You are here

Seminars

2017

  • Deep learning techniques - overview

    by: Yann-Aël Le Borgne
    Machine Learning Group, Université libre de Bruxelles, Belgium
     
    When:
    9 May 2017 from 4PM
     
    Where: La Plain Campus, Forum A
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    The presentation will provide a brief recall of neural networks (perceptron and multi-layer perceptrons, gradient descent, backpropagation) and will then cover in more details how convolutional and recurrent networks work, discuss parallelisation strategies, and give an overview of deep learning toolboxes. The presentation will be given as part of the course 'Statistical foundations of Machine Learning' (INFO-F-422, 1st year Master).

2016

  • Big Data Science with Applications in Genomics

    by: Alexander Schliep
    Department of Computer Science, BioMaPS Institute for Quantitative Biology, Rutgers, The State University of New Jersey, USA
     
    When:
    20 April 2016 from 3PM
     
    Where: La Plain Campus, Building NO, 8-th floor, local (Rotule) P.2NO8.08
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    The rapid advances in the quantification of the natural and the technical world in which we live, the quantification of ourselves and our actions requires effective responses both in method development and in education, particular with respect to computational thinking for non-CS majors. I will present relevant aspects of my work. High-throughput sequencing (HTS), a technology to unravel genomic sequences on a large scale, is pervasive in clinical and biological applications such as cancer research and basic science, and is expected to gain enormous momentum in future personalized medicine applications. To address this deluge of data we developed compressive genomics methods which operate directly on reduced representations of the data and enable the use of advanced statistics even on very large data sets. For identifying Copy Number Variants (CNV) our approach accelerated full Bayesian methods substantially over the non-Bayesian state-of-the-art. Computational Thinking: With CS is becoming an important tool outside of the classical technical and natural sciences field, e.g. in the digital humanities, the question arises of how we can improve education in computational and algorithmic. I will present some work on algorithm animations and outline future work in social and mobile learning with effective scaffolding.
  • Scalable solutions for time series tasks

    by: Mourad Khayati
    Exascale lab, University of Fribourg, Switzerland
     
    When:
    13 January 2016 from 3PM
     
    Where: La Plain Campus, Building NO, 5-th floor, local P.2NO5.06
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Time series data arise in a variety of domains, such as environmental, telecommunication, financial, and medical data. This sheer amount of data emanates from different sources such as sensors. Due to technical problems, these sensors might either stop working for a while yielding missing values, or start malfunctioning yielding anomalies. The state-of-the-art techniques to solve these two problems often suffer from the lack of scalability. In this talk, I will discuss two scalable solutions for time series tasks. More specifically, I will first introduce an accurate recovery solution of missing values that scales up to long time series. The proposed solution is based on a matrix decomposition technique called Centroid Decomposition. Second, I will describe an online anomaly detection solution that scales out on a real-time big data environment. The introduced approach accurately detects anomalies caused by gradual and abrupt changes using respectively the entropy and the correlation. Finally, I will discuss future research directions for this work.

2015

  • Introduction to convolutional neural networks

    by: Laura Mannie-Corbisier
    Solvay Brussels School of Economics and Management, ULB
     
    When:
    16 September 2015 from 11:00
     
    Where: La Plain Campus, Building NO, 9th floor, local NO9.06
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Convolutional neural networks (CNN) are part of deep learning methods. Its main characteristic is its convolutional layer aiming at valorizing spatial information. CNN make features extraction step nearly irrelevant, reducing the need for domain-specifc prior knowledge. Like other neural networks, CNN can be computed through gradient-descent (backward propagation algorithm). One of the breaking-through papers using backward propagationon CNN was written in 1998 by LeCun Y., Bottou L., Bengio Y., Hafner P. ( Gradient-based learning applied to document recognition). CNN also suffer from more general issues such as learning slowdown and overfitting.However, in contrast to fully-connected neural networks, CNN has less free parameters thanks to weight-sharing. We finish by introducing unsupervized context of neural networks.
  • Just In Time Classifiers For Recurrent Concepts

    by: Giacomo Boracchi
    Politecnico di Milano, Italy
     
    When:
    16 September 2015 from 12PM
     
    Where: La Plain Campus, Building NO, 9th floor, local NO9.06
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Many machine-learning techniques make the assumption that training and testing data are sampled from the same probability distribution. Unfortunately, in an increasing number of real-world learning scenarios data arrive in a stream, and the probabilistic properties of the data generating process might be changing with time, violating the above assumption. Any algorithm or model that does not account for such change is almost certainly going to fail when data are sampled from a drifting or changing distribution, namely when data are affected by concept drift. Approaches for learning under concept drift can be divided in two main learning strategies: i) undergoing continuous adaptation to match the recent concept (passive approach), or ii) steadily monitoring the data stream to detect concept drift and eventually react (active approaches). In this talk, I will present Just In Time (JIT) Classifiers, a family of viagraforlife.net classifiers that implement an active approach to handle concept drift. In particular, JIT classifiers monitor the data-generating process by means of change-detection tests, and build representations of the encountered concepts that are then handled by suitable operators to identify recurrent concepts.
  • When is undersampling effective in unbalanced classification tasks?

