Track Talk Th19

Machine Learning: The Problems We Don’t Speak Of!

Martin Karsberg

Elias Sonnsjö

16:30-17:15 CEST Thursday 9th June

As machine learning is becoming a more natural part of software solutions, it also presents new challenges for the software testing community. When everything is black-box and when the data used to train the function is more interesting to test that the functionality itself… How will this affect the way we test? Based on work done in two research projects funded by the Swedish government and European union, some of the mist surrounding the topic will be dispelled.

The presentation will cover some of the challenges when it comes to testing non-coded (trained) functionality, along with an introduction to common problems one might face, and suggestions on overcoming some of the difficulties. As a way to tackle these challenges, a possible process of qualifying a trained functionality will be presented.

Using safety critical ML functionality within autonomous drive as a starting point, the presentation will cover topics such as:

  • Basic terms to keep track of
  • Challenges when it comes to optimization of functionality
  • Defining requirements
  • Use of simulators in software testing
  • Use of test levels