September, 12

15.30

55.309

Invited Research Seminar

By Esteban Maestre, McGill University

Virtual Acoustic Rendering by stave wave synthesis

Abstract: 

We present State Wave Synthesis (SWS), a framework for the efficient rendering of sound traveling waves as exchanged between multiple directional sound sources and multiple directional sound receivers in time-varying conditions. We introduce a mutable state-space system modal 
formulation through which input/output matrices change size and coefficients in time-varying conditions, and perceptually motivated designs are possible. SWS enables the accurate simulation of frequency-dependent source directivity and receiver directivity, provides means for simulating frequency-dependent attenuation of propagating waves, and allows an alternative realization through which state variables are treated as propagating waves.

Biography:

B.Sc. and M.Sc. degrees in Electrical Engineering from Universitat Politècnica de Catalunya, Barcelona, in 2000 and 2003; and the D.E.A. Master and Ph.D. degrees in Computer Science and Digital Communication from Universitat Pompeu Fabra, Barcelona, in 2006 and 2009. 
Lecturer at UPC, Department of Electrical Engineering, in between working as a Junior Researcher at Philips Research Laboratories Aachen.
Researcher (Music Technology Group) and a Lecturer (Department of Information and Communication Technologies) at Universitat Pompeu Fabra, working on diverse topics in Music Technology, 2004-2013.
I spent close to four years at the Center for Computer Research in Music and Acoustics, Stanford University. Also visiting Researcher at the Department of Mathematics, Universidad Federico Santa María, Santiago, Chile.
Through a Marie Curie IOF fellowship, I spent the years between 2014 and 2017 pursuing my research at the Computational Acoustic Modeling Lab / Center for Interdisciplinary Research in Music Media and Technology of McGill University, and also at the Music Technology Group of Universitat Pompeu Fabra.
Since early 2017, I am a Research Associate at the Computational Acoustic Modeling Lab, McGill Univeristy, where I also lecture on Music and Audio Computing.
My research interests include audio digital signal processing, gesture control and numerical schemes for efficient sound synthesis by physical models, computational modeling of music performance, cognitive aspects of sensory-motor integration in music performance, and artificial reverberation.
https://ccrma.stanford.edu/~esteban/bio.html

September, 10

15.00

55.309

Invited Research Seminar

By Caroline Essert

Modelling the surgical knowledge for computer-assisted preoperative planning

Abstract: 

Minimally invasive surgery, and more particularly percutaneous surgical interventions, are very popular alternatives to open surgery. However, the main difficulty with such interventions is the lack of visibility on the surgical sites. A thorough preoperative planning is essential for the best efficacy and safety of the procedure, to select the most optimal tools placement while preserving the surrounding sensitive organs. However, an optimal plan is usually complex to imagine and requires a high expertise. In this talk, we will show how a computer model of the surgical knowledge coupled with patient-specific images is used to assist the surgeon to choose the best intervention strategy before performing the surgery.

Biography:

Caroline Essert received her PhD of Computer Science in 2001 and her Habilitation in 2011 from the University of Strasbourg / ICUBE lab, where she is an associate professor since 2003. Her main research interests are preoperative surgery planning applied to abdominal surgery and neurosurgery, computer-assisted surgery and navigation, virtual reality and haptics for surgery. She obtained a number of national and international grants on that topics, and authored or co-authored more than 50 scientific articles. She initiated the PILOT software project for preoperative surgical planning. She has been involved in the organization of conferences in the fields of medical imaging or computer graphics for many years. In 2019, she is a co-Program Chair for MICCAI, a member of the PC of IPCAI and CARS, and is a member of the organizing committee of CARS in Rennes. In 2014, she co-organized the international conference Eurographics in Strasbourg, and will be the general chair of MICCAI 2021. She is a member of the international boards of the MICCAI Society and the ISCAS Society. At a national level, she has been a committee member of the French National Research Agency (ANR), and she is currently a member of the Computer Science section of the French National Council of Universities (CNU).

Host: Miguel Angel González Ballester

July, 23

12.00

55.410

Invited Research Seminar

By Gabriela Ferraro

Extracting Normative Rules from Legal Texts: Challenges (short presentation)

Laws and regulations govern our daily lives, and we often require expert’s consult to deal with them. Having an automated system capable of interpreting and reasoning about laws and regulations from legal documents would indeed ease the burden required with dealing with the subject. However, a required first step in order to automatise the process, is to extract such laws and regulations in a formalised format that a machine can interpret and reason about. To do so, it is necessary to extract normative rules from sentences written in natural language.
In this talk, I will present some of the NLP related challenges of extracting normative rules from sentences, and explore promising research avenues to approach this problem.

Transfer Learning for Hate Speech Detection in Social Media (long presentation)

In today's society more and more people are connected to the Internet, and its information and communication technologies have become an essential part of our everyday life. Unfortunately, the flip side of this increased connectivity to social media and other online content is cyber-bullying and -hatred, among other harmful and anti-social behaviors. Models based on machine learning and natural language processing provide a way to detect this hate speech in web text in order to make discussion forums and other media and platforms safer. The main difficulty, however, is annotating a sufficiently large number of examples to train these models. In this paper, we report on developing automated text analytics methods, capable of jointly learning a single representation of hate from several smaller, unrelated data sets.
We train and test our methods on the total of 37,520 English tweets that have been annotated for differentiating harmless messages from racist or sexists in the first detection task, and hateful or offensive tweets in the second detection task.
Our most sophisticated method combines a deep neural network architecture with transfer learning. Its prediction correctness is the macro-averaged F1 of 78% and 72%  in the first and second task, respectively. This method enables generating an interpretable two-dimensional text visualization - called the Map of Hate - that is capable of separating different types of hate speech and explaining what makes text harmful. These methods and insights hold a potential for not only safer social media, but also reduced need to expose human moderators and annotators to distressing online-messaging.

Biography:

Gabriela Ferraro is a research scientist in Natural Language Processing and Computational Linguistics at the Commonwealth Science and Industrial Research Organization in Australia, and an Adjunct Research Fellow in the College of Engineering and Computer Science at the Australian National University.
She received a BA in Computer Science from Champagnat University (Argentina) in 2004, and a Master degree in Science of Language and Applied Linguistics from Universitat Pompeu Fabra (Spain) in the following year. In 2006, she joined the TALN Research Group from the same university and received a PhD under the supervision of Leo Wanner in 2012. In 2013, she joined the Machine Learning Research Group at NICTA for four years, and the Australian National University, where she is been a lecturer ever since.

Host: Leo Wanner

July, 17

15.00

Auditorium

By Ulises Cortés (UPC) and  Darío Garcia-Gasulla (BSC)

AI & Health @ High Performance Artificial Intelligence

Abstract:

Artificial intelligence (AI) is having a significant impact on Health care. As a matter of fact, AI in Health care is redefining the medical care field and all its functions. We bring a presentation of the eHealth care-related research topics and challenges that we have been exploring, at the HPAI & KEMLg, along the last years. We present an analysis on how the use of the AI techniques - in particular, Deep Learning- to improve the management of the data generated by the eHealth activity, permitting to take more advanced decisions on the diagnosis, treatment and supervision of the patients. 
We stress the need for responsible use of AI technology, stressing that the potential harms related to the use of AI technology must be transparent to all involved.

Biography:

Dr.Ulises Cortés is a Full-Professor and Researcher of the Technical  University of Catalonia (UPC) since 1982 (tenured since 1988 and habilitated as Full-Professor since 2006) working on several areas of Artificial Intelligence (AI) in the Computer Science (formerly Software Department) including knowledge acquisition for and concept formation in knowledge-based systems, as well as on machine learning and in autonomous intelligent agents. Since 1989 Professor Cortés and his group have been applying their work in Artificial Intelligence to Environmental Sciences in special to Wastewater Treatment Plants with the financial support of CICyT and CIRIT and the European Union. Professor Cortés advised 22 PhD. Thesis, two of them have been awarded with the ECCAI AI Dissertation Award, and more than 20 Master Thesis in the field of Artificial Intelligence and has more than 50 papers in international journals.

Dr. Darío Garcia-Gasulla is a Senior Researcher at BSC, where he has worked since 2015. Before that, he was an assistant researcher for 4 years  at  the  KEMLg  group (UPC),  participating  in  several  research  projects  related to Artificial Intelligence (particularly in Knowledge Representation and Reasoning, Machine Learning and Data Mining). His PhD thesis tackled the large-scale graph-mining problem, bringing together AI and HPC topics. As a Post Doc, he is now leading several research projects in Deep Learning and its internal representations, and Graph Analytics.

Host: Jérôme Noailly

July, 10

15.30

55.309

Invited Research Seminar

Facebook News Feed Integrity

By Lluís Garcia-Pueyo

Abstract: 

Facebook goal with News Feed is to show people the stories that matter most to them, every time they visit Facebook. News Feed is a personalized, ever-changing collection of photos, videos, links, and updates from the friends, family, businesses, and news sources they've connected to on Facebook. People on Facebook value meaningful, informative stories, and accurate and authentic content. Integrity teams within Facebook are responsible to maintain and enforce Community Standards that reflect our collective values for what should and should not be allowed on the platform. In this talk, I'll give an overview of how the News Feed works, and how we enforce Integrity in the News Feed, as well as the challenges in combining maximizing relevance while minimizing bad experiences. 

Biography:

June, 20

15.30

52.S29

PhD Research Seminar 

Impact of machine intelligence in healthcare

By Sergio Sánchez-Martínez

Abstract:

This talk aims at advancing the scientific understanding of machine learning (ML) related to healthcare and at studying the impact of ML algorithms on humans, focusing on clinical decision-making. The talk will be articulated around the essential building blocks to achieving the high-level task of clinical decision-making, namely data acquisition, feature extraction, interpretation and decision support. For each of these blocks, the speaker will provide a concise review of state-of-the-art applications, followed by the challenges still to overcome and the potential benefits of their application in clinical practice. At the end, there will be a discussion on the main problems to tackle when creating algorithms to analyze clinical data and also implementation challenges, such as which interaction paradigms we should use, or the competences medical doctors should have. 

Biography:

Sergio Sánchez-Martínez, Postdoctoral Research Fellow at Institut d’Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS).

June, 18

15.30

55.309

Invited Research Seminar 

Exact Time Series Motif Discovery using a Commercial GPU Cluster: How to 
Execute More than One Quintillion Pairware Comparisons in a Single Day

By Philip Brisk

Abstract 

The ability to discover motifs, conserved (repeated) patterns in time series, is arguably the most important computational primitive in time series data mining. Time series motifs are useful in their own right, as they provide insight to domain scientists about the behavior or physical  phenomena that the time series characterizes. Time series motifs can also be used as inputs into and are also used as inputs into classification, clustering, segmentation, visualization, and anomaly detection algorithms. In recent years, the Matrix Profile has emerged as a promising way to represent highly similar subsequences within a larger time series, and allows the efficient exact computation of the top-k motifs. Many scientific domains, such as astronomy and seismology exhibit an insatiable appetite to consider ever-large time series data 
sets. To meet the needs of scientists, the Matrix Profile can be computed in a rapid and scalable manner by deploying on commercial GPU clusters in the cloud; using this framework, it is possible to achieve throughput as high as one quintillion exact pairwise time series comparisons in a single day. For the first time, it has been possble to perform exact motif discover on more than one year's worth of continuous earthquake data in a single run; this has led to the discovery of what may be subtle precusor earthquakes that have previously escape attention, along with other novel seismic irregularities that domain scientists are presently studying.

Biography

Philip Brisk received the B.S., M.S., and Ph.D., all in Computer Science, from UCLA in 2002, 2003, and 2006 respectively. From 2006-2009, he was a postdoctoral scholar at EPFL in Lausanne, Switzerland. Since 2009, he has been with the University of California, Riverside; he has been promoted to Professor effective July 1, 2019. Dr. Brisk's research interests lie at the intersection between processor architecture, VLSI/CAD, compilers, FPGAs, and reconfigurable computing; most recently, he has been applying these principles to the design and analysis of biological instruments. He is a Senior Member of the ACM and IEEE, and is presently an Associate Editor of the IEEE Transactions of Computer-Aided Design on Integrated Circuits and Systems (TCAD) and Integration: The VLSI Journal. 

June, 17

15.30

55.309

Invited Research Seminar 

Spatial representations and high level descriptors for symbolic music analysis

By Louis Bigo 

Abstract

This talk focus on two computational approaches to symbolic music analysis. The first approach includes the notion of pitch/chord space, that we use to reformulate in spatial terms different musical tasks as style recognition and harmonic transformations. Musical spaces are formalized with topological structures named simplicial complexes. Elementary musical objects (for example pitches or chords) are represented by simplices that are organized according to a neighborhood relationship which translates a musical property, for example consonance. Such spatial formulation of musical structures enables the application of geometrical and topological operations, such as translations, rotations, embeddings and filtrations, that bring original approaches to music composition and analysis.
The second approach focus on the extraction of high level musicological features that can serve automated analysis of classical forms. We will discuss the detection of cadences and the retrieval of section boundaries in sonata form with different machine learning methods.

Biography

I am associate professor (maître de conférence) in computational musicology at University of Lille in the Algomus team (Algorithmic musicology) at CRIStAL laboratory since 2016. My research interests include mathematical models and machine learning for automatic music generation, analysis and classification. I received my Ph.D in Computer Science on the topic of symbolic music representations and spatial computing in 2013 at IRCAM (French Institute for Research and Coordination in Acoustics/Music) and LACL, University Paris 12. I joined the European project Learning To Create on music and machine learning from 2014 to 2016 in the Music Informatics Group at the University of Basque Country in San Sebastian, Spain.

Host: Xavier Serra

June, 12th

15.30

52.217

PhD Research Seminar 

Profiling and automated decision making under the General Data Protection Regulation

By  Antoni Rubí-Puig

Abstract:

The session aims at discussing the rules on profiling and automated individual decision-making under the General Data Protection Regulation and at assessing their adequacy from the standpoint of the Fairness, Accountability, Transparency and Ethics (FATE) framework. Our focus will be on profiling, that is, the automated processing of personal data with the goal of evaluating certain personal aspects relating to an individual, such as for instance analyzing or predicting her performance at work, economic situation, health, personal preferences, interests, willingness to pay, reliability, behavior, location or movements.

Progresses in technology and the use of big data analytics, artificial intelligence and machine learning have made it easier to elaborate profiles. Profiling can be in the advantage of individuals and organizations as it may lead to increased efficiencies and cost savings. However, it also involves significant risks for individuals’ rights and freedoms, not only to privacy. Rules in the GDPR aim at finding a balance between those advantages and risks but the final trade-off may not be the most adequate one for the FATE framework.

Biography:

Antoni Rubí-Puig,  Associate Professor in Civil Law and Associate Director for Research at the Law Department at UPF (research group).

June, 4th

15.30

55.309

PhD Research Seminar

By Simone Tassani

Abstract

The course will start with a brief digression over the several implications that bad statistics have today over the scientific society and why every researcher should know the basic concepts behind a statistical analysis.

Biography

Dr. Simone Tassani is senior researcher of the Multiscale and Computational Biomechanics and Mechanobiology (MBIOMM) team in the SIMBiosys group.

