The CEDRIC-Lab at Conservatoire National des Arts et Métiers (Paris, France) is inviting applications for a fully funded PhD position in the cross-field between Automatic Control and AI. The successful candidate will join a vibrant research community and work on the project entitled "Set-Valued Observers For Neural Networks: Applications To Their Supervision".
The thesis aims to develop systematic and unified set-valued algorithms that quantify the safety and robustness of neural networks by determining the over-approximation
of their bounded outputs under the assumption that (i) the uncertainties are unknown but bounded, and (ii) all possible neural network inputs are within a bounded set. For (ii), there are two possibilities to interpret this bound: (1) the first way is that the noisy measurement is contained within a bound centered at the non-noisy one and vice versa, (2) the non-noisy measurement is contained within a bound centered at the noisy one. The thesis pursues three concrete objectives:
- Objective 1: To develop a set-valued algorithm computing the tight bounded output of neural networks with bounded input based on simple operations. In the same spirit of what is done in propagating the estimation covariance of a Kalman Filter, evaluating the robustness of a neural network estimator quantifies how much the estimator can reduce the influence of the uncertainty on the prediction.
- Objective 2: To develop methods for the safety of the neural networks based on the set-valued algorithm from Objective 1. Through the use of geometric sum, tight thresholds are acquired, and an alarm is generated when anomalies occur.
- Objective 3: To develop a proof-of-concept demonstrator system that uses the methods and assurance evidence from Objectives 1-2. For a possible use case scenario, we construct an AC servo motor experimental platform. The experimental platform consists of a host computer, a NVIDIA Jetson Nano, an AC servo motor and its servo system. The NVIDIA Jetson Nano is at the heart of the platform. It receives the control instructions from the host computer and plays the role of an overseer and recipient of real-time measurement data such as torque, position, and velocity of the AC servo motor. In addition, the AC servo system will respond to control instructions from the NVIDIA Jetson Nano and drives the servo motor to implement corresponding actions.
For more details:
Requirements:
- A Master's degree in Electrical Engineering, Computer Science, Applied Mathematics, Automatic Control or a related field.
- Excellent academic performance, with a strong background in control theory, optimization, and machine learning.
- Proficiency in programming languages such as Python or Matlab.
- Good communication skills in English, both oral and written.
- Ability to work independently and as part of a team.
To apply for the PhD position, please prepare a single PDF file containing the following documents:
- A detailed CV of the candidate, including educational qualifications, academic achievements, research experience, and any relevant publications.
- A cover letter outlining your motivation for applying for the position, your research interests, and your relevant experience and skills.
- The marks of your master's degree, including your ranking and proof of obtaining the diploma giving you the right to register for a doctorate.
- Letters of recommendation from two academic referees who can attest to your research abilities and potential (if available).
- Your TOEIC score or equivalent English language test (if available), to demonstrate your proficiency in English.
Please email your application as a single PDF file to ngoc-thach.dinh@lecnam.net by the deadline of May 28, 2023. Shortlisted candidates will be contacted for an interview. If you have any questions about the application process or the position, please do not hesitate to contact us. We look forward to receiving your application.
