16 September 2008 


Geometrical Approaches to Active Learning
 

Prof. Dr. rer. nat. Klaus Obermayer 
TU Berlin and Bernstein Center for Computational Neuroscience
 


Learning from examples is a key property of automous agents, and - consequently - many learning algorithms for inductive learning have been developed over the years in the machine learning field. In my contribution I will focus on a particular class of algorithms referred to as "optimal experimental design" or "active learning".Learning machines, which employ these algorithms, request examples which are maximal "informative" for learning a predictor rather than "passively" scanning their environment. There is a large body of empirical evidence, that active learning algorithms - though computationally more expensive - are more efficient in terms of the required number of training examples. Hence active learning should be preferred if training examples are costly to obtain. In my contribution I will present an approach to active learning for binary classification, which is based on the geometrical concept of a "version space". Using an adapted Riemannian metric for the data space, I construct a space of classifiers, where the distance between classifiers is given by their probability of disagreement. Using isometries induced by charts of the sphere, I then relate general binary classification problems to binary classification problems on the hypersphere. Active learning then proceeds by optimally subdividing the "version space", defined as a submanifold of the hypersphere whose classifiers are consistent with the training data. I provide hard upper bounds on the generalization error and provide evidence for an exponential (rather than polynomial) decrease of the generalization error with the number of training examples. Finally, I will discuss the performance of practical algorithms for subdividing version space using kernel methods and Monte-Carlo techniques. The presentation covers joint work with F. Henrich and A. Paus

About Telekom Innovation Laboratories

LogoDTGerman

As one of the world's leading telecommunications and information technology service providers, Telekom Innovation Laboratories is setting international standards.

Contact Us

contact_us

Telekom Innovation Laboratories at Ben-Gurion University of the Negev - P.O.B. 653 Beer Sheva, 84105, Israel
Phone: +972 8 6428120/21
e-mail : This email address is being protected from spambots. You need JavaScript enabled to view it.