Developing a robust machine learning code is hard, but debugging it is even harder! When developing machine learning (ML) solutions for complex problems such as autonomous driving we tend to focus on the hard research problems and develop theoretical models whose implementation we take for granted. In practice, however, bugs always creep in. How do you detect bugs when working with multidimensional arrays containing millions of parameters? How do you identify sources of error when building dynamic computational graphs? In this talk, I will provide an overview of the existing methods for debugging ML code and showcase the first of its kind visual 3D debugger which makes debugging deep learning models a breeze.
