Woonam Lim (ETH Zürich)
Virasoro constraints in sheaf theory and vertex algebras abstract
Abstract:
In enumerative geometry, Virasoro constraints were first conjectured in Gromov-Witten theory with many new recent developments in the sheaf theoretic context. In joint work with A. Bojko and M. Moreira, we rephrase the sheaf-theoretic Virasoro constraints in terms of primary states coming from a natural conformal vector in Joyce’s vertex algebra. This shows that Virasoro constraints are preserved under wall-crossing. As an application, we prove the conjectural Virasoro constraints for moduli spaces of torsion-free sheaves on any curve and on surfaces with only (p,p) cohomology classes.
10:15 • EPF Lausanne, Salle GR A3 32
Julian Demelo (Universität Basel)
Title T.B.A.
14:15 • Universität Basel, Spiegelgasse 5, SR 05.002
Gabriele Rembado (Universität Bonn)
Wild mapping class groups abstract
Abstract:
The standard mapping class groups are fundamental groups of moduli spaces/stacks of pointed Riemann surfaces: they thus encode much information about the topology of the deformations of such surfaces.Recently this story has been extended to wild Riemann surfaces, which generalise pointed Riemann surface by adding local moduli at each marked point -- the irregular classes. The new parameters control the polar parts of meromorphic connections with wild/irregular singularities, defined on principal bundles, and importantly provide an intrinsic viewpoint on the `times\' of isomonodromic deformations.
In this talk we will explain how to compute the fundamental groups of (universal) spaces of deformations of irregular classes, related to cabled versions of braid groups, which thus play the role of `wild\' mapping class groups. This is joint work with P. Boalch, J. Douçot and M. Tamiozzo. If time allows we will sketch a relation with bundles of irregular conformal blocks in the Wess-Zumino-Witten model, in joint work with G. Felder (past) and G. Baverez (in progress).
15:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 43
Dr. Léo Bigorgne (Université de Rennes 1)
Modified scattering for small data solutions to the Vlasov-Maxwell system abstract
Abstract:
Abstract: We will capture the asymptotic behavior of the solutions to the Vlasov-Maxwell system arising from sufficiently small and regular data. Our analysis is based on vector fields methods, allowing us to exploit the null structure of the equations, and the Glassey-Strauss decomposition of the electromagnetic field. In particular, we will see that the electromagnetic field approaches, for large time, a solution to the vacuum Maxwell equations. Due to the long-range effects of the Lorentz force, the Vlasov field converges along logarithmic corrections of the linear characteristics.
16:00 • UZH Zentrum, Building KOL, Room G209
Holger Rauhut (RWTH Aachen)
The implicit bias of gradient descent for learning linear neural networks abstract
Abstract:
Deep neural networks are usually trained by minimizing a non-convex loss functional via (stochastic) gradient descent methods. Unfortunately, the convergence properties are not very well-understood. Moreover, a puzzling empirical observation is that learning neural networks with a number of parameters exceeding the number of training examples often leads to zero loss, i.e., the network exactly interpolates the data. Nevertheless, it generalizes very well to unseen data, which is in stark contrast to intuition from classical statistics which would predict a scenario of overfitting. A current working hypothesis is that the chosen optimization algorithm has a significant influence on the selection of the learned network. In fact, in this overparameterized context there are many global minimizers so that the optimization method induces an implicit bias on the computed solution. It seems that gradient descent methods and their stochastic variants favor networks of low complexity (in a suitable sense to be understood), and, hence, appear to be very well suited for large classes of real data.Initial attempts in understanding the implicit bias phenomen considers the simplified setting of linear networks, i.e., (deep) factorizations of matrices. This has revealed a surprising relation to the field of low rank matrix recovery (a variant of compressive sensing) in the sense that gradient descent favors low rank matrices in certain situations. Moreover, restricting further to diagonal matrices, or equivalently factorizing the entries of a vector to be recovered, leads to connections to compressive sensing and l1-minimization.After giving a general introduction to these topics, the talk will concentrate on results by the speaker on the convergence of gradient flows and gradient descent for learning linear neural networks and on the implicit bias towards low rank and sparse solutions.
16:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room F 3
Corinna Ulcigrai (Universität Zürich and ETH-ITS)
Abstract:
Renormalization is a key technique in dynamics, which provides the mathematical tools to explain many (approximate) self-similarity phenomena. In this colloquium we will discuss two problems: the study of diffusion phenomena in the Ehrenfest model in mathematical physics and the description of symbolic sequences of very low complexity, which often appear also in the theoretical computer science literature. In both, we will explain some results which were proved using renormalization on geometric surfaces and offer an insight on how they are obtained.
16:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room E 5
Shai Evra (Hebrew University of Jerusalem)
Abstract:
It is a classical result that the modulo map from SL_2(Z) to SL_2(Z/qZ), is surjective for any integer q. The generalization of this phenomenon to other arithmetic groups goes under the name of strong approximation, and it is well understood. The following natural question was recently raised in a letter of Sarnak: What is the minimal exponent e, such that for any large q, almost any element of SL_2(Z/qZ) has a lift in SL_2(Z) with coefficients of size at most q^e? A simple pigeonhole principle shows that e > 3/2. In his letter Sarnak proved that this is in fact tight, namely e = 3/2, and call this optimal strong approximation for SL_2(Z). The proof relies on a density theorem of the Ramanujan conjecture for SL_2(Z).
In this talk we will give a brief overview of the strong approximation, a quantitative strengthening of it called super strong approximation, and the above mentioned optimal strong approximation phenomena, for arithmetic groups. We highlight the special case of p-arithmetic subgroups of classical definite matrix groups and the connection between the optimal strong approximation and optimal almost diameter for Ramanujan complexes. Finally, we will present the Sarnak-Xue density hypothesis and describe recent ongoing works on it relying on deep results coming from the Langlands program.
This talk is based on ongoing joint works with B. Feigon, M. Gerbelli-Gauthier, H. Gustafssun, K. Maurischat and O. Parzanchevski.
17:00 • Universität Basel, online Seminar
Prof. Dr. Łukasz Delong (SGH Warsaw School of Economics)
The use of autoencoders for training neural networks with mixed categorical and numerical features abstract
Abstract:
We focus on modelling categorical features and improving predictive power of neural networks with mixed categorical and numerical features in supervised learning problems. The goal is to challange the current dominant approach in actuarial data science with a new architecture of a neural network and a new training algorithm. The proposal is to use a joint embedding for all categorical features, instead of separate embeddings for each categorical features, and initalize the parameters of the neural network with parameters trained with a denoising autoencoder in a unsupervised learning problem, instead of random initialization. In other words, we propose a special initialization strategy for taining neural networks. We illustrate our ideas with experiments on a data set with insurance claim numbers. We demonstrate that we can achieve a predictive power on an independent test set higher than in the current approach. Moreover, we also show with experiments that our special initialization strategy performs better than other initialization strategies (with PCA, MCA, GLM) which can be alternatively proposed in our setting. Finally, we investigate the predictive power vs the bias of the predictions.
17:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 43