PDE and applications seminar | Metz

Résumé à venir

Résumé à venir

Résumé à venir

Résumé à venir

Résumé à venir

Kernel-based methods are central to modern data analysis, supporting nonlinear regression, forecasting, and advanced function approximations and extensions. Here, we introduce two multiscale kernel constructions, adapted for three distinct application settings, each centered
around function approximations and extension schemes.

First, a multiscale iterative scheme is used to enhance coarse-grid finite-difference computations, enabling accurate fine-grid solutions. This same iterative scheme is also employed in a spatiotemporal kernel regression model, which fuses information from multiple spatial locations. This approach yields accurate forecasts across diverse real-world datasets, including solar energy productions, epidemiological trends, and fire-events dynamics.
Finally, a second construction uses high-order multiscale kernels, which are constructed via combinations of scaled Gaussians. This approach preserves more spectral energy in the leading modes and enhances Nyström-type extensions and out-of-sample embeddings in manifoldlearning settings.
Altogether, these contributions offer a unified multiscale perspective. By capturing data geometry with kernels and transferring information across scales and modalities, the framework enables a robust yet simple scheme for approximations and extensions. This approach is readily applicable to a wide range of scientific problems.

Résumé à venir