In’Tro (May 30th 2023, 1.30pm): On the Random Subset Sum Problem and Neural Networks – Emanuele Natale (COATI)

Abstract:
The Random Subset Sum Problem (RSSP) is a fundamental problem in mathematical optimization, especially in understanding the statistical behavior of integer linear programs.
Recently, the theory related to this problem has also found applications in theoretical machine learning, providing key tools for proving the Strong Lottery Ticket Hypothesis (SLTH) for dense neural network architectures. In this talk, I will give a brief overview of this research direction and present my recent joint work that pushes the application of RSSP further by providing a proof of the SLTH for convolutional neural networks.

Bio:
Before accepting a CNRS position in 2018 and joining the I3S Lab and INRIA d’Université Côte d’Azur, Emanuele Natale has been a fellow of the Simons Institute for the Theory of Computing in the Brain and Computation Program and a postdoctoral fellow at the Max Planck Institute for Informatics. In 2019, He has received the Best Italian Young Researcher in Theoretical Computer Science award by the Italian Chapter of the European Association of Theoretical Computer Science, from which he also received the Best PhD Thesis in Theoretical Computer Science in 2017. In 2016, he has been a recipient of the Best Student Paper Award at the European Symposium on Algorithms.
Emanuele Natale’s research originally focused on the mathematical analysis of simple distributed probabilistic algorithms that allow multi-agent systems to solve global coordination tasks, with applications spanning machine learning, sociology and theoretical biology. More recently, his research interests have shifted mainly to neuroscience and machine learning, with a focus on the role of sparsification in neural networks.

The presentation will be in English and streamed on BBB

In’Tro (February 10th 2023, 1.30pm): AI for species identification that explains like an expert. – Diego Marcos (ZENITH)

Title: AI for species identification that explains like an expert.
Abstract: In this project the aim is to enable large scale biodiversity monitoring by citizen scientists by developing Explainable Machine Learning methods that reason like a taxonomist, explicitly detecting relevant traits on a specimen’s image and reaching a species identification conclusion based on them. This will help scientists obtain valuable data from rare or undescribed species, make use of low quality real-world images and make it easier for everyone to become an amateur naturalist, thus raising awareness about biodiversity and the rapid pace at which we are losing it. We will jointly use Computer Vision (CV) and Natural Language Processing (NLP) methods to extract the relevant visual features and model taxonomic descriptions.
First, we will leverage the vast amount of structured textual species descriptions that are available online, such as in Wikipedia, to train a first NLP model, starting with a pre-trained transformers-based model, that will be used to discriminate between text belonging to species descriptions. This first model will be used to further increase the amount of textual descriptions by parsing additional websites that contain species descriptions.
In a second step, the textual descriptions will be analyzed in terms of part-of-speech in order to understand the different life stages (e.g. egg, hatchling, immature, female adult, male adult, etc.) and parts (e.g. leaf, stem, flower, bark, fruit, etc.) that are being described, extract their corresponding attributes, and identify relative descriptions in case another species is mentioned for comparison.
The last step requires developing a method for linking this knowledge graph to a Computer Vision model. To do this we will leverage the millions of images annotated with the species names, belonging to tens of thousands of species, that are freely available at the Global Biodiversity Information Facility (GBIF).

The presentation will be in English and streamed on Webex.

In’Tro (12 décembre 2022, 1.30pm): How? And Why? in biomedical data assimilation and modelling – Irène Balelli (EPIONE)

Replay Irène’s presentation

Summary

In this talk I will discuss the need of questioning the “How?” and the “Why?” when building data assimilation methods and predictive models, especially in the medical context, for more interpretable and trustable algorithmic prescriptions. This will be the occasion to present and discuss some of my works, embedded in this philosophy.

Biography

I am a research scientist (ISFP – Inria Starting Faculty Position) at Centre INRIA d’Université Côte d’Azur, in the EPIONE team. I am interested in developing mathematical and statistical models for computational biomedicine. Current research interests include Bayesian learning, Mechanistic modeling, Federated learning, Population dynamics, In-silico trials and Causality.

