Trading Convexity for Scalability
186 views · Published 6 September 2016 · 48:26 · Indexed 16 November 2026
Channel: Microsoft Research · 2016 · Science & Technology
Joint work with Ronan Collobert, Fabian Sinz, and Jason Weston. Convex learning algorithms, such as Support Vector Machines (SVMs), are often seen as highly desirable because they offer strong practical properties and are amenable to theoretical analysis.  However, in this work we show how non-convexity can provide scalability advantages over convexity.  We show how concave-convex programming can be applied to produce (1)~faster SVMs where training errors are no longer support vectors, and (2)~much faster Transductive SVMs.