Seminars are usually held in the main seminar room (Euler Room - A002) of the Building Euler.
If you wish to receive the seminar announcements, please send an email to sympa2@listes.uclouvain.be.
The subject line should be 'subscribe big-data yourname' where yourname is your first and last name.

Monday 13 Jan, 11h00
Is the search for non trivial new physics phenomena at the LHC an impossible mission ?
Vincent Lemaître (UCL)

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 the smoking gun of a mechanism suggesting the existence of a dramatic phase transition that altered the structure of the vacuum around 10 ^ -10 seconds after the Big Bang. However, several important and fundamental questions remain unanswered and the discovery of the Higgs has opened a new era in the search for physics beyond the "Standard Model" of fundamental interactions. In this seminar, we will try to better define what "observation" means and which data analysis techniques have been/are being put in place to observe possible new physics phenomena at the LHC in the next decade.

Tuesday 7 Jan, 14h
How to extract the oscillating components of a signal? A wavelet-based approach compared to the Empirical Mode Decomposition
Adrien Deliège, Université de Liège

Researchers are often confronted with time series that display pseudo-periodic tendencies with time-varying amplitudes and frequencies. In that framework, a classic Fourier analysis of the data may be of limited interest, especially if the objective is to derive components from the signal that capture the non-stationary behaviour of the oscillating factors. In this talk, we present two powerful tools designed to extract amplitude modulated-frequency modulated (AM-FM) components from a given signal. The first one is the renowned Empirical Mode Decomposition (EMD); we explain the technique, its main benefits, limitations and major practical uses. Then, we introduce the continuous wavelet transform and the equations that justify its relevance in the present context. We propose an algorithm based on the wavelet-induced time-frequency representation of a signal to extract its main components. The performances of this method are compared with the EMD on various AM-FM signals exhibiting different particularities. After briefly broaching the problem of edge effects, we investigate whether the wavelet-based procedure can be used in the domain of time series forecasting. For that purpose, we study the El Nino Southern Oscillation index and develop a model aimed at predicting the long-term trends of the signal. Its predictive skills are tested in several ways and exposed in the final part of the talk.

Thursday 15 Dec, 9h00
Belgian Network Research Meeting, BeNet 2016
Universite Catholique de Louvain

It is our pleasure to announce the sixth edition of the *Belgian Network Research Meeting*, BeNet 2016, to be held at the Université catholique de Louvain on Thu, Dec 15th, 2016. Please circulate this call for participation among your connections in Belgium and feel free to join the BeNet mailing list (info at the end of this call and on the website)
https://be-net.github.io/2016/
This annual meeting is a place for the Belgium-based researchers interested in the network paradigm to meet and communicate their results, regardless of the discipline: sociology, economics, geography, communication, history, biology, physics, medicine, informatics, mathematics, statistics, etc.
Previous editions were held in Namur (UNamur, 2015), Brussels (ULB, 2014), Leuven (IIS/KUL, 2013), Antwerpen (UA, 2012) and Brussels (VUB, 2011). See https://be-net.github.io for the respective websites.
We welcome the submission of short research abstracts. Authors of selected contributions will be invited to present and discuss their work during the meeting. Poster presentations are also encouraged. Please take into account the interdisciplinary nature of this event: tutorial presentations will be favoured over detailed reports.
Important Dates
October 30, 2016: Abstract submission and registration deadline
October 15, 2016: Notification of acceptance
December 15, 2016: Conference

Wednesday 23 Nov, 19h30
First DataBeers Brussels
DataBeers Brussels

Databeers gathers all those with an interest in data-based stories in an informal and relaxed event, together with the best social lubricant: beer. A central part of each meeting is a selection of 4 to 5 short (<7min) and entertaining talks about data aimed at the general public and of course free beer! Successfully born in Madrid, the event has spread across the world in over 10 cities, and it is finally arriving to Brussels. If you like data and beers (or just data), come and join us! Details about the program and speakers are coming soon! Registation is free but mandatory!

Thursday 17 Nov, 11h
Modelling the Human Mental Lexicon via Percolation, Markov Chains and Multiplex Network
Massimo Stella, Institute for Complex Systems Simulation, University of Southampton, UK

