ML Testing: The Need of a New Way of Thinking
When testing functionality based on Machine learned, or trained functionality, the focus of your testing changes. The code it self stops to be interesting, to some degree, and instead focus need to be elsewhere. Based on personal experience and research projects this talk will highlight the importance of testing your data, why independent testing is vital and how some "old school" tools can help when thinking and planing test activities.
This talk will use trained functionality intended for autonomous driving as base but will also touch more general problems faced when testing Machine Learning functionality. Key takeaways include: why we need to think slightly different when testing ML, importance of knowing your test data, and how we can reuse the concept of test levels to showcase our testing activities.