    by: Andrea Dal Pozzolo
    Université Libre de Bruxelles, BE
     
    When:
    2 September 2015 from 4PM
     
    Where: La Plain Campus, Building NO, 8th floor, 2NO8.08
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    A well-known rule of thumb in unbalanced classification recommends the rebalancing (typically by resampling) of the classes before proceeding with the learning of the classifier. Though this seems to work for the majority of cases, no detailed analysis exists about the impact of undersampling on the accuracy of the final classifier. This presentation aims to fill this gap by proposing an integrated analysis of the two elements which have the largest impact on the effectiveness of an undersampling strategy: the increase of the variance due to the reduction of the number of samples and the warping of the posterior distribution due to the change of priori probabilities. In particular we will propose a theoretical analysis specifying under which conditions undersampling is recommended and expected to be effective. It emerges that the impact of undersampling depends on the number of samples, the variance of the classifier, the degree of imbalance and more specifically on the value of the posterior probability. This makes difficult to predict the average effectiveness of an undersampling strategy since its benefits depend on the distribution of the testing points. Results from several synthetic and real-world unbalanced datasets support and validate our findings.
  • Context-sensitive Ordinal Regression Models for Human Facial Behaviour Analysis

    by: Ognjen Rudovic
    Imperial College London, UK
     
    When:
    8 July 2015 from 11:30
     
    Where: La Plain Campus, Building NO, 8th floor, 2NO8.08
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Enabling computers to understand human facial behaviour has the potential to revolutionize many important areas such as clinical diagnosis, marketing, human computer interaction, and social robotics, to mention but a few. However, achieving this is challenging as human facial behaviour is a highly non-linear dynamic process driven by many internal and external factors, including ‘who’ the observed subject is, ‘what’ is his current task, and so on. All this makes the target problem highly context-sensitive, resulting in the changes of dynamics of human facial behaviour, which, in turn, is critical for interpretation and classification of target affective states (e.g., intensity levels of emotions or pain). In this talk, I will propose several extensions of the Conditional Ordinal Random Fields (CORF) model that are able to learn spatio-temporal and context-sensitive representations of human facial behaviour useful in various tasks of facial analysis. In particular, I will show how the proposed CORF models can be used for problems such as intensity estimation of facial expressions of emotion, intensity estimation of facial action units and facial expressions of pain. I will also demonstrate the performance of the models on the task of classification of facial expressions of persons with autism spectrum condition. Finally, I will discuss other potential applications of the models proposed and further challenges in modelling of human facial behaviour.

2014

  • An hands-on introduction to probabilistic programming

    by: Manuel Pegalajar Cuéllar
    Department of Computer Science and Artificial Intelligence, the University of Granada, Spain
     
    When:
    30 October 2014 at 3PM (until 5PM)
     
    Where: La Plain Campus, Building NO, 8th floor, 2NO8.08
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    This talk is aimed to provide an introduction to what Probabilistic Programming is, its current research topics opened and challenges. Also, we study R-Stan, a R interface to the probabilistic programming language Stan with introductory examples. I would like that the seminar will be practical as much as possible. People are encouraged to come with their laptops to follow the experiments carried out during the presentation, and also to test the exercises.
  • Nonlinear forecasting of macroeconomic variables using three automated model selection techniques

    by: Timo Terasvirta
    Aarhus University, Danemark
     
    When:
    25 September 2014 at 12:30 (until 13:30)
     
    Where: La Plain Campus, Building NO, 7th floor, 2NO7.08
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    In this talk, the focus is on the forecasting performance of a well-defined class of flexible models, the so-called single hidden-layer feedforward neural network models. A major aim of our study is to find out whether they, due to their flexibility, are as useful tools in economic forecasting as some previous studies have indicated. A central problem with these models is how to specify their structure and estimate the parameters. Recently, White (2006) presented a solution that amounts to converting the specification and nonlinear estimation problem into a linear model selection and estimation problem. This leads to a situation that is somewhat atypical, at least in time series econometrics, in which the number of variables may vastly exceed the number of observations. We compare three methods of model selection capable of handling this problem. One is White's QuickNet, and the other two are the Marginal Bridge Estimator (MBE), well known to statisticians and microeconometricians, and Autometrics, popular among time series econometricians. We consider multiperiod forecasts. There are two main ways of generating them. One is to specify and estimate a single model and generate the forecasts recursively from it. It is also possible to build a separate model for each forecast horizon and use it for obtaining the forecasts directly for the horizon in question. We compare the performance of these alternatives, when the set of available models consists of linear autoregressive, neural network and nonparametric ones.
  • When R meets the Hadoop and Spark ecosystem

    by: Eric Charles
    Datalayer
     
    When:
    23 june 2014 at 12:00 (until 13:00)
     
