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

Previous seminars

7/06/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Alex Olshevsky
Distributed Optimization: A Tutorial

This talk will provide a gentle introduction to the field of distributed optimization. The core question is how a collection of nodes interconnected in a network can cooperate to compute a minimizer of a convex function, provided each of them knows only a piece of this function. We will sketch the proof of a few of the results that have been obtained over the past decade and conclude with a list of open questions

29/05/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Samuel Branders (Tools4Patient)
Leveraging historical data for high-dimensional regression adjustment, a machine learning approach

The amount of data collected from patients involved in clinical trials is continuously growing. All those patient's characteristics are potential covariates that could be used to improve study analysis and power. At the same time, the development of computerized systems simplifies the access to huge amount of historical data. However, it is still difficult to leverage those data when dealing with small clinical trials, such as in Phases I and II. Their restricted number of patients limits the possible number of covariates included in the analysis. The purpose of this talk is to present how machine learning can overcome this problem by taking advantage of historical data with larger sample sizes. Our approach is to pre-specify the combination of the baseline covariates by building a "meta-covariate". In small studies, using this meta-covariate alone will limit the loss of degrees of freedom while making the best uses of all generated data. Two advantages of fitting the covariates on independent data are to free the modeling from the study constraints and to limit the risk of overfitting. Those are of particular interest with complex data, i.e non-normal distribution or in the presence of non-linearities. To demonstrate the benefit of the methodology, we discuss several questions: - What are too many covariates? Can we go beyond the simple rule-of-thumb 1 variable for 10 patients? - What should be the minimum performance of the ML model in this context? - When should we use machine learning?

08/05/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Thibault Helleputte (DNAlytics)
Data science for precision medicine

The field of precision medicine aims at developing new decision-support tools for the healthcare professionals, in order to improve patients health, healthcare system sustainability, and potentially better position some pharmaceutical products. These tools can be developed by combining a broad range of health-related data and mathematical modeling strategies, namely machine learning approaches. In this talk, we will discuss how some very well defined machine learning approaches for predictive model training, feature selection and performance evaluation survive the reality of biology, medicine and clinical applications.

24/04/2018 (14:00) [Lieu : Maxwell building, Shannon room]
Patrick Thiran (EPFL)
Locating the source of diffusion in large-scale and random networks

We will survey some results on the localization of the source of diffusion in a network. There have been significant efforts in studying the dynamics of epidemic propagations on networks, and more particularly on the forward problem of epidemics: understanding the diffusion process and its dependence on the infecting and curing rates. We address here the inverse problem of inferring the original source of diffusion, given the infection data gathered at some of the nodes in the network. Indeed, because of the large size of many real networks, the state of all nodes in a network cannot in general be observed. We show that it is fundamentally possible to estimate the location of the source from measurements collected by sparsely placed observers. We present a strategy that is optimal for arbitrary trees, achieving maximum probability of correct localization, and describe efficient implementations for arbitrary graphs. When propagation times are (close to) deterministic values, the smallest number of sensors needed to exactly localize the source is the double metric dimension of the network. We compute tight bounds for this quantity in a number of topologies that include Erdös-Renyi random graphs, and we comment on its implications for the detectability of a source in actual networks. This is a joint work with Elisa Celis, Pedro Pinto, Brunella Spinelli and Martin Vetterli.

17/04/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Vincent Moens (COSY, Institute of Neuroscience, UCL)
The Hierarchical Adaptive Forgetting Variational Filter

A common problem in Machine Learning and statistics consists in detecting whether the current sample in a stream of data belongs to the same distribution as previous ones, is an isolated outlier or inaugurates a new distribution of data. We present a hierarchical Bayesian algorithm that aims at learning a time-specific approximate posterior distribution of the parameters describing the distribution of the data observed. We derive the update equations of the variational parameters of the approximate posterior at each time step for models from the exponential family, and show that these updates find interesting correspondents in Reinforcement Learning (RL). In this perspective, our model can be seen as a hierarchical RL algorithm that learns a posterior distribution according to a certain stability confidence that is, in turn, learned according to its own stability confidence. Finally, we show some applications of our generic model, first in a RL context, next with an adaptive Bayesian Autoregressive model, and finally in the context of Stochastic Gradient Descent optimization.