Previous to that, I received my Ph.D. from Université Paris 13, Sorbonne Paris Cité in 2016, with a dissertation titled Mathematical foundations of antibody affinity maturation, where I focused on the development of a mathematical framework based on graphs, to model antibody affinity maturation of B-cells.
In 2017 I have joined the SISTM team (INSERM U1219 Bordeaux Population Health and INRIA) as a postdoc. As part of the EBOVAC European consortia, I focused on the mechanistic modeling of the immune response to a prime-boost vaccination strategy against Ebola virus, developed by Janssen.
In 2020 I have joined the EPIONE team at Centre Inria d’Université Côte d’Azur as a postdoc, where I got interested in the development of Bayesian learning methods in a federated setting, with a particular attention to health applications.

irene balleli

irene balleli

 

The presentation will be in English and streamed on BBB.

In’Tro (26 septembre 2022, 1.30pm): All that glitters is not gold: Decision Making Under Noisy Observations – Samir M. Perlaza (NEO)

Replay Samir’s presentation

 

In this short talk, the impact of noisy observation of data in decision making is studied through the lenses of information and game theories. The conclusions of this talk are general and span several applications in engineering, including supervised machine learning, data transmission over networks, and state estimation in electrical smart grids.

Samir M. Perlaza is a permanent member of the scientific staff at INRIA, the French Institute for Research in Computer Science and Applied Mathematics; an associate member of the Department of Mathematics (Laboratory GAATI) of the University of French Polynesia; and a visiting research collaborator in the Department of Electrical and Computer Engineering at Princeton University. He received the M.Sc. and Ph.D. degrees from École Nationale Supérieure des Télécommunications (Telecom ParisTech) in 2008 and 2011, respectively. From 2008 to 2011, he was also a research engineer at France Télécom – Orange Labs (Paris, France). He has held long-term academic appointments at the Alcatel-Lucent Chair in Flexible Radio at Supélec; Princeton University and the University of Houston.

Dr. Perlaza’s research interests are in the areas of information theory, game theory, data sciences, and their applications in wireless networks, power systems, and artificial intelligence. Among his publications in these areas is the recent book ‘‘Advanced Data Analytics for Power Systems’’ (Cambridge University Press, 2021).

Dr. Perlaza has served as an Editor of the IEEE TRANSACTIONS ON COMMUNICATIONS, the IET Smart Grid Journal, and Frontiers in Communications and Networks.

Recognition of his work includes the Alban Fellowship and the Marie Sklodowska-Curie Fellowship, both from the European Commission.

 

The presentation will be in English and streamed on BBB.

In’Tro (13 juin 2022, 1.30pm) : Flat optics for future daily life applications – Mahmoud Elsawy (ATLANTIS)

Replay Mahmoud’s presentation

 

Abstract

Modern real-life technology demands miniaturization and boosting the performance of conventional optical components. Recently, the field of flat optics “Metasurfaces” has received remarkable attention owing to its capability of controlling all the light properties in a short propagation distance with high resolution. Metasurfaces are optical components consisting of sub-wavelength spaced nanostructures which introduce highly resolved phase, amplitude, and polarization changes on the incoming wavefront. Owing to the versatility and the capabilities of metasurfaces, many exotic and peculiar applications ranging from subwavelength optical microscopy to augmented reality have been demonstrated. Yet, the strong light-matter interactions at the nanoscale require advanced modelling techniques to boost the performance towards the industrial level.
In this talk, I will present our recent activities in the field of metasurfaces. The underlying physical mechanism will be introduced associated with the modelling challenges. Besides, the optimized devices ranging from passive beam deflectors to broadband metalenses together with the perspective trends and opportunities as dynamical control and light-emitting metasurfaces will be highlighted.

Biography

Mahmoud Elsawy is permanent researcher (ISFP) at Atlantis project-team, Inria, Sophia Antipolis. His research focuses on numerical modelling and optimization of metasurfaces and complex nanophotonic devices. He received his PhD degree at the University of Aix-Marseille, France in 2017 with a specialization in Optics, Photonics, and image processing. The topic of the PhD was related to modelling and improvements of complex nonlinear plasmonic waveguides that can be fabricated and characterized experimentally.

