Postdoc position in machine learning in Grenoble

The computer science and applied mathematics labs in Grenoble are currently seeking candidates for an open postdoctoral position in the field of machine learning.
Title: Learning welding prediction and classification models from various, heterogeneous sources
Duration: 2 years
Partners: LIG UGA, LJK UGA, Total
Starting date: November/December 2016 if possible
Description: Machine learning methods are meeting an increasing success in various domains, such as marketing with customer behavior prediction, health with patient diagnosis and industry with the optimization of industrial processes.
The present project fits within a general problem addressed by Total on trying to predict, from various characteristics (or parameters/variables), different properties (as mechanical properties under stress) of welding in pipelines. The parameters can take various forms (quantitative or qualitative, ordinal or non-ordinal, real or Boolean) and are highly heterogeneous. They however need to be combined in order to obtain good prediction and one of the main challenges of this project is precisely to find the best way to combine different parameters for enhanced prediction and classification. In parallel, it is of course important to determine whether the different parameters are correlated or not, and to make use of possible correlations in the prediction/classification tasks. The developed method will have to be well adapted to large scale, heterogeneous datasets that are common to many different domains; it will furthermore be applied to the prediction of weld properties from paramters of the welding process.
During the project, the successful candidate will have to address the following points:
1. Study correlations between variables of many different types and extend existing models/methods to integrate all data types as well as their dependencies. The dataset collected by Total for studying welding in pipelines is unique by the diversity of the variables it relies on (product names, physical measures, manual annotations, …). This diversity constitutes a major challenge for all existing data analysis and machine learning methods. We will also try, whenever possible, to quantify the uncertainty associated with the representation of each data type;
2. In addition to the above-mentioned datasets, physical phenomena (as welding) are often described via equations that display relations between variables; they are also subject to simulations aimed at assessing their future evolution. One of the goals of the project will be to study how one can couple machine learning and physical equations and simulations to improve the accuracy of the prediction. This is a promising line of research that can bring together communities that do not usually work together;
3. Provide tools to help experts understand the results obtained by the models developed.
This will include:
● Working with a team of computer scientists and mathematicians
● Developing new machine learning/data analysis models
● Implementing and testing the models developed
Required skills: Ph.D. or equivalent experience in computing, modeling, machine learning, statistics and applied mathematics
Application: The application should include a CV mentioning the publications, and any relevant documents. Candidates are encouraged to provide contact information to reference persons. Please send your application in one single pdf to Marianne.Clausel-AT-imag.fr and Eric.Gaussier-AT-imag.fr