Charles Bricout

Deep generative models are becoming more and more expressive, and at the same time computation-hungry, which can limit their applicability in several scenarios where speed and energy-consumption are fundamental. One key objective of the PhD will be understanding how and when we can use tractable probabilistic models to support or replace intractable deep generative models (or parts of them) and their inference routines. Possible research directions will include designing novel hybrid diffusion and flow matching variants that can scale better while certifying their sustainable consumption. Furthermore, these modeling advancements will be applied in inference-heavy scenarios where data is limited, such as Bayesian optimization and simulation based inference in the context of computational sciences applications such as material design and climate modeling.