The training of classifiers is a central Machine Learning task. However, with increasing data set sizes, standard such procedures become increasingly time intensive for practitioners because increasing amounts of labels for the training data have to be provided. Semi-supervised learning has consequently moved into the focus of many research groups. A typical strategy is to adapt standard deep learning approaches to the problem. Here, I will discuss an alternative strategy using deep mixture models instead of standard deep neural networks. With the help of the deep mixture approach, I will highlight some principal problems for learning from data with few labels. A particular focus will be the question how much the time required to obtain a good classifier can be reduced. I close with a discussion on the current state-of-the-art in the field of semi-supervised learning and an outlook.
Hinweis: Die Veranstaltung richtet sich an Personen mit Vorkenntnissen.