קולוקוויום בביה"ס למדעי המחשב - Fundamental limits of modern machine learning and how to get around them
Yair Carmon - Stanford
This talk presents new computational and statistical barriers in machine learning, along with the algorithmic developments that they inspire.
The computational barriers arise in nonconvex optimization: we prove lower bounds on the (oracle) complexity of finding stationary points using (stochastic) gradient methods, showing that gradient descent is unimprovable for a natural class of problems. We bypass this barrier by designing an algorithm that outperforms gradient descent for a large subclass of problems with high-order smoothness. Our algorithm leverages classical momentum techniques from convex optimization using a “convex until proven guilty” principle that we develop.
The statistical barrier is the large amount of data required for adversarially robust learning. In a Gaussian model, we prove that unlabeled data allows us to circumvent an information theoretic gap between robust and standard classification. Our analysis directly leads to a general robust self-training procedure; we use it to significantly improve state-of-the-art performance on the challenging and extensively studied CIFAR-10 adversarial robustness benchmark.
Yair Carmon is a PhD student at Stanford University working with Prof. John Duchi and Prof. Aaron Sidford. His research focuses on understanding and overcoming the fundamental limits of machine learning; specific interests include nonconvex optimization, efficient algorithms for matrix-structured data, and robust machine learning. Yair received an M.Sc. from the Technion in 2015, under the supervision of Prof. Shlomo Shamai and Prof. Tsachy Weissman. Between 2009 and 2015 he held several positions as an algorithm engineer and research team leader, working on communications, signal processing and computer vision. His awards include the Stanford Graduate Fellowship, the Numerical Technologies Fellowship, and membership in the Technion Excellence Program.