讲座信息
9.04 | Learning to optimize graphs
2019.09.02
演讲者 Fabrizio Costa
头衔职位 Lecturer
时间 2019 年 9 月 4 日(周三)上午 10:00-11:00
地点 复旦大学邯郸校区老逸夫楼 407
联系人 朱山风,zhusf@fudan.edu.cn

演讲简介

Graphs offer a powerful yet intuitive way to encode information. Given a problem of interest, domain experts have often no problem in identifying the key elements (nodes) and their corresponding relations (edges). Several methods that operate directly on graph representations have been developed in the Machine Learning literature, ranging from artificial neural networks for structured data to graph kernels, from inductive logic programs to probabilistic graphical models. These approaches have been used to tackle supervised and unsupervised problems, but only recently there has been an interest in addressing generative problems, that is the problem where the output is a collection of newly constructed graphs that exhibit some desired property inferred directly from given examples.

The applications of graph learning techniques are far-reaching since they are a way to achieve “design by example”. In the bio-sciences one can formulate the problem of drug design as a graph learning task given a sample of toxic and non-toxic molecular graphs. In the domain of computer games, one could design various assets (e.g. characters or game levels) using a measure of user satisfaction and automatically generate an endless stream of ever challenging environments.

In this lecture we will present some recent work that combines ideas from graph grammars and graph kernels with Bayesian optimization to address the interesting problem of learning to optimize graphs.

关于讲者

Fabrizio received his PhD in Computer Science from the University of Firenze, Italy. Before joining Exeter University as a Lecturer in Data Analytics (2017), he worked at the Universita' degli Studi di Firenze, Italy, at the Katholieke Universiteit Leuven, Belgium and at the Albert Ludwig University of Freiburg, Germany. His research in Constructive Machine Learning (i.e. the development of data driven procedures to generate structured instances endowed with desired properties) has been awarded the best paper prize at the 2013 workshop on Constructive Machine Learning at the Neural Information Processing Systems Conference (NIPS) and he was a co-organizer of the workshops on Constructive Machine Learning held at the International Conference on Machine Learning (ICML) in Lille in 2015 and at NIPS in Barcelona in 2016 .
© 2019 复旦大学计算机科学技术学院 地址:上海市张衡路825号 Tell:+86-21-51355555 Fax:+86-21-51355558 Emall:cs_school@fudan.edu.cn
复旦大学计算机科学技术学院
扫一扫了解学院