The Departement of Mathematical Engineering organizes a series of seminars. The seminars are held in the Euler lecture room, Building EULER, av. Georges Lemaître 4-6, Louvain-la-Neuve (Parking 13).

If you wish to receive the seminar announcements by email, please send an email to Etienne Huens.

Master students can take this seminar for credit in either of the two semesters; see LINMA2120 for more information.

Seminars to come

27/11/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Erik Bekkers (Eindhoven University of Technology)


12/12/2018 (?) [Lieu : Bât. Euler (room A.002)]
Colin Jones (EPFL)

TBA Note the unusual day (Wednesday).

Previous seminars

13/11/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Thanh Son Nguyen (UCLouvain)
Parametric model order reduction and interpolation on matrix manifolds as a tool

Large-scale control systems appear frequently as mathematical models for many practical fields such as heat conduction, electrical circuits. Simulation of such systems requires solving equations whose order can reach dozen of thousands. Besides, in many cases, these systems depend on parameter which makes the simulation more challenging. In this talk, first we will give an introduction to model order reduction of parameter-dependent linear time-invariant systems as well as a review of interpolation-based methods. Then we will explain why in some cases, we have to do it in the framework of interpolation on matrix manifolds. Finally, we will present some numerical examples for illustrative purpose.

06/11/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
New PhD students in INMA
Welcome Seminar #2

In this seminar, new PhD students in INMA will introduce themselves and their research topic: "Compressive Learning : learning from 'too much' data" (Vincent Schellekens); "Signal processing for direct imaging of stellar systems" (Benoît Pairet); "Optimal interference nulling for large arrays of coupled antennas" (Valentin Hamaide); "Algorithms for smart content marketing" (Mridul Seth).

23/10/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Laurent Demanet (Departments of Mathematics and EAPS, MIT)
Interferometry and convex programming

A number of hard estimation questions in science and engineering can be expressed as recovering a rank-1 matrix from linear measurements. One example that I will cover in this talk is "interferometry", a trick in signal processing and imaging consisting in taking cross-correlations to attenuate incoherent noise. I show that recovery for inverse problems involving interferometric combinations is sometimes possible with varying levels of convex relaxations, and depends on the underlying graph of measurements being an expander. I also show an example of "interferometric thinking" to blind deconvolution, where it enables the design of new sparse regularizers.

16/10/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
New PhD students in INMA
Welcome Seminar #1

In this seminar, new PhD students in INMA will introduce themselves and their research topic: "Nonnegative matrix factorization in infinite-dimensional feature spaces : a parameterized approach" (Cécile Hautecoeur); "Analysis and Control of Interconnected Systems" (Ayoub Ben Ayed); "Decentralized optimization in open multi-agent systems" (Charles Monnoyer); "MILP-based Algorithm for the Global Solution of Economic Dispatch Problems with Valve-Point Effects" (Loïc Van Hoorebeeck).

09/10/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Jean-François Cardoso and Pierre Ablin (CNRS - INRIA)
Fast and invariant learning for Independent Component Analysis

Independent Component Analysis (ICA) is a widely used unsupervised data exploration technique. It models a set of observed signals as a linear mixtures of statistically independent sources and aims at recovering blindly those underlying sources. Here `blind' means that sources are recovered from the signals by applying a separating matrix which is totally inconstrained, making ICA applicable in a large variety of tasks. After a brief introduction, we stress the multiplicative structure of the ICA problem: the parameter space is the multiplicative group of GL(n) of square invertible matrices. Learning algorithms which exploit this structure are `equivariant' and are `rewarded' for doing so in terms implementation, optimization and performance, as we shall see. In a second part, we introduce a fast quasi-Newton equivariant algorithm for ICA, the Preconditioned ICA for Real Data (Picard) algorithm. It exploits the specific structure of the problem to compute cheap Hessian approximations, and then refines them using L-BFGS, a classical optimization algorithm. It shows state of the art convergence speed when applied to real datasets. Interestingly, it can be straightforwardly constrained to work on the rotation manifold, a constraint often imposed in ICA.

