Thursday 21st May 2015
Magdalen Grove Auditorium, Magdalen College, Oxford
2.30 p.m. Professor Sonia Petrone, Department of Decision Sciences, Bocconi University, Milan, Italy
Title: Extensions of exchangeability for bivariate evolutionary phenomena
Abstract: Exchangeability plays a fundamental role in Bayesian statistics, and it is a main tool for studying the limiting behavior of some evolutionary stochastic processes, such as urn processes. However, in some applications, exchangeability is too restrictive, as it implies stationarity. Moving from results by Kallenberg  and Berti, Pratelli and Rigo , we revise a notion of stochastic dependence that generalizes exchangeability while preserving some of its predictive properties. Roughly speaking, such condition can be regarded as exchangeability without stationarity.
In the talk we present a novel extension to bivariate processes, that provides a weaker form of partial exchangeability, and preserves some main properties. In particular, we prove limit theorems for the empirical and predictive distributions. We illustrate the results in some examples, including extensions of known classes of prior distributions for Bayesian nonparametrics, and discuss possible applications in competing networks. This is a joint work with Sandra Fortini and Polina Sporysheva.
3.30 p.m. Tea, coffee and biscuits
4.00 p.m. Professor Bernhard Schoelkopf, Max Planck Institute for Intelligent Systems, Tuebingen, Germany
Title: Toward Causal Machine Learning
Abstract: In machine learning, we use data to automatically find dependences in the world, with the goal of predicting future observations. Most machine learning methods build on statistics, but one can also try to go beyond this, assaying causal structures underlying statistical dependences. Can such causal knowledge help prediction in machine learning tasks? We argue that this is indeed the case, due to the fact that causal models are more robust to changes that occur in real world datasets. We touch upon the implications of causal models for machine learning tasks such as domain adaptation, transfer learning, and semi-supervised learning.
We also present an application to the removal of systematic errors for the purpose of exoplanet detection. Machine learning currently mainly focuses on relatively well-studied statistical methods. Some of the causal problems are conceptually harder, however, the causal point of view can provide additional insights that have substantial potential for data analysis.
5.00 p.m. Drinks Reception
Please book your place at this event by emailing firstname.lastname@example.org
A map showing the location of the Grove Auditorium at Magdalen College can be found here.