    Where: La Plain Campus, Building NO, 8th floor, 2NO8.08 (Rotule Room)
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Corporates are beginning to extract business value from their Big Data based on open source solutions such as Hadoop and Spark. R and its academic ecosystem is also needed as basis and first step to model the cases. This technical solutions and unexplored areas will described (how can we distribute at scale a standalone R model) and we will give a few real-life examples based on the projects done at Datalayer implemented at Belgian companies.
     
    Speaker:
    Eric CHARLES is the founder of DATALAYER that provides in Belgium development services based on the HADOOP and SPARK ecosystem. He worked in London on BIG DATA projects. Eric is also APACHE Member and Committer. You can contact him via email (eric@datalayer.io) or on Twitter (@echarles).

2013

  • Towards Smart Rehabilitation: Proactive Sensing for Remote and Automatic Medical Evaluation

    by: Manuel Pegalajar Cuéllar
    Department of Computer Science and Artificial Intelligence, University of Granada, Spain
     
    When:
    5 November 2013 at 10:00 (until 11:00)
     
    Where: La Plain Campus, Building NO, 8th floor, 2NO8.08 (Rotule Room)
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    The recent advances in hardware, and more specifically sensor technologies, allows us to acquire accurate information of human behaviour and activities in daily life environments. This fact has attracted the attention of a high number of researchers from very different knowledge areas, such as computer science, biomechanics, or medical and health care, between others, that join forces in multidisciplinar teams to develop new systems and devices aimed to improve life quality. In this presentation, we outline the goals and developments of the project "Towards Smart Rehabilitation: Proactive Sensing for Remote and Automatic Medical Evaluation". We explore the use of depth sensor cameras to acquire motion-based behaviours from users, and develop new models for performance evaluation of physical activity. As a result, our work leads to improve automation evaluation of diagnosis exercises in physiotherapy, and to provide physiotherapists with a tool to acquire and analize motion data using cheap solutions that do not use markers. The presentation ends with a performance demonstration of the prototype under development.
  • A gradient boosting approach to the Kaggle load forecasting competition.

    by: Souhaib Ben Taieb
    Machine Learning Group (MLG), Belgium
     
    When:
    28 March 2013 at 12:30 (until 13:30)
     
    Where: La Plain Campus, Building NO, 5th floor, Solvay Room
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    We describe and analyze the approach used by Team TinTin (Souhaib Ben Taieb and Rob J. Hyndman) in the Load Forecasting track of the Kaggle Global Energy Forecasting Competition 2012. The competition involved a hierarchical load forecasting problem for a US utility with 20 geographical zones. Eight in-sample and one out-of-sample weeks for 21 separate time series need to be backcast and forecast, respectively. The electricity demand for the next day is forecasted using a separate model for each hourly period. We use component-wise gradient boosting to estimate each hourly model with univariate penalized regression splines as base learners. The models allow for the electricity demand to change with time-of-year, day-of-week, time-of-day, and on public holidays with the main predictors being current and past temperatures as well as past demand. Our model ranked fifth out of 105 participating teams.
  • Machine learning in DG Sanco - EC : database clean up, fraud detection and web semantic.-- Slides

    by: Philippe Loopuyt and Eric Ngantchjon
    European Commission
     
    When:
    9 January 2013 at 3.15PM (until 4.15PM including questions) 
     
    Where: La Plain Campus, Building NO, 8th floor, Rotule room
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    The last years, DG Sanco has used some Machine learning techniques for diverse projects. With the increase of data received from different actors and the tremendous number of applications having related information, there is a need to define common entities and clean up databases with duplicate records. An application for the detection of duplicate records and the the generation of clusters of similar records has been developped. The algorithms are based on Levenshtein distance as text metric and K-Means as clustering. An expert system application has been developed in order to help the BIP (Border Inspection Point) to improve the random control for the trade of animals and products among Europe and also between MS and third countries. Based on historical values of fraudulent consignments, patterns are found and predictive models are built to check future consignments. Predictive models are built with KXEN, a machine learning software based on the 'Vapnik-Chervonenkis' theory. With the web semanic project, the objective is to publish public data in a semantic format; furthermore, there are more and more external users asking for an API allowing them to get the public data in an automatic way in such a way that it will avoid the long process of manually updating their own system. The web semantic responds efficiently to these requirements; it is a global concept allowing the existence of links on distributed data spread over the web. Besides there are some protocols used to query and access these data. But one of the big challenges of web semantic is to find a way to set automatically (or at least semi-automatically) the links with unstructured data, and this leads to the use of technics such as text mining, taxonomy and ontology.
     