20/03/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Duncan Callaway (University of California, Berkeley)
Distributed Energy Resource Control and Network Optimization

In this talk I'll discuss theory, simulation and experimental work aimed at understanding how distributed energy resources (DERs; such as batteries, flexible loads and photovoltaic generators) impact distribution networks and how these impacts can be managed. I will also briefly discuss a new approach to engaging DERs in providing transmission level services. The core theory focus of the talk will be the development and analysis of an approach to utilize extremum seeking control for decentralized coordination of DER in real time.

13/03/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Thibaut Lienart (University of Oxford)
Approximate Bayesian inference in Machine Learning: where stats meets optimisation

In this talk I will discuss some recent work on distributed approximate Bayesian inference for Machine Learning where one is interested not only in finding good parameters for a given model (e.g.: the weights of a Neural Network) but also to model the uncertainty around these parameters and thereby model the uncertainty in the predictions made with the model. This is of tremendous importance in modern applications of ML such as self-driving cars, healthcare applications etc. I will show how approximate Bayesian inference converts a hard integration problem into a simpler optimisation problem and discuss some approaches that have been used to tackle the optimisation. Finally, I will give a critical view on the field, hopefully leading to a discussion.

06/03/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Arnaud Latiers (Actility)
Electricity Demand Response in practice - why simplicity matters in a complex system

Demand Response has become an essential part of the Belgian Electric system operations. This tools allows consumers to participate in the balancing of the grid, or in the value definition in electricity markets (demand curve, elasticity). It brings value at all stages in the energy system : adequacy, security, efficiency. While the technical potential is enormous, market potential relatively lower. One of the biggest barriers to accessing more flexibility in the current market is the fact that more complex models must be put in place to access the unexploited flexible capacity. Furthermore, services themselves are becoming less user friendly, which makes contractualisation more difficult. In this talk, we will highlight how simplicity at all stages is the key for success. Quite amazingly, it starts often by simpler mathematical formulation of the problem...

27/02/2018 (16:30) [Lieu : CORE, b-135]
Jenny Benois-Pineau (University of Bordeaux)
Increasing stability of training of Deep CNNs with Stochastic Gradient Descent method. Application to Image classification tasks

[Note the unusual time and location.] See

20/02/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Bart De Moor (KU Leuven)
Back to the roots: Multivariate polynomial optimization by numerical linear algebra

Finding some or all of the roots of a set of multivariate polynomials has numerous applications. This century old problem is at the foundation of algebraic geometry sometimes called the 'queen of mathematics'. In this discipline, typically, numerical computations are done symbolically, and only quite recently, it was realized that the multivariate polynomial rooting problem can be tackled using the machinery of numerical linear algebra, because of the fundamental insight that it is equivalent to an eigenvalue problem. In order to find some or all roots, we deploy tools like the QR- and singular value decomposition, and (possibly large scale) iterative eigenvalue solvers. We will discuss several eigenvalue decomposition based algorithms to calculate the global minimum of multivariate polynomial optimization problems. We will illustrate our approach with the numerical solution of two open problems in system theory: Calculating from (noisy) data linear dynamic models that are least squares optimal, and finding the least squares global optimum of the H2 model reduction problem.

13/02/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Ignacio Aravena (UCL)
An Asynchronous Distributed Algorithm for Solving Stochastic Unit Commitment

We present an asynchronous algorithm for solving the stochastic unit commitment (SUC) problem using decomposition. The algorithm is motivated by large differences in run times observed among subproblems, which can result in inefficient use of distributed computing resources by synchronous parallel algorithms. Dual iterations are performed asynchronously using a block-coordinate subgradient descent method which allows performing block-coordinate updates using delayed information, while candidate primal solutions are recovered from the solutions of scenario subproblems using heuristics. The asynchronous algorithm is implemented following the master-slave paradigm and using MPI to handle communication among processes. We conduct numerical experiments using a high performance computing cluster for two-stage SUC instances of (i) the Western Electricity Coordinating Council (WECC) system with up to 1000 scenarios and of (ii) the Central Western European (CWE) system with up to 120 scenarios. The algorithm provides solutions to all problems within 2% of optimality in at most 23 minutes for WECC and 98 minutes for CWE, and solutions within 1% of optimality in at most 63 minutes for WECC and 133 minutes for CWE. Moreover, we find that an equivalent synchronous parallel algorithm would leave processors idle up to 80.4% of the time, an observation which underscores the need for designing asynchronous optimization schemes in order to fully exploit distributed computing on real world applications.

