Special talk by Kasper Green Larsen on Understanding how much training data is needed for popular machine learning models

Machine learning models are being used to solve and automate ever more tasks, ranging from spam identification to image recognition, product recommendations and many more. But when can we trust that the excellent performance of a machine learning model on known training data carries over to new unknown data? This question is at the core of learning theory. Learning theory gives a theoretical framework for answering questions like: “How much training data do we need?” and “What properties of the training data and machine learning model improves its performance?”. In this talk, I will introduce the basics of learning theory and highlight some of my recent contributions in understanding the performance of popular machine learning models such as Support Vector Machines and Boosting algorithms. Finally, I will briefly highlight some of my other contributions in machine learning, data structures, lower bounds and cryptography and argue that many of the techniques developed in these seemingly disjoint areas, often carry over and yield surprising new connections.