    Speaker:
    Philippe Loopuyt, Head of Unit Information Systems, Directorate General Health and Consumers, European Commission. Eric Ngantchjon, graduated in Electrical(Telecommunication) Engineering in 1998 (Polytech-Mons), and Statistics-Operations Research in 2008 (ULB). He is an experienced Software Architect with strong background in applied statistics. He developed business critical solutions on inventory control, statistical analysis, risk assessment, fraud detection, web semantic and text mining.
  • Applications of Machine Learning and Soft Computing techniques to human behaviour inference and chemical sensors modelling -- Slides

    by: Manuel Pegalajar Cuéllar, 
    Department of Computer Science and Artificial Intelligence,
    University of Granada (Spain)
     
    When:
    3 December 2012 at 2PM (until 3PM including questions) 
     
    Where: La Plain Campus, Building NO, 8th floor, Rotule room
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Machine Learning and Soft Computing are two key areas in contemporary Artificial Intelligence. Their techniques have been applied to a large number of real applications, obtaining promising results and solving problems where other traditional methods have failed. For instance, this is especially the case when noisy data is present. In this talk, we show our current advances in two areas: Disposable optical sensor modelling and, what will be out main emphasis, human behaviour inference and recognition.
    Regarding the first area, we describe different ML techniques used, ranging from classic non-linear regression to neural networks, expert systems and multi-objective optimisation to minimise the size of the sensor. In relation to the second area, we show our approaches for adaptive models of human behaviours using discrete response sensors, and discuss the recent work with active sensors such as cameras and accelerometry sensing in order to learn about locomotion habits and gestures. In the end, we also discuss open problems and some possible ways of collaboration with MLG.
     
    Speaker:
    Manuel P. Cuéllar graduated in Computer Engineering in 2003. He finished his PhD on time series prediction, parameter identification and neural networks in 2006. He is currently an associate Professor with the Department of Computer Science and Artificial Intelligence at the University of Granada (Spain). His main interests are neural and social networks, evolutionary optimisation and fuzzy systems, although he has also worked in multivariate image analysis and real-time control tasks. His current work encompasses different research areas including real-time learning, chemical parameter identification, medical imaging, development of disposable optical sensors and intelligent systems for ambient assisted living. He has contributed with more than 40 papers in conferences and research journals and has collaborated in 6 research projects under competitive application from the government of Spain.
 
2012
  • The role of Self Organizing Dynamic Agents for Decentralized Optimization in Smart Grids -- Slides

    by: Prof. Alfredo Vaccaro, 
    Electric Power Systems at the Department of Engineering,
    University of Sannio
     
    When:
    September 13, 2012 at 16:00 (until 17:00 including questions) 
     
    Where: La Plain Campus, Building NO, 8th floor, Rotule room
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    In this Talk we propose a decentralized and self-organizing solution framework aimed at solving Optimal Power Flow (OPF) problems in a distributed scenario. In particular we will demonstrate that, under some hypothesis, the solution of the OPF can obtained by computing proper weighted averages of the variable of interests. To compute these global quantities we propose the deployment of a network of dynamic agents solving a distributed average consensus problem. This bio-inspired solution strategy exhibits several advantages over traditional client server-based paradigms as far as less network bandwidth, less computation time, easy to extend and reconfigure are concerned. These features make the overall computing architecture highly scalable, self-organizing and distributed and thus a potential candidate for addressing the economic dispatch analysis in smart grids.
     