06/02/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Frédéric Crevecoeur (UCL)
Robust and adaptive control of reaching movements in human

Neural plasticity and flexible representations of reach dynamics allows humans and animals to adapt to a wide range of disturbances applied to the limb such as robotic force fields, Coriolis force fields, and even changes in gravity. This ability has been almost exclusively described across trials, characterizing the gradual improvement of movement performances through learning curves. However, it remains unknown how we control movements in the presence of unexpected disturbances. Two candidate hypotheses are motivated by engineering techniques: robust (Hoo) and adaptive control can be used to mitigate the impact of model errors. Our experimental results show that healthy humans may implement these two strategies. We found that the presence of uncertainty in reach dynamics evoked changes in control consistent with a robust control model. Furthermore, within movement feedback corrections for unexpected dynamics evoked online adaptation of the ongoing motor command. These two components of movement execution may provide a powerful framework for linking control over time with learning over trials.

16/01/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Sasa V. Rakovic
Robust model predictive control

Model predictive control (MPC) is an advanced control technique that employs an open-loop online optimization in order to take account of system dynamics, constraints and control objectives and to obtain the best current control action. Robust MPC (RMPC) is an improved MPC form that is robust against the bounded uncertainty. RMPC employs a generalized prediction framework that allows for a meaningful optimization of, and over, the set of possible system behaviours effected by the uncertainty. A real intricacy in RMPC arises due to the facts that the exact RMPC provides strong structural properties but it is computationally unwieldy, while the conventional MPC is not necessarily robust even though it is computationally convenient. The seminar focuses on novel RMPC methods, developed through my research investigations and collaborations, that address effectively the fundamental challenge of reaching a meaningful compromise between the quality of guaranteed structural properties and the associated computational complexity. In particular, the talk discusses flexible and efficiently optimizable parameterizations as well as tube MPC, which lead to synthesis methods that are theoretically sound (i.e., they guarantee a-priori strong structural properties) and computationally efficient (i.e., they have a manageable computational complexity that is close enough to that of the conventional MPC synthesis).

12/01/2018 (14:00) [Lieu : Bât. Euler (room A.002)]
Kapil Ahuja (IIT Indore, India)
Inexact Linear Solves for Model Reduction

Dynamical systems arise in varied fields of science and technology. Dimension of such systems is usually large. Hence, simulations using them is cumbersome. Model reduction of a large dynamical system gives us a smaller dynamical system that mimics the input-output behavior of the original dynamical system.
With the increase in the dimension of the original dynamical system, the model reduction process itself becomes a hard task. The reason behind this is that most model reduction algorithms require solution of linear systems of equations, which traditionally are solved by direct methods that scale badly with increase in the dimension. This is the main motivation for using iterative methods for solution of linear systems of equations arising in model reduction algorithms.
Iterative methods are inexact in nature. That is, they solve a linear system up to a certain tolerance. Hence, studying stability of model reduction algorithms with respect to the error introduced by inexact solves is an important task. In this talk, I will discuss this kind of backward stability of popular model reduction algorithms for first-order bilinear and second-order linear dynamical systems.

19/12/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Nikos Deligiannis (VUB)
Data recovery from parsimonious observations: From compressed sensing to deep learning

The problem of recovering high-dimensional data from parsimonious, low-dimensional measurements is linked to a plethora of applications ranging from big data collection in smart environments to the enhancement of multimodal imaging sensors. The problem has been studied from various angles from the signal processing and machine learning communities, including compressed sensing - which enables the acquisition of a sparse (or compressible) signal using far less measurements than the classical Shannon-Nyquist theory - and matrix completion - which is essentially matrix factorisation with low-rank constraints. In the first part of the talk, we will present approaches to regularise such problems using constraints that encode the similarity of the data we wish to recover with observed, correlated (multimodal) data; while in the second part, we will present deep learning formulations that address the problem. Special focus will be put on learning good designs for the data collection mechanisms so as to reduce the associated complexity of the task. Finally, we will present several applications of our frameworks in multimodal image recovery, computer vision, and data recovery from internet-of-things devices measuring the air pollution.

