Alessandro Leite (Inria Saclay – LISN, Paris-Saclay University) animera un Seminar@SystemX sur le thème « Causal Knowledge Discovery through Large Language Models: Challenges and Opportunities », le 10 octobre 2024 de 14h à 15h.
>> Lien de connexion au webinaire <<
Résumé (en anglais)
Understanding the data-generating process is critical in many disciplines, including physics, econometrics, epidemiology, and data analytics. In recent years, various methods have been developed to approximate complex data distributions by transforming latent variables, typically sampled from simple distributions (e.g., Gaussian), into more complex ones resembling the distribution of observational data. However, these methods often lack interpretability, limiting their use in out-of-sample prediction and a deeper understanding of the underlying processes. Causal inference, on the other hand, provides a robust framework for understanding how changes in one variable affect another, allowing one to answer what-if questions. Despite its promise, inferring causal relationships from observational data remains challenging, often relying on assumptions rarely met in real-world scenarios and, thus, requiring validation from human experts. In this talk, we will explore the critical intersection of causality and deep generative models, particularly the potential of large language models (LLMs) to advance causal discovery. We may bridge the gap between statistical association and causal understanding by integrating LLMs into the causal inference pipeline. Moreover, we will discuss the strengths and limitations of such an approach in this context, present empirical results, and highlight their potential to enhance research and application across data-driven fields.
Biographie (en anglais)
Alessandro Leite is an associate researcher with the TAU team at Inria Saclay and the LISN (Interdisciplinary Laboratory on Numerical Sciences) at Paris-Saclay University. His recent research focuses on causal inference, explainable AI, and uncertainty quantification, contributing to understanding complex data-driven systems.
