Stéphane Rivaud
Associate Professor (Maître de conférences) · Université Paris-Saclay · LISN & Inria Saclay · A&O Team
Associate Professor (Maître de conférences) · Université Paris-Saclay · LISN & Inria Saclay · A&O Team
I am an Associate Professor (Maître de conférences) at Université Paris-Saclay, in the A&O Team (LISN & Inria Saclay). My research focuses on functional optimization for neural network training, with a strong emphasis on efficiency. I am also deeply interested in agentic tools for scientific research and education.
Lectures, practical sessions, and supervision of applied data science student projects
Course design and delivery for 50+ students (lectures & practical sessions) · Course materials
Foundations of statistical learning and data analysis for undergraduate students
Practical sessions · Course materials
Practical sessions
Lectures
Lectures
Oral Examiner
Distillation-based neural network growth. Studied how architecture growth can be leveraged to recover internal representations using feature distillation methods coupled with model growth.
Certified Functional Gradient Descent. Formulated function-space criteria for neural network training and model growth. Co-authored CAGE-NAS (NeurIPS 2026 AXIOM).
Transformer Architecture Growth
NAS Unseen Challenge 2025. Co-supervised with Stella Douka. Applied neural network growth methods to the NAS Unseen Data Challenge benchmark during a 3-month research internship.
Parallelizable training in deep learning through local and distributed approaches. Published 2 papers at ICLR and ICML. [Thesis]
Bird Vocalization Classification with Few-Shot Learning
Integration of Expert Knowledge in Generative Modeling: Application to Music Production
Comprehensive web-based dashboard for exploring bird population monitoring data from Martinique's 12-year monitoring program. Features interactive visualizations with statistical analysis including Mann-Kendall trend tests, linear regression for population trajectories, and k-means clustering for observer network analysis. The dashboard processes 114,495 observations across 47 species, providing real-time filtering, habitat-specific biodiversity trends, and species population trajectories.