Programme

The full programme is now available: click here

The poster spotlight information is now available: click here

The proceedings are now available: click here

Invited lectures

Ricardo Silva

Department of Statistical Science/CSML, UCL (UK)

Structured Copula Models in Supervised and Unsupervised Learning

It is not hard to understand what a copula function is: a cumulative distribution function with uniformly distributed marginals. But this seemingly simple concept has interesting consequences. Copulas provide an alternative way of representing multivariate dependencies: it allows one to encode the marginal and the dependence structure of a joint distribution separately, a type of representational modularity that complements other popular tools in machine learning such as graphical models. For instance, one could build flexible nonparametric representations for univariate marginals while using parametric forms for the potentially much more daunting task of representing associations in high-dimensional spaces - and such a parametric structure will impose no constraints on how marginals are represented. In this talk, we will first provide a brief introduction to copulas. We then proceed to exploit this concept in the context of graphical models. First, we show how to build high-dimensional copulas using a mixture of trees representation, which we then learn with a Bayesian approach. Second, we will discuss how copulas can be exploited in structured prediction models, where there is a residual association structure among a set of labels or other outcome variables of interest that remain dependent even after observing the predictive features (joint work with Robert Gramacy (Chicago), Charles Blundell and Yee Whye The).

Peter Challenor

National Oceanography Centre (UK)

Climate, models and uncertainty

Anthropogenic climate change is one of the most important challenges facing society. The associated scientific problems are difficult, urgent and fascinating. Predicting what future climates might look like, under different greenhouse gas emission scenarios, involves the use of very large computer codes to solve complex numerical models of the Earth system. We would like to have a measure of the uncertainty on such predictions. In this talk I will concentrate on the problem of how we might estimate the uncertainty on predictions of the climate from complex numerical models. I will discuss both the single and multi-model case.

Tijl De Bie

University of Bristol (UK)

Structured output prediction: from biology to music

In this talk I will survey a number of recent approaches to the task of learning to predict structured output labels, and I will discuss recent developments such as the adaptation of these methods for semi-supervised learning scenarios. The talk will be motivated and illustrated by practical use-cases from biological sequence analysis, music annotation, and more.

 

 

 

 

 

 

About KERMIT

KERMIT (acronym for Knowledge Extraction, Representation and Management by means of Intelligent Techniques) is an interdisciplinary team of (bio-)engineers, computer scientists and mathematicians, initiated by Bernard De Baets in 2000. It draws upon intelligent techniques resulting from the cross-fertilization between the fields of computational intelligence and operations research. The activities are organized in three highly interwoven research threads: knowledge-based modeling, predictive modeling and spatio-temporal modeling.

Academic Partners

BeneLearn 2012 | 24-25 May 2012 | Gent