In’Tro (2 mai 2022, 1.30pm) : Computational Brain Connectivity Mapping – Samuel Deslauriers-Gauthier (ATHENA)

Watch Samuel’s presentation

Title

Computational Brain Connectivity Mapping

Abstract

In the first part of this talk, I will present my background, research interests, and the project that allowed me to join Inria: the ERC CoBCoM. In the second part, I will provide an overview of my current research, which is on the estimation and quantification of whole brain networks. I will start by introducing the tools of the trade: magnetic resonance imaging (MRI), diffusion MRI, functional MRI, electroencephalography, and magnetoencephalography. Then, I will discuss how these imaging techniques provide complementary views of the brain and how we can combine them to gain new insights about the dynamics of the brain. Finally, I will conclude by discussing the open problems and challenges of non-invasively characterizing a network composed of 80 billion neurons connected by 160 000 km of axons.

Bio

Samuel Deslauriers-Gauthier obtained a master from the l’École de Technologie Supérieure de Montréal in 2010 and a doctorate in signal processing from the Nanyang Technological University of Singapour in 2014. After a postdoctoral fellowship at l’Université de Sherbrooke, he joined Inria in 2017 as a Starting Research Position in the ERC Advanced Grant of Rachid Deriche and in September 2021 he obtained a Inria Starting Faculty Position. His research interests are centered around signal modelling and processing in medical imaging, more specifically in electroencephalography, magnetoencephalography, magnetic resonance imaging, and electromyography. His recent work has focused on recovering the information flow in the white matter of the brain from multi-modal imaging and understanding the link between the brain’s structure and function.

 

The presentation will be in English and streamed on BBB.

In’Tro (7 mars 2022, 1.30pm) : Reasoning with Ontologies on Knowledge Graphs – David Carral (GRAPHIK)

Watch David’s talk

Summary

In a couple of days, I will be giving you a short intro presentation, which is structured in three parts:

  1. First, I will tell you a bit about my career and research interests. More precisely, I will talk about the researchers and institutions with whom and where I have worked, respectively; as well as the venues where I usually publish.
  2. Second, I will describe and provide a high-level motivation for my research. The goal provide an accessible introduction to the work I do and try to convince you of its usefulness.
  3.  Finally, I will  also provide you with some links and materials to know more about my work (in case you’re interested!). I will also leave my contact information so you can ask me any questions directly.

Slides of the presentation

Biography

Hi! My name is David Carral and I am a CRCN researcher working in the GraphIK Inria team, which is based in Montpellier. Broadly speaking, I am interested in the study of logical languages (mostly first-order logic, existential rules, and Description Logics) and their theoretical/computational properties. Moreover, I am also interested in the implementation of efficient reasoning algorithms for these logical languages. For more information about what I am doing these days, have a look at my website: https://www-sop.inria.fr/members/David.Carral/

 

 

 

 

 

The presentation will be in English on BBB

In’Tro (7 février 2022, 1.30pm): What else is leaked when eavesdropping federated learning? – Chuan Xu (COATI)

Watch Chuan’s presentation

Abstract

Federated learning (FL) offers naturally a certain level of privacy, as clients’ data is not collected at a third party. However, maintaining the data locally does not provide itself formal privacy guarantees. An (honest-but-curious) adversary can still infer some sensitive client information just by eavesdropping the exchanged messages (e.g., gradients). In this talk, we will present a new model reconstruction attacks for federated learning, where a honest-but-curious adversary reconstructs the local model of the client. The success of this attack enables better performance of other known attacks, such as the membership attack, attribute inference
attacks, etc. We provide analytical guarantees for the success of this attack when training a linear least squares problem with full batch size and arbitrary number of local steps. One heuristic is proposed to generalize the attack to other machine learning problems. Experiments are conducted on logistic regression tasks, showing high reconstruction quality, especially when clients’ datasets are highly heterogeneous (as it is common in federated learning).

Short bio

Chuan Xu joined Université Côte d’Azur (UCA) as an associate professor (« maître de conférences ») in Sept. 2021 and she is a member of the I3S laboratory and of project-team COATI. Before that, she was a postdoctoral researcher working in the NEO team at Inria Sophia Antipolis from 2018 to 2021. She received her PhD in Computer Science from Université Paris-Saclay in Dec. 2017. Her research interests include : Distributed machine learning, privacy in federated learning and self-stabilizing distributed algorithms.

