We are glad to invite you to a seminar by Prof. Quentin Paris, titled “On High-Dimensional Linear Prediction”.
In this talk, we present performance bounds for an adaptation of the lasso estimator in the context of semi-supervised learning. In other words, the setting under consideration is that of regression with random design with a partially labelled learning sample. We will discuss non asymptotic oracle inequalities, in probability and in expectation, which highlight the interplay between the mis-specification of the linear model and the approximate sparsity level. Our results quantify the positive impact of a large amount of unlabelled data. A second part of the talk will discuss a (more computationally friendly) alternative to the lasso procedure, called exponentially weighted aggregate, which replaces the minimization step by an averaging approach.
Since September 2014, Quentin Paris is Assistant Professor at the HSE. From 2013 to 2014, he was a postdoctoral fellow at the CREST laboratory in Paris. In 2013, he obtained his PhD in Statistics from the Ecole Normale Superieure de Cachan.
If you like to participate and for further information or questions, please Liliya Abaimova.