|Time:||July 14, 2021, 2:00 p.m. (CEST)|
|Download as iCal:||
With this ML Session series we intend to provide individual and independent lecture sessions on ML related topics.This time Bernard Haasdonk will talk about "Kernel Methods".
In this presentation we address different machine learning and pattern/data analysis tasks and how they can be solved by kernel methods. Based on a given choice of kernel, many geometric operations with data can be represented in the corresponding feature space. This especially motivates classification by support-vector-machines (SVM). Similarly, regression is possible by SVR. Feature extraction can be obtained by Kernel Principal Component Analysis (KPCA). A multitude of further kernel methods exists that solve suitable tasks. It can be demonstrated that incorporation of prior knowledge, e.g. invariances, improves the generation ability of the resulting models. Some theoretical considerations (statistical learning theory, regularization, representer theorem) explain, why these methods are frequently successful.