ResearchBib Share Your Research, Maximize Your Social Impacts
Sign for Notice Everyday Sign up >> Login

Two Marie-Curie PhD students at INRIA - Ecole Normale Supérieure

Country/Region : France

Website : http://macsenet.eu

Description

The machine learning group at INRIA - Ecole Normale Supérieure in Paris is recruiting two PhD students starting this summer.
The funding comes from the Marie Sklodowska-Curie Innovative Training Networks (ITN) MacSeNet (see more details at http://macsenet.eu/). This means that the Phd student will receive a competitive salary and collaborate with other members of the network.
To be eligible for this position funded by EU, the candidate must not have spent more than 12 months in France in the 3 years prior to starting. However, any nationality can apply.
The two projects will be undertaken within the Computer Science Department of Ecole Normale Supérieure, located in downtown Paris, within the INRIA/CNRS/ENS project-team SIERRA (http://www.di.ens.fr/sierra/).
The two PhD projects are as follows:
(1) Non-linear adaptive sensing/learning
In recent years, non-linear features and measurements have shown great empirical promises in machine learning and signal processing. The aim of this PhD project is to develop new methodologies adapted to non-linear measurements and predictors, with a particular focus on (a) methods based on convex reformulation of neural network training, (b) computationally efficient methods adapted to large-scale problems typically found in applications, and (c) principled adaptivity to the learning capacity of the underlying problems.
(2) Robust unsupervised learning
The aim of this PhD project is to develop new robust unsupervised learning methods suitable for the needs identified in other disciplines and by industrial partners. Several approaches to unsupervised learning exist; we plan to follow our earlier work on matrix factorization approaches, with a particular focus on robustness. This is both an optimization and a statistical problem, as the usual matrix factorization approaches are (a) based on non-convex optimization and (b) have several hyperparameters which may impact performance. The two problems are often treated separately. For example, Bayesian approaches will provide elegant solutions to the adaptivity of hyperparameters but typically put little emphasis on provable computational and stability issues, while research on convex relaxations often ignores the statistical issues related to hyperparameters and the automatic adaptivity of model to data. Combining these two approaches should bring the best of both worlds.
Application deadline is May 1st, 2015. Application details may be found at http://macsenet.eu/#1|0 and http://macsenet.eu/#1|1 .

Last modified: 2015-03-28 09:41:41