 

 

 

The presentation will be in English and streamed on BBB.

In’Tro (10 janvier 2022, 1.30pm) : Causal Inference with incomplete data – Julie Josse (Montpellier)

Slides from the presentation

 

Abstract

Causal inference aims to estimate the effect of an intervention on an outcome. It can be used to evaluate the effect of a public policy, a medical treatment, an advertisement, etc.
Randomized controlled trials (RCTs) are the gold standard in the field but they often suffer from sampling bias because they are not representative of the target population. We will present estimators that correct for the distributional shift and predict the effect of a treatment in a population using external observational data. We will also discuss how to handle missing information.

Abstract

Her first employment was in the statistics department of an Agronomy University (Agrocampus Ouest) where she was trained to « the French data analysis school » and had the opportunity to work closely with researchers from other departments and increases her interest in transversal studies. In the meantime, she prepared her PhD which was defended in 2010 and rewarded by the French Statistical Society as the best PhD in applied statistics. She has specialized in missing data, visualization and the nonparametric analyses of complex data structures. Her work was rewarded by a Marie Curie European Union grant in 2013 to increase her research potential and to spend a year at Stanford University. She spent a year as a researcher in INRIA before joining Polytechnique in 2016 as a Professor of Statistics. At Polytechnique, she was responsible of a master in data-sciences for business in collaboration with HEC. She has been a visiting researcher at Google Brain Paris, for a year (2 days a week) in 2019. In september 2020, she join Inria as an advanced researcher to set-up a team in data-science for health. She has published over 50 articles and written 2 books in applied statistics. Her experience on dealing with incomplete data is recognized by the community: she organized an ICML workshop, the MissData conference, created the Rmistatic website and she is often invited to give lectures to share her experience. Her vocation is to push methodological innovation to bring useful application of her research to the user in particular in bio-sciences and health. Her current research focuses on causal inferences techniques for personalized medicine. She leads a project with the Traumabase group dedicated to the management of polytraumatized patients to help emergency doctors making decisions. Julie Josse is dedicated to reproducible research with the R statistical software: she has developed packages including FactoMineR, denoiseR, missMDA to transfer her work, she is a member of the R foundation and of Rforwards to increase the participation of minorities in the community.

 

The presentation will be in English and streamed on BBB

In’Tro (13 décembre 2021, 13h30) : An introduction to Topological Data Analysis – Mathieu Carrière (DataShape)

Watch Mathieu’s presentation

 

Abstract

Topological Data Analysis (TDA) is a growing field of research at the intersection of data science and computational geometry and topology. It has encountered key successes in several different applications (ranging from cancer subtype identification in bioinformatics to shape recognition in computer vision, just to name a few), and become the landmark product of several companies in the recent years. Indeed, many data sets nowadays come in the form of point clouds embedded in very large dimensional spaces, yet concentrated around low-dimensional geometric structures that need to be uncovered. Unraveling these structures is precisely the goal of TDA, which can build descriptors that can reliably capture geometric and topological information (connectivity, loops, holes, curvature, etc.) from the data sets without the need for an explicit mapping to lower-dimensional space. This is extremely useful since the hidden, non-trivial topology of many data sets can make it very challenging to perform well for classical techniques in data science and machine learning, such as dimensionality reduction.

In this talk, I will provide a global overview of TDA, by introducing its main descriptors and by presenting the theoretical guarantees that they enjoy. I will also show how they can be efficiently computed in practice with the dedicated, open-source library GUDHI, and describe some applications where TDA proved useful.

Short bio

I did my PhD at Inria Saclay in the DataShape team, under the supervision of Steve Oudot, and a postdoc of two years in the Rabadán Lab, at the Department of Systems Biology of Columbia University, under the supervision of Raúl Rabadán. My research focuses on topological data analysis (TDA) and statistical machine learning (ML), with an application to bioinformatics and genomics. I contributed to the analysis of topological descriptors and their use in ML methods such as kernel-SVM or deep learning. My favorite languages are C++ and Python, but I also know a bit of R, Matlab and Java. I am also very familiar with Scikit-Learn and TensorFlow.

 

 

 

The presentation will be in English and streamed on BBB