This month we have Mehryar Mohri, a professor at NYU's Courant Institute of Mathematical Sciences, presenting "Successful Algorithms for Learning Kernels." More details on Mehryar and his research can be found here: http://cs.nyu.edu/~mohri/
Title: Successful Algorithms for Learning Kernels Kernel methods combined with large-margin algorithms such as SVMs are widely used in statistical learning. However, in the standard framework of these methods, the choice of an appropriate kernel, which is critical to the performance achieved, is left to the user. An exciting idea contemplated since the beginning has been to use a labeled sample to "learn the kernel", as well as a hypothesis based on that kernel. But, experiments and a large body of literature on the topic over the last decade or so have proven it to be surprisingly difficult to outperform even the simple scheme of a fixed combination of base kernels. This talk reports the results of extensive experiments with a new learning kernel algorithm demonstrating consistent and significant performance improvements over this simple scheme. We describe this algorithm, give several justifications for its use, and more generally discuss a novel and promising theoretical foundation for learning kernels. [This is joint work with Corinna Cortes and Afshin Rostamizadeh]
This month O'Reilly provided us with a ticket to Strata, happening next month. We'll do the raffle for the ticket at the beginning before we get the talk started.