June, 6th

10.30

55.309

Invited Research Seminar 

The value of Aligning Learning Analytics with a Theory-Grounded Learning Design 

By Korah Wiley   

Abstract

Learning analytics can provide teachers with information about students’ engagement with and performance on learning tasks. When these tasks are components of a theory-grounded learning design, the associated analytics have the potential to provide teachers with insight into the learning process. In this talk, I will describe the design of such theory-grounded learning analytics. Additionally, I will present findings on how teachers used these learning analytics to develop instructional interventions. I will also present findings regarding the effect of these interventions on student learning.

Biography

Korah Wiley is a doctoral candidate in the Graduate Group in Science and Mathematics Education (SESAME) program at the University of California-Berkeley and a National Science Foundation CADRE Fellow. Prior to pursuing her doctorate, she earned a Bachelor of Science in Biochemistry, a Master of Science in Molecular Cancer Biology, and taught for over 10 years as a biology instructor at the North Carolina School of Science and Mathematics. Her time teaching inspired her current research interest to support equitable STEM classroom teaching and learning. Currently, in her dissertation work, Korah investigates how to develop learning analytic solutions that promote implementation by teachers and the development of coherent science knowledge in students.

Host: Davínia Hernandez-Leo

May,31

15.30

55.309

PhD Research Seminar

Human behaviour and machine intelligence

By Emilia Gómez

Biography

Emilia Gomez, lead scientist of the HUMAINT project at the Centre for Advanced Studies, Joint Research Centre, European Commission, and head of the MIR (Music Information Research) lab of the Music Technology Group (MTG) at UPF.

May,30

15.30

55.309

PhD Research Seminar

By Simone Tassani

Abstract

The course will start with a brief digression over the several implications that bad statistics have today over the scientific society and why every researcher should know the basic concepts behind a statistical analysis.

Biography

Dr. Simone Tassani is senior researcher of the Multiscale and Computational Biomechanics and Mechanobiology (MBIOMM) team in the SIMBiosys group.

May,29

15.30

55.309

Multi-modal hearing devices and their evaluation

By Volker Hohmann

Abstract

Spatial filtering and decomposition of sounds into acoustic source objects is increasingly investigated for speech enhancement in hearing aids. However, with increasing performance and availability of these "space aware" hearing aid algorithms, knowledge of the user’s personal listening preferences and knowledge of the attended source becomes crucial.

Here we present an algorithm, which combines information of the subject’s eye gaze with an acoustic analysis of the sound source positions to identify the attended source(s) from a mixture of sources. Gaze direction is recorded by electrooculography (EOG) combined with a head tracking system, which could in principle be integrated in hearing aids. The spatio-temporal distribution of source positions is estimated from input signals to a binaural hearing device. Beamformer techniques are then used to enhance the source (or the sources) that are estimated as being attended. By using (electro-)physiologic sensor signals to estimate attention and behavior, the hearing device becomes part of the active communication loop between environment and behaving subject.

Such closed-loop hearing devices require “subject-in-the-loop” evaluation methods, which simulate realistic interactive environments to reproduce natural behavior. This means that audio-visual signals are required that also include behavioral simulation of the sources, e.g., simulated lip movement, or simulated conversational gaze behavior. The current status of our developments towards such an audiovisual interactive lab are presented in this talk.

Project funded by DFG SFB 1330 project B1 and EU H2020 MSCA ITN No
675324 (ENRICH).

Biography

Volker Hohmann received the Physics degree (Dipl.-Phys.) and the doctorate degree in Physics (Dr. rer. nat.) from the University of Göttingen, Germany, in 1989 and 1993. From 1993-2012 he was a faculty member of the Physics Institute, Oldenburg University, Germany, and was appointed full professor at the Faculty for Medicine and Health Sciences at Oldenburg University in 2012. His research expertise is in acoustics and digital signal processing with applications to signal processing in speech processing devices, e.g., hearing aids. He is a consultant with the Hörzentrum Oldenburg GmbH and with HörTech gGmbH Oldenburg. He was a Guest Researcher at Boston University, Boston, MA, (Prof. Dr. Colburn) in 2000 and at the Technical University of Catalonia, Barcelona, Spain (Prof. Dr. C. Nadeu) in 2008. Prof. Hohmann received the Lothar-Cremer prize of the German acoustical society (DEGA) in 2008 and the German President’s Award for Technology and Innovation in 2012.

Host: Federico Sukno

May,23

15.30

55.309

PhD Research Seminar

Automation of personal data and consumer law enforcement using AI

By Francesca Lagioia

Abstract

Despite the European Consumer Law and EU General Data Protection Regulation (GDPR) are in place, and despite enforcers’ competence for abstract control, Terms of Services and Privacy Policies of online services  still often fail to comply with regulations. Artificial intelligence and in particular machine learning methods can be used for automating the legal evaluation of both terms of services and privacy policies, in order to empower the civil society representing the interests of consumers. 

Biography

Francesca Lagioia, postdoctoral research fellow at Interdepartmental Centre for Research in the History, Philosophy, and Sociology of Law and in Computer Science and Law (CIRSFID) at the University of Bologna, and Research Associate at the European University Institute (EUI)

Invited Research Seminar 

Medical image analysis, artificial intelligence & tech transfer

By Sir Michael Brady

Host: Miguel Angel González Ballester

Abstract

After a very brief summary of my research career, I discuss several topics in medical image analysis and their transfer to clinical practice, via a number of start-up companies that I have founded. My work has always been driven by clinical applications. It has also been guided by a number of principles: (i) the huge variation in the appearances of medical images, with the need for algorithms in clinical practice to work reliably and robustly implies modelling the physics of image acquisition; (ii) AI, including machine learning, can play a valuable role in encoding and mobilising prior knowledge about medical images; (iii) variation in signal characteristics necessitates the definition of what constitutes an image “feature”; and (iv) since information is intrinsically uncertain, representations need to be developed of probability density functions. My talk illustrates these points with particular reference to breast cancer (mammography) and to liver disease (MRI).

Biography

Professor Sir Michael Brady, FRS FREng FMedSci HonFIET FInstP FBCS PhD, holds BSc, MSc and PhD degrees in mathematics from the University of Manchester and the Australian National University in Canberra. After 10 years as senior lecturer at the University of Essex, Prof. Brady joined the Artificial Intelligence Lab at MIT (1980-1985), where he developed fundamental contributions to artificial intelligence, mobile robotics and computer vision. He then became a Professor at the Department of Engineering Science in Oxford, of which he became director, and later joined the Department of Oncology, where he is Professor of Oncological Imaging. Prof. Brady has supervised 115 PhD students, published over 750 articles and filed 26 patents. He holds 9 honorary doctorates and important distinctions, including the IEEE Third Millenium Medal and the Henry Dale Prize from the Royal Institution. He is a member of the Royal Society and the Academie des Sciénces, amongst other prestigious institutions, and was knighted as Sir Brady in 2003. In addition to academic work, he has founded numerous startups, including Perspectum Diagnostics, Volpara Solutions, Mirada Medical Limited, ScreenPoint, Optellum, and Guidance Navigation Holdings, and was director of ISIS Innovation in Oxford from 1993 until 2011.

15.30 h

55.003

Invited Research Seminar 

PET/MR: GE's perspective 

By Gaspar Delso

Abstract

The combination of MR and PET data has received increasing attention in recent years, in fields such as clinical oncology and neuroimaging research. In comparison to currently used PET/CT scanners, PET/MR scanners offer improved soft-tissue contrast, reduced ionizing radiation and a wider range of contrast mechanisms, such as functional, spectroscopic and diffusion tensor imaging. The design of integrated PET and MR detectors has required a remarkable engineering effort, due to several compatibility and interference challenges, culminating in the commercialization and world-wide adoption of clinical PET/MR scanners. In this talk, we review the SIGNA PET/MR scanner, briefly describing its technical design, current performance and ongoing R&D work.

Biography

Gaspar Delso is a Biomedical Engineering PhD with a long experience in medical imaging, having published on key topics such as hardware attenuation, truncation correction and system performance. After completing his Thesis on multimodal non-rigid registration at the Universitat Politècnica de Catalunya and École Nationale Supérieure des Télécommunications in 2016, he accepted a scientific position with Philips Medical Systems, where he worked in the development of novel denoising algorithms for three-dimensional medical imaging. In 2007 he joined the Technische Universität München, where he collaborated with Siemens in the development of the first clinical PET/MR scanner. His paper describing the performance evaluation of this system has been referenced in over 600 publications. In 2012 he moved to Zurich, where he worked for GE Healthcare at the UniversitätsSpital Zürich in the development of the next generation of clinical PET/MR. In 2016, he accepted a Senior Scientist position with GE Healthcare at the University of Cambridge, collaborating with the UK Dementias Platform (DPUK) and developing the next generation of attenuation correction software for the SIGNA PET/MR system. He has recently moved to Barcelona, where he is collaborating with the Hospital Clínic in the development of novel cardiac MR applications. Dr. Delso has co-authored over 200 scientific publications in peer-reviewed journals and conferences, as well as several patents and book chapters.

Host: Jérôme Noailly

15.30 h

Auditorium

Invited Research Seminar

The Fourth Phase of Water: Beyond Solid, Liquid, and Vapor

By Gerald H.Pollack

Abstract

Everyone knows that water has three phases: solid, liquid and vapor. But we have recently uncovered a fourth phase. This phase occurs next to water-loving (hydrophilic) surfaces. It is surprisingly extensive, projecting out from the surface by up to millions of molecular layers. And, its properties differ markedly from those of bulk water. Of particular significance is the observation that this fourth phase is charged; and, the water just beyond is oppositely charged, creating a battery that can produce electrical current. We found that light charges this battery. Thus, water can receive and process electromagnetic energy drawn from the environment in much the same way as plants. Absorbed electromagnetic (light) energy can then be exploited for performing work, including electrical and mechanical work. Recent experiments confirm the reality of such energy conversion. This energy-conversion framework seems rich with implication. Not only does it provide an understanding of how water processes solar and other energies, but also it may provide a foundation for simpler understanding natural phenomena ranging from weather and green energy all the way to issues such as the origin of life, transport, and osmosis. The talk will present evidence for the existence of this novel phase of water — how come nobody’s seen it before? — and will consider the potentially broad implications of this phase for health. The book dealing with this subject is now available http://www.amazon.com/The-Fourth-Phase-Water-Beyond/product-reviews/0962689548

Biography

Gerald Pollack received his PhD in biomedical engineering from the University of Pennsylvania in 1968. He then joined the University of Washington faculty and is now professor of Bioengineering. He is also Founding Editor-in-Chief of the journal, WATER, convener of the Annual Conference on the Physics, Chemistry and Biology of Water, and Executive Director of the Institute for Venture Science. His interests have ranged broadly, from biological motion and cell biology to the interaction of biological surfaces with aqueous solutions. His 1990 book, Muscles and Molecules: Uncovering the Principles of Biological Motion, won an “Excellence Award” from the Society for Technical Communication. His 2001 book, Cells, Gels and the Engines of Life, and his newest book, The Fourth Phase of Water: Beyond Solid, Liquid, and Vapor won that Society’s “Distinguished Award,” their highest distinction. The latter book went on to receive the World Summit Excellence Award. Pollack received an honorary doctorate in 2002 from Ural State University in Ekaterinburg, Russia, and was more recently named an Honorary Professor of the Russian Academy of Sciences, and foreign member and Academician of the Srpska Academy. He received the Biomedical Engineering Society’s Distinguished Lecturer Award in 2002. In 2008, his colleagues chose him as the recipient of his university’s highest annual distinction: the UW Faculty Lecturer Award. Pollack is a Founding Fellow of the American Institute of Medical and Biological Engineering and a Fellow of both the American Heart Association and the Biomedical Engineering Society. He received an NIH Director’s Transformative R01 Award. He was the 2012 recipient of the Prigogine Medal for thermodynamics of dissipative systems, and in 2014 he received the Scientific Excellence Award from the World Academy of Neural Therapy, as well as the Dinsdale Prize from the Society for Scientific Exploration. He has presented two TEDx talks on water. In 2015, he won the Brandlaureate Award, previously bestowed on notables such as Nelson Mandela, Hillary Clinton and Steve Jobs. In 2016 he was awarded the 1st Emoto Peace Prize. And, he appears briefly in the 2016 Travis Rice sports-action film, The Fourth Phase, named after his recent book.

Host: Jérôme Noailly

10:30 h

Room 55.309

Invited Research Seminar 

Network based approaches for drug discovery in osteoarthritis  

By Michael Neidlin

Abstract

In order to obtain a mechanistic understanding of complex multivariate diseases, various experimental and numerical techniques from systems biology can be used. More specifically, pathway and network based approaches provide an interesting option to unravel molecular mechanisms and pinpoint pharmaceutical treatment strategies using whole-genome sequencing data. In this talk co-expression network analysis and network based drug discovery methods will be presented and applied to osteoarthritis (OA), a complex multi-tissue disease with high prevalence in the western world.

Biography

Michael Neidlin is a postdoctoral researcher from the Biomedical Systems Laboratory at the National Technical University of Athens. He graduated as a mechanical engineering from RWTH Aachen University in Germany where he wrote his dissertation at the Institute of Applied Medical Engineering in 2016. During his work he is focusing on the biomechanics and systems biology of cartilage degradation in order to leverage drug discovery for osteoarthritis. He is interested in applying engineering approaches to in vitro experiments in order to better understand the dynamics in biological systems.

 Host: Jérôme Noailly 

15:30 h

Room 55.309

Invited Research Seminar

Compositionality and Automated Hierarchical Skill Discovery Using KL Control

By Andrew Saxe

Abstract

Hierarchical architectures are critical to the scalability of reinforcement learning methods. Most current hierarchical frameworks execute actions serially, with macro-actions comprising sequences of primitive actions. We propose an alternative to these control hierarchies based on concurrent execution of many actions in parallel. Our scheme exploits the guaranteed concurrent compositionality provided by the linearly solvable Markov decision process (LMDP) framework, which naturally enables a learning agent to draw on several macro-actions simultaneously to solve new tasks. We introduce the Multitask LMDP module, which maintains a parallel distributed representation of tasks and may be stacked to form deep hierarchies abstracted in space and time. Next we turn to the problem of learning an appropriate hierarchical decomposition of a domain into subtasks. In the compositional setting afforded by the LMDP, the subtask discovery problem can be posed as finding an optimal low-rank approximation of the set of tasks the agent will face in a domain. We use non-negative matrix factorization to discover this minimal basis set of tasks, and show that the technique learns intuitive multilevel hierarchical decompositions in a variety of domains. Our method has several qualitatively desirable features: it is not limited to learning subtasks with single goal states, instead learning distributed patterns of preferred states; it learns qualitatively different hierarchical decompositions in the same domain depending on the ensemble of tasks the agent will face; and it may be straightforwardly iterated to obtain genuinely deep hierarchical decompositions.
Joint work with Adam Earle and Benjamin Rosman.

