The recent performance gain makes now possible the practical use of machine learning designed components in complex systems as exemplified by the recent demonstrations of so-called "autonomous cars". However, in this kind of application, the computer system ("Autopilot") able to drive the vehicle is only considered to be an assistant: when the artificial system fails, it is the driver's responsibility to detect the system abnormal behavior and to ultimately take control of the vehicle.
However, for critical functions such as automatic landing, automated medical diagnosis or "see and avoid" capability in complex dynamic environments, abnormal behaviors may not be detected sufficiently early and may contaminate other components of the system, potentially leading to disastrous consequences.
One question that naturally arises to improve the usability of machine learning based processes is to provide a way to guarantee and even control a given level of performance. This is a difficult question for such processes which are effective in their empirical domain of expertise – the learning database – but are often unable to state why they are so.
The post-doctoral study will address the question of assessing the good behavior of systems involving components parametrized by machine learning techniques. The emphasis will be on deep learning approaches that are now the current state of the art in many signal processing and computer vision applications.
There will be several possible directions of research to achieve this objective:
It is expected from the candidate to address theoretical issues that will be validated on computer vision functions, typically object recognition and semantic segmentation.
Keywords: Machine Learning, Computer Vision, Deep Learning, Safety of algorithms, Statistical tests, Adversarial networks.
ONERA, a central player in aeronautics and space research, employs approximately 2,000 people. As a government expert, ONERA prepares tomorrow’s defense, meets future aerospace challenges and contributes to the competitiveness of the aerospace industry. It masters all of the disciplines and technologies in the field (http://www.onera.fr/en).
The postdoc will be integrated in a team of about 20 permanent researchers working in computer vision in the Information Processing and Systems Department located at Palaiseau near Paris. Activities of the team are divided equally into research and applications.
The work is expected to address rather fundamental issues but can make use of the various experimental facilities available in the laboratory (cameras, robot, drone...) to validate the results. The candidate will be involved in two ONERA research projects: DELTA addressing deep learning techniques applied to aerospace problems (https://delta-onera.github.io/), and SUPER addressing the question of safety of computer vision algorithms for aerospace applications.
The candidate should have a strong interest in machine learning and computer vision. Given the type of addressed issues, a theoretical background in statistics is also welcome. Of course, a good programming experience is expected (C++, Python, Deep Learning frameworks, GPU).
Send a motivation letter describing your research interests and your match to the proposed position, your CV and publication list, and names and email address of 3 references to Stéphane HERBIN (firstname.lastname@example.org).
(c) GdR 720 ISIS - CNRS - 2011-2018.