Colloquium Computer Science Prof. B. Hammer
|When:||Th 28-03-2019 16:00 - 17:00|
Interpretable models, learning with reject option, and learning with drift
Machine learning technologies have revolutionized many domains such as vision or language processing, yet many models and in particular the majority of mathematical substantiations are restricted to the classical setting of batch learning (i.e. data are given prior to training), stationary distributions (i.e. data characteristics do not change during their lifetime), and optimization of the classificatio error (rather than optimizing strategies, when to best abstain from a classification in unclear cases). In the talk, we will have a glimpse on machine learning technologies which, by design, provide interpretability and open up avenues how to extend them to reject options, how to transfer known models tonew settings, and how to learn for possibly non-stationary streaming data.