    Speaker:
    Alfredo Vaccaro (M?01, SM?09) received the M.Sc. degree with honours in Electronic Engineering in 1998 from the University of Salerno, Salerno, Italy. From 1999 to 2002, he was an Assistant Researcher at the University of Salerno, Department of Electrical and Electronic Engineering. Since March 2002, he has been an Assistant Professor in electric power systems at the Department of Engineering of the University of Sannio, Benevento, Italy. His special fields of interest include soft computing and interval-based method applied to power system analysis and advanced control architectures for diagnostic and protection of distribution networks. Prof. Vaccaro is an Associate Editor and member of the Editorial Boards of IET Renewable Power Generation, the International Journal of Renewable Energy Technology, the International Journal of Reliability and Safety and the International Journal on Power System Optimization. He is the Director of the bureau of the Research Centre on Pure and Applied Mathematic at University of Sannio and the Rector Delegate for Technological Innovations.
  • Credal classification -- Slides

    by: Dr. Giorgio Corani, 
    Dalle Molle Institute for Artificial Intelligence (IDSIA),
    Switzerland
     
    When:
    June 11, 2012 at 11:00 (until 12:00 including questions) 
     
    Where: La Plain Campus, Building NO, 8th floor, Rotule room
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Bayesian networks are important tools for uncertain reasoning in AI. Typically, they are based on a single prior distribution, which is updated through a likelihood yielding a posterior. They are often used for classification. Credal networks generalize Bayesian networks, letting prior probabilities vary in a set (eg., interval). This provides a more realistic model of expert knowledge and returns more robust inferences.  Credal classifiers, being based on a set of priors, can identify prior-dependent instances, in which the most probable class varies with the prior. On such instances, credal classifiers return a set of classes (indeterminate classification) rather than a single one, thus preserving reliability. Extensive experiments show that traditional Bayesian classifiers undergo a severe drop of accuracy on prior-dependent instances, over which instead credal classifiers preserve reliability thanks to indeterminate classifications.
     
    Speaker:
    G. Corani obtains in 2005  the PhD in Information Engineering at Politecnico di Milano. During the PhD he spends a visiting period at the Machine Learning Group of the Université de Bruxelles.  Since 2006 he is researcher at IDSIA, Switzerland. His research interests include probabilistic graphical models, data mining, imprecise probabilities.
  • Cooperative decision-making in cell regulation" 

    by: Kim van Roey, 
    Gibson Team, European Molecular Biology Laboratory (EMBL) Heidelberg, 
    Meyerhofstraße 1, 69117 Heidelberg, Germany
    http://www.embl.de/research/units/scb/gibson/members/?s_personId=CP-60012106
     
    When:
    March 30, 2012 at 12:30 (until 14:00 including questions) 
     
    Where: La Plain Campus, Building NO, 5th floor, Solvay room
    Université Libre de Bruxelles
    Boulevard du Triomphe
    1050 Bruxelles
    Campus access : http://mlg.ulb.ac.be/access
     
    Abstract: 
    Cells must continuously monitor and integrate the variety of signals they perceive in order to generate appropriate responses. This requires reliable and robust signal transduction, which is mediated by an intricate and interlinked network of pathways and processes that are tightly regulated. Assembly of the dynamic macromolecular complexes that modulate these pathways often depends on multiple transient, low-affinity interactions that are context-dependent, highly cooperative and easily tuneable. Such interactions provide the dynamic plasticity that is required for proper cell signalling and underlie the ability of proteins to act as switchable regulatory modules. This raises the question, how do proteins integrate available information to correctly make decisions? This talk addresses the role of intrinsically disordered protein regions, and more specifically short linear motifs, in cooperative decision-making and briefly introduces our current efforts to computationally describe cooperative interactions.
  • Structure, Unstructure and Alternative Splicing

Dr. Philip Kim, The Donnelly Centre for Cellular and Biomolecular Research, Banting and Best Department of Medical Research, Departments of Molecular Genetics and Computer Science University of Toronto 

Wednesday 22 February 2012 from 16:00 to 17:30 (+ questions) - ULB La Plaine campus, Building NO, Salle Solvay (5th floor in the rotule)

Abstract:

Many protein interactions, in particular those in signaling networks, are mediated by peptide recognition domains. These recognize short, linear amino acid stretches on the surface of their cognate partners with high specificity. Residues in these stretches are usually assumed to contribute independently to binding, which has led to a simplified understanding of protein interactions. Conversely, in large binding peptide data sets different residue positions display highly significant correlations for many domains in three distinct families (PDZ, SH3 and WW). These correlation patterns reveal a widespread occurrence of multiple binding specificities and give novel structural insights into protein interactions. For example, a new binding mode of PDZ domains can be predicted and structurally rationalized for DLG1 PDZ1.

While protein structure is very important for peptide binding domains, the regions they bind are usually unstructured (intrinsically disordered). These regions are widespread, especially in proteomes of higher eukaryotes, and have been associated with a plethora of different cellular functions. Aside from general importance for signaling networks, they are also important for such diverse processes as protein folding or DNA binding. Leveraging knowledge from systems biology can help to structure the phenomenon. Strikingly, disorder can be partitioned into three biologically distinct phenomena: regions where disorder is conserved but with quickly evolving amino acid sequences (“flexible disorder”), regions of conserved disorder with also highly conserved amino acid sequence (“constrained disorder”) and, lastly, non-conserved disorder. I will also introduce new efforts to map protein interactions affected by alternative splicing.