12/12/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Michaël Fanuel (UCL)
Positive semi-definite embedding for dimensionality reduction and out-of-sample extensions

Dimensionality reduction of data is often an important preprocessing step in statistics or machine learning. Several non-linear techniques exist, for instance, Diffusion Maps or, more generally, manifold learning. In this context, we will discuss a recent proposal to reduce the dimension of Euclidean data by relying on methods developed for graph embedding. More specifically, we discuss a numerical method phrased as a Semi-Definite Program for estimating a distance (semi-metric) on the data cloud. As an interesting feature, this positive semi-definite "matrix" can be extended to a "kernel" function thanks to an out-of-sample formula. Indeed, if any new data point arises, it may also be embedded at a low computational cost thanks to the extension formula. Inspired by this property, we will also briefly discuss an infinite dimensional analogue of this discrete problem in the context of machine learning.

6/12/2017 (17:00) [Lieu : BARB92]
Mattia Gazzola (Mechanical Science and Engineering, UIUC National Center for Supercomputing Applications)
Design of soft machines

[This is a joint seminar with IMMC. Note the usual date and place.] We introduce an approach based on minimal theoretical modeling, direct numerical simulations and evolutionary optimization for the characterization and design of arbitrary muscoloskeletal architectures. Applications range from slithering and swimming biolocomotion strategies to artificial muscles and bio-hybrid systems.

21/11/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Christophe Delaere (UCL)
Search for new rare physics processes: towards a "deeper" understanding of LHC data?

The Large Hadron Collider (LHC) allows us to probe sub-protonic scales where new elementary particles and / or new fundamental interactions could appear. The recent observation of the Higgs particle in 2012 was only made possible through the use of advanced multivariate analysis techniques. While more data is being collected, the collaboration are developing new approaches to data analysis, both for the low-level "event reconstruction" and for the high-level signal extraction, using deep learning when appropriate. In this seminar, we will try to better define what "observation" means and illustrate with examples the data analysis techniques explored to observe possible new physics phenomena at the LHC in the next decade.

14/11/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Noémie Jaquier (Ecole Polytechnique Fédérale de Lausanne)
Improving learning in robotics by exploiting the structure and geometry of the data: application to prosthetic hands

In many sensing and control applications, data are represented in the form of multidimensional arrays with particular geometric properties such as symmetries. Considering the underlying structure and geometry of the data can be beneficial in many robotics applications. This is particularly important in the context of robot learning from demonstration, when a small set of demonstrations is available, i.e. when the robots are required to reproduce tasks with generalization to new situations.
I will present different regression techniques and their extension to models that can efficiently exploit the specific structure and geometry of the acquired datapoints. The approaches will be illustrated in the application of controlling a prosthetic hand using tactile and electromyography to predict hand movements.

7/11/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Damien Scieur (Inria Paris)
Integration Methods and Accelerated Optimization Algorithms

We show that accelerated optimization methods can be seen as particular instances of multi-step integration schemes from numerical analysis, applied to the gradient flow equation. In comparison with recent advances in this vein, the differential equation considered here is the basic gradient flow and we show that multi-step schemes allow integration of this differential equation using larger step sizes, thus intuitively explaining acceleration results.

