

High-dimentional classification with deep neural networks: decision boundaries, noise, and margin
By Philipp Petersen


New dynamical low-complexity approximations for the Schrödinger equation
By Virginie Ehrlacher
By Martin Frank
Appears in collection : Jean Morlet Chair 2021- Conference: Kinetic Equations: From Modeling Computation to Analysis / Chaire Jean-Morlet 2021 - Conférence : Equations cinétiques : Modélisation, Simulation et Analyse
The dynamical low-rank approximation is a low-rank factorization updating technique. It leads to differential equations for factors in a decomposition of the solution, which need to be solved numerically. The dynamical low-rank method seems particularly suitable for solving kinetic equations, because in many relevant cases the effective dynamics takes place on a lower-dimensional manifold and thus the solution has low rank. In this way, the 5-dimensional (3 space, 2 angle) radiation transport problem is reduced, both in computational cost as well as in memory footprint. We show several numerical examples.