Seminar on Computational Learning and Adaptation

  Improving the Query-Time "Bang Per Buck" of Support Vector Machines: Towards Robust Nonlinear Models for the (Amortized) Price of Linear Ones

Dennis DeCoste
Jet Propulsion Laboratory / Caltech
http://www-aig.jpl.nasa.gov/home/decoste

Support vector machines (and other kernel machines) have recently become very popular in machine learning research, especially for high-dimensional classification tasks, due to their ability to robustly find good nonlinear models. However, a key tradeoff is much (e.g. 10-1000 fold) higher query times, relative to other common machine learning methods, such as simple linear methods (e.g. Fischer discriminants), decision trees, and neural networks. Given that improvements over such alternatives are often only a few percentages in classification accuracy, this raises serious "bang per buck" questions concerning the practical utility of kernel machines, especially in embedded (e.g.resource-limited spacecraft) or real-time (e.g. robotic control) applications. In this talk, I will present new methods that address this dilemma by "compiling" a kernel machine into a query-time form which enables "easier" queries to be incrementally handled by fast subsets of the full kernel machine -- invoking the full machine only on the hardest queries for which it is truely needed to retain its high classication rate.

Return to the seminar schedule