“What is learnable?” is a fundamental question in learning theory. The talk will address this question for deep learning, in which a hypothesis class is defined using a network architecture and learning is performed by a stochastic gradient oracle. Join work with Eran Malach.