Imprecision in Machine Learning and AI

I this note we consider five different relevant problems in AI and machine learning. We argue that possible solutions to such problems might be achieved by replacing the probability distributions in the systems with sets of them. Such a robust approach is based on the so-called impreciseprobabilistic framework. The proposed solutions provide a persuasive justification of the imprecise framework. The problems we consider are: • proper treatment of missing data, • reliable classification, • sensitivity analysis, • feature selection, • elicitation of qualitative expert knowledge. Before reporting a separate discussion for each problem, let us briefly resume the general ideas characterising imprecise-probabilistic methods.