Diyora Salimova (ETH Zürich)
Deep neural networks in the numerical approximation of PDEs abstract
Abstract:
In recent years deep artificial neural networks (DNNs) have very successfully been used in numerical simulations for a numerous of computational problems including, object and face recognition, natural language processing, fraud detection, computational advertisement, and numerical approximations of partial differential equations (PDEs). Such numerical simulations indicate that DNNs seem to admit the fundamental flexibility to overcome the curse of dimensionality in the sense that the number of real parameters used to describe the DNN grows at most polynomially in both the reciprocal of the prescribed approximation accuracy and the dimension of the function which the DNN aims to approximate in such computational problems. In this talk I present our recent result which rigorously proves that DNNs do overcome the curse of dimensionality in the numerical approximation of Kolmogorov PDEs with constant diffusion and nonlinear drift coefficients.
10:30 • Université de Genève, Section Mathématiques, Centre Acacias II, Salle 623
Richard Hain (Duke University)
The Johnson homomorphism and its generalizations abstract
Abstract:
The Torelli group of a compact orientable surface is the subgroup of its mapping class group that consist of those isotopy classes of diffeomorphisms that act trivially on the homology of the associated closed surface. Torelli groups are important in topology (the study of homology 3-spheres) and algebraic geometry (their connections to the Ceresa cycle and tautological rings of moduli spaces of curves).
The Johnson homomorphism and its generalizations are tools for understanding Torelli groups. In this talk I will survey the work of many mathematicians (Morita, Enomoto, Sato, ...) on generalizations of Johnson\'s original homomorphism, including recent work of Kawazumi and Kuno on the completed Goldman--Turaev Lie bialgebra and its relation to computing the images of generalized Johnson homomorphisms.
Hodge theory (equivalently, Galois actions) provides additional structure that can be used to understand the Johnson homomorphism and its image. This leads to the definition of "arithmetic Johnson homomorphisms". I will explain how Oda\'s conjecture (proved by Takao), the work of Francis Brown, and the Turaev cobracket combine to constrain the image of the arithmetic Johnson homomorphism.
10:30 • Université de Genève, Villa Battelle, 7 route de Drize, 1227 Carouge
Bart De Moor (KU Leuven)
More Mathematics Mandatory in Data Science abstract
Abstract:
The tsunami of data and the increasing availability of massive computing power create many opportunities for AI in general, and for data science in particular. However, in many instances, the results delivered by machine learning algorithms are difficult to ‘understand’ or ‘explain’. Black boxes as they are, we basically do not know when and how they are successful, and when and why not. The whole field is pervaded, if not swamped, by hundreds of ‘heuristic tricks’, such as tampering with the neuronal activation function (sigmoid, RELU (rectified linear unit),…), experimenting with the number of neurons and hidden layers, selecting optimization methods (numerical ones, such as back-propagation or Gauss-Newton, or randomized ones, such as genetic algorithms), specifying user choices (initial guesses, step sizes, convergence tests, early stopping and/or regularization measures, seeds and cross-over/mutation rates, ...), etc.Therefore, machine learning is still more of an art than a science, and many results published in the literature do not even satisfy the minimal scientific requirement of ‘reproducibility’. Admittedly, this situation is not new. Even in the (‘classical’) field of system identification or statistical modeling of linear time-invariant finite dimensional dynamical linear systems, the origins of which predate those of deep learning by many years, heuristics prevail. Of course, there is a certain systematic methodology to tackle data-based identification problems:
1. Collect data;
2. Select a pre-specified model class that is parametrized by unknown parameters;
3. Choose an appropriate approximation criterion;
4. ‘Solve’ the resulting nonlinear optimization problem by numerical algorithms that output `optimal\' parameters;