2/11/2017 (11:00) [Lieu : Bât. Euler (room A002)]
Michael Schaub (MIT and University of Oxford)
Graph embeddings and community detection -- a control theoretic perspective

Community detection, the task of partitioning a network into groups of nodes similar to each other according to some criterion, has received enormous attention in the past decade. Many different notions of what constitutes a good community exist in the literature, some based on finding groups with high edge-density, a small cut between the different groups, or by assuming a particular (generative) model for the structure of the network as a whole. In this talk we focus on a dynamical notion of community structure: we think of our system under study as a network of coupled dynamical units, and aim to find groups of nodes which influence the system in a coherent way for a particular time-scale. Previous work, focusing on diffusion processes, has shown that such a perspective can indeed be fruitful, and that notions such as Modularity, spectral clustering and various other well-known graph partitioning heuristics can be recovered using this point of view. In this talk we reconsider such a dynamical approach towards community detection using a more control-theoretic perspective. This enables us not only to consider more general system descriptions, such as networks with both positive and negative edges, but also reveals interesting connections to ideas used in dimensionality reduction and graph embeddings. More importantly, however, it also allows us to connect to concepts considered in Control Theory, such as controllability and observability Gramians, thereby offering a fresh perspective on the problems of community detection and graph embeddings.

17/10/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
First-year PhD students (UCL)
Welcome seminar

In this seminar, the new PhD students in INMA will introduce themselves and their research topic: Dynamic networks in neuroscience (Benjamin Chiêm), Global Riemannian optimization for tensor decomposition (Guillaume Olikier), Cyber-Physical Systems control : new optimization techniques (Guillaume Berger), Behavioural and modelling study of rapid control during complex movements in humans (Antoine De Comité), Finger pad mechanics during dexterous object manipulation (Félicien Schiltz), Predictive maintenance for proton therapy machines (Boris Dehem).

10/10/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Pierre Pinson (Technical University of Denmark)
Exploring community-based and peer-to-peer electricity markets

The deployment of distributed renewable generation capacities, new ICT capabilities, as well as a more proactive role of consumers, are all motivating rethinking electricity markets in a more distributed and consumer-centric fashion. After motivating the design of various forms of consumer-centric electricity markets, we will focus on two alternative constructs (which could actually be unified) consisting in community-based and peer-to-peer electricity markets. The mathematical framework for these markets will be described, with focus on negotiation and clearing algorithms in a distributed and decentralized setup. Opportunities and challenges related to these markets, both mathematical and related to real-world applications, will be discussed.

26/09/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Ricardo Castro-Garcia (KU Leuven)
Impulse response constrained LS-SVM modelling for MIMO Hammerstein system identification

Hammerstein systems are commonly found in many different areas. From chemical processes to power amplifiers, this type of structures have been used for decades in the system identification community. In this presentation we introduce a new method for modeling MIMO Hammerstein systems through an impulse response constrained LS-SVM formulation. One of the main advantages of this method is its flexibility concerning the class of problems it can model and that no previous knowledge about the underlying nonlinearities is required except for very mild assumptions.

11/09/2017 (14:00) [Lieu : Bât. Euler (room A.002)]
Sewoong Oh (University of Illinois at Urbana-Champaign)
Achieving budget-optimality with adaptive schemes in crowdsourcing

Crowdsourcing platforms provide marketplaces where task requesters can pay to get labels on their data. Such markets have emerged recently as popular venues for collecting annotations that are crucial in training machine learning models in various applications. However, as jobs are tedious and payments are low, errors are common in such crowdsourced labels. A common strategy to overcome such noise in the answers is to add redundancy by getting multiple answers for each task and aggregating them using some methods such as majority voting. For such a system, there is a fundamental question of interest: how can we maximize the accuracy given a fixed budget on how many responses we can collect on the crowdsourcing system.
We characterize this fundamental trade-off between the budget (how many answers the requester can collect in total) and the accuracy in the estimated labels. In particular, we ask whether adaptive task assignment schemes lead to a more efficient trade-off between the accuracy and the budget.
Adaptive schemes, where tasks are assigned adaptively based on the data collected thus far, are widely used in practical crowdsourcing systems to efficiently use a given fixed budget. However, existing theoretical analyses of crowdsourcing systems suggest that the gain of adaptive task assignments is minimal. To bridge this gap, we investigate this question under a strictly more general probabilistic model, which has been recently introduced to model practical crowdsourced annotations. Under this generalized Dawid-Skene model, we characterize the fundamental trade-off between budget and accuracy. I will present a novel adaptive task assignment scheme that matches this fundamental limit. This allows us to quantify the fundamental gap between adaptive and non-adaptive schemes, by comparing the trade-off with the one for non-adaptive schemes.