Biography

Dr. Andrew Saxe is a Postdoctoral Research Associate in the Department of Experimental Psychology, University of Oxford working with Christopher Summerfield and Tim Behrens. He was previously a Swartz Fellow at Harvard University with Haim Sompolinsky. He completed his PhD in Electrical Engineering at Stanford University, advised by Jay McClelland, Surya Ganguli, Andrew Ng, and Christoph Schreiner. His dissertation received the Robert J. Glushko Dissertation Prize from the Cognitive Science Society. His research focuses on the theory of deep learning and its applications to phenomena in neuroscience and psychology.

Host: Vicenç Gómez

15:30 h

Room 55.003

Invited Research Seminar

Divergence based motivation for online EM and for combining hidden variable models

By Manfred Warmuth

Abstract

Expectation-Maximization (EM) is the fall back method for parameter estimation of hidden (or latent) variable models. In each iteration the EM algorithm forms an upper-bound for the negative log-likelihood of the data and then updates to the minimizer of this upper bound. While EM is naturally a batch algorithm, online variants are strongly desirable when processing data streams or large datasets. We introduce a versatile online variant of EM. Our motivation is based on the relative entropy divergences between two joint distributions over the hidden and visible variables. We show that for the typical mixture models where EM is applied such as mixtures of exponential families, Hidden Markov Models, Kalman Filters, the joint density is again an exponential family distribution. Therefore the EM upper-bound is essentially a negative log likelihood of the joint and it is natural to use the divergence between the joints between old and new parameters as an inertia term for motivating our online EM update.

The resulting update is more widely applicable than previous versions. In particular it leads to an online update for Kalman filters. Additionally, the finite sample form of the inertia term provides a framework to apply the online EM algorithm when the updates do not have a closed form. Experimentally, sweeping the data with an online update converges much faster than the vanilla batch update. Our divergence based methods also lead to a simple way to combine hidden variable models and this immediately gives efficient algorithms for the distributed setting.

Joint work with Ehsan Amid, UC Santa Cruz.

Host: Gergely Neu

PhD Research Seminar

Workshop: Getting your PhD but keeping your sanity

By Scientists Dating Forum.

As part of the PhD seminars and as a complement to the activities during the 7th PhD workshop (April 4th) this workshop will be offered by Scientists Dating Forum. Previous registration is required.
 
Abstract

Recent studies suggest that PhD students suffer a mental health issues at alarming levels—around six times the level of the general population—issues which impact not only productivity and program outcomes but also the long-term well-being of the developing researcher.

Research centers have begun to address the issue, but it resists a simple policy solution. Indeed, the issue of mental health and well-being in PhD programs has the characteristics of a “wicked problem”. For example, it lacks a definitive solution, every problem is unique, it can be explained in a number of ways, and it tightly intertwined with a number of interconnected systems, also experiencing issues.

In light of these challenges, this workshop takes a collaborative, action-oriented approach to generate awareness and empower PhD students. Based on methods developed in psychology, organizational studies, and design thinking, it provides a means of understanding and addressing the challenges PhD students face in a way that allows for the variety of forms these are likely to take.

Participants will gain increased awareness of the complexity of the issue and propose their own solution in a hands-on, supportive setting.

Invited Research Seminar

Data processing under uncertainty and tractability limitations 

By Santiago Mazuelas

Invited Research Seminar

Learning of musical structures for automatic improvisation systems 

By Ken Déguernel

Abstract

Invited Research Seminar

Design for effective orchestration and pedagogical interventions aligned with learning analytics in technology-enhanced learning ecosystems: Notes of a sabbatical year

By Yannis Dimitriadis

PhD Research Seminar

Sounds of Science: share your ideas with sounds

Invited Research Seminar

Shaping the City with Data Science

By Eduardo Graells-Garrido

Abstract

Cities are growing at a fast rate, and transportation networks need to adapt accordingly. To design, plan, and manage transportation networks, domain experts need data that reflect how people move from one place to another, at what times, for what purpose, and in what mode(s) of transportation. However, traditional data collection methods are not cost-effective or timely. For instance, travel surveys are very expensive, collected every ten years, a period of time that does not cope with quick city changes, and using a relatively small sample of people. Considering this context, In this talk I give an overview of how we can infer patterns of transportation and mobility in a city by analyzing non-traditional data sources. I focus mainly on Mobile Phone Data, but I also mention Social Media and Volunteered Geographic Information. As result, the audience will learn about individual- and city-level insights related to transportation, mobility, and visualization. http://bit.ly/ShapingTheCity

Biography

Eduardo Graells-Garrido: Profesor Investigador en el Instituto de Data Science de la Universidad del Desarrollo, Facultad de Ingeniería, y fellow de Telefónica I+D Chile. Obtuvo su doctorado en la Universitat Pompeu Fabra de Barcelona, a través de una estadía en los laboratorios de Yahoo! Research. En su tesis investigó los sesgos de comportamiento utilizando computación centrada en las personas y análisis de redes sociales. Sus temas de investigación actuales son informática urbana, transporte computacional, y ciencia social computacional.

Host: Carlos Castillo

PhD Research Seminar

Statistical course and Design of Experiments 

By  Simone Tassani

Abstract

Tutorial - Data driven-text simplification

By  Sanja Štajner, Horacio Saggion

Abstract

In this tutorial, we aim to provide an extensive overview of automatic text simplification systems proposed so far, the methods they used and discuss the strengths and shortcomings of each of them, providing direct comparison of their outputs. We aim to break some common misconceptions about what text simplification is and what it is not, and how much it has in common with text summarisation and machine translation. We believe that deeper understanding of initial motivations, and an in-depth analysis of existing TS methods would help researchers new to ATS propose even better systems, bringing fresh ideas from other related NLP areas. We will describe and explain all the most influential methods used for automatic simplification of texts so far, with the emphasis on their strengths and weaknesses noticed in a direct comparison of systems outputs. We will present all the existing resources for TS for various languages, including parallel manually produced TS corpora, comparable automatically aligned TS corpora, paraphrase- and synonym- resources, TS-specific sentence-alignment tools, and several TS evaluation resources. Finally, we will discuss the existing evaluation methodologies for TS, and necessary conditions for using each of them.

Previous registration necessary: https://www.upf.edu/web/mdm-dtic/tutorial-data-driven-text-simplification

Host: Aurelio Ruiz

Invited Research Seminar

Entity-centric information access for high-end semantic applications

By  Simone Paolo Ponzetto

Abstract

In the past years a great deal of work across different communities, ranging from Natural Language Processing (NLP) and Information Retrieval (IR) all the way through the Semantic Web (SW) and Digital Libraries (DL), has focused on entities as central tools to build semantic models of ever increasing complexity. Entity Linking, for instance, has arguably brought `classic' sense disambiguation tasks, historically confined to the domain of computational semantic analysis, to a much larger audience, including real-world applications like semantic search, producing as a consequence full-fledged, entity-aware and entity-centric document search models.

All in all, this is good news, since these lines of research all seem to point out the benefits of text understanding for high-end applications. Even more important is the fact that work on entity-centric models prompts, perhaps even requires, lines of research from related, yet far apart research communities like NLP, IR, SW and DL to synergistically come together.

In this talk, I will elaborate on this topic of entity-centric information access and explore a variety of different high-end applications that leverage entities to make sense and bring order within large datasets of textual data. These include ranking entities for Web search queries, building event-centric corpora from Web archives, as well as using entity-aspect linking to predict relevant entities for user queries, events and tweets. 

Biography

Simone Paolo Ponzetto is Professor of Information Systems at the University of Mannheim and member of the Data and Web Science Group, where he leads the Natural Language Processing and Information Retrieval group. His main research interests lie in the areas of knowledge acquisition, text understanding, and the application of natural language processing methods for research in the digital humanities and computational social sciences.

Host: Horacio Saggion

PhD Research Seminar

Statistical course and Design of Experiments 

By  Simone Tassani

Abstract

Invited Research Seminar

MobInsight - Approach to Interpretation of Urban Dynamics through Crowdsourced Local Knowledge

By Souneil Park

Abstract

Collective urban mobility embodies the residents' local insights on the city. Mobility practices of the residents are produced from their spatial choices, which involve various considerations such as the atmosphere of destinations, distance, past experiences, and preferences. The advances in mobile computing and the rise of geo-social platforms have provided the means for capturing the mobility practices; however, interpreting the residents' insights is challenging due to the scale and complexity of an urban environment, and its unique context. In this talk, I will present MobInsight, a platform for localized interpretations of urban mobility that reflect various aspects of the urbanism. MobInsight provides a scalable solution for localized mobility interpretation; first, by extracxting urban contextual knowledge from diverse online sources such as geo-social media, open government data, second, by developing tailored machine learning and visualization techniques which augments the urban contextual knowledge with large-scale mobility data. The talk will mainly go through the techniques in detail, but also cover various examples and applications. 

Biography

Souneil Park is currently at Telefonica Research, working in the areas of social computing, HCI, and information visualization. His recent research interest is in expanding our understanding of the urban dynamics, by applying NLP/AI techniques (used for semantic analysis) to the data that does not explicitly carry semantics, such as mobility and sensory data. Prior to joining Telefonica, he worked on analyzing online media use and developing technologies to enhance people's ability to comprehensively understand, reason, and make well-informed decisions. 

Host: Carlos Castillo

Invited Research Seminar

The Hearing Device of the Future

By Waldo Nogueira

Abstract

Hearing is the ability to perceive sounds by detecting vibrations and is one of the most important senses, as humans rely on speech to communicate in daily life. Recent advances in engineering, computing and neuroscience are contributing to novel approaches to recover the hearing sense based on devices such as hearing aids, cochlear implants and central auditory prostheses including the auditory nerve implant and the auditory midbrain implant. These new approaches are based on stimulating the hearing system at different stages, from the periphery to the central auditory system, recovering binaural hearing with cochlear implants and the use of closed-loop systems that record cortical activity and use it to actively control the signal processing of the hearing device.

During the seminar I will present novel contributions in the field of signal processing, sound coding and electrophysiology within the common topic of designing the hearing device of the future.

Biography

Waldo Nogueira received his Dipl.-Ing. and Dr.-Ing. degree from the Polytechnic University of Catalonia and the Leibniz University of Hannover in 2003 and 2008, respectively. In 2008 he joined the R&D labs of Advanced Bionics in Belgium and Germany. In 2011 he was a Post-Doc at the Pompeu Fabra University in Barcelona. Since 2013 he is Junior Professor at the Hannover Medical School within the Cluster of Excellence Hearing4all.

Invited Research Seminar

Learning, Representations, and Planning in AI

By Blai Bonet

Abstract

Planning, or the use of predictive models for generating goal-directed behaviors is a key capability of flexible, intelligent agents.The research challenges in planning are of two types. One is algorithmic:how to plan is presence of complex models with many variables.The other is representational: how to learn the models from experience. In this talk, I will review these central problems in AI, from the perspective of generalized planning, where plans represent general  strategies for solving a whole family of problem, and which connects with current work in learning. I'll look at the formulation of generalized planning, at the representation challenge that it raises, and at the structural learning methods that are being developed.

Biography

Blai Bonet is a professor in the computer science department at Universidad Simón Bolívar, Venezuela. He received a Ph.D. degree in computer science from the University of California, Los Angeles. His research interests are in the
areas of automated planning, heuristic search, and knowledge representation. Blai has received two ICAPS Influential Paper Awards for pioneering work on automated planning, and other best paper awards. He is a co-author (with Hector Geffner) of the book titled "A Concise Introduction to Models and Methods for Automated Planning." Blai Bonet is an Associate Editor of Artificial Intelligence and the Journal of Artificial Intelligence Research. He served as Conference co-Chair of ICAPS-12 and as Program co-Chair of AAAI-15, and he is a member of the AAAI Executive Council.

Host: Ralph Andrzejak

PhD Research Seminar

Statistical course and Design of Experiments 

By  Simone Tassani

Abstract

The course of statistics aims to introduce a number of tools for master/Ph.D. students and post docs. The presented tools will play a role in planning many kind of studies, properly analyse the results and understand if data analysed by other researchers are or not reliable.

Invited Research Seminar

A Theory of Regularized Markov Decision Processes 

by Matthieu Geist

Abstract

Many recent successful (deep) reinforcement learning algorithms make use of regularization, generally based on entropy or on Kullback-Leibler divergence. We propose a general theory of regularized Markov Decision Processes that generalizes these approaches in two directions: we consider a larger class of regularizers, and we consider the general modified policy iteration approach, encompassing both policy iteration and value iteration. The core building blocks of this theory are a notion of regularized Bellman operator and the Legendre-Fenchel transform, a classical tool of convex optimization. This approach allows for error propagation analyses of general algorithmic schemes of which (possibly variants of) classical algorithms such as Trust Region Policy Optimization, Soft Q-learning, Stochastic Actor Critic or Dynamic Policy Programming are special cases. This also draws connections to proximal convex optimization, especially to Mirror Descent.

Biography

Matthieu Geist obtained an Electrical Engineering degree and an MSc degree in Applied Mathematics in Sept. 2006 (Supélec, France), a PhD degree in Applied Mathematics in Nov. 2009 (University Paul Verlaine of Metz, France) and a Habilitation degree in Feb. 2016 (University Lille 1, France). Between Feb. 2010 and Sept. 2017, he was an assistant professor at CentraleSupélec, France. In Sept. 2017, he joined University of Lorraine, France, as a full professor in Applied Mathematics (Interdisciplinary Laboratory for Continental Environments, CNRS-UL). Since Sept. 2018, he is on secondment at Google Brain, as a research scientist (Paris, France). His research interests include machine learning, especially reinforcement learning and imitation learning.

Host: Gergely Neu

Invited Research Seminar

Deep Reinforcement Learning with demonstrations  

by Olivier Pietquin

Abstract

Deep Reinforcement Learning (DRL) has recently experienced increasing interest after its success at playing video games such as Atari, DotA or Starcraft as well as defeating grand masters at Go and Chess. However, many tasks remain hard to solve with DRL, even given almost unlimited compute power and simulation time. These tasks often share the common problem of being "hard exploration tasks". In this talk, we will show how using demonstrations (even sub-optimal) can help in learning policies that can reach human level performance or even super-human performance on some of these tasks, especially the remaining unsolved Atari games or human-machine dialogues.

