Giuseppe Marra (DIISM, University of Florence and Siena)

Jun 12, 2019 – 11:00 AM

DIISM, Artificial Intelligence laboratory (room 201), Siena SI

Description

Markov Logic Networks (MLN) are a very well-known example of statistical relational model. In order to build good models, MLNs ask the user to define in advance a set of first-order logic formulas describing the relational properties of the domain. However, in most of the cases, such properties are only partially known or not at all, requiring MLNs to make use of external rule learners with, arguably, low effectiveness. In this seminar, we propose Neural Markov (Logic) Networks, where relational properties are expressed as neural networks over fragments of the domain. Such neural networks are trained together with the probabilistic model, allowing us to get rid of external rule learners. Thus, Neural MLNs allow us to learn simultaneously both the structure and the weights of the probabilistic model. We show the foundations of the theory of Neural MLNs, together with some applications to graph generation and knowledge base completion.