipline is highly desirable.