Language is a complex system, with a hierarchical set of units interacting on several levels. For instance, sentences are made of interconnected words, which themselves can be thought of as correlated sequences of sounds (i.e. sequences of phonemes). At the level of individual words, psycholinguists conjecture that the interactions among them are encoded within the human mind in the so-called human mental lexicon (HML), i.e. a mental dictionary where words are stored together with their linguistic data. Empirical research has shown that the interaction patterns among words have an impact in learning, memorising and retrieving words from the HML, hence the meaningfulness of considering the whole structure of such relationships through the complex network paradigm.
In (1, 2) we proposed a quantitative framework for phonological networks (PNs), where nodes represent words and links represent phonological similarities (i.e. two phonetic transcriptions having edit distance one). Our null models, based on site percolation and Markov processes, suggested the presence of additional constraints in the assembly of real words, such as (i) the avoidance of large degrees, (ii) the avoidance of triadic closure, and (iii) the avoidance of large non-percolating clusters, in order to avoid word confusability.
In (3) we extended previous analyses of the HML by adopting a multiplex network framework, including (i) phonological similarities, (ii) synonym relationships and (iii) empirical free word associations. We exploited this multi-layered structure for investigating the interplay between phonological and semantic relationships in influencing word acquisition. We proposed a novel toy model of lexicon growth driven by the phonological level, in which real words are inserted according to different orderings and they can be either accepted or rejected for memorization. Our model showed that when similar sounding words are preferentially learned, the lexicon grows according to the multiplex structure, while when novel learned words sound different from the already learned ones, features of the semantic layers and frequency become predominant, instead.
(1) M. Stella and M. Brede, Patterns in the English language: phonological networks, percolation andassembly models, JSTAT, P05006 (2015).
(2) M. Stella and M. Brede, Investigating the Phonetic Organisation of the English Language via Phonological Networks, Percolation and Markov Models, accepted in Proceedings of ECCS2014, Lecture Notes in Computer Science, Springer (2015).
(3) M. Stella and M. Brede, Mental Lexicon Growth Modelling Reveals the Multiplexity of the EnglishLanguage, Proceedings of the 7th Workshop on Complex Networks, Springer (2016).

Tuesday 15 Nov, 14h
Three PhD topics
Cyril De Bodt, Shuyu Dong, Dounia Mulders (UCL)

Three PhD students will explain their topics: "User-centered machine learning" (Cyril De Bodt), "Dictionary learning for graph signals" (Shuyu Dong) and "Human pain connectome identification" (Dounia Mulders)

Thursday 10 Nov, 11h
Assessing dynamic changes of the US financial market: insights from networks science
Yérali Gandica (UNamur)

Drawing on the recent and buoyant literature inferring financial interconnectedness from market data by means of various time series techniques [Billio et al. (2012), Diebold and Yilmaz(2015) and Geraci and Gnabo (2016)], we propose in this communication an in-depth analysis of the US financial market and its dynamic, using tools coming from network science. The financial system analyzed consists in a large set of 154 banks, brokers/dealers, insurance and real state companies listed in the Standard & Poor’s 500 index for the period 1993 - 2014. Looking at the individual, sectoral, community and system wide levels, we show that network science’s tools are able to support well-known features of financial markets such as the dramatic fall of connectivity amid Lehmann Brother collapse. In addition, we also unveil several important new patterns within US financial institutions interconnectedness. Overall, our results improve our understanding of the US financial landscape and may have important implications for risk monitoring as well as macroprudential policy design.

Tuesday 8 Nov, 14h
Mining Patterns in Data
Siegfried Nijssen (UCL)

Pattern mining algorithms are algorithms that search for patterns in large collections of data. While initially developed for the discovery of frequently occurring structures in binary databases, nowadays they can be applied on many types of data: on graph databases, numerical databases, sequence databases, time series and heterogeneous databases; they can not only find frequent structures, but also identify interesting subgroups, accurate classifiers and good clusterings. In this talk, I will discuss number of different pattern mining techniques, with a focus on techniques that find patterns in supervised data; these are techniques that can be used to find associations between one particular attribute in a database and the remaining other attributes. I will discuss algorithms, their computational complexity and remaining challenges.

Wednesday 26 Oct, 14h at room Paul Otlet (building Réaumur, top floor)
Interpolating between Random Walks and Optimal transportation routes
Guillaume Guex (UCL)

In recent articles about graphs, different models proposed a formalism to find a type of path between two nodes, the source and the target, at crossroads between the shortest-path and the random-walk path. These models include a freely adjustable parameter, allowing to tune the behavior of the path towards randomized movements or direct routes. In this talk, a natural generalization of these models is presented, namely a model with multiple sources and targets. In this context, source nodes can be viewed as locations with a supply of a certain good (e.g. people, money, information) and target nodes as locations with a demand of the same good. An algorithm is constructed to display the flow of goods in the network between sources and targets. With again a freely adjustable parameter, this flow can be tuned to follow routes of minimum cost, thus displaying the flow in the context of the optimal transportation problem or, by contrast, a random flow, known to be similar to the electrical current flow if the random-walk is reversible. Moreover, a source-target coupling can be retrieved from this flow, offering an optimal assignment to the transportation problem. This algorithm is described in the first part of the talk and then illustrated with case studies

Thusday 20 Oct, 11h
Twitter datasets: what, how, why?
Adeline Decuyper, Martin Gueuning, Corentin Vande Kerckhove, Renaud Lambiotte

In this working session the researchers mentioned above will explain and discuss their experience of Twitter datasets and what they (plan to) do with them. This not a regular seminar with slick slides and roaring results but more like a brainstorming, work-in-progress session, feel free to come if you have interest in Twitter or similar data and want to discuss it.

