Annonce

Les commentaires sont clos.

M2 internsip: Runtime Monitoring of Deep Neural Networks

26 Octobre 2022


Catégorie : Stagiaire


With the increasing use of Machine Learning in critical autonomous systems, runtime monitors are essential components to detect prediction errors and keep the system in a safe state during operations.

Master 2 internship proposal
ANITI CertifAI-Towards the certification of ML-based systems
J. Guérin, K. Delmas & J. Guiochet

 

Internship data

Title Runtime Monitoring of Deep Neural Networks
Keywords Machine Learning Safety, Fault Tolerance, Computer Vision
Technologies Programming, Python, Pytorch, Scikit-learn, Git
Location LAAS-CNRS, Toulouse, France
Level Master 2 (MSc)
Duration/Dates 4-6 months
Student profile Computer Science, Artificial Intelligence, Mathematics/Statistics
Salary ~600 euros/month

Supervisors
Joris Guérin (Espace-DEV),
Kévin Delmas (ONERA),
Jérémie Guiochet (LAAS-CNRS)

Internship description

With the recent progress in machine learning (ML) research, deep neural network (DNN) architectures are now used to address safety-critical tasks, such as self-driving cars, surgical robots, and drones landing. Online fault tolerance approaches, or runtime monitors, are a promising research direction to improve the safety of such systems. A runtime monitor is a component aiming to identify and reject unsafe data encountered at inference time.

In the past two years, our team has conducted many researchabout runtime monitoring of DNNs. Our recent efforts include the introduction of new evaluation criteria, the release of a new benchmarking library, and the development of new statistical models for efficient monitoring of DNN internal layers activations.In particular, we showed that different layers of the DNN represent data differently and require specific statistical modeling for efficient monitoring.

In this context, the intern is expected to participate to this ongoing effort by contributing to (atleast) one of the following research directions:
- Propose and test new safety monitoring approaches for DNNs.
-Propose and test new ways to optimize the monitoring approach for a specific DNN layer.
-Propose and test new ideas for combining results of an ensemble of monitoring approaches, applied to different layers of the monitored DNN.
The intern will also contribute to maintain and extend the benchmarking library. Possible extensions can consist in adding new ML tasks, new datasets, new DNN architectures, and new recent monitoring approaches from the literature.

Application process

To apply, please send yourCV,cover letter, andtranscripts(marks) of the last 2 years to
joris.guerin@ird.fr, kevin.delmas@onera.fr and jeremie.guiochet@laas.fr