Kernel-based Learning from Heterogeneous Information Sources

Abstract

In the idealized setting of machine learning, we learn a classifier separating two distributions. The reality is, however, much more complex: in practice, we may face thousands of classes or structured outputs, unlabeled and noisy data, dependent data or non-stationary distributions, as well as multiple data representations or learning tasks. In my talk, I give an overview of my research on these challenging topics.

I illustrate my mode of operation - which ranges from fundamental theory via practical algorithms to applications - by means of my key research topic: "Kernel-based Learning from Heterogeneous Information Sources". Heterogeneous information sources occur in many application areas, especially when complementary features or multiple sensors are deployed. A prime example for such an application domain are the life sciences. I propose a kernel-based approach for synthesis of heterogeneous information sources, for which I present theoretical bounds and practical algorithms that are efficient and effective in life science applications and beyond.

http://www2.informatik.hu-berlin.de/~kloftmar/