skip to content
 

The role of invariance in learning from random graphs and structured data

Presented by: 
Peter Orbanz Columbia University
Date: 
Thursday 20th October 2016 - 14:00 to 15:00
Venue: 
INI Seminar Room 2
Abstract: 
Graphon models can be derived from the concept of exchangeability, which has long played an important role in (Bayesian) statistics. Exchangeability is, in turn, a special case of probabilistic invariance, or symmetry. This talk will be an attempt to explain, in as non-technical a manner as possible, why and how invariance is useful in statistics. I will cover some general results, discuss how different notions of exchangeability fit into the picture, and how invariance can be regarded as a consequence of assumptions on the process by which the data was sampled. All of this ultimately concerns the problem: What can we learn about an infinite random structure if only a finite sample from a single realization is observed?


University of Cambridge Research Councils UK
    Clay Mathematics Institute London Mathematical Society NM Rothschild and Sons