Program‎ > ‎

2014-06-19: Chris Holmes and Pierre Pudlo

publié le 7 avr. 2014 à 07:47 par Nicolas Chopin   [ mis à jour : 16 juin 2014 à 01:06 ]
Changes of plan: Talks will be in Amphithéatre Darboux, inside IHP (not some other place outside, as initially announced).

* at 3pm, Chris Holmes, Oxford University, will talk about:

Robust statistical decisions via re-weighted Monte Carlo samples

Large complex data sets typically demand approximate models at some level of specification. In such situations it is important for the analyst to examine the robustness of conclusions to approximate predictions. Recent developments in optimal control and econometrics have established formal methods for linear quadratic state space models (see e.g. Hansen and Sargent 2008) by considering the local-minimax outcome within an information divergence (Kullback-Leibler) neighbourhood around the approximating model. Here we show how these approaches can be extended to arbitrary probability models using Monte Carlo methods. We derive theoretical results establishing the uniqueness of the Kullback-Leibler criteria, as well as Bayesian non-parametric methods to sample from the space of probability distributions within a fixed divergence constraint.

* at 4.15pm, Pierre Pudlo, Université de Montpellier 2, will talk about:

ABC and machine learning

Since its introduction in the late 1990’s, the ABC method has been analysed from several perspectives, from a pure practical one to a non-parametric one. We develop a new vision on how generic machine learning tools like the random forests due to Breiman (2001) can be used to run model selection in the complex models covered by ABC techniques. Our perspective radically alters the way model selection is operated as we do not compute posterior probabilities for the models under comparison, which cannot be reliably estimated, but propose instead to compute the performances of the selection method. As an aside, we argue that random forest methods can be adapted to the settings of interest, with a recommendation on sparse implementations of the random forest tree construction, using subsampling and reduced reference tables.