  • Modular protein interaction domains - Evolution, Selectivity and Complexity

by Dr. Piers D. Nash The Ben May Department for Cancer Research, The University of Chicago Chicago, IL, USA
Tuesday 14 February 2012 from 11:00 to 12:30 (+ questions) -  room : D.005 (VUB campus, building D, lower floor)
AbstractModular protein interaction domains (PIDs), such as the SH2 domain, are a common feature of many proteins, particularly those involved in cellular signal transduction.  The SH2 domain recognizes phosphotyrosine modified peptide sequences, and in doing so couples tyrosine kinases to downstream signaling networks.  We have examined the evolution of SH2 domains and find that they expand rapidly with the emergence of multicellularity and subsequent expansions concomitant with leaps in organismal complexity within the animal lineage.  Increasing connectivity within and between SH2 proteins may underlie more highly interconnected and robust signal transduction networks. Yet the rapid evolutionary expansion of SH2 domains comes at some cost to selectivity so that the extant SH2 domains explore only a small region of the available peptide ligand sequence space.  The ability of PIDs to nucleate highly selective interactions is essential for signal fidelity yet relies on limited peptide sequence information.  For instance, SH2 domains may appear to have simple binding motifs characterized by a few residues surrounding a phosphotyrosine (eg. pY-X-X-P/L).  We have recently shown that by reading both permissive and non-permissive residues and longer regions of adjacent sequence, the SH2 domain is able to make use of a wider information channel to prescribe selective interactions.  This results in a complex language for SH2 domain-peptide interactions in which the SH2 domain is readily able to distinguish physicochemically similar amino acids. Thus, despite evolutionary constraints, individual SH2 domains have distinct recognition profiles and exhibit a remarkable degree of selectivity.
Speaker: Dr. Piers D. Nash is a world-renowned scientist investigating protein-protein interactions involved in signal transduction, and the molecular mechanisms by which cells respond to external cues.   After completing a postdoctoral position in the lab of Tony Pawson in Toronto, he became Assistant Professor in The Ben May Department for Cancer Research and a Scientist of the Comprehensive Cancer Center at The University of Chicago. His current work focuses on understanding the SH2 domain at a systems level and investigating the role of ubiquitination in controlling endocytosis and modulating signal transduction. 
 
2011
 
  • "Ensemble learning for real-world classification."
    by Nima Hatami, Department of Electrical and Electronic Engineering, University of Cagliari, Italy -- Slides
    Wednesday 18 November 2012 from 12:30 to 13:30 (+questions) - room: NO7.07 - NO building
     
    Abstract: Most real-world classification problems are too complicated to be tackled by a single expert. An alternative approach is to use ensemble of experts inspired by Divide-and-conquer principle which has proven to be efficient in many of these cases. A complex problem is first divided into some simpler sub-problems, each of them assigned to an expert. The final solution of the problem obtained by consensus of experts, is proven to be more effective and efficient. This talk will cover the application of different multiple-classifier systems to some real-world classification problems e.g. gene expression cancer classification, face recognition and text categorization.
     
  • "Representing Cooperative Interactions in Bioinformatics."
    Thursday 22 december 2011 from 12:30 to 13:30 (+questions) - room: NO6.07 - NO building Cancelled
     
    Abstract: Cells must continuously monitor external and internal cues, integrate the variety of signals they perceive, and translate these inputs into proper outputs. This requires reliable and robust signal transduction, which is mediated by intricate and interlinked networks of pathways and processes that are tightly regulated. Assembly of the dynamic macromolecular complexes that modulate these pathways depends on multiple transient, low-affinity interactions, many of which are regulated by post-translational modifications. These distinct binding events are highly cooperative, affecting each other either positively or negatively. Such cooperative interactions provide the dynamic plasticity that is required for proper cell signaling. However, despite the central importance of cooperativity in these systems, it is missing from all current formalisms for describing molecular interactions. This talk addresses our current efforts to computationally describe cooperative interactions.
     
  • "Cartification: from Similarities to Itemset Frequencies."
    by Bart Goethals, Professor, Department of Mathematics and Computer Science, University of Antwerp, Belgium -- Slides
    Thursday 17 November 2011 from 14:30 to 15:30 (+questions) - Rotule NO8 - NO building
     
    Abstract: Suppose we are given a multi-dimensional dataset. For every point in the dataset, we create a transaction, or cart, in which we store the k-nearest neighbors of that point for one of the given dimensions. The resulting collection of carts can then be used to mine frequent itemsets; that is, sets of points that are frequently seen together in some dimensions. Experimentation shows that finding clusters, outliers, cluster centers, or even subspace clustering becomes easy on the cartified dataset using state-of-the-art techniques in mining interesting itemsets.
     
