Place: Large lecture room.
Affiliation: Department of Computer Science, TU Dortmund. Germany.
Machine learning methods are capable of achieving impressive performances on a variety of pattern recognition tasks today. This is mostly due to two factors.
First, large annotated data collections are available and, second, increasingly complex models can be estimated reliably on this data.
However, the necessary annotations can only be provided with considerable human effort such that such supervised learning strategies have an inherent limitation.
Semi-supervised learning methods attempt to relax the requirements on the availability of labelled data by making use of some smaller amount of annotated data in combination with large amounts of unlabelled data for training purposes.
In this presentation the basic concepts behind semi-supervised learning will be introduced.
Afterwards, a novel group of semi-supervised learning methods developed in our group will be presented that rely on a combination of the multi-view principle and ensemble classification techniques.
The two basic methods employ clustering in a multi-view setting or cast the semi-supervised learning problem into a retrieval framework.
Recently, we further developed these methods into a partitioning-based semi-supervised learning technique that outperforms existing techniques.
The capabilities of our proposed approaches to semi-supervised learning will be exemplified with experimental results on character recognition tasks and on two challenging scene categorization benchmarks.