5. Validate the resulting model on validation/test sets or by assessing the quality of predictions.
6. Re-iterate the whole loop when necessary.
It is in Step 4 that still plenty of heuristics are required. Typically, the optimization problem is nonlinear and provably non-convex. The nonlinearities originate in the assumptions that are made on the data model (e.g. additive misfit on the observations, or unobserved additional inputs), and the fact that products between unknowns occur (e.g. between unknown model parameters and unobserved inputs as in ARMAX models, or between unknown model matrices and states in subspace identification methods). The nonlinearities can be dealt with depending on the identification framework one deploys: by invoking asymptotic orthogonality and numerical linear algebra tools in subspace identification, by using the machinery of instrumental variables in errors-in-variables approaches, or by implementing iterative nonlinear least squares algorithms like in Prediction-Error-Methods. Despite hundreds of person-years of experience, and thousands of published papers, books and software suites, all of these ‘solutions’ contain plenty of heuristics. The main message of this talk is that we have to be braver: There is a lot of old and new mathematics out there that we collectively ignore, but that could provide us with a deeper understanding of our identification approaches. Specifically, for the problem of identifying LTI models from observed data, the central observation is that all nonlinearities involved are multivariate polynomial. With least squares as an approximation criterion, we end up with a multivariate polynomial optimization problem. From algebraic geometry, we know that such problem is fundamentally equivalent to a large-scale eigenvalue problem. We will show how the set of first order optimality conditions comprises a multi-parameter eigenvalue problem (MEVP), the solution of which is surprisingly little studied in the literature, requiring ‘new’ mathematics: Such a MEVP can be solved by recursively building up a quasi-block-Toeplitz block Macaulay matrix, the null space of which is block multi-shift invariant (a property studied in operator theory). By then applying multi-dimensional (exact) realization theory in that null space (new multi-dimensional system theory), we can find the optimal parameters from the eigenvectors and -values of a specific large-scale matrix. This ‘solution’ as described, satisfies the very scientific definition of the word, because a set of a priori minimal assumptions on data and model leads to a sequence of mathematically verifiable and reconstructable steps that uniquely characterize the optimal solution to be an eigenvalue problem. Even if heuristics would still be needed to compute the optimal solution because of the mere size of the problem, we now know that we are looking for a minimizing eigenvalue-eigenvector pair of a matrix constructed from the data and chosen model class. Can we learn something about the challenges posed by deep learning from this very special case of linear system identification? The answer is a resounding ‘yes’, and we will provide some first indications why our mathematical framework as outlined above might also be applicable to neural networks. The take home message is that more - old and new - mathematics is mandatory to understand and explain deep learning. The endeavor is difficult, challenging, time consuming, requires patience and endurance, and the research involved is definitely high-risk high-gain. The journey will take us into the realms of mathematical disciplines such as one- and multidimensional system theory, algebraic geometry, operator theory and numerical linear algebra and optimization. For sure, diving deeper in the mathematical depts of neural networks is largely a-typical in the current massive wave of heuristic deep learning activities. It will require a creative understanding of mathematics, that is more profound than the acquired know-how that most practitioners of AI and deep learning currently possess. Yet, there is no alternative if we want to turn the toolbox of deep learning heuristics into real science.
12:30 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 19.1
Pierre-Olivier Goffard (ISFA, Université de Lyon 1)
For a few bitcoins more: Fraud risk assessment within blockchain transactions abstract
Abstract:
The probability of successfully spending twice the same bitcoins is considered. A double-spending attack consists in issuing two transactions transferring the same bitcoins. The first transaction, from the fraudster to a merchant, is included in a block of the public chain. The second transaction, from the fraudster to himself, is recorded in a block that integrates a private chain, exact copy of the public chain up to substituting the fraudster-to-merchant transaction by the fraudster-to-fraudster transaction. The double-spending hack is completed once the private chain reaches the length of the public chain, in which case it replaces it. The growth of both chains are modeled by two independent counting processes. The probability distribution of the time at which the malicious chain catches up with the honest chain, or equivalently the time at which the two counting processes meet each other, is studied. The merchant is supposed to await the discovery of a given number of blocks after the one containing the transaction before delivering the goods. This grants a head start to the honest chain in the race against the dishonest chain.
14:00 • Université de Lausanne, Quartier UNIL - Dorigny, Salle Extranef 125
Dr. Pankaj Vishe (Durham University)
Rational points on global fields and applications abstract
Abstract:
We present analytic methods for counting rational points on complete intersections defined over function field F_q(t). The main ingredient is obtaining a version of Kloosterman refinement over function field for complete intersections.
14:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 43
Prof. Dr. Gérard Iooss (LJAD, Nice)
Existence of quasipatterns in the superposition of two hexagonal periodic patterns abstract
Abstract:
Let us consider a quasilattice, spanned by the superposition of two hexagonal lattices in theplane, differing by a rotation of angle ß. We study bifurcating quasipattern solutions of theSwift-Hohenberg PDE, built on such a quasilattice, invariant under rotations of angle π /3. /3. Fornearly all ß, this is a small divisor problem. We prove that in addition to the classicalhexagonal patterns, there exist two branches of bifurcating quasipatterns, with equalamplitudes on each basic lattice.
14:15 • EPF Lausanne, Salle MA B1 11
Ulrik Brandes (ETH, Departement Geistes-, Sozial- und Staatswissenschaften )
Network science as the dual of statistics abstract
Abstract:
If statistics is concerned with the collection, management, analysis,presentation, and interpretation of data, then so is network science.The distinction lies neither in a universal network theory nor thecomplex systems of certain application domains but (i) an incidence structure relating the units of observation and (ii) a corresponding focus on dependencies among observations. While not mainstream, this view clears away some of the smoke and mirrors obscuring the close relationship between statistics and network science. Following this rather principled line of thought, I will introduce the positional approach to network analysis as a means to capitalize on it.
15:15 • ETH Zentrum, Rämistrasse 101, Zürich, Building HG, Room G 19.1
Hélène Ruffieux (EPF Lausanne)
Large-scale variational inference for Bayesian joint regression modelling of high-dimensional genetic data
17:30 • EPF Lausanne, Salle ELA 2