Biography

Olivier Pietquin obtained an Electrical Engineering degree from the Faculty of Engineering, Mons (FPMs, Belgium) in June 1999 and a PhD degree in April 2004. In 2011, he received the Habilitation à Diriger des Recherches (French Tenure) from the University Paul Sabatier (Toulouse, France). Between 2005-2013, he was a professor at the Ecole Superieure d'Electricite (Supelec, France), and subsequently joined the University Lille 1 as a full professor in 2013. In 2014, he has been appointed at the Institut Universitaire de France as a junior fellow. He is now on leave with Google, first at Google DeepMind in London and, since 2018, with Brain in Paris. His research interests include spoken dialog systems evaluation, simulation and automatic optimisation, machine learning (especially direct and inverse reinforcement learning), speech and signal processing. Title: Deep Reinforcement Learning with demonstrations Abstract: Deep Reinforcement Learning (DRL) has recently experienced increasing interest after its success at playing video games such as Atari, DotA or Starcraft as well as defeating grand masters at Go and Chess. However, many tasks remain hard to solve with DRL, even given almost unlimited compute power and simulation time. These tasks often share the common problem of being "hard exploration tasks". In this talk, we will show how using demonstrations (even sub-optimal) can help in learning policies that can reach human level performance or even super-human performance on some of these tasks, especially the remaining unsolved Atari games or human-machine dialogues. Web: https://ai.google/research/people/105812 http://www.lifl.fr/~pietquin/

Host: Gergely Neu

PhD Research Seminar

Facial Analysis for Emotions, Interaction and Beyond

By Federico Sukno

Abstract

In this talk I will present our research in facial analysis over the last 4 years. I will start by motivating the interest in facial analysis for diverse applications, briefly covering the more traditional ones related to identity recognition for law enforcement and to the automatic recognition of facial expressions. The latter has become especially relevant in the last few years, given its importance for the understanding of human behavior and for advanced human-computer interaction. This will be showcased from specific research work done in our group in automatic head pose estimation, emotion recognition (both from discrete and dimensional approaches) and automatic lip reading.

In the second part of the talk, I will discuss other emerging applications that go beyond the traditional analysis of identity and expressions to target extracting more subtle information from the face. Some of this information, however, might be not apparent or it might even be hidden to us, and it could only be recovered by means of specialized techniques. This will be showcased from applications in photoplethysmography, facial asymmetries, craniofacial dysmorphologies and human factors.

Biography

Dr Federico Sukno is a Ramón y Cajal Fellow at UPF. He received the degree in electrical engineering at La Plata National University (Argentina, 2000) and the Ph.D. degree in biomedical engineering at Zaragoza University (Spain, 2008). His research activity has been framed in the field of image analysis with statistical models of shape and appearance, targeting diverse applications, most of which related to facial analysis. He is the author or co-author of more than 60 peer-reviewed publications, including 21 journal publications (from which 16 are Q1 journals and 10 are in the top-ranked journals in the fields of artificial intelligence and medical imaging, e.g. IEEE T Pattern Anal, IEEE T Med Imaging, IEEE T Image Process, IEEE T Cybernetics, Int J Comput Vision, Med Image Anal, Pattern Recogn). He has participated in 4 national and 5 international research projects (FP6, FP7, H2020 and Welcome Trust), as well as in 5 technology transfer projects acting as coordinator or PI in several of them. He has been a Marie Curie and a Ramon & Cajal fellowships in 2012 and 2015, which constitute two highly-competitive and prestigious individual grants in the EU and the Spanish systems, respectively.

PhD Research Seminar

euCanSHare: Bypassing the long and winding road towards “true” big data in biology, medicine and beyond

By Karim Lekadir

Abstract

The big data revolution continues to have a transformative effect on research and innovation in a wide range of scientific and societal domains. In computer vision, for example, databases such as ImageNet now include tens of millions of images from tens of thousands of semantic categories, leading every year to important methodological advances and technological applications. However, in other domains such as in biomedicine, the promise of big data is faced with ethical/legal, operational and financial constraints, which have made it a very hard challenge to establish large-scale research databases covering multiple data types and populations. In this talk, I will first present the euCanSHare H2020 project, which aims at addressing the lack of large heterogeneous databases (including biological, imaging and clinical data), by developing the information technology tools and the data science algorithms that will enable to integrate and co-analyse multiple smaller databases, thus totalling an unprecedented 1,000,000 records. I will also describe the computational challenges and investigated solutions to enable automated and robust large-scale, multi-type and multi-cohort data analysis like never before. I will list emerging opportunities that the euCanSHare project will offer for personalised medicine and translational research. I will conclude with future perspectives in integrative data science at UPF for bypassing the long and winding road towards “true” big data in biology, medicine and beyond.

Biography

Dr Karim Lekadir is a Ramon y Cajal researcher at the Barcelona Centre for New Medical Technologies, Universitat Pompeu Fabra, Barcelona. He received a PhD in Computing from Imperial College London and was a postdoctoral researcher at Stanford University, USA. His algorithm developed during his PhD for cardiac functional quantification has been FDA/CE marked and is used in more than 250 clinical centres worldwide. He participated in several EU projects in the field of computational biomedicine, including the euHeart project for computational modelling of personalised interventions in cardiology. Through his work on statistical shape modelling using partial least squares, he finished in the first position of the MICCAI 2015 Challenge on myocardial infarct classification. His current research focuses on the development of data science, machine learning and image computing approaches for the integrative analysis of large-scale biomedical data. He is currently the Project Coordinator of the euCanSHare H2020 project (2018-2022) funded by the European Commission (6 million Euros), leading a consortium of 16 institutions to address data sharing and big data approaches in cardiovascular personalised medicine. He is an Associate Editor of the IEEE Transactions on Medical Imaging and a Guest Associate Editor on the Frontiers Special Issue on Artificial Intelligence and Cardiac Imaging.

PhD Research Seminar

Ethics working with social media

By Carlos Castillo

Abstract

Invited Research Seminar

Music Recommendation - Current state-of-the-art and challenges

By Fabien Gouyon

Abstract

In this talk I will present an overview of recommendation systems in the music streaming industry. We will go over the practical case of playlist generation and station continuation. I'll also highlight what I think are interesting challenges in this field. Let's make it interactive and discuss your own questions!

Biography

Fabien Gouyon leads Pandora’s data science research team in Europe, working on Music Information Retrieval, Recommender Systems, and Natural Language Processing, and applying research to personalized music recommendation. Before joining Pandora, he received a PhD in Computer Science from Universitat Pompeu Fabra and was a co-founder of Barcelona Music and Audio Technologies (BMAT), worked in the Austrian Research Institute for Artificial Intelligence in Vienna, and started and led the Sound and Music Computing Group while teaching at the University of Porto.

Host: Xavier Serra

PhD Research Seminar

Paradigmatic Research Challenges in IoT Systems Engineering 

By  Schahram Dustdar

Abstract

This talk explores the research challenges in the domain of IoT from multiple angles and reflects on the urgently needed collective efforts from various research communities to collaborate on those. Our approach fundamentally challenges the current thinking and understanding of scientific, technological, and political paradigms in tackling the engineering of IoT systems. We discuss technical paradigms and research challenges in the domains of Cloud and Edge Computing as well as the requirements of people in such systems. We will explore how these novel approaches impact application composition utilizing AI and Edge Computing.

Biography

Schahram Dustdar is Professor of Computer Science heading the Distributed Systems Group at the Technical University of Vienna. From 2004-2010 he was also Honorary Professor of Information Systems at the Department of Computing Science at the University of Groningen (RuG), The Netherlands. From 1999 - 2007 he worked as the co-founder and chief scientist of Caramba Labs Software AG in Vienna (acquired by Engineering NetWorld AG), a venture capital co-funded software company focused on software for collaborative processes in teams. Caramba Labs was nominated for several (international and national) awards: World Technology Award in the category of Software (2001); Top-Startup companies in Austria (Cap Gemini Ernst & Young) (2002); MERCUR Innovationspreis der Wirtschaftskammer (2002).

From Dec 2016 until Jan 2017 he was a Visiting Professor at the University of Sevilla, Spain and from January until June 2017 he was a Visiting Professor at UC Berkeley, USA. He is co-Editor-in-Chief of the new ACM Transactions on the Internet of Things as well as Editor-in-Chief of Computing (Springer). He is an Associate Editor of IEEE Transactions on Services Computing, IEEE Transactions on Cloud Computing, ACM Transactions on the Web, and ACM Transactions on Internet Technology, as well as on the editorial board of IEEE Internet Computing and IEEE Computer. Dustdar is recipient of the ACM Distinguished Scientist award (2009), the IBM Faculty Award (2012), an elected member of the Academia Europaea: The Academy of Europe, where he is chairman of the Informatics Section, as well as an IEEE Fellow (2016).

Invited Research Seminar

The forces that drive cells

By Xavier Trepat

Abstract

Cells exert, sense, and respond to physical forces through an astounding diversity of mechanisms. I will discuss some of these mechanisms in the context of tissue growth, migration and division. I will present maps of cell-cell and cell-extracellular matrix (ECM) forces during cell migration and division in a variety of epithelial models, from the expanding MDCK cluster to the regenerating zebrafish epicardium. These maps revealed that migration and division in growing tissues are jointly regulated. I will also present direct measurements of epithelial traction, tension, and luminal pressure in three-dimensional epithelia of controlled size and shape. By examining epithelial tension over time-scales of hours and for nominal strains reaching 1000%, we establish a remarkable degree of tensional homeostasis mediated by superelastic behavior.

Biography

Xavier Trepat received a BSc in Physics in 2000 and a B.Sc in Engineering in 2001. In 2004 he obtained his PhD from the Medical School at the University of Barcelona. He then joined the Program in Molecular and Integrative Physiological Sciences at Harvard University as a postdoctoral researcher. In 2008 he became a Ramon y Cajal researcher at the University of Barcelona and the Institute for Bioengineering of Catalonia (IBEC), and in January 2011 he became an ICREA Research Professor. His research at IBEC focuses on integrative tissue dynamics and cytoskeletal mechanics.

Host Ralph Andrzejak

Room 55.410

Invited Research Seminar

Integrative machine learning methods for the analysis of heterogeneous data in prospective neurodegeneration studies

By Marco Lorenzi, INRIA Researcher

Abstract

Machine Learning and Biophysical Modelling: Learning by Heart

By Maxime Sermesant, INRIA Researcher

Abstract

Room 52.119 

PhD Research Seminar

Software development best-practices for reproducible research

By  Alastair Porter

Abstract

In software development it is considered a best practice to test code, include documentation, use source code management tools, and make frequent backups. A lot of the time technical research tends to eschew these best practices, resulting in missing data, hard to reproduce results, and wasted time. For researchers who haven't worked in or studied software engineering roles, it can often be confusing to know where to start, or how these best practices improve code quality and save time. In this talk I will show some examples why software engineering best practices are a valuable part of technical research and how to start applying them if you do not know what tools and resources are available.

Room 52.s29 

Invited Research Seminar

Introduction to computational musicology with hands-on case studies for non-western music traditions

By Barış Bozkurt , Assoc. Prof. of Electrical Engineering at Izmir Democracy University, Turkey.

Abstract

The main goal of this tutorial is to present an overview for computational analysis applied on non-Western music traditions, mostly within the context of the CompMusic project (https://compmusic.upf.edu/). The target audience is researchers and students interested in computational musicology. The tutorial will start by presenting some of the relevant problems and challenges for analysis of non-Western music traditions that have been studied from an MIR perspective. We will present resources created during the CompMusic project, that are openly available to the community and demonstrate accessing non-Western music data and processing these data (focusing on intonation analysis) with publicly available tools. We will consider research corpora from various music traditions: Hindustani (North India), Carnatic (South India) and Turkish-makam (Turkey).  This session would serve as a quick start for students and researchers without prior experience in analysis of non-Western music and will provide them a good entry point for further investigation. 

Biography

Barış Bozkurt is an Assoc. Prof. of Electrical Engineering at Izmir Democracy University, Turkey. He is a collaborator of the Music Technology Group at the Universitat Pompeu Fabra in Barcelona since 2011. He has obtained his PhD degree (on speech analysis) in 2005 from Faculte Polytechnique De Mons, Belgium and also worked in speech industry after his PhD. Since 2007, he has been teaching in Electrical Engineering and Computer Science departments, and carrying research in the fields of audio signal processing and computational musicology.

Host:  Xavier Serra

Room 55.309

Invited Research Seminar

Fusion of images

By Enric Meinhardt-Llopis, École Normale Supérieure de Paris-Saclay

Abstract

The problem of image fusion consists in producing an ideal image of an object from several images of the object that are incomplete, deformed, blurry, noisy and with strong illumination changes.  This general problem contains as particular cases the problems of image registration, denoising, deblurring and inpainting.  We showcase three applications (1) the virtual periscope, that allows to see through the surface of turbulent water; (2) museum photography, to obtain clean and complete images of artwork from several incomplete and slanted photos; (3) reconstruction of asteroid surfaces, to combine inconsistent 3D models of an asteroid obtained from parallax and from shape-from-shading into a single coherent surface.  The fusion of N images always becomes a trivial problem in the limit when N goes to infinity; however in practice we are interested in the limit when N goes to 1.  We will see that, under rather mild hypotheses, the fusion of a small number of images is still a well-posed problem. 

Biography

Enric Meinhardt-Llopis is an associate professor at the École Normale Supérieure de Paris-Saclay.  After completing his studies in mathematics at the Universitat Politècnica de Catalunya, he obtained his PhD degree from the Universitat Pompeu Fabra in 2011, under the direction of Professor Vicent Caselles.  Now he is working in the image processing group of Professor Jean-Michel Morel, mostly in the analysis and exploitation of optical and radar satellite images. 

Host:  Coloma Ballester

Room 55.410

PhD Research Seminar

Introduction to Copyright in ICT

By Brisa Burriel Fuster, UPF. 

Abstract

We'll share the most relevant concepts of Copyright Law, and learn from popular cases, so we can have an idea of what is important to consider in our day to day.

Biography

I work as R&D and IT Legal Adviser at the UPF with a strong focus on Technology Transfer from the Public to the Private sector. My aim is to help organizations reaching the highest standards of valorization and exploitation of research results through IPRs awareness activities, as well as designing and implementing effective processes and best practices.

January, 9th  

11:00h

Room 55.309

Invited Research Seminar

Asynchronous Provably-Secure Hidden Services (CT-RSA 2018)

By Philippe Camacho , R&D engineer for Enuma Technologies

Abstract

The client-server architecture is one of the most widely used in the Internet for its simplicity and flexibility. In practice the server is assigned a public address so that its services can be consumed. This makes the server vulnerable to a number of attacks such as Distributed Denial of Service (DDoS), censorship from authoritarian governments or exploitation of software vulnerabilities. In this work we propose an asynchronous protocol for allowing a client to issue requests to a server without revealing any information about the location of the server. In addition, our solution reveals limited information about the network topology, leaking only the distance from the client to the corrupted participants. We also provide a simulation-based security definition capturing the requirement described above. Our protocol is secure in the semi-honest model against any number of colluding participants, and has linear communication complexity. Finally, we extend our solution to handle active adversaries. We show that malicious participants can only trigger a premature termination of the protocol, in which case they are identified. For this solution the communication complexity becomes quadratic. To the best of our knowledge our solution is the first asynchronous protocol that provides strong security guarantees.

Biography

Philippe Camacho got his PhD at the University of Chile and is currently working as a R&D engineer for Enuma Technologies, a blockchain company. His research interests include authenticated data structures, anonymity protocols and cryptocurrencies.

Invited Research Seminar

Convex optimisation and applications

By Victor Valls, Trinity College Dublin

Abstract

Convex optimisation is a subfield of mathematical optimisation that plays a central role in many areas, such as automatic control, finance, or statistical learning. In this talk, I will first give a non-technical introduction to convex optimisation, which will cover the basic concepts, some of its algorithms, and canonical applications in machine learning and image processing. In the second part of the talk, I will focus on recent applications of convex optimisation to the areas of networking and control. In particular, on how to solve max-flow type problems with stochastic constraints while making discrete actions. Some specific applications of the latter include energy minimisation in wireless networks, scheduling in queueing systems, and traffic signal control with constrained decisions. 

