PUB - Publikationen an der Universität Bielefeld

Machine Learning (ML) methods are increasingly used and outperform humans in many specified and well-defined tasks. Considerable research focuses on optimizing the performance of such methodologies. However, the nature of application areas poses further challenges. For example, in critical domains, a false model behavior poses the risk of fatal mistakes. This is particularly relevant in traffic or medicine. In the latter, the data frequently contains sensitive information which should be preserved. Further, much data are recorded on distributed devices with limited computational power, like smartphones and peripheral devices. Hence, models of low complexity are required. As due to technical, legal, or strategic constraints, data transfer is limited employing intelligent mechanisms is crucial. This requires the consideration of further aspects beyond mere accuracy, namely privacy, robustness, efficiency, and distribution of the data itself.
In this thesis, I address these additional aspects, namely privacy, robustness, efficiency, and distribution of the data for prototype-based classifiers. In particular, I focus on Generalized Learning Vector Quantization (GLVQ) models and their variation to metric adaptations. I show that the original GLVQ model bears the risk of revealing private information of samples present during the training. I propose three versions of training schemes provably obeying privacy. Further, I propose a novel reject option scheme for GLVQ models. Thereby increasing the robustness of the model is achieved. To reduce the complexity of a model and obtain a sparse representation of feature vectors, I apply regularization to the GLVQ scheme. Finally, I propose a methodology fusing model parameters of several models trained on distributed data sets.