  • "Unraveling networks of co-regulated genes on the sole basis of genomesequences."
    by Sylvain Brohée, Post-doc, ULB -- Slides
    Wednesday 21 September 2011 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building
     
    Abstract: With the growing number of available microbial genome sequences, regulatorysignals can now be revealed as conserved motifs in promoters of orthologousgenes (phylogenetic footprints). A next challenge is to unravel genome-scaleregulatory networks. Using as sole input genome sequences, we predicted cis-regulatory elements for each gene of the yeast Saccharomyces cerevisiae bydiscovering over-represented motifs in the promoters of their orthologs in 19 Saccharomycetes species. We then linked all genes displaying similar motifs intheir promoter regions and inferred a co-regulation network including 56919 links between 3171 genes. Comparison with annotated regulons highlights thehigh predictive value of the method: a majority of the top-scoring predictionscorrespond to already known co-regulations. We also show that this inferrednetwork is as accurate as a co-expression network built from hundreds oftranscriptome microarray experiments. Furthermore, we experimentally validated14 among 16 new functional links between orphan genes and known regulons. Thisapproach can be readily applied to unravel gene regulatory networks fromhundreds of microbial genomes for which no other information is availableexcept the sequence. Long-term benefits can easily be perceived whenconsidering the exponential increase of new genome sequences. 
 
  • "Predicting structured-output from protein sequence"
    by Andrea Passerini, Assistant Professor, Università degli Studi di Trento
    Friday 2 September 2011 from 14:00 to 15:00 (+questions) - NO7.07 - NO building
     
    Abstract: Recent advances in high-throughput sequencing techniques are drastically increasing the amount of biological sequences available for further study. On the other hand, experimentally determining their three-dimensional structure is an expensive and time-consuming process. In this scenario, automatic approaches to sequence analysis are crucial in order to fill this gap and devise information on their biological function. 
    I will present machine learning techniques for predicting protein structural features from sequence. The talk will focus on challenging problems where the desired output is a discrete structure, e.g. a graph connecting certain residues in the sequence. I will first discuss the prediction of disulphide bridges, i.e. covalent bonds between pairs of cysteines, which help stabilizing protein 3D structure and have a relevant structural and functional role. This task can be effectively addressed with a nearest-neighbour approach in the space of candidate configurations. I will then introduce the problem of metal binding site prediction, whose characteristics prevent the application of this method. I will present a search-based structured-output technique relying on an online strategy learning to discriminate between correct and incorrect moves. The advantages and drawbacks of these algorithms will be discussed together to their applicability to other structured-output problems. 
     
  • "Efficient prediction of patterns for context-aware embedded systems"
    by Yves Vanrompay, researcher in the Embedded and Ubiquitous Systems taskforce of the Distrinet research group in the department of computer science of the Katholieke Universiteit Leuven
    Monday 11 April 2011 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building
  • "Machine learning and Web Mining"
    by Doru Tanasa, full time faculty at the International University of Monaco, part-time R&D engineer for Up&Net.
    Thursday 7 April 2011 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building
  • "Beyond Space For Spatial Networks"
    by ARenaud Lambiotte, Imperial College, UK
    Friday 25 February 2011 from 14:00 to 15:00 (+questions) - A2.122 - A building
  • "Automatic Recognition of Multiparty Human Interactions using Dynamic Bayesian Networks"
    by Alfred Dielmann, Research Scientist at the Instiute Telecom ParisTech, Paris, France
    Thursday 24 February 2011 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building
  • "Statistical and relational learning for understanding enzyme function"
    by Elisa Cilia, University of Trento, Italy
    Friday 4 February 2011 from 12:00 to 13:00 (+questions) - Rotule NO8 - NO building
  • "Predictive Network Inference in Colon Cancer"
    by Catharina Olsen, PhD student, MLG, ULB
    Thursday 27 January 2011 from 12:30 to 13:30 (+questions) - NO7.08 - NO building

2010

  • "Automated analysis of biological oscillator models using mode decomposition"
    by Tomasz Konopka, Postdoc, Service de Biosystèmes, Biomodélisation et Bioprocédés (3Bio)
    Thursday 28 October 2010 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building

  • "Statistical issues in the development of clinically useful biomarkers in oncology from microarrays"
    by Stefan Michiels (Bordet, ULB)
    Friday 22 October 2010 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building

