Events

我将讨论与弗朗西斯科共同工作Camilli and Jean Barbier concerning an information-theoretical analysis of a two-layer neural network trained from input-output pairs generated by a teacher network with matching architecture, in overparametrized regimes. Our results come in the form of bounds relating i) the mutual information between training data and network weights, or ii) the Bayes-optimal generalization error, to the same quantities but for a simpler (generalized) linear model for which explicit expressions are rigorously known. Our bounds, which are expressed in terms of the number of training samples, input dimension and number of hidden units, thus yield fundamental performance limits for any neural network (and actually any learning procedure) trained from limited data generated according to our two-layer teacher neural network model. The proof relies on rigorous tools from spin glasses and is guided by ``Gaussian equivalence principles'' lying at the core of numerous recent analyses of neural networks. With respect to the existing literature, which is either non-rigorous or restricted to the case of the learning of the readout weights only, our results are information-theoretic (i.e. are not specific to any learning algorithm) and, importantly, cover a setting where all the network parameters are trained.

We present a new and relatively simple method for proving large data local well-posedness in low regularity Sobolev spaces for general quasilinear Schrödinger equations of the form \begin{equation*} \begin{cases} &i\partial_tu+g^{jk}(u,\overline{u},\nabla u,\nabla\overline{u})\partial_j\partial_k u=F(u,\overline{u},\nabla u,\nabla\overline{u}),\hspace{5mm} u:\mathbb{R}\times\mathbb{R}^d\to\mathbb{C}^m, \\ &u(0,x)=u_0(x), \end{cases} \end{equation*} assuming only non-degeneracy of the metric, nontrapping and mild regularity/decay of the initial data. As a consequence, we remove the uniform ellipticity assumption from the main result of Marzuola, Metcalfe and Tataru (Arch. Ration. Mech. Anal. 2021) and substantially weaken the regularity/decay assumptions from the pioneering works of Kenig, Ponce, Rolvung and Vega. This is based on joint work with Ben Pineau (UC Berkeley).

Following a brief introduction to general relativity and the Einstein field equations, I will discuss two interrelated recent developments in the mathematical study of black holes and some of the mathematical techniques involved. In the first part, I describe the stability properties of various exact black hole solutions under regular small perturbations, namely the relaxation of an out-of-equilibrium black hole to a stationary state; this has been the subject of intense recent activity by a number of research groups, and is now understood in a fair amount of detail. In the second part, I will discuss work in progress towards the construction of singular perturbations of spacetimes via the insertion of small black holes: this aims to provide the first rigorous examples of spacetimes describing the merger of two black holes with extreme mass ratios.