|
Support Vector and Kernel learning methods have gained considerable popularity in the machine learning community in recent years. I will give an introduction to the background of these methods, while also high-lighting their relation with Statistical Learning Theory.
Support Vector Machines (SVM) are the best known group of algorithms that successfully exploit these methods. Typically these algorithms are formulated as a quadratic programming problem. I will discuss a new and simpler formulation of the SVM for the case of classification, called DirectSVM. The formulation exploits the property of linear independence of the support vectors, for constructing a geometrically more intuitive algorithm.
I will also discuss applications to visual 3D object recognition and detection. In particular, I will address the problem of incorporating domain knowledge/prior information for learning such object recognizers. In contrast to e.g. segmentation or feature extraction methods that encode such prior information, I will instead propose a quite naive learning approach to 3D object recognition, with prior knowledge conveyed by pedagogically selected training data only. Concretely, a pedagogical support vector learning approach to background-invariance will be proposed.
PhD Thesis:
Danny Roobaert. Pedagogical Support Vector Learning: A Pure Learning Approach to Object Recognition