MLCon 2017 Archive

Deep Learning is these days often considered a "holy grail" when it comes to lending algorithms, machines, and systems intelligence. While fully automatic and autonomous machine learning is on its way, present solutions require a software designer's and engineer's understanding of the underlying principles and possibilities. This talk focuses on the essentials of making Deep Learning work in a broad range of applications and settings. This includes often encountered limitations such as limited knowledge on an optimal problem domain representation, limited availability of annotated data, or the better understanding of internal decision-making processes. Tricks and hacks presented include end-to-end learning from raw data by convolutional layers, coping with sparse data by generative adversarial topologies, active, reinforced, semi-supervised, and transfer learning, and ways to interpret a trained network. Hints are further given on the design of a suited network topology, and tools in the field. This leaves to the listener's creativity how to best exploit Deep Learning in the next thrilling real-life use-case.