(Machine) Learning to Do More with Less

by Marat Freytsis (University of Oregon)

Thursday 3 Aug 2017, 13:15 → 14:15 Europe/Zurich

4/2-037 - TH meeting room (CERN)

I will discuss some aspects of "weakly supervised" training, recently introduced into the high energy literature as a way to apply machine learning for classification in the absence of event-by-event truth-level information. Originally advocated as a way to allow for data-driven training, I will focus on the resulting performance under distortions and uncertainties with respect to the input, a concern for both the experimental and theoretical understanding of the training data. Surprisingly, the resulting performance is much more robust with respect to such errors than traditional fully supervised approaches, and this behavior is simple enough that some of its aspects can be understood analytically. Along with some examples, I will explain how these results can help reduce the opacity of the conventional machine-learning black box.