CITI seminar – Rémi Bardenet (CNRS) – 14/11 at 12:15

Speaker: Rémi Bardenet (CNRS) is a recipient of a 2021 CNRS bronze medal and PI of the ERC Starting Grant Blackjack (https://rbardenet.github.io/).

Date: 14/11/2023

Time: 12h15

Place: Amphi Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: Monte Carlo integration with repulsive point processes

Abstract: Joint work with Adrien Hardy, Ayoub Belhadji, Pierre Chainais, Diala Hawat, and Raphaël Lachièze-Rey.
Monte Carlo integration is the workhorse of Bayesian inference, but the mean square error of Monte Carlo estimators decreases slowly, typically as 1/N, where N is the number of integrand evaluations. This becomes a bottleneck in Bayesian applications where evaluating the integrand can take tens of seconds, like in the life sciences, where evaluating the likelihood often requires solving a large system of differential equations. I will present recent results on variance reduction and fast Monte Carlo rates using interacting particle systems. The underlying idea is that to integrate a function with a handful of evaluations, one should evaluate the function at well-spread (random) locations, where “well-spread” means “so that one can benefit from the smoothness of the target function”.

journée de rentrée CITI

Aujourd’hui a lieu notre journée de rentrée du CITI, au programme présentation des nouveaux arrivants, interventions “Mixité des genres”, “Ecoanxiété”, et “Burn out”, présentation de la cellule DDRS du laboratoire, présentation des divers chantiers au niveau du CITI, échanges et moments de cohésion au musée Confluence avec entre autre un escape game.

CITI seminar – Alexandre Proutière (KTH) – 11/10 at 15:30

Speaker: Prof. Alexandre Proutière (KTH)

Date: 11/10/2023

Time: 15h35

Place: Room TD-C Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: Radio Network Optimization: A Bandit Approach

Abstract: In this talk, we demonstrate how to efficiently solve radio network optimization problems using a bandit optimization framework. We mainly consider the problem of controlling antenna tilts in cellular networks (so as to reach an efficient trade-off between network coverage and capacity). We start with the design of algorithms learning optimal antenna tilt control policies at a single base station, and formalize this design as a Best Policy Identification (BPI) problem in contextual Multi-Arm Bandits (MABs). We then consider coordinated antenna tilt policies at several interfering base stations, and formalize the design of algorithms learning such policies as a multi-agent MAB problem. In both settings, we derive information-theoretical performance upper bounds satisfied by any algorithm, and devise algorithms approaching these fundamental limits. We illustrate our results numerically using both synthetic and real-world experiments.

This is a joint work with Filippo Vannella (KTH / Ericsson Research) and Jaeseong Jeong (Ericsson Research). The talk is based on the following papers:

https://arxiv.org/pdf/2201.02169.pdf (IEEE Infocom 2022)
https://proceedings.mlr.press/v202/vannella23a/vannella23a.pdf (ICML 2023)
Statistical and computational trade-off in multi-agents multi-armed bandits (to appear in NeurIPS 2023)

PhD Defence: “Spatio-temporal Data Analysis for Dynamic Phenomenon Monitoring Using Mobile Sensors”, Ichrak Mokhtari, Amphi Chappe Building, 6th of June 2023 at 10 AM

The defense will take place on Tuesday 6th June at 10 AM in the Heidi Lamarr building (Amphi Chappe), Insa-Lyon, Villeurbanne.

Title

Spatio-temporal Data Analysis for Dynamic Phenomenon Monitoring Using Mobile Sensors

Abstract

Monitoring air pollution in emergencies (industrial accidents, terrorist attacks, volcanic eruptions, etc.) is of utmost importance given the dramatic effects that the released pollutants can cause on both human health and the environment. In these situations, the pollution plume is strongly dynamic leading to a fast dispersion of pollutants in the atmosphere. Thus, the need for real-time response is very strong and a solution to get a precise mapping of pollution dispersion is needed to mitigate risks.

This thesis focuses on the monitoring of air pollution in emergencies using a fleet of drones, with three main areas of investigation: 1) the spatiotemporal prediction of pollution plume evolution; 2) the optimal planning of drones trajectories to improve pollution mapping; and 3) the development of a generic solution for dynamic pollution monitoring. Through this work, we
propose a spatio-temporal Deep Learning model for multi-point forecasting of pollution concentrations, and we built upon several uncertainty quantification techniques to make it more trustworthy. Furthermore, we examine and identify the main challenges related to the underlying phenomena as well as its emergency context, and we suggest a new systemic approach for monitoring dynamic air pollution based on aerial sensing, that combines Deep Learning approaches, with Data Assimilation techniques, while relying at the same time on adequate path planning
strategies. The framework is then extended to address the data scarcity issues encountered in