  • "Deep Web mining and knowledge mining using machine learning"
    by Lu Jiang (Erasmus Mundus Exchange Program fellowship)
    Friday 01 October 2010 from 15:00 to 16:00 (+questions) - Rotule NO8 - NO building

  • "The Bag-of-Frames approach to music genre classification: Challenges and Limitations"
    by Miguel Lopes (PhD student, Universidade do Porto)
    Wednesday 25 August 2010 from 16:30 to 17:30 (+questions) - Rotule NO8 - NO building

  • "Solving Non-Convex Lasso Type Problems With DC Programming"
    by Romain Herault (INSA de Rouen, France)
    Thursday March 4, 2010, 12:30PM (+questions) - room: NO7.08 - NO building
    Abstract: We propose a novel algorithm for addressing variable selection (or sparsity recovering) problem using non-convex penalties. A generic framework based on a DC programming is presented and yields to an iterative weighted lasso-type problem. We have then showed that many existing approaches for solving such a non-convex problem are particular cases of our algorithm. We also provide some empirical evidence that our algorithm outperforms existing ones. Based on the article:
    G. Gasso, A. Rakotomamonjy, S. Canu, Recovering sparse signals with non-convex penalties and DC programming, IEEE Trans. Signal Processing, Vol 57, no.12, pp 4686-4698, 2009.

  • "Game Tree Search Strategies for Computer Poker"
    by Boris Iolis (ULB graduate student)
    Thursday 14 January 2010 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building

2009

  • "Biomarker selection from microarray data: a transfer learning approach"
    by Pierre Dupont (UCL, Belgium)
    Friday 13 November 2009 from 14:30 to 15:30 (+questions) - Rotule NO8 - NO building

  • "Network Inference based on Information Theory Applied to Microarray Data"
    by Patrick E. Meyer (ULB, Machine Learning Group)
    Thursday 5 November 2009 from 12:30 to 13:30 (+questions) - Rotule NO8 - NO building

  • "The role of cooperative sensor networks in smart grids"
    by Alfredo Vaccaro (U. Sannio, Italy)
    30 April 2009 from 15:30 to 16:30 (+questions) - Rotule NO8 - NO building

2008

  • "Exploratory Analysis of Functional Data via Clustering and Segmentation"
    by Fabrice Rossi (Telecom ParisTech)
    Tuesday 16 December 2008 from 14:00 to 15:00 (+questions) - Rotule NO8 - NO building

  • "Stochastic self-similar processes and large scale structures"
    by Marta Chinnici, PhD (University of Napoli "Federico II")
    Monday 10 November 2008 from 10:30 to 11:30 (+questions) - Salle de séminaire NO8 - NO building

  • "Data mining with SAS"
    by Hadrien Polastro (ULB graduate student)
    Wednesday 1 Ocotber 2008 from 10:00 to 11:30 - Rotule NO8 - NO building

  • Computer Science Department Seminar
    "Distributed Indexing and Querying in Sensor Networks using Statistical Models"
    by Arnab Bhattacharya from Indian Institute of Technology, Kanpur
    Thursday 17 July 2008 from 12:30 @ rotule NO8 (NO building)

  • "q-Nested Partial Correlation Graphs for Genetic Network Inference"
    by Kevin Kontos from ULB/MLG
    Friday 4 July 2008 @ 11:00 - Rotule NO8 - NO building

  • "Proposals for Trustable Visualization of High-Dimensional Data"
    by Abhilash Miranda and Gianluca Bontempi from ULB/MLG
    Tuesday 01 May 2008 @ Indian Institute of Technology, Kanpur

  • "Trends in Dimensionality Reduction for Visualization"
    by Abhilash Miranda from ULB/MLG
    Thursday 13 March 2008 @ 14:30 - Rotule NO8 - NO building

  • "Customer Intelligence and Data Mining"
    by Martine George from ING Belgium
    Tuesday 12 February 2008 @ 12:00 - UB4.136 - Solbosch

  • "An Introduction to Entropy Estimation"
    by Catharina Olsen from ULB/MLG
    Wednesday 6 February 2008 @ 14:30 - Rotule NO8 - NO building

  • "Spatial Data Mining: Exemples d'Application à la Détection/Prédiction du Changement"
    by Hussein Atoui from ULB/MLG
    Tuesday 29 January 2008 @ 14:00 - Rotule NO8 - NO building

  • "Hierarchical Visualization using Mixture of PCAs"
    by Abhilash Miranda from ULB/MLG
    Tuesday 22 January 2008 @ 14:30 - Rotule NO8 - NO building

2004-2007

Theme by Danetsoft and Danang Probo Sayekti inspired by Maksimer