XRCE organises public scientific seminars on a regular basis which you are welcome to attend. These seminars are an occasion to exchange with researchers from various backgrounds and to broaden scientific expertise. You can subscribe to our seminar RSS Feed for dates, speakers and topics.
Many real world dynamical systems such as industrial processes,activities of human users, bio-medical systems, etc. can beinterpreted as following different regimes. For instance an industrial process can follow a set of nearly deterministic laws in normaloperation (a normal regime), but show markedly different behaviorafter the failure of certain parts (a failure regime). A switchinglinear dynamical system (SLDS) is a probabilistic model thatexplicitly models such regime changes.
The conditional independence structure of the SLDS is extremelysimple, namely a chain. This allows for a straightforward recursiveinference algorithm. However, the model consists of both multinomialand Gaussian variables, and local integrals are over conditionalGaussian distributions. These integrals have analytic solutions, butsince the conditional Gaussian family is not closed undermarginalization, the size of the messages -- the information that ispassed along the graph in a recursive algorithm -- grows exponentiallywith the problem size.
Motivated by the expectation propagation framework from Tom Minka, I describe how a Kikuchi free energy with weak marginalizationconstraints, i.e. constraints that only enforce equality on expectedsufficient statistics, can be used to formulate an approximateinference algorithm.
If we restrict the switching linear dynamical system in an appropriateway, we obtain a model that can detect change points in a dynamicsystem. I will show that for this change point model the Kikuchi freeenergy approach results in an interesting class of approximateinference algorithms where a trade-off can be made betweencomputational complexity and accuracy.Slides (558.23 kB)
( web e-mail ), Maitre Assistant at Equipe connexionniste, Laboratoire d'Informatique de Paris 6, Paris, France will give a talk:
A Study of the Area Under the ROC Curve with application to Automatic Text Summarisation
Many real-life applications are concerned with the ranking of objects instead of their classification. This is for example the case in Information Retrieval, where for a given query we are interested in retrieving documents from a fixed collection w.r.t to this query. In my speech I am interested in a particular case of ranking, the bipartite ranking, where the objects are assumed to come from two categories, positive and negative, and where the goal is to learn a scoring function by optimising the Area Under the ROC (AUC) such that positive examples get higher scores than negative ones. I'll particulary show that solving bipartite ranking problems under a classification setting is suboptimal. I'll present a data-dependant generalisation error bound for the AUC and show that under this setting kernel function classes have strong generalisation guarantees provided that the weights of the functions are small. Finally I'll provide empirical evidence on the automatic text summarisation task that the ranking setting is better suitted for the problem than a classification setting.Slides (436.23 kB)
( web e-mail ), Head of "Information: Signals, Images, Systems" Research Group at School of Electronics and Computer Science, University of Southampton, Southampton, U.K. will give a talk:
Learning Complex Mappings using an SV Strategy
An approach to learning mappings between complex spaces is presented that reduces the learning complexity to that of a one-class SVM. Results are presented using the approach for a hierarchical classification task.Slides (442.98 kB)