talk@freiburg.ai: Frank Hutter, University of FreiburgTowards Automated Deep Learning
The performance of most deep learning methods heavily depends on the chosen network architectures and their hyperparameters. In this talk, I will discuss methods for effective optimization in this combined space, thereby paving the way to fully automated end-to-end deep learning. Next to competition-winning AutoML systems, I will discuss BOHB, a robust and efficient multi-fidelity hyperparameter optimization system that is parallelizable and applicable to tuning a wide range of deep learning methods, as well as recent advances in neural architecture search. I will end with an application of AutoML to the problem of learning to design RNA using deep reinforcement learning (RL), in which we used BOHB to jointly tune the RL agent's state representation, its policy network's architecture, and its hyperparameters, yielding a clear new state-of-the-art in RNA design.
(Kopieren und einfügen in Ihre Host-Anwendung - weitere Infos hier)