Biography

Victor Valls is a research fellow at Trinity College Dublin (Ireland) working in the group of Prof. Iosifidis. He obtained his degree and MSc in engineering from Universitat Pompeu Fabra (Barcelona) in 2011 and 2012 respectively, and PhD in applied mathematics from Trinity College Dublin in 2017. His research interests are in the area of mathematical optimisation, with applications in networks, control and machine learning. He has been recently awarded an MSCA global fellowship (FANC; 795244) with Yale University (USA) and Trinity College Dublin, which he will start in 2019. 

Host: Boris Bellalta

PhD Research Seminar

Reproducibility in research.

By Aurelio Ruiz

Abstract

Biography

Invited Research Seminar

Tangible Interaction and Cultural Forms: Supporting Learning in Informal Environments

By Dr. Michael Horn, Northwestern University

Abstract

Designers who create computer-based learning experiences for places like museums, out-of-school programs, and homes face a number of challenges related to the informal nature of such settings. Designs must generally function on their own without the support of teachers or curriculum while at the same time engaging a diverse audience, supporting productive social interaction, and activating appro- priate prior knowledge and skills. In this talk, I present an approach to the design of informal learning experiences based on tangible interaction. The term tangible refers to a variety of human–computer interaction techniques that move beyond computer screens and create opportunities for people to interact with digital systems using their bodies and physical artifacts. I argue that tangible interaction creates unique opportunities for designers to shape objects and situations to evoke cultural forms of literacy, learning, and play. To illustrate these arguments, I describe design cases that colleagues and I have created to support learning in museums and homes.

Biography

Michael Horn is an Associate Professor of Computer Science and Learning Sciences at Northwestern University where he directs the Tangible Interaction Design and Learning (TIDAL) Lab. Michael serves as the Program Coordinator for the Learning Sciences PhD Program at Northwestern and is co-Founder of the new Joint PhD Program in Computer Science and Learning Sciences. Michael's research explores the use of interactive technology in the design of innovative learning experiences. He takes a cautious but optimistic stance towards technology in a process that tightly couples research and design. His work has been exhibited at museums around the world including the California Academy of Sciences (San Francisco), the Museum of Science (Boston), the Field Museum (Chicago), and the Computer History Museum (Silicon Valley). Michael's research on tangible programming has contributed to the commercial products: Osmo Coding (https://www.playosmo.com/en/coding/) and Kibo Robotics (http://kinderlabrobotics.com/kibo/). Michael earned his Ph.D. in Computer Science at Tufts University working in the Human-Computer Interaction Lab and the Developmental Technologies research group. He received his undergraduate degree in Computer Science from Brown University and has worked as a software engineer for several companies including Classroom Connect and iRobot Corporation.

Host: Narcís Parés

PhD Research Seminar

UPF- Library for PhD students 

By Gemma Alvarez

Host:Aurelio Ruiz

Invited Research Seminar

How do people explore virtual environments?  

By Belén Masía, Universidad de Zaragoza

Abstract

Virtual reality (VR) systems provide a new medium that has the potential to have a significant impact on our society. The experiences offered by these emerging systems are inherently different from radio, television, or theater, opening new directions in research areas such as cinematic VR capture, interaction, or content gener ation and editing. However, the behavior of users who visually explore immersive VR environments is not well understood, nor do statistical models exist to predict this behavior. Yet, with unprecedented capabilities for creating synthetic immersive environments, many important questions arise. How do we design 3D scenes or place cuts in VR videos? How do we drive user attention in virtual environments? Can we predict visual exploration patterns? How can we efficiently compress cinematic VR content?

To address these and other questions from first principles, it is crucial to understand how users explore virtual environments. In this talk, I will go through our work on analyzing user behavior in cinematic VR scenarios.

Biography

Dr. Belen Masia is an Assistant Professor in the Computer Science Department at Universidad de Zaragoza, and a member of the Graphics & Imaging Lab of the I3A Institute. Prior to that, she was a postdoctoral researcher at MPI Informatik, and a member of the Max Planck Center for Visual Computing and Communication.

PhD Research Seminar

"How do agile practice support organizing a PhD"  

Invited Research Seminar

Automatic Assessment of Student Music Performances

By Alexander Lerch, Georgia Institute of Technology


Abstract

While there is large interest in Music Information Retrieval tasks focusing on extracting score-like information such as pitches, chords, and rhythmic properties, less attention is being paid to performance information such as (micro-)tempo, intonation, and dynamics.
This presentation will look at Music Performance Analysis from the perspective of assessing student music performances. The possibilities and limitations of current data-driven approaches to the assessment of middle and high school students will be discussed based on real-world results on a large dataset of graded student recordings.

Biography

Alexander Lerch is Assistant Professor at the Georgia Tech Center for Music Technology, where he leads the music informatics group. He is working on new artificially intelligent technologies for accessing, producing, and listening to music. His main research field is called Music Information Retrieval (MIR), an emerging interdisciplinary field. He is Co-Founder of the company zplane.development, a research-driven technology provider for the music industry. His book “An Introduction to Audio Content Analysis,” published by Wiley/IEEE Press, is used as course text book in multiple institutions.

More: https://www.alexanderlerch.com/

PhD Research Seminar

Introduction to ethical issues in research

PhD Research Seminar

Computational Cluster at DTIC

Invited Research Seminar

General-purpose audio tagging using deep learning approaches  

November 13th

16:00h 

Room 55.003

Invited Research Seminar

From biomechanics to sound: status and challenges in the numerical simulation of voice 

by Oriol Guasch, GTM Grup de recerca en Tecnologies Mèdia

Abstract

Muscle activations position the articulators, which define a vocal tract geometry and posture the vocal folds. Air emanating from the lungs induces self-oscillations of the vocal folds, which result in aeroacoustic sources and the subsequent propagation of acoustic waves inside the vocal tract. There, many things could happen. For instance, the air could resonate to generate vowels, or, at constrictions, airflow may be accelerated to create turbulent sounds such as fricatives. The vocal tract walls are flexible and react to the inner acoustic pressure. Also, articulators can change the vocal tract geometry to generate vowel-vowel utterances or syllables. Sound is finally radiated from the mouth. Numerical methods have revealed as a very powerful tool for the realistic simulation of all those complex phenomena. However, attempting unified 3D numerical simulations of all the above processes, which involve coupling of a biomechanical model and the mechanical, fluid and acoustic field interactions, may seem unwise. Unified approaches have been shunned for their daunting complexity and high-performance parallel computation requirements. This situation now seems to be changing. In this talk, we briefly review recent approaches towards 3D realistic voice simulation that unify, at least to some extent, some of the involved physical fields. Remaining challenges will be highlighted. 

Biography

Oriol Guasch holds a five-year degree in Physics from the University of Barcelona and a PhD in Computational Mechanics and Applied Mathematics from the Polytechnic University of Catalonia. He worked for several years in industry but definitely moved to academics in 2007. He is currently a full professor at the Department of Engineering, La Salle, University Ramon Llull, where he heads the research on acoustics within the GTM (Grup de recerca en Tecnologies Mèdia). His research lines involve computational acoustics and aeroacoustics with special emphasis on numerical voice production, graph theory applied to mid and high frequency vibroacoustics, theoretical and experimental transmission path analysis, parametric array technology to generate audible sound from ultrasonic collimated beams and acoustic black holes in beams, plates and ducts. He has authored a significant number of papers in first rank international scientific journals and conferences, where he has organized and chaired many special sessions. He has also participated and been the principal investigator of several national and international projects, and recently led the scientific coordination of the EUNISON (Extensive Unified-Domain Simulation of the Human Voice) project, EU-FET no 308874.

November  9th

12:00 h

Room 52.213

November 9th

 9:30 h

 Auditorium

Campus Poblenou

Jornada:  Innovació tecnològica a les ciutats

Jornada destinada al debat i exposició sobre tecnologies emergents i el seu impacte en les ciutats i el ciutadà. La jornada consta de dues taules rodones i una xerrada inspiracional on investigadors, empreses, emprenedors, centres tecnològics i l'administració pública adreçaran el tema des de diversos angles i iniciatives.

Organitza: UPF, Escola d'Enginyeria, Departament de Tecnologies de la Informació i les Comunicacions i UPF Ventures

Incripcions

November  8th

15:30 h

Room 55.410

PhD Research Seminar

Scientist Dating Forum

By Ryan Armstrong & Carla Conejo González

Abstract

The Scientists Dating Forum (SciDF) will present to the community in DTIC their approach to promote that researchers are more actively involved in society. We first got to know SciDF during the Falling Walls Lab qualifying round we co-organised in 2017, and their interests to impact on a broad number of topics, ranging from general outreach to science diplomacy, are quite well aligned with several of the priorities of the María de Maeztu transversal actions, that include opening up scientific work and setting the means that guarantee that excellent science also results in positive impacts in society.

Who is SciDF?

Scientists Dating Forum (SciDF) is a science-based association created in 2016 in Barcelona. SciDF aims to strengthen the interactions of scientists with politics, society and economics by fostering dialogue and community building.

We are a multidisciplinary young team integrated by researchers, communication specialists and people from private/public sector that share an interest in science. We are working to give to science a stronger voice in politics and economics, and strengthen its position in society. We contribute on a voluntary basis showing that the scientific community is committed to reinvent this sector and open it up.  

What does SciDF do?

SciDF wants to show how scientists through their own self-initiative try to close gaps between science and other sectors. More than 80 young and dynamic professionals from over 50 entities have been involved in the SciDF, and over 1000 people so far have attended SciDF events. These events consist of round table discussions and debates in bars (SciDF-Bars) and at institutes (SciDF-Talks), social activities for scientists and collaborations with existing institutions and programs.

We have also organized activities with other entities, such as the YoMo festival and the Festa de la Ciencia (Barcelona), and AsBioMad After Science event in Madrid.

SciDF has mobilized the scientific community of Barcelona by organizing the March for Science Barcelona 2017. We gave also a voice to science in the elections of 2017 in Catalonia: first, debating on how the independence scenario could affect research and scientists, and next, by asking all political parties to present their position on 10 different scientific policy topics, elaborating an exclusive summary for science-based vote.

SciDF opens science!

We believe that a scientific community that motivates society and politicians to participate in science, through dialogue and constructive discussions, makes more critical citizens and politicians.

Biography

Ryan Armstrong is a researcher in the field of management and organizations. His research interests lie primarily in two areas. First, in exploring, describing, and developing methodology for addressing complex challenges in the applied social sciences (and sometimes the natural ones). Second, in employing these methods to contribute to robust, actionable theory to help navigate the challenges brought on by the 4th Industrial Revolution: Increased technological, social, and ecological change, automation, Big Data, and uncertainty. He has published and presented on the topics of strategic management, applied philosophy, and interdisciplinarity. He is the outgoing PhD Representative for the University of Barcelona Business PhD program and for the British Academy of Management, Performance Management group. Since 2012 Ryan has worked as a consultant for organizations in Europe and the USA to support himself during his graduate education, working in strategic management, analysis, information systems, and the use of visual methods for managing complexity. Since July 2018 he has served as Treasurer for Scientists Dating Forum.

Carla Conejo González is the Vice-president of Scientist Dating Forum. She works at Fundació Catalunya La Pedrera, where she is the Science Projects Leader of the Area of Knowledge, Education & Research. BS in Human Biology and MS in Pharmaceutical and Biotechnological Industry at Universitat Pompeu Fabra with a degree specialization in neurobiology and research internships in the Center for Genomic Regulation in Barcelona and in the Dipartamento di Scienze Biomediche e Neuromotorie at the Università di Bologna in Italy. Scientific advisor, content researcher and creator of the TV program Quèquicom of the Canal33 (TV3). In the field of science advocacy, she is also a board member of WhatIf – Passionately Courious and has volunteered at MAGMA – Association for Promoting Youth Research, both initiatives aimed at promoting science education and dissemination by connecting science to society and facilitating access to research. Carla is also one of the founders of Associació Babu, a non-governmental humanitarian organization aimed at developing programs of health and psychological assistance worldwide.

November  8th

11:00 h

Room 52.019

Invited Research Seminar

An introduction to colour management and workflow from set to post-production

By Ariana Bonavia, DIT and Dailies Colourist

Abstract

The purpose of a colour Management workflow, is to provide colour consistency throughout the entire workflow. From set to postproduction and final deliveries. Ideally on any production we would like to maintain colour accuracy regardless of the cameras, hardware and software we use. The process will need to be standardized and capable to communicate with different systems and users. In order to translate correctly the director´s vision to the final screen.

Biography

Invited Research Seminar

Optimal Real-Time Coding of Markov Sources 

Invited Research Seminar

In silico trials: how modeling and simulation and a changing regulation combined can accelerate the development and approval of new drugs and medical devices. 

October 25th  

PhD Research Seminar

Gender and Science: the importance of intersectional feminism in academia

PhD Research Seminar

Introduction to PhD Seminars 

"Deep learning for medical image reconstruction, super-resolution, classification and segmentation "

"Deep learning for image quality control and data labelling using attention maps"

Invited Research Seminar

Patch-based MRI Analysis:From voxel to knowledge

Invited Research Seminar

Patch-based MRI Analysis:From voxel to knowledge

 July 4

16:00h

55.003

Invited Research Seminar

Identifying Idiomatic Language with Distributional Semantic Models

July 3

11.00 h

Room 51.100

Invited Research Seminar

"TumbleBit: An Untrusted Bitcoin-Compatible Anonymous Payment Hub"

July 3

12:30 h

Room 51.100

Invited Research Seminar

"Post-quantum cryptography from supersingular isogeny problems?" 

By Christophe Petit 

Abstract

We review existing cryptographic schemes based on the hardness of computing isogenies between supersingular isogenies, and present some attacks against them. In particular, we present techniques to accelerate the resolution of isogeny problems when the action of the isogeny on a large torsion subgroup is known, and we discuss the impact of these techniques on the supersingular key exchange protocol of Jao-de Feo. 

Biography

Christophe Petit is a Lecturer in Computer Security in the University of Birmingham's School of Computer Science, and a member of the Security and Privacy research group. Previously he was a Research Fellow at the University of Oxford's Mathematical Institute, where he helped founding the Cryptography group,and prior to that he worked at Université catholique de Louvain's Department of Electrical Engineering and at University College London's Computer Science Department.

