Seminar on Computational Learning and Adaptation

  Real-Time Statistical Learning

Stefan Schaal
Computational Learning and Motor Control Laboratory
Computer Science and Neuroscience, HNB-103, Univ. of Southern California, Los Angeles, CA 90089
ATR Human Information Systems, 2-2 Hikaridai, Seika-cho, Soraku-gun, 619-02 Kyoto, Japan
sschaal@usc.edu
http://www-clmc.usc.edu

Real-time modeling of complex nonlinear dynamic processes has become increasingly important in various areas of information technology, including the on-line prediction of dynamic processes observed by visual surveillance, user modeling for advanced computer interfaces and game playing, and the learning of value functions, policies, and models for learning control, particularly in the context of high-dimensional movement systems like humans or humanoid robots. To address such problems, we have been developing special statistical learning methods that meet the demands of on-line learning, in particular the need for low computational complexity, rapid learning, and scalability to high-dimensional spaces. In this talk, we introduce a novel algorithm for regression learning that possesses all the necessary properties. The algorithm combines the benefits of nonparametric learning with local linear models with a new Expectation-Maximization algorithm for finding low-dimensional projections in high-dimensional spaces; it can be regarded as a nonlinear and probabilistic version of partial least squares regression, and also as a method of probabilistic backfitting. Variational Bayesian inference allows us to derive computationally cheap regularization against overfitting. We demonstrate the applicability of our methods in traditional benchmark datasets, synthetic examples that have thousands of dimensions, and in various applications in humanoid robotics, illustrated by video presentation.

Return to the seminar schedule