Speaker: Niao He, University of Illinois at Urbana-Champaign
Abstract: The era of Big Data is introducing big challenges for machine learning to accommodate to rapidly growing data and increasingly complex models. Many traditional learning methods fail to work due to the lack of scalability or theoretical guarantee. We show simple optimization algorithms such as stochastic gradient descent, when combined with newly developed techniques, could make a huge difference. We discuss some of our recent works that advance several important sub-domains of machine learning, including kernel machines, Bayesian inference, and reinforcement learning. For all these cases, we developed simple new algorithms that allow to train bigger models, learn better yet faster, come with provable guarantees, and improve significantly over previous state-of-the-arts. These advances are proven to be useful in a wide range of machine learning applications on large-scale real world datasets.