Tusday 18 Oct, 16h30h, (CORE, room b-135)
Randomized Algorithms for Convex Optimization
Sebastian Stich, CORE, Université catholique de Louvain

In the last years, randomized optimization algorithms, and especially coordinate descent methods attracted more and more attention in the optimization community. These methods often replace the efficient ̶ but often relative expensive ̶ iterations of the classical methods with much cheaper ̶ but less efficient ̶ updates. Consequently, such methods can be applied to problems of very big size. This application is sometimes also theoretically justified by very attractive worst-case efficiency guarantees.
In this talk, we will introduce randomized analogues of the classical primal-, dual and accelerated schemes and provide the corresponding complexity estimates on unconstrained convex optimization problems. We also discuss important applications where the randomized schemes dominate the classical methods.
If time allows, we will also present two randomized variable metric schemes that can be viewed as analogues of second order methods like Newton’s method and BFGS.

Tusday 18 Oct, 14h
Network analysis through invariant subspaces: symmetries, graph partitions, and dynamics
Michael T. Shaub (UCL, Namur)

Many tools used in network analysis are grounded in the spectral decomposition of a matrix Q associated to the network structure, such as the adjacency matrix A, or the (normalized) Laplacian L; or products of such matrices. From an abstract viewpoint, we may thus interpret all these algorithms as assessing a dominant invariant linear subspace with respect to an appropriately chosen network descriptor matrix.
By limiting the subspace to be one-dimensional, we can recover centrality measures including the famous PageRank algorithm, the Katz centrality, and Kleinberg's HITS score (which is based on the SVD of A). Considering dominant subspaces of larger sizes naturally leads to spectral clustering, and other graph partitioning algorithms.
However, instead of looking only at the dominant invariant subspaces, in this talk we will be concerned with a different class of invariant subspaces, namely, those invariant subspaces that can be spanned by the indicator vectors of a graph partition. Such invariant subspaces are naturally related to so called (external) equitable graph partitions, which are closely associated to the inherent symmetries in the network. As we will show, by analysing these invariant subspaces and the associated graph partitions, we can gain insights about various kinds of cluster synchronization behaviours in consensus dynamics and networks of coupled oscillators, such as Kuramoto oscillators.
Joint work with N. O’Clery, Y.Billeh, R. Lambiotte, J.-C. Delvenne, and M.Barahona

Tusday 11 Oct, 14h
Networks and Quantum Mechanics?
Mauro Faccin (UCL)

How can we conjugate Quantum Mechanics and Network Science? The first is a well established field in Physics while the latter is newborn with a fast-growing interdisciplinary community. In this seminar I will describe some representative cases where these two fields meet and overlap.These examples take inspiration from biophysics to technological applications and apply tools from Network Science to characterize complex systems which obey the laws of Quantum Mechanics. I will show how, in some cases, both fields can benefit from each other when our aim is to speedup a particle moving on a graph or compute the result of a boolean function, to simulate a big system or just distinguish quantum from classical systems.

Tusday 04 Oct, 14h
Semidefinite programming relaxations for matrix completion, inverse scattering and blind deconvolution
Augustin Cosse (UCL)

The talk will discuss three instances of semidefinite programming relaxation. In the first part, the talk will close the line of work on rank one matrix completion by introducing a stable algorithm based on two levels of semidefinite programming relaxation. For this algorithm, we first certify recovery of the matrix encoding the unknowns in the absence of noise, at the information limit, through the construction of a dual (sum of squares) polynomial. In passing, this dual polynomial also provides a rationale for the use of the trace norm in semidefinite programming. The dual polynomial is then used to derive a stability estimate for the noisy version of the problem. In the second part, we introduce a fast algorithm for inverse scattering which leverages the traditional Adjoint State method used in geophysics by lifting the search space. The algorithm is based on a first level of semidefinite programming relaxation encoded through a low rank factorization which guarantees its scalability. Numerical experiments on 2D community models are used to highlight a modest increase with respect to the basin of attraction of traditional least squares waveform inversion. A geometric intuition for this improvement is provided. Finally, in the last part of the talk, we discuss how blind deconvolution can be solved with high probability through nuclear norm relaxation by considering multiple input signals. Recovery is certified through the construction of a dual certificate. A candidate certificate is first constructed through the golfing scheme. Relying on the recent line of work on blind deconvolution, this candidate certificate is then shown to satisfy the conditions from duality theory for the recovery of the rank one matrix encoding the impulse response and input signals, through the Orlicz version of the Bernstein concentration bound.

Useful links : UCLouvain | ICTEAM | INMA