"Human error": How can we help people build models that do what they expect
It's never been easier to train machine learning models. With excellent open source tooling, lower compute techniques, and incredible educational material online, really anybody can start to train their own models today. Yet, Anna Roth explains, when domain experts try to transfer their expertise to an ML model, the results can be unpredictable.
Talk Title | "Human error": How can we help people build models that do what they expect |
Speakers | Anna Roth (Microsoft) |
Conference | O’Reilly TensorFlow World |
Conf Tag | |
Location | Santa Clara, California |
Date | October 28-31, 2019 |
URL | Talk Page |
Slides | Talk Slides |
Video | |
It’s never been easier to train machine learning models. With excellent open source tooling, lower compute techniques, and incredible educational material online, really anybody can start to train their own models today. Yet when domain experts try to transfer their expertise to an ML model, the results can be unpredictable. The same model can be astonishingly good and then make errors that make absolutely no sense to the human trying to teach the machine. Motivated by a series of real stories (mostly in computer vision), Anna Roth discusses both human and technical factors and suggests some future directions.