Presentation on YouTube.

Classical continuous optimization generally studies the dynamics in parameter space. We argue that, in several cases, it is more appropriate to analyze these problems in function space. Using as examples standard problems in reinforcement learning and supervised learning, we show both new theoretical results as well as novel practical algorithms.

Bio: Nicolas Le Roux got his PhD in 2008 from University of Montreal where he worked with Yoshua Bengio on neural networks in general and their optimisation in particular. Since then, he worked on generative models of images, large-scale convex optimisation and stochastic variance reduction methods, for which he was a co-recipient of the Lagrange prize in 2018, and reinforcement learning. He managed multiple research teams at Criteo, Google and now Microsoft. He holds a CIFAR AI Chair and is an adjunct professor at UdeM and McGill.