June 14

15:30 h

Room 51.100

PhD Research Seminar

Design of Experiment - decreasing the time of a study by increasing its reliability

By Simone Tassani

June 13

15:30 h

Room 51.100

PhD Research Seminar

Design of Experiment - decreasing the time of a study by increasing its reliability

By Simone Tassani

June 8

15:30 h

Room 55.309

Invited Research Seminar 

Designing expressive engagement with electronic and hyper instruments

By Hans Leeuw

June 7

15:30 h

Room 55.309

Invited Research Seminar

Reflecting on the PELARS project and Multimodal Learning Analytics

By Daniel Spikol, PhD

Associate Professor, Faculty of Technology and Society, Department of Computer Science and Media Technology, Malmö University, Sweden

Abstract

The talk presents key results and lessons learnt from the Practice-based Experiential Learning Analytics Research and Support (PELARS) project. PELARS was a three-year project that developed a learning analytics system to investigate small group learning for open-ended engineering tasks. The aim is to share the results and challenges of the project to further the dialogue about how to increase progress for multimodal learning analytics (MMLA). The theories and different methods of the project will be discussed that include the different sensors used to capture, record, and analyse students physical interactions and how we used this data to create models to understand aspects of collaboration. To investigate the multimodal data, we started with a simple grading of the student's final products and progressed to a richer framework for assessing student non-verbal collaboration with different machine learning strategies. Some of the results illustrate that body location and movement are strong features for further investigating collaboration. However, challenges remain across the sustainability and ethics of how this data can be used to support the learners and teachers in creative, open-ended activities.

Biography

June 6

15:30 h

Room 52.S27

PhD Research Seminar

Design of Experiment - decreasing the time of a study by increasing its reliability,

By Simone Tassani

May 24

15:30h

Room 55.309

PhD Seminars

How to do good experiments online / Good practices for online experiments

By:

• Xavier Favory - Freesound Datasets https://datasets.freesound.org/
• Maria Rauschenberger - Lessons learned from conducting a medium-scale online experiment to distinguish a person with and without dyslexia. https://mariarauschenberger.com/projects/
• Davinia Hernández-Leo / Patricia Santos - Integrated Learning Design Environment  http://ilde2.upf.edu/

Abstract

User experiments can be crucial for researchers who need to evaluate methods involving humans, or gather human-generated data. The Web has created an alternative to expensive lab experiments: online and crowdsourcing experiments are valuable resources for research which can bring many advantages. However, online experiments introduce different challenges which need to be addressed carefully in order to be accurately performed. This seminar aims at sharing good practices when performing online-based experiments as well as crowdsourcing research. First this talk will give an overview of the ways in which online experiments can be used. Then, the following researchers from the department will present their work and share valuable advices:

• Xavier Favory will present Freesound Datasets, a collaborative platform for the creation of open audio datasets.
• Maria Rauschenberger will draw your attention to  possible pitfalls and lessons learned from conducting a online experiment for children at the age of 7 till 12 for different languages with a web-game.
• Davinia Hernández-Leo will present the Integrated Learning Design Environment 2 (ILDE2), a community platform for learning design.

Background material

Making better use of the crowd. NIPS 2016, ACL 2017, and KDD 2017

Instructor: Jenn Wortman Vaughan

http://www.jennwv.com/projects/crowdtutorial.html  

May 10

15:30 h

Room 55.309

PhD Research Seminar

The experience by PhD students (2)

By:

- Adrià Arbués, Image Processing Group

Data Science in Sports and experience during the 1st year in the PhD program

- Merljin Blaauw, Music Technology Group

Singing synthesis and his motivation for doing a phd after years of industrial research projects, spinoff, etc

- Olga Slizovskaia, Music Technology Group and Image Processing Grou

Multimodal video analysis for music information retrieval. You can read part of her experience over the past 2 years in the blog of the MdM program

May 8

15:30 h

Room 55.309

Invited Research Seminar

Learning Deep Networks from ranked data

By Joost van de Weijer 

Abstract

In the second part of the talk I will discuss two applications of the decomposition of fully connected layers into several layers.  Deep Neural Networks trained on large datasets can be easily transferred to new domains with far fewer labeled examples by a process called fine-tuning. However, networks designed to be optimal for the source task are often prohibitively large for the target task. We show how layer decomposition can be used to effectively compress the network. In addition, we will show how layer decomposition can be used for the task of learning without forgetting where the task is to learn new tasks without forgetting previously learned tasks. We will show how this can be used to improve the diagonal assumption underlying the Fisher information matrix for the Elastic Weight Consolidation algorithm. The work in this talk is presented at ICCV 2017 and CVPR 2018.

Biography

Joost van de Weijer is  the leader of the Learning and Machine Perception Team (LAMP) and a senior researcher at the Computer Vision Center (Universitat Autònoma de Barcelona). His main research interest is in applying color imaging theory to the field of computer vision: color constancy, color naming, object recognition, recoloring algorithms, color feature detection and color feature extraction.

May 7

15:30 h

Room 55.309

Invited Research Seminar

Efficient Sound Synthesis By Recursive Digital Filters In Parallel Form: Virtual Musical Instruments And Artificial Reverberation

By Esteban Maestre

Computational Acoustic Modeling Laboratory, McGill University

Abstract

Recursive digital filters in parallel form offer a number of advantages. Besides their suitability for parallel implementation, improved numerical precision and parametric control are some of their attractive features. This talk will cover some recent progress in using frequency-domain optimization techniques to design high-order recursive parallel filters suited for the efficient simulation musical instruments and reverberation via modal synthesis.  

The presentation will first introduce novel methods to render musical instrument sound by combining the digital waveguide and modal synthesis frameworks. Through a modal analysis technique based on constrained optimization of parallel filter coefficients, measured resonator immitances are faithfully reproduced via an efficient modal formulation that allows their realization as lumped elements within a digital waveguide model. Sound examples will be provided for wind and string instruments.

Then, the problem of designing a modal reverberator to match a measured room impulse response will be considered. The modal reverberator architecture expresses a room impulse response as a parallel combination of resonant filters, with the pole locations determined by the room resonances and decay rates, and the zeros by the source and listener positions. The method first estimates the pole positions in an iterative process involving a series of constrained pole position optimizations in overlapping frequency bands. With the pole locations in hand, the zeros are fit to the measured impulse response using least squares. Example models for a medium sized room will be provided.

Host: Dr.Xavier Serra

April 26

15:30 h

Room 55.309

PhD Research Seminar

Experiences by PhD students

By:

April, 6

12:30 h

Room 24.009 (Ciutadella Campus)

Invited Research Seminar

Neural Mechanisms for Perceiving Object Motion During Self-Motion

By Greg DeAngelis

Abstract

Many organisms have highly evolved neural circuits for processing visual motion cues. However, most laboratory studies of visual motion perception are performed under highly simplified conditions in which there is no self-motion such that object motion in the world directly maps onto retinal image motion. Under many natural conditions, however, we must judge object motion during self-motion, which greatly complicates the problem. Thus, the brain needs to parse the complex pattern of retinal image motion into components that correspond to object motion and self-motion. In addition, to compute object motion in world coordinates, the brain must estimate self-motion and factor it into computations of object motion.  I will describe two studies that make important progress into understanding the visual and multi-sensory mechanisms by which the brain computes object motion during self-motion.  I will show that neural activity in macaque area MT reflects the operation of a “flow parsing” mechanism (which has been previously established in human psychophysics) that discounts global optic flow resulting from self-motion.  I will also show that neural activity in area VIP reflects flexible computation of object motion in either world- or head-centered coordinates.  Together, these studies begin to reveal critical neural processes that are involved in perceiving object motion under more natural conditions in which self-motion also occurs.

April, 5

15:30 h

Room 52.219

PhD Research Seminar

Introduction to Software Licensing 

By Malcolm Bain, idLaw partners

Abstract

When using and developing open source software, there are a number of considerations to take into account that go beyond the pure technical aspects. Malcolm Bain will introduce and review the legal aspects of using and developing open source software, with a focus on licensing, license choice and license compliance obligations with distributing products that are or embed open source components.  

Biography

Malcolm Bain is an English solicitor and Spanish lawyer, specialising in Information Technology and Intellectual Property law, and co-founder of id law partners (now part of BGMA, a Barcelona based law firm). He has a wide experience representing clients on both sides of IT transactions, and advises on licensing, software contracts, technology transfer, copyright, privacy and trademark issues. He has participated in various R+D projects and written and lectured on many aspects of IT law, e-commerce and internet regulation.

April, 3

15:00 h

Room 55.410

Invited Research Seminar

One size doesn't fit all: towards adaptive and personalized text simplification

By Joachim Bingel

Abstract

Biography

Joachim Bingel is a final-year PhD student at the University of Copenhagen, supervised by Anders Søgaard. His primary research interests include adaptive models for text simplification and multitask learning, as well as the combination of these. Joachim has further published papers in corpus linguistics and the usage of cognitive data to improve NLP models. He has a Master's degree from the University of Heidelberg and has previously worked as a researcher at the Institute for the German Language in Mannheim.

Host

Monika Dominguez

March, 27th

15:30 h

Room 55.309

Invited Research Seminar

Conditional Risk Minimization for Stochastic Processes

By Alexander Zimin

Abstract

We study the learnability of stochastic processes with respect to the conditional risk, i.e. the existence of a learning algorithm that improves its next-step performance with the amount of observed data. We introduce a notion of pairwise discrepancy between conditional distributions at different times steps and show how certain properties of these discrepancies can be used to construct a successful learning algorithm. In addition to theoretical results that establish criteria for learnability for many classes of stochastic processes, we provide a practical algorithm that can be efficiently implemented and applied to real data.

Biography

Alexander is a PhD student at IST Austria working in the group of Christoph Lampert. He obtained his Master degree from Central European University in 2013 under supervision of Laszlo Gyorfi and Gergely Neu. His research is focused on various machine learning scenarios, which involve dependent data. Previously he has also worked on online learning and reinforcement learning.

Host

Gergely Neu

March, 19

10:00

55.410

Invited Research Seminar

The influence of cellists' postural movements on their musical expressivity

By Joselyn Roze, Université Aix-Marseille

Abstract

The musician's body is often deeply forgotten in the traditional instrumental pedagogy. In the case of bowed-string instrumentalists, the attention is required to be paid on the fingers' dexterity at the left hand or the bow accuracy at the right hand, like preliminary and mandatory conditions for musical expressivity. However, just by observing the play of great violinists or cellists, we can realize that they use postural movements more or less consciously, which are far away from their purely instrumental gestures involved in the sound production. What would be thus the function of these corporeal movements ? If they visually contribute to make a performance more lively, we actually don't know their influence on the perceived expressive acoustical features, such as the musical phrasing or the sound quality.

In my thesis, we try to bring some response elements by investigating the cellists' ancillary movements and their influence on the musical expressivity. On the gesture side, we focus on the chest and head displacements, which are the most salient visually but also the most essential for the cellist's motor planning. On the acoustical side, we focus on the two principal means of producing musical expressivity : the rhythmic and the timbre. This investigation takes place within various areas implying different perspective levels : Instrumental pedagogy, musician's posture, cellist's physiology, and more globally musical embodiment. The influence of postural movements on the musical expressivity is assessed by the means of four postural conditions affecting the cellists' sensorimotor mechanisms : A standard playing condition and three gradual immobilization constraints (mental, physically immobilized by the chest, and finally by the chest and the head). Within each postural session, the bow playing mode (short détaché or large legato) has been chosen as a studying factor of the crossed relationships between sound and motion.
The first part of the thesis highlights the relationships between the cellist's postural movements and their rhythmical expressivity. It appears that the coordination between posture and movement of a cellist is optimal with a good postural-kinetic capacity, which allows to keep the balance by resisting to bow expansions. The natural symbiosis between the musician and his cello also turns out to be a key factor for the metric and rhythmic cohesion between trunk ancillary gestures and instrumental bowing. The rhythmical unbalance caused by the absence of postural movements is particularly repeatable during passages requiring a good motor coordination between the two arms. The
discomfort felt is assessed stronger in the bowing mode détaché, which coincides with a bigger uniformity in speed modulations and inter-note timbre colors. The second part of the thesis is dedicated to the detailed exploration of the timbre for a note frequently increased in harshness within the postural constraint by chest and head. This acoustical degradation is generally linked with a weaker bow speed and a narrower gesture of the instrumental right arm. By tracing back to the central body parts, we highlighted the importance of the spine torsion and the right shoulder opening for the sound quality. The whole of these works support flexibility in the postural movement freedom for producing a balanced and expressive play.

Biography

Graduated from the french engineer school in Electronic and Informatics (EFREI), Jocelyn Rozé worked eight years as a consultant for an information-technology service company (ALTEN). Around the end of this period, he discovered the cello and developed a passion for playing this instrument. Jocelyn then tried to redirect and choose a professional career likely to merge his technical skills to his instrumental passion. He found this opportunity by starting a PhD thesis at the Laboratory of Mechanic and Acoustic (LMA) of Marseille on the topic entitled : « The influence of cellists’ postural movements on their musical expressivity ». He obtained his doctorate degree in 2017 and now he continues a post-doctoral research on the same context to refine and promote his thesis results.

March, 14

12:30

Auditorium

March, 12

15:30

Room 55.410

March, 8

15:30

Room 55.309

PhD Research Seminar

The power of ordinary people in the Web - Studying quality and inequalities in User Generated Content

By  Diego Sáez-Trumper

Host: Aurelio Ruiz

Abstract

Internet utopia came with the promise of democratize the access to knowledge, and allows to create, share and receive free information. But after 26 years of the releasing of the first Web Browser, how is people producing, receiving, and propagating information in the Internet? In this talk, I will introduce my work on the usage of massive data processing (a.k.a Data Science) for studying digital prints of human behavior, as well as discuss how Machine Learning and Artificial Intelligence can be used to empower people on the digital era, specially in  the context of a free knowledge community, where we want to use those techniques not to replace, but to support human judgment. 

Biography

March, 7

15:30

Room 55.410

Phd Research Seminar

Extended Minds and Machines 

By Karina Vold 

Abstract

Traditional cognitive science subscribes to the computational theory of mind, which says that our thoughts are realized by neural computations, or symbol manipulations, carried out by our brains. Philosophers Clark and Chalmers (1998) roughly accept the computational theory of mind but argue that when we use tools, such as smartphones and tablets, they can become seamlessly integrated into our cognitive processes such that computations in the tools are just as essential to our cognition as computations in our brain: smartphones extend our cognition. The ‘extended mind thesis’ is an increasingly popular view in philosophy of mind and cognition. It maintains not only that technologies can compensate for biological deficiencies, but also that technologies can augment our minds and enhance our biologically bound cognitive capacities, making us smarter and more capable agents (e.g. Sutton 2007; Menary 2007).

 Given our cognitive reliance on technology, recent advancements in artificial intelligence may be cause for concern. The outcomes of algorithms can have adverse effects on human-decision making: take the case of biased risk-assessment algorithms used to predict recidivism rates and inform judges in parole decisions. Furthermore, machines now have goals, or desired outcomes, of their own and depending on what these goals are, humans might be a good means to achieve those outcomes. The more we build machines to have human-like general intelligence the more likely it seems that machines may move beyond their intended, or original, hardware base to make use of tools to complete their tasks, just as we have. This might not be a bad thing, but one concern is whether us humans will be their tools of choice. Thus, we might need to protect humans from being used, nudged, or manipulated by machines.

Biography

 Karina Vold is a Research Associate at the Leverhulme Centre for the Future of Intelligence and Research Fellow in the Faculty of Philosophy at the University of Cambridge where she works on the Agency and Personhood Project. She specializes in philosophy of mind and cognitive science and is currently interested in cognitive extension, machine agency, and consciousness.

