SC²S Colloquium - October 8, 2010

Date:	October 8
Room:	02.07.023
Time:	14:00 pm, s.t.

Narek Melik-Barkhudarov: Spatially Adaptive Semi-Supervised Learning (MA)

Conventional classification techniques rely on labeled data only to train. Semi-supervised learning allows to take advantage of the large amounts of unlabeled data usually available in classification problems. The implementation of semi-supervised learning on adaptive sparse grids allows the algorithm to be used for substantially high-dimensional datasets. Since distances between data points are used extensively throughout semi-supervised learning, computational problems as well as questions of meaningfulness of euclidean distances arise. A number of approaches to these questions is discussed, which allow to improve performance for high-dimensional data.

The thesis is concerned with the implementation of semi-supervised learning using the SG++ framework, and the results for classification on several publicly available real-life datasets.