While deep learning has produced exceptional results in many areas of application over the last decade, the general understanding of the methodology lacks behind: Given a new use case, it is unclear what (ansatz; model; hyperparameters; …) will work, or what was finally responsible for the breakthrough once it works.

In this talk, I will highlight lessons learned from observing deep learning models in two distinct domains (computer vision for e.g. defect recognition, and reinforcement learning for logistics planning), and draw practically applicable conclusions from them for future cases. A highlight will be first results from a rescheduling application for trains: using deep reinforcement learning, the trains learned a communication language of their own to negotiate their own schedules.

Download the slides for this talk.Download ( PDF, 9674.74 MB)