IEEE ITW 2020, Riva del Garda, Italy
Inductive bias (reflecting prior knowledge or assumptions) lies at the core of every learning system and is essential for allowing learning and generalization, both from a statistical perspective, and from a computational perspective. What is the inductive bias that drives deep learning? A simplistic answer to this question is that we learn functions representable by a given architecture. But this is not sufficient neither computational (as learning even modestly sized neural networks is intractable) nor statistically (since modern architectures are too large to ensure generalization). In this tutorial we will explore these considerations, how training humongous, even infinite, deep networks can ensure generalization, what function spaces such infinite networks might correspond to, and how the inductive bias is tightly tied to the local search procedures used to train deep networks.