Ecole Polytechnique Federale de Lausanne

The random initialisation of Artificial Neural Networks (ANN) allows one to describe, in the functional space, the limit of the evolution of ANN when their width tends towards infinity. Within this limit, an ANN is initially a Gaussian process and follows, during learning, a gradient descent convoluted by a kernel called the Neural Tangent Kernel.

Connecting neural networks to the well-established theory of kernel methods allows us to understand the dynamics of neural networks, their generalization capability. In practice, it helps to select appropriate architectural features of the network to be trained. In addition, it provides new tools to address the finite size setting.

{\bf This seminar is at ground floor!}

\\

An important question in geometry and analysis is to know when two $k-$forms

$f$ and $g$ are equivalent. The problem is therefore to find a map $\varphi$

such that%

\[

\varphi^{\ast}\left( g\right) =f.

\]

We will mostly discuss the symplectic case $k=2$ and the case of volume forms

$k=n.$ We will give some results when $3\leq k\leq n-2,$ the case $k=n-1$ will

also be considered.

\\

The results have been obtained in collaboration with S. Bandyopadhyay, G.

Csato and O. Kneuss and can be found, in part, in the book below.\bigskip

\\

\newline

Csato G., Dacorogna B. et Kneuss O., \emph{The pullback equation for

differential forms}, Birkha\"{u}ser, PNLDE Series, New York, \textbf{83} (2012).

Following the framework of Formaggia and Perotto (Numer.

Math. 2001 and 2003), anisotropic a posteriori error estimates have been

proposed for various elliptic and parabolic problems. The error in the

energy norm is bounded above by an error indicator involving the matrix

of the error gradient, the constant being independent of the mesh aspect

ratio. The matrix of the error gradient is approached using

Zienkiewicz-Zhu error estimator. Numerical experiments show that the

error indicator is sharp. An adaptive finite element algorithm which

aims at producing successive triangulations with high aspect ratio is

proposed. Numerical results will be presented on various problems such

as diffusion-convection, Stokes problem, dendritic growth.

Following the framework of the heterogeneous multiscale method, we present a numerical method for nonlinear elliptic homogenization problems. We briefly review the numerical, relying on an efficient coupling of macro and micro solvers, for linear problems. A fully discrete analysis is then given for nonlinear (nonmonotone) problems, optimal convergence rates in the H¹ and L² norms are derived and the uniqueness of the method is shown on sufficiently fine macro and micro meshes.

Numerical examples confirm the theoretical convergence rates and illustrate the performance and versatility of our approach.

We consider monochromatic planar electro-magnetic waves propagating through a nonlinear dielectric medium in the optical regime.

Travelling waves are particularly simple solutions of this kind. Results on the existence of guided travelling waves will be reviewed. In the case of TE-modes, their stability will be discussed within the context of the paraxial approximation.