Francesco Giannini (DIISM, University of Siena)
Jun 21, 2018 – 9:30 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI
In a learning from constraints problem, the prior knowledge can be expressed by logical formulas and then converted into real valued functions. Hence the satisfaction of a certain formula F can be enforced by requiring the minimization of a certain decreasing transformation of the function corresponding to F. Since in general, the functions corresponding to logical formulas have to be evaluated on a large amount of data yielding a [0,1]-value per grounding, the choice of an opportune fuzzy aggregator of such evaluations plays a special role. In particular, we are interested in the relations between t-norms, a special class of fuzzy aggregators, and typical machine learning loss functions. Some preliminary results are discussed as well as a general approach to generate loss functions defined by different fuzzy aggregators.