Successful Deep Learning Requires Prior Knowledge

Abstract

Deep learning is nowadays widely being marketed as an approach able to learn useful behaviours from scratch – as opposed to methods that require extensive manual feature engineering. While this is true, it is also a view, which tends to downplay the degree to which successful deep learning approaches rely on prior knowledge. The talk will give an overview of how some of the best-performing deep learning methods incorporate prior knowledge and then proceed to discuss the same idea in the context of deep reinforcement learning. It will also present some work of the Laboratory of Artificial Intelligence of the University of Žilina (LUIZA).