Florentin Coeurdoux

A transient spectral effect driven by anisotropy and noise reveals the hidden dynamics of early stopping in gradient descent training.

Uncover hidden relationships in your diffusion model's learned distribution by tracing paths between samples with a new string method framework, revealing realistic morphing sequences and protein transition pathways.

Skip the GANs: this kernel method generates complex data like financial time series and images without any training, just by solving linear systems.

Escape the MCMC slowdown: this diffusion-based approach efficiently samples maximum entropy distributions by directly guiding moments, offering a practical alternative for high-dimensional uncertainty quantification.