M1/M2 Internship in Computer Science / Artificial Intelligence / Machine Learning (M/F)

To face these dangers, works and tools are constantly emerging to increase the trust one can have in AI systems. Our team develops some of those tools: CAISAR and PyRAT.

The main goal of this internship is to implement a method to determine when AI system can be considered to be Abstract Safe, i.e. we can prove that the system's misclassifications respect a formal hierachy. For example, with a hierarchy that distinguish "animal" classes and "vehicule" classes, a system that misclassifies a dog as "a cat" but not as "a car" could be considered to be abstract safe.

Durée du contrat (en mois)

normally 3 or 4 months; can be up to 6 months, This internship is designed for Master 1 students (duration 3 to 4 months), but can be extended (duration 6 months) for Master 2 students.

The main objective of this internship is to introduce the notion of Hierarchical Classification in CAISAR, and use it to verify AI systems with PyRAT. The main steps are:

• to define a file format to represent the hierarchy of classes
• to modify how the provers are called by CAISAR and their outputs to determine when the AI system is abstract safe
• to experiment on at least one case study (e.g. CIFAR-10)
• if the intership is exended, there will also be a theoretical part (Abstract Interpretation framework)

CAISAR is an open-source platform that focuses on the characterization of AI systems' Robustness and Safety. In order to ensure the safety of an AI system, this platform can call several provers including PyRAT, a Python tool based on Abstract Interpretation techniques also developed at CEA in the AISER team. Those two tools are under active development, as new features are added to improve their accuracy and the expressivity of their specification language. CaiSAR is written in Ocaml, PyRAT is written in Python3.

This work will have contributions to the field of automata learning and to neural networks verification. The internship will likely conclude by publishing a paper (workshop, conference) depending on the quality of the work to be carried.

Moyens / Méthodes / Logiciels

Abstract Interpretation / Ocaml / Python3

The candidate will work at the crossroads of formal methods and machine learning. As it is not realistic to be expert in both fields, we encourage candidates that do not meet the full qualification requirements to apply nonetheless.

Minimal Requirements

• Master 1 student or equivalent (2nd engineering school year) in computer science or applied mathematics
• knowledge of at least one programming language
• ability to work in a team
• fluent in French or English

Preferred Requirements

• knowledge of OCaml and Python
• notions of abstract interpretation and/or formal methods

We strive to provide an inclusive and enjoyable workplace. We are aware of discriminations based on gender (especially prevalent on our fields), race or disability, we are doing our best to fight them.

Entité de rattachement Le CEA est un acteur majeur de la recherche, au service des citoyens, de l'économie et de l'Etat. Il apporte des solutions concrètes à leurs besoins dans quatre domaines principaux : transition énergétique, transition numérique, technologies pour la médecine du futur, défense et sécurité sur un socle de recherche fondamentale. Le CEA s'engage depuis plus de 75 ans au service de la souveraineté scientifique, technologique et industrielle de la France et de l'Europe pour un présent et un avenir mieux maîtrisés et plus sûrs. Implanté au cœur des territoires équipés de très grandes infrastructures de recherche, le CEA dispose d'un large éventail de partenaires académiques et industriels en France, en Europe et à l'international. Les 20 000 collaboratrices et collaborateurs du CEA partagent trois valeurs fondamentales, Through the recent developments of AI, their use has become even more widespread, both in domestic and industrial settings. Nevertheless, studies are flourishing showing the dangers that such AI can bring, whether in terms of safety, privacy or fairness.