02/10/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Adrien Scheuer (UCLouvain)
Multi-scale modelling of fibre suspensions: Particle inertia, confined flows and data-driven approach

Suspensions of fibres and non-spherical particles are encountered in many fields ranging from engineering to biology, e.g. papermaking, composite manufacturing, pharmaceutical applications, red blood cells, food-processing and cosmetics industries, etc. Predicting the evolution of the orientation state of the particles is crucial to estimate the rheology of the suspension, that is its flow behaviour, as well as the final properties of the material. Jeffery’s theory, describing the kinematics of a single particle immersed in an homogeneous flow of Newtonian fluid, lays the foundation for almost every models used today. Coarser representations, built upon this work, have been introduced later to describe statistically the orientation state of the particles, either using a probability density function, or even moments of this function (Advani-Tucker orientation tensors). The assumptions underlying Jeffery’s model are however quite restrictive to predict reliably what happens in fibre suspensions flows encountered in industrial processes. In this thesis, we first revisit this model, studying the impact of particle inertia and of confinement (wall effects) on the particle kinematics. In each case, we propose a multi-scale approach, but given the challenges to upscale the microscopic description to the macroscopic scale, we then came up with an innovative approach based on data-driven simulations to circumvent upscaling issues and inaccuracies introduced by macroscopic closure approximations. Finally, we developed efficient numerical methods to simulate fluid flows in thin geometries, considering, within the Proper Generalized Decomposition (PGD) framework, an in-plane/out-of-plane separated representation of the solutions of the incompressible Navier-Stokes equations.

25/09/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Elarbi Achhab (Université Chouaïb Doukkali)
Compensator Design for a class of Infinite-Dimensional Semilinear Systems

In this talk, we consider a class of partially observed infinite-dimensional semi-linear systems. First, we address the problem of the design of an exponential Luenberger-like observer for this class of systems. Then, using the state estimation result stated, a stabilizing compensator for this class of systems will be designed. A compensator is an auxiliary system which has as its input the output of the initial system and as its output the input of the original one. In our setting, the obtained control law is showed to exponentially stabilize around a desired equilibrium profile the system under consideration. Finally, the main result is applied to a non isothermal chemical plug flow reactor model. The approach is illustrated by numerical simulations.

18/09/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Antoine Godin (Agence Française de Développement)
The Stock-Flow Consistent approach or the importance of disequilibrium

Stock-Flow Consistent (SFC) models are sectoral macroeconomic models that combine two fundamental insights: first, the economy is ruled by imbalances and its reactions to these imbalances, and second, the economy is a dynamical multi-layered network of financial relationships. By combining these two aspects, SFC models are a powerful tool to understand modern financialised economies in modelling feedback loops between exchanges of good and services, financial flows, and wealth. The GEMMES research project at the French Agency for Development aims at contributing to the international debate related to climate change (both for the adaptation and mitigation aspects). Given the importance and the pervasive aspect of the transition to a low carbon economy, it is fundamental to be able to grasp how it will re-shape, create or steer imbalances and hence might lead to unsustainable dynamics such as financial crisis. This seminar will go through three different modelling approaches, all using the insight of Stock-Flow Consistency, to show the importance of disequilibrium and imbalances in economics

11/09/2018 (16:30) [Lieu : Core(room b.135)]
Ion Necoara (Univ.Pol.Bucharest)
Stochastic algorithms for convex feasibility and convex minimization

In this talk we present stochastic first-order methods for solving convex feasibility problems or convex minimization problems with many constraints. First, for the convex feasibility problem (that is finding a point in the (in)finite intersection of convex sets) we propose several equivalent stochastic reformulations, such as stochastic (non)smooth optimization problem, stochastic fixed point problem, or stochastic intersection problem. Based on these reformulations and on new characterization of the conditioning parameters we introduce a general random projection algorithmic framework with an over-relaxed stepsize, which generates either new algorithms or extends to random settings many existing alternating projection schemes. We also derive (sub)linear convergence rates for this general algorithm that depend explicitly on the conditioning parameters and on the number of projections computed at each iteration. Then, we extend this stochastic algorithmic framework to convex minimization problems subject to (in)fi nite intersection of constraints. For this algorithm we also derive convergence rates in terms of the expected quadratic distance from the iterates to the optimal solution for smooth strongly convex objective functions, which in the best case is of order O(1/k). We also provide necessary and sufficient conditions for linear convergence of this stochastic method. Finally, we give examples of several functional classes satisfying our new conditions and discuss several applications of these results.

4/09/2018 (14:00) [Lieu : Bât. Euler (room A.207)]
Sebastian Stich (EPFL)
Communication Efficient Variants of SGD for Distributed Computing

Nowadays machine learning applications require stochastic optimization algorithms that can be implemented on distributed systems. The communication overhead of the algorithms is a key bottleneck that hinders perfect scalability. In this talk we will discuss two techniques that aim reducing the communication costs. First, we discuss quantization and sparsification techniques that reduce the amount of data that needs to be communicated. We present a variant of SGD with k-sparsification (for instance top-k or random-k) and show that this scheme converges at the same rate as vanilla SGD. That is, the communication can be reduced by a factor of the dimension of the whilst still converging at the same rate. In the second (and shorter) half of the talk we discuss strategies that tackle the communication frequency instead of the communicated data. In particular, we compare local SGD (independent runs of SGD in parallel) with mini batch SGD.