March, 1st 

13:30

Room 52.223

Phd Seminar : Software development best-practices for reproducible research

By  Alastair Porter

Abstract

In software development it is considered a best practice to test code, include documentation, use source code management tools, and make frequent backups. A lot of the time technical research tends to eschew these best practices, resulting in missing data, hard to reproduce results, and wasted time. For researchers who haven't worked in or studied software engineering roles, it can often be confusing to know where to start, or how these best practices improve code quality and save time. In this talk I will show some examples why software engineering best practices are a valuable part of technical research and how to start applying them if you do not know what tools and resources are available.

March, 1st

15:30

Room 55.309

Invited Research Seminar

Cyber-physical resilient systems

By Joaquin García-Alfaro

Abstract

Cyber-physical systems consist of upgraded infrastructures in which physical, network and software components interact with each other, in a continuous and dynamic way. This upgrade introduces new threats to the resilience of traditional infrastructures. Known incidents include the sabotage of critical energy facilities, disruption of hospitals and financial services, and successful hacking of navigation systems. The problem is drawing a great deal of attention, since it can have serious effects to businesses, governments and society at large. During this talk, we will focus on disruptive attacks hidden as unintentional component malfunction. Such attacks must be handled by jointly addressing protection at both cyber and physical domains. We will argue about the necessity of novel solutions expanding results from research communities including information security, control theory and automation engineering. Application domains will include industrial scenarios, autonomous vehicles, ambient intelligence and Internet-of-Things.

Biography

Full Professor at Institut Mines-Telecom (Telecom SudParis & Université Paris-Saclay). Double Ph.D.degree in Computer Science from Universitat Autonoma de Barcelona (UAB) & Université de Rennes. Research Habilitation from Université Pierre et Marie-Curie (Paris Sorbonne VI). He holds two engineering awards from UAB, a doctoral fellowship award and a graduate award from the "la Caixa” savings bank foundation. His research interests are situated in several domains of network & system security, with a special emphasis on areas related to the management of formal policies, analysis of threats, enforcement of mitigation and evaluation of countermeasures.   

February,22nd 

15:30

Room 55.309

Phd Research Seminar

Ethical issues and data protection, CIREP-UPF.

February,22nd 

10:30-12:30

Room 55.309

Course (6 hours: 9th, 15th and 22nd of February)

Phd Seminar : Statistical course and Design of Experiments

By Simone Tassani

February,15th 

15:30

Room 55.410

Phd Research Seminar

Scientific Dissemination, Online Repositories, and Author's Rights

February,15th 

10:30 - 12:30

Room 55.309

Course (6 hours: 9th, 15th and 22nd of February)

Phd Seminar: Statistical course and Design of Experiments

By Simone Tassani

The course of statistics aims to introduce a number of tools for master/Ph.D. students and post docs. The presented tools will play a role in planning many kind of studies, properly analyse the results and understand if data analysed by other researchers are or not reliable.

The first part of the course will introduce General Linear Modelling and its most common applications:

Linear regression

T-test

F-test

Analysis of Variance (ANOVA)

February,9th 

10:30 - 12:30

Room 55.309

Course (6 hours: 9th, 15th and 22nd of February) 

Phd Seminar: Statistical course and Design of Experiments

By Simone Tassani

The course of statistics aims to introduce a number of tools for master/Ph.D. students and post docs. The presented tools will play a role in planning many kind of studies, properly analyse the results and understand if data analysed by other researchers are or not reliable.

The course will start with a brief digression over the several implications that bad statistics have today over the scientific society and why every researcher should know the basic concepts behind a statistical analysis.

February,8th 

15:30

Room 55.410

Phd Research Seminar

Adapting research capabilities to industry

February,8th 

15:00

Room 55.309

Invited  Research Seminar

Sound Source Localization in Wireless Acoustic Sensor Networks

By Maximo Cobos

Host: Emilia Gómez

Abstract

With the rapid evolution in the design and manufacture of electronic circuits, wireless nodes that integrate sensors and communication interfaces of different types have become economic resources for the design and deployment of innovative monitoring systems. In this context, wireless sensor networks (WSNs) have spread in fields as diverse as security, industry and systems oriented to care and well-being. When nodes incorporate acoustic transducers and the processing involves the manipulation of acoustic signals, such networks are commonly known as wireless acoustic sensor networks (WASNs). Inference of location information has always been an attractive research problem, where traditional (wired) microphone arrays have been employed to identify the number and location of active sound sources. However, practical limiations such as limited processing power, lack of synchronization and battery life of the nodes restrict the direct application of classical approaches within a WASNs framework. The seminar will cover a brief introduction to this research field, as well as an overview of recent approaches to tackle this problem.

Biography

Maximo Cobos received the Ph.D. degree in telecommunications from Universitat Politècnica de València in 2009. His Ph.D. dissertation was awarded with the Ericsson Best Thesis Award on Multimedia Environments from the Spanish National Telecommunications Engineering Association (COIT). He completed with honors his Ph.D studies under the University Faculty Training program (FPU) and was awarded with a "Campus of Excellence" post-doctoral fellowship to work at the Institute of Telecommunications and Multimedia Applications (iTEAM). Since 2011 he has been with the Computer Science Department of the Universitat de València, where he is now an Associate Professor. His work is focused on the area of digital signal processing for audio and multimedia applications, where he has published more than 90 technical papers in international journals and conferences. Dr. Cobos is a Senior Member of the Institute of Electrical and Electronics Engineers (IEEE), a Full Member of the Acoustical Society of America (ASA) and a member of the Audio Signal Processing Technical Committee of the European Acoustics Association (EAA).

February,8th 

14:30

Room 55.309

Invited  Research Seminar

Multidimensional and multi-level learning of music structure for machine improvisation

by Emmanuel Vicent

Host:Emilia Gómez

Abstract

Current musical improvisation systems generate unidimensional musical sequences by recombining their musical contents. However, considering several dimensions (melody, harmony...) and several temporal levels are difficult issues. In this talk, I will summarize the work of my PhD student Ken Déguernel on the combination of probabilistic approaches with formal language theory in order to address those issues. First, I will present a system able to follow the contextual logic of an improvisation modelled by a factor oracle whilst enriching its musical discourse with multidimensional knowledge represented by interpolated probabilistic models. Then, this work is extended to create another system using a belief propagation algorithm representing the interaction between several musicians, or between several dimensions, in order to generate multidimensional improvisations. Finally, we propose a system able to improvise on a temporal scenario with multi-level information modelled with a hierarchical grammar. We also propose a learning method for the automatic analysis of hierarchical temporal structures. All systems have been evaluated by professional musicians and improvisers during listening sessions.

Biography

Emmanuel Vincent is a Senior Research Scientist with Inria (Nancy, France). He received the Ph.D. degree in music signal processing from IRCAM (Paris, France) in 2004 and worked as a Research Assistant with the Centre for Digital Music at Queen Mary, University of London (London, U.K.) from 2004 to 2006. His research focuses on statistical machine learning for audio, with application to source localization and separation and noise-robust speech recognition, among others. He is a founder of the series of Signal Separation Evaluation Campaigns (SiSEC) and CHiME Speech Separation and Recognition Challenges.

February,7th 

12:30

Room 55.410

Invited  Research Seminar

The Face of Emotion: A computational model of the production and visual perception of facial expression of emotion

By  Aleix M. Martinez

Host: Xavier Binefa

Abstract

We now have computer vision algorithms that can successfully detect low-level image features (such as, edges), recover the 3D structure and motion of objects, and provide a semantic label for them (e.g., a face, or John’s face). But how about higher level, abstract concepts like emotions? This talk will introduce the first algorithms to successfully identify the emotion categories people regularly use to communicate emotion. I will first summarize our research uncovering the image features used by the human visual system to recognize emotion in faces. I will then explain how these results can be used to define computer vision systems that can work “in the wild” (i.e., outside controlled, in-lab conditions). In doing so, I will show the novel finding that people regularly use more than 35 distinct facial expressions (not the six propound by Darwin). We will see that the major problem in computer vision is that of landmark detection, not recognition, as is typical of most modern systems.

Biography

February, 1st 

15:30 

Room 55.309

Phd Research Seminar

Communication skills in science: the research in 4 minutes competition

Invited Research Seminar

Scale-invariant unconstrained online learning

By Wojciech Kotłowski

Abstract

We consider a variant of online convex optimization in which both the instances (input vectors) and the comparator (weight vector) are unconstrained. We exploit a natural scale invariance symmetry in our unconstrained setting: the predictions of the optimal comparator are invariant under any linear transformation of the instances. Our goal is to design online algorithms which also enjoy this property, i.e. are scale-invariant. We start with the case of coordinate-wise invariance, in which the individual coordinates (features) can be arbitrarily rescaled. We give an algorithm, which achieves essentially optimal regret bound in this setup, expressed by means of a coordinate-wise scale-invariant norm of the comparator. We then study general invariance with respect to arbitrary linear transformations. We first give a negative result, showing that no algorithm can achieve a meaningful bound in terms of scale-invariant norm of the comparator in the worst case. Next, we compliment this result with a positive one, providing an algorithm which "almost" achieves the desired bound, incurring only a logarithmic overhead in terms of the norm of the instances.

Biography

Wojciech Kotłowski is an assistant professor at Poznan University of Technology, Poland. From 2009 to 2012, he was a post-doctoral researcher in Centrum Wiskunde & Informatica (Amsterdam, Netherlands) in the group of Peter Grünwald. He obtained his Ph.D. degree in Computer Science from Poznan University of Technology in 2009 under supervision of Roman Słowiński. His main research interests are in the theory of machine learning, particularly in the online learning withadversarial data.

February, 1st 

10:30

Room 55.309

Research Seminar

Towards Green Web Search

By Ana Freire

Abstract

Web search engines have to deal with a rapid increase of information, demanded by high incoming query traffic. This situation has driven companies to build geographically distributed data centres housing thousands of computers, consuming enormous amounts of electricity and requiring a huge infrastructure around. Google reported the use of 4,400 GWh of electricity (equivalent to the amount of energy consumed by 367 US households). At this scale, even minor efficiency improvements result in large financial savings.

This seminar will introduce the concept of Green Web Search, by first exploring several (still few) approaches for developing sustainable search engines data centres. In the second part of the talk I will focus on a machine-learning-based model for click prediction to reduce the energy consumed while processing queries. Models like this are environmentally and economically beneficial, as they reduce pollution and lead to cost savings.

January 25th

15:30 

Room 55.410

January 25th

10:00

Room 55.309

Research Seminar

Learning, inference and control in complex systems

By Vicenç Gómez

Machine Learning is the study of data-driven methods capable of mimicking, understanding and aiding complex information processing tasks. In the first part of this seminar, I will review some of my contributions to this field in two basic research areas: probabilistic graphical models and reinforcement learning. In particular, I will first focus on efficient approximate inference algorithms for estimating the partition function of a graphical model, and second, on the equivalence between probabilistic inference and optimal control computation through the class of linearly-solvable Markov decision processes.

In the second part of this seminar, I will talk about applied research. I will first explain how we can design practical algorithms for robot control using the previous principles. Then, I will show how we can improve brain-computer interfaces using reinforcement learning, and, finally, I will describe how we can better understand online human activity using data-driven methods applied to complex social networks.

Since this research seminar is included in the ML tenure-track evaluation position, I will end up with a brief statement about my future research/academic plans.

January 24th

15:30 

Room 55. 309

PhD Research Seminar

Building Machine Learning Models to Value Heterogeneous Transactions in a Multi-Dimensional Customer-Vendor Relationship

by Gabriel Silberman

Abstract

January, 18th 

15:30

Room 55.410

PhD Research Seminar

Reproducibility in research.

By Aurelio Ruiz

Abstract

Biography

January, 18th 

10:00

Room 55.309

Research Seminar

Practical Spectral Learning Algorithms for Compositional Latent State Models

By Ariadna Quattoni

The main topic of this talk will center around the problem of learning compositional latent state models. Many problems in natural language understanding (NLU) can be framed as learning compositional functions over structured input/output domains, for example: language modeling, sequence tagging, parsing, relation extraction, etc. More specifically, I will focus on learning non-deterministic weighted automaton (WFAs) and grammars for these problems.

Spectral learning algorithms for WFAs are based on using a non-parametric representation and exploit a link between recurrence relations satisfied by a function and low-rank constraints on certain matrices constructed from function outputs. They have nice theoretical and algorithmic properties but despite this it has been challenging to obtain competitive results in real language modeling tasks. This is mainly because the classical spectral algorithms have practical limitations that make it hard to scale them to the requirements of real applications. In the first part of the talk I will describe recent work on overcoming some of these limitations that results in significant improvements.

The second part of the talk will be about future research directions. I will touch on another critical aspect of making learning compositional latent state models practical for NLU applications: namely the need for robust active learning algorithms specifically designed for structured domains. Most NLU applications have unique information needs, and obtaining the large collections of annotated data necessary to achieve good performance under the supervised approach might be too expensive for many use-case scenarios. Therefore, we need a collaborative machine learning paradigm that breaks the distinction between annotation and training so that models can be built with a minimal amount of supervision. A key idea that we can borrow from spectral learning is that the problem of active learning of compositional latent state models can be reduced to a form of active low-rank matrix completion.

January, 11th 

15:30

Room 55.309

PhD Research Seminar

"How to prepare a good abstract"

By Bart Bijnens

December,14th 

11:30

Room: 52.321

Phd Research Seminar

IEEE Author's Workshop

By Ana Baiges, Federico Peña and Jean Gardy Germain

Services and resources that the UPF Library puts at your disposal for scientific publishing.

Registration Form

December,14th 

11:00

Room 55.410

Invited Research Seminar

Rational data processing in complex networks.

By Santiago Mazuelas

Host: Xavier Binefa

Abstract

Data has been shown to be an extremely useful resource in many fields including communications, information technology, weather forecasting, and economics. Data processing techniques developed under paradigms such as machine learning, data science, and Bayesian inference are enabling many critical applications. In each specific application, data can reduce the uncertainty on the consequences of possible actions and hence improve decision making. For instance, certain key words in an email reduce the uncertainty regarding email usefulness and, hence, improve spam filtering.

Rational decision making that is consistent with preferences among consequences requires a probabilistic processing of data.  Such probabilistic framework for data processing encounters several technical difficulties including 1) modeling of complex data, 2) approximation of intractable computations, and 3) efficiency in high dimensions. In this talk, I will provide an overview of such probabilistic framework and describe techniques to overcome the above-mentioned difficulties. In addition, I will show the practical benefits of the techniques presented through the case study of network localization and navigation.

Keywords: probabilistic processing, machine learning, positioning, distributed processing, information coupling, graphical models, belief propagation, density estimation, Kalman filters, Particle filters, Belief condensation filters

Biography

December,12th 

15:30

Room 55.309

Invited Research Seminar

Creating music at Jukedeck

By Katerina Kosta

Host: Emilia Gómez

Abstract

November,30th 

15:30

Room 55.309

Invited Research Seminar

Competition policy and consumer outcomes - Assessing the impact of mobile consolidation on innovation and quality

By Xavier Pedrós  

Host: Miquel Oliver

Abstract

November,29th 

15:30

Room 55.309

PhD Research Seminar

A brief 18 year career in music understanding

By Brian Whitman (formerly co-founder of The Echo Nest, Principal Scientist of Spotify)

Host: Xavier Serra

Abstract

November,28th 

15:30

Room 55.309

Invited Research Seminar

Analyzing and Exploiting User-generated Listening Data for Listener Modeling and Music Recommendation

By Markus Schedl (Department of Computational Perception, Johannes Kepler University, Linz )

Host: Xavier Serra 

Abstract

November,16th 

15:30

Room 55.309

PhD Research Seminar

Gender and Science Journal Club

By Irene Torres , Adrián Ponce and Elisa Ruiz 

Host: Aurelio Ruiz

Abstract

November,16th 

15:30

Room 52.s29

Invited Research Seminar

Challenges in Uncertainty Quantification in Medical Image Registration

By Hervé Delingette (INRIA in Sophia Antipolis, Francia)

Host: Miguel Angel Gonzalez Ballester

Abstract

Medical Image Registration is the key ingredient in  many aspects of computer-aided decision medical support systems. For instance it is a standard processing step in many pipelines for medical imaging, abd computer assisted surgery. As such registration would benefit from the development of principled strategies to analyze its output and quantify its uncertainty. In this talk, a generic Bayesian method for non rigid image registration will be presented leading towards an estimation of registration uncertainty. The accuracy of this uncertainty is then evaluated with  an (asymptotically) exact inference scheme based on reversible jump Markov Chain Monte Carlo (MCMC) sampling. Challenges for a tractable uncertainty estimation will be reported at this occasion.

November,15th 

15:30

Room 55.410

Invited Research Seminar

The Bioengineering Initiative at Children’s National Health System: Device Development for the Pediatric Environment

By Kevin Cleary (Children's National Medical Centre, Washington DC)

Host: Miguel Angel Gonzalez Ballester

Abstract

November,14th 

15:30

Room 55.309

Invited Research Seminar

Parsing and Grammatical Inference as Constraint Solving

By Dr. Veronica Dahl

Host: Horacio Saggion

Abstract

We describe devised a new, linguistically motivated model of parsing and grammatical inference- the Womb Grammar Model (WGM)-, which allows the application of resources and tools of one language to another, making it possible to study under-resourced languages and dialects, or specialized uses of a same language, e.g., tweeter or SMS (short message service). Our model can particularly help linguists keep up with the rate of grammar discovery needed in our Babel-ish world: just as human wombs can generate any race from an appropriately fertilized egg, our linguistic "womb" can generate, from a language’s known grammar (or from a more abstract, “universal” grammar), the grammar of a different language, given only a set of positive, correct, and representative input sentences of that language, plus its associated lexicon. Our model has received peer validation as a model of parsing and of syntax inference with good semantics integration possibilities. Several experimental implementations  have supported applications to efficient, failure-based parsing, second language tutoring, and language acquisition. Visualization interactions were also studied. A strong proof-of-concept was provided through my student Ife Adebara’s thesis on inferring the grammar of a subset of Yoru`ba ́ noun phrases from that of an English noun phrase grammar. This work demonstrates that the WGM works well even with source languages not closely related to the target.

Biography

Verónica Dahl is an Argentine/Canadian computer scientist who is recognized as one of the 15 founders of the field of logic programming. She has contributed over 100 scientific publications in the fields of computational linguistics, deductive knowledge bases, computational molecular biology and web based virtual worlds. She has received numerous scientific awards-  such as the Calouste Gulbenkian Award for Science and Technology-, and a few literary awards as well. Her greatest ambition is to help bridge the gap between the formal and the humanistic sciences, in the hopes that this will be conducive to an overall more balanced world. She is presently Professor Emeritus at Simon Fraser University. Her research is supported by NSERC.

November,13th 

15:30

Room 55.309

PhD Research Seminar

Women in AI, Engineering and Computer Science  

By Dr. Veronica Dahl

Host: Aurelio Ruiz

Abstract

Women constitute a majority in the present world, yet are grossly underrepresented in many science. We discuss the societal causes of this curious phenomenon focusing of female underrepresentation in Engineering and in Computing Science -in particular in AI-, from a multicultural perspective. We contrast the specific cases of Argentina, Spain and Canada, trying to shed light upon the causes of female underrepresentation. As well, we discuss possible actions that could heal these causes and could help the integration of women into these fields.

Biography

Verónica Dahl is an Argentine/Canadian computer scientist who is recognized as one of the 15 founders of the field of logic programming. She has contributed over 100 scientific publications in the fields of computational linguistics, deductive knowledge bases, computational molecular biology and web based virtual worlds. She has received numerous scientific awards-  such as the Calouste Gulbenkian Award for Science and Technology-, and a few literary awards as well. Her greatest ambition is to help bridge the gap between the formal and the humanistic sciences, in the hopes that this will be conducive to an overall more balanced world. She is presently Professor Emeritus at Simon Fraser University. Her research is supported by NSERC.

November, 9th 

15:30

Room 54.006

PhD Research Seminar

Mendeley reference manager and research network

By Gemma Álvarez

Host:  Aurelio Ruiz

Abstract

During the presentation we will talk about:

Mendeley as a reference manager:

​- How to import your documents and references into your ​own library

​- How to generate citations and bibliographies with MS Word and LibreOffice Plug-ins

Mendeley as a research network:

- Connect and collaborate with researchers

- Your professional researcher profile

IMPORTANT: Previous tasks to do before the session (we recommend you to bring your on laptop. Please fill this form if you plan to attend):

1. Create your Mendeley account.

2. Access the upgraded Mendeley Institutional Edition (MIE) as a UPF's member to get the advantages.

3. Download the Mendeley Desktop version into your device.

November, 9th 

12:30

Auditorium

DTIC-MdM Research Seminar

Open Debate Machine Learning and Human Behaviour

Within the context of the HUMAINT project and the World Science Day for Peace and Development (November 10th), we will have the opportunity to start an open discussion about aspects linked to the relations between technological development and human behaviour, related to  some of the research carried out at our ICT Department.

Program:

• Introduction by Emilia Gómez and the new JRC project HUMAINT:"machine learning and human behaviour"
• Vladimir Estivill - Ethical decision making in robots
• Carlos Castillo - algorithmic bias in machine learning, related to the project awarded by Data Transparency Lab
• Xavier Serra, cultural bias in computational models, linked to the work at the ERC project CompMusic
• Open discussion.

Followed by some snacks at the hall of the auditorium to continue discussions (to make an estimation of attendants, please leave your name and surname in this doodle)

October,26th 

15:30

Room 55.410

PhD-MdM Research Seminar

DKPro Core - With Reusable and Interoperable Components Towards Reproducible Experiments

By Dr. Richard Eckart de Castilho

Host: Jens Grivolla

Abstract

Vast amounts of information are stored in unstructured formats such  as audio, video, or plain text. The field of Natural Language Processing (NLP) addresses   this by providing not only methods but also many ready made tools to analyze  text and to structure it in many ways ranginge from simple segmentation to complex  semantic analysis. However, it is often difficult to incorporate these tools into new experimental setups because the complexities of installation, configuration,  data transformation and interoperability in general are often underestimated.  This negatively affects the reproducibility of such experiments.  This talk will give an introduction to doing NLP with DKPro Core and the DKPro   eco-system with a focus on the approaches employed to promote and facilitate   reproducible experiments and the development of NLP-based applications.

Biography

About the speaker: Dr. Richard Eckart de Castilho is a senior research and technical lead at the Ubiquitous Knowledge Processing Lab, TU Darmstadt, the lead developer of DKPro Core,   member of the Apache UIMA project, and maintainer of the Apache uimaFIT library [1].  [1] https://www.ukp.tu-darmstadt.de/people/senior-staff/richard-eckart-de-castilho/

October,25th 

16:30

Room 55.309

Invited Research Seminar

Decoding Neural Drive for New Generation of Neurorehabilitation Systems

By Ivan Vujaklija 

Host: Laura Becerra-Fajardo

Abstract

The generation of a movement is the combination of discrete events (action potentials) generated in the brain, spinal cord, nerves, and muscles. These discrete events, discharged in the various parts of the neuromuscular system, constitute the neural code for movements. Recording and interpretation of this code provides the means for decoding the motor system. Currently we are able to detect and process the available information content only partially and this limitation impedes us to establish widely applicable man-machine interface systems. I will present the developments done within the ERC funded project “DEMOVE”, and recently at the Bioengineering Department at Imperial College London on advanced electrode systems for in-vivo electrophysiological recordings in humans and new computational methods/models for extracting functionally significant information on human movement. The focus of these developments is on providing the link between the cellular mechanisms and the behaviour of the whole motor system in order to establish clinically viable and effective neurorehabilitation technologies.

Biography

Ivan Vujaklija received the engineering degree in Electrical Engineering and Computer Science from the University of Belgrade, Serbia, in 2011, and the M.Sc. degree in Biomedical Engineering from the University of Lübeck, Germany in 2013. He obtained his PhD degree in human medical sciences at the University of Göttingen, Germany in 2016 while working as a research assistant at the Institute of Neurorehabilitation Systems at the University Medical Center Göttingen, Georg-August University, Germany. From 2012 until 2014 he worked for Ottobock Healthcare GmbH, Duderstadt, Germany, one of the world leading prosthetic manufacturers. In 2014 and 2015 he was a research fellow at Arizona State University and Medical University of Vienna respectively. In 2017 he moved to the Department of Bioengineering at the Imperial College London to work as a research associate. His research interests include bio-signal processing, advanced control algorithms, robotics, translational neurorehabilitation and neural control of movement.

October,19th 

15:30

Room 55.309

PhD Research Seminar

Introduction to PhD Seminars

By Aurelio Ruiz Garcia 

September,14th 

15:30

Room 55.309

Invited Research Seminar

Large Research Infrastructures and managing enormous amounts of data: The TOMCAT experience at the Swiss Synchrotron

By Anne Bonnin on behalf of the TOMCAT team

Host Bart Bijnens

Abstract

Nowadays, Tomographic Microscopy is a well-known technique used to visualize and study in 3D internal structure and material properties of a variety of samples. Thanks to the photon density reached by third generation synchrotron radiation large-scale facilities, the high-brilliance and high–coherence of synchrotron light enables the detection of micron-sized details in millimetre-sized samples and within few minutes.

The TOMCAT (TOmographic Microscopy and Coherent rAdiology experiments) Beamline of the Swiss Light Source offers such unique opportunity, thus providing routinely fast, non-destructive, high resolution, quantitative investigations on a broad kind of materials. With the recent advanced technologies, 3D datasets can even be acquired within seconds involving the acquisition of hundreds of data to cover cm-sized samples. Indeed, the acquisition rate can reach up to 8GB per second of image data. Therefore, the task of collecting, transferring, storing, sharing and analysing such amount of tomographic big data leads to the need of new robust solutions. Within the context of a selected big scale projects, we will present in this talk the current status of the data management from acquisition to storage and the future challenges that X-PCI implies.

Biography

Anne Bonnin studied physics and material science at the University Lyon 1. Graduated of a master in theoretical physics and a master in applied physics, she then received her PhD in Physics from the National Institute of Applied Sciences of Lyon (INSA-Lyon, France).

During her PhD, she studied the physical properties of materials in order to discriminate them by means of dual-energy tomography, diffraction and Compton scattering. After graduating with the highest honors, she spent four years in the X-ray imaging group at the European Synchrotron Radiation Facility (Grenoble, France), first in collaboration with the CEA-Cadarache on a project to developed diffraction tomography (XRD-CT) at the ID22Ni beamline (now ID16b) and then on projects in paleontology and material science at the ID19 beamline. In April 2014, she joined the X-ray tomography group at the TOMCAT Beamline (Villigen, Switzerland) to be in charge of the TOMCAT nanoscope development. As local contact, she collaborates with international researchers in material sciences, paleontology and biomedical applications.

Specialized in X-ray imaging (micro and nano-tomography, single-distance phase-retrieval) and powder diffraction (XRD-CT), her research interests focus on the development of new imaging methods to characterize materials and understand their behavior. Finally, she leads the Heart Imaging Project at PSI, which is part of an international collaboration aiming at better describing and understanding the cardiac remodelling processes in rat and human hearts.

September,13th 

15:00

Room 55.309

Invited Research Seminar

By Justin Salamon

Host Emilia Gomez

Abstract

Internet of Things (IoT) sensor networks, and particularly acoustic sensor networks, open up new opportunities in large-scale, data-driven analyses of acoustic environments. Examples include (but are not limited to) monitoring noise sources in urban environments or monitoring the migration of bird species based on their vocalizations. In both cases, it is necessary to identify and classify specific sound events in an audio signal based on their source, be it jackhammers and car honks or different species of migrating warblers. In this talk I'll present our work on machine listening algorithms for source identification in acoustic environments as part of the Sounds of New York City (SONYC) and BirdVox projects. Finally, I will also give a brief overview of the tools/resources we have been developing for sustainable MIR research including JAMS, mir_eval, Scaper and Massage.

Biography

Justin Salamon is a Senior Research Scientist at New York University's Music and Audio Research Lab (MARL) and Center for Urban Science and Progress (CUSP). He obtained a B.A. (Hons.) in Computer Science from the University of Cambridge, UK, in 2007, and an M.Sc. (2008) and Ph.D. (2013) in Computer Science from the Music Technology Group (MTG) of Pompeu Fabra University (UPF), Barcelona, Spain, specializing in sound and music computing. He was a visiting researcher at the Sound Analysis-Synthesis research team of the Institut de Recherche et Coordination Acoustique/Musique (IRCAM), Paris, France. His research interests include digital signal processing, machine learning and data mining, with applications to machine listening, music information retrieval, bioacoustics and environmental sound analysis. For further information and a full list of publications please see: www.justinsalamon.com

September,7th 

15:00

Room 55.309

Invited Research Seminar

Language Modelling for Automatic Music Transcription

By Andrew McLeod

Host Sergio Oramas

Abstract

Abstract: Automatic music transcription (AMT) is defined as the process of converting an acoustic music signal into some form of human- or machine-readable musical notation. Advances in automatic music transcription performance have been disappointingly slow, with accuracy still falling well below that of human experts at the task. In this presentation, I argue that this problem cannot be sufficiently solved without the use of some sort of music language model, trained on symbolic music data. To that end, I first present an HMM-based voice separation model which works on symbolic music data, both quantized and performed, and also show results of integrating it with an acoustic pitch-detection model for a cappella voice separation from an audio signal. I also discuss work on metrical structure (time signature) detection based on rhythmic analysis using a lexicalized PCFG, arguing that such a grammar can capture the long-range dependencies of musical rhythms. Finally, in future work, I look ahead to the integration of models such as these into a full AMT system with an acoustic model.

Biography

Andrew McLeod is a PhD student under the advisement of Mark Steedman at the University of Edinburgh, School of Informatics, Institute for Language, Cognition and Computation (ILCC), set to graduate in early 2018. He received a Bachelors and Masters in Mathematics and Computer Science from Emory University in the USA in 2013. His PhD subject is automatic music transcription, specifically working with symbolic music data to create a music language model based on music theory for the task.