HIT Featured Talks: Luciano Serafini
Torre Archimede - Padua
08.12.2016
Learning and reasoning with knowledge data: A review of my favourite approaches.
Automatic text simplification: confluence of challenges for Psychology and Computer Science
Time: 15:30
Venue: 1BC45, Torre Archimede
Speaker: Luciano Serafini
Affiliation: Fondazione Bruno Kessler, Trento, Italy
Organization: Alessandro Sperduti
Abstract:
Hybrid domains are domains where objects are organised in a structure (e.g., a labelled graph) and some of the components of such a structure is associated to a set of numerical attributes or features (e.g., the vertex of a graph are associated with numeric features, and the arcs are associated to weights). In these domains, structural properties and numerical properties are tightly connected and they cannot be managed separately. On the one hand, logical approaches provide excellent tools to describe some known structural properties of a domain and to automatically infer via deductive reasoning new true properties about the structures that logically follows from them. On the other hand, machine learning techniques, such as regression, kernels, support vector machines, neural networks and graphical models, are quite useful and flexible methodologies to infer, via inductive reasoning (aka learning), new numerical and structural properties from the numerical attributes/featur!
es associated to the domain. Since it's beginning Artificial Intelligence dream has been to find a satisfactory integration of these two forms of inference. Along the years many proposals have been done, but they never get to the state of being mature enough. It's only in the recent years, that researchers were looking to suitably combine Logical reasoning and machine learning in hybrid domains, and they developed a number of frameworks which seem to be promising for the solution of such a key AI challenge. In this talk I'll revise some of them. In particular, I'll give an overview of the main principles, which is at the base of all the modern approaches, and I'll briefly present some of the emerging approaches.
