Keywords: uncertainty theories, information and data fusion, learning.
Topic: The internship concerns the problem of merging uncertain information together to obtain better information, and how to learn such merging. Information will be collected through a game with a purpose.
Context: Many applications involve extracting information about uncertain quantities from experts or human sources: image annotation, medical diagnosis, expert opinions in risk analysis. To model such kinds of uncertainty, it is desirable to use theories more expressive than probabilities and intervals.
Although these theories are well developed, there are surprisingly many questions related to information fusion that remains, and few benchmark data that allow to empirically answer these questions. Such empirical studies are essential to validate theoretical studies and to test new proposals, as well as to discover new theoretical issues.
Goal: The internship primarily aims at answering pending questions (both practical and theoretical ones) within information fusion. There are indeed many issues remaining with such a setting, for instance:
How to assess the quality of the answers provided by the user when the true value of the quantity is known (possibly uncertainly as well)?
Whether the inclusion of quality assessment in the merging process of multiple-user answers has some impact on the quality of the obtained result?
An additional question is also to know if an efficient information fusion rule can be learnt from data, or said differently if in our case we can learn a fusion rule that will beat any of the players (using those player answers). To fulfill this purpose, we are currently developing a game with a purpose such that player uncertain answers will feed a database. The database will then be used to perform information fusion tasks in order to answer the previously mentioned questions. The candidate will also participate to the game development and evolution.
Outputs: The proposals aims at exploring information fusion problem from a new perspective, using the recent idea of game with a purpose. It will lead to the establishment of a first benchmark that will allow testing information fusion rules. Information fusion is at the core of many settings, such as autonomous vehicles or robot swarms (where multiple sensors are at work) or expert opinions in risk analysis. Beyond this benchmark, the internship will address questions still unanswered in the information fusion setting, such as the interest of integrating source quality to the fusion process, or in which cases it is possible to learn an efficient fusion schemes.
Expected profile: The candidate is expected to have an interest towards in at least one of the field of uncertainty modeling, machine learning or information fusion. Previous experience in one of these fields, as well as in programming, is a plus.
(Msc. internship of six months in Computer Science, possibly followed by a thesis grant if the candidate proves her/himself to be excellent).
Salary: 430€ / month.
Laboratory: Heudiasyc (HEUristique et DIAgnostic des SYstèmes Complexes, Heuristics and diagnostics for complex systems), a joint research unit between the Université de Technologie de Compiègne and the CNRS (attached to the INS2I). Heudiasyc operates in the field of Information, Technology and Communication Sciences (STIC), namely computers, automation, robotics, decision making and image processing. Heudiasyc’s activity is based on the synergy between upstream research and final research to meet the major challenges of today’s society (safety, mobility and transport, environment and health), working hand in hand with business partners, in particular industrial companies. Several platforms and demonstrators, developed within the laboratory, illustrate this desire to bring fundamental research closer to the complexity of its real-life applications. The candidate will work within the DI (Decisions, Images) team.
Contacts: Sébastien Destercke, Nadia Ben Abdallah, Mylène Masson {sebastien.destercke; nadia.ben-abdallah; mmasson}@utc.fr
--
====================================
Sebastien Destercke, Ph. D.
CNRS researcher in computer science.
Université de Technologie de Compiegne
U.M.R. C.N.R.S. 7253 Heudiasyc
Centre de recherches de Royallieu
F-60205 Compiegne Cedex
FRANCE
Tel: +33 (0)3 44 23 79 85
Fax: +33 (0)3 44 23 44 77
====================================