Rapid Uncertainty Computation with Gaussian Processes and Histogram Intersection Kernels

TitleRapid Uncertainty Computation with Gaussian Processes and Histogram Intersection Kernels
Publication TypeConference Paper
Year of Publication2012
AuthorsFreytag, A., Rodner E., Bodesheim P., & Denzler J.
Other Numbers3377

An important advantage of Gaussian processes is the abilityto directly estimate classication uncertainties in a Bayesian manner. Inthis paper, we develop techniques that allow for estimating these uncertaintieswith a runtime linear or even constant with respect to thenumber of training examples. Our approach makes use of all trainingdata without any sparse approximation technique while needing only alinear amount of memory. To incorporate new information over time, wefurther derive online learning methods leading to signicant speed-upsand allowing for hyperparameter optimization on-the-y. We conductseveral experiments on public image datasets for the tasks of one-classclassication and active learning, where computing the uncertainty is anessential task. The experimental results highlight that we are able tocompute classication uncertainties within microseconds even for largescaledatasets with tens of thousands of training examples.

Bibliographic Notes

Proceedings of the 11th Asian Conference on Computer Vision, Daejeon, Korea. Recipient of a best paper honorable mention award.

Abbreviated Authors

A. Freytag, E. Rodner, P. Bodesheim, and J. Denzler

ICSI Research Group


ICSI Publication Type

Article in conference proceedings