From Probabilistic Testing to Certifiable AI: Large Language Models and Neuro-Symbolic Reasoning for Verifiable Autonomous Systems
- James Xi Zheng
16 juin 2025
Digitéo Moulon - Bât. 660, Amphi Shanon

Learning-enabled Cyber-Physical Systems (LE-CPS), such as autonomous vehicles anddrones, face critical safety and reliability challenges due to the stochastic nature of deep neural networks. In our FSE’22 and TSE’23 studies, we conducted an in-depth investigation into industry testing practices, uncovering significant gaps between current testing techniques and the needs for regulatory assurance. To address these challenges, we introduced two pioneering approaches: test reduction for ROS-based multi-module autonomous driving systems, and scenario-based construction for checking traffic rule compliance in autonomous driving systems.

Growing Tiny Networks: Spotting Expressivity Bottlenecks and Fixing Them Optimally - Manon Verbockhaven
13 mai 2025
Nano-Innov, bâtiment 862, salle 1073

Machine learning tasks are generally formulated as optimization problems, where one searches for an optimal function within a certain functional space. In practice, parameterized functional spaces are considered, in order to be able to perform gradient descent. Typically, a neural network architecture is chosen and fixed, and its parameters (connection weights) are optimized, yielding an architecture-dependent result. This way of proceeding however forces the evolution of the function during training to lie within the realm of what is expressible with the chosen architecture, and prevents any optimization across architectures. Costly architectural hyper-parameter optimization is often performed to compensate for this. Instead, we propose to adapt the architecture on the fly during training. We show that the information about desirable architectural changes, due to expressivity bottlenecks when attempting to follow the functional gradient, can be extracted from backpropagation. To do this, we propose a mathematical definition of expressivity bottlenecks, which enables us to detect, quantify and solve them while training, by adding suitable neurons. Thus, while the standard approach requires large networks, in terms of number of neurons per layer, for expressivity and optimization reasons, we provide tools and properties to develop an architecture starting with a very small number of neurons. As a proof of concept, we show results~on the CIFAR dataset, matching large neural network accuracy, with competitive training time, while removing the need for standard architectural hyper-parameter search.

Automatically Assessing Gender Biases in Autoregressive Language Models - Fanny Ducel
6 mars 2025
Nano-Innov, Bâtiment 862, salle 1073

Les modèles de langue génératifs (aussi dits auto-régressifs), sont de plus en plus utilisés aussi bien dans la communauté scientifique, dans les entreprises que par le grand public. Toutefois, ces systèmes contiennent et amplifient de nombreux biais stéréotypés, qui peuvent contribuer à la stigmatisation et la discrimination de certaines catégories de population. Cette présentation commencera avec quelques définitions et enjeux liés aux biais stéréotypés des modèles de langue, puis un passage en revue de l’état de l’art sur les corpus, méthodes et métriques utilisés pour identifier, atténuer et évaluer ces biais. Ensuite, je présenterai une expérience menée sur la génération de lettres de motivation par des modèles de langue en français et en italien, qui met en évidence de forts biais genrés, invisibilisant le féminin et renforçant des associations stéréotypées entre certains domaines professionnels et un genre. Finalement, je partagerai les résultats d’une expérience connexe menée sur les biais de genre dans des cas cliniques générés en français.