|
Neuromathématiques (séminaire du Centre d’analyse et de mathématique sociales (CAMS), UMR 8557 CNRS, EHESS Paris)| title | Neuromathématiques (séminaire du Centre d’analyse et de mathématique sociales (CAMS), UMR 8557 CNRS, EHESS Paris) |
|---|
| start_date | 2026/03/10 |
|---|
| schedule | 14h30-16h30 |
|---|
| online | no |
|---|
| location_info | salle D2.2 & en ligne |
|---|
| summary | We present a functional framework for shallow neural networks based on reproducing kernel Banach spaces. This approach enables a nonparametric treatment of neural networks, in direct analogy with kernel methods. A representer theorem shows that finite networks are optimal for empirical risk minimization. Estimation and approximation error bounds can then be derived in linear function spaces. As a byproduct, universality results and approximation bounds can be proved, showing that neural networks can adapt to latent structure in the problem. Further, we derive complexity estimates based on the Rademacher complexities of RKBS balls, independent of network size. Time permitting we will discuss extension to deep networks. |
|---|
| responsibles | Sarti, Citti, Petitot, Nadal, Ribot |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2026/03/04 08:40 UTC |
| |
|