Comments on 'Learning from imprecise and fuzzy observations: Data disambiguation through generalized loss minimization' by Eyke Huellermeier

摘要

Eyke Huellermeier provides a very convincing approach to learn from fuzzy data, both about the model and about the data themselves. In the process, he links the shape of fuzzy sets with classical loss functions, therefore providing strong theoretical links between fuzzy modeling and more classical machine learning approaches. This short note discusses various aspects of his proposal as well as possible extensions. I will first discuss the opportunity to consider more general uncertainty representations, before considering various alternatives to the proposed learning procedure. Finally, I will briefly discuss the differences I perceive about a loss-based and a likelihood-based approach.