Thursday 09 Feb 2017: Constructive Machine Learning
Fabrizio Costa - University of Exeter
Harrison 170 14:30-15:30
Complex entities can be modeled using graph data structures in a natural and expressive way. Several types of discriminative systems that can deal with graphs in input are known in machine learning literature (e.g. recursive neural networks, graph kernels, graphical models, etc), however, there are not many generative approaches that can output structures belonging to a desired distribution or class. Constructive machine learning (CML) studies the problem of sampling a distribution defined over graphs using data driven generative models. Systems for CML are of great interest as they can be used to address a vast class of design problems in a variety of domains, ranging from de novo drug synthesis (biology), to automatic algorithm generation (computer science) and level or character generation (games).
In this talk I will introduce some initial ideas on how to formulate generative problems in structured domains. Finally I'll present a cursory list of my recent works in the constructive machine learning field.