Creativity in Generative AI
About this Event
29 Oxford Street, Cambridge, MA 02138
Widely Applied Mathematics Seminar
November 6 at 3pm
Pierce Hall 209
Creativity in Generative AI
Speaker: Mattieu Wyart, W. H. Miller Professor, Johns Hopkins University
Is AI creative? Generative AI such as chatGPT or diffusion models can create new texts or images from a finite training set of examples. I will argue that AI can achieve this magical by learning how compose observed low-level elements into a new whole. I will discuss the type of correlations the model can exploit to do so, how many data are needed for that, and how it relates to a hierarchical construction of latent variables. The analysis is based on the introduction of synthetic languages, and comparison with experiments performed on modern AI architectures trained on real text and images.
Bio: Matthieu Wyart studied physics, mathematics and economics at the Ecole Polytechnique in Paris where he obtained in 2001 his degree in physics and, the following year, the Diploma of Advanced Studies in Theoretical Physics, with highest Honors at the Ecole Normale Supérieure, Paris. In 2006 he obtained a doctoral degree in Theoretical Physics and Finance at the SPEC, CEA Saclay, Paris with a thesis on electronic markets. He then moved to the United States, to Harvard, Janelia Farm, and Princeton before joining in 2010 New York University as Assistant Professor, where he was promoted Associate professor in 2014. In July 2015, he was appointed Associate Professor of Theoretical Physics in the School of Basic Sciences at EPFL, and promoted to full professor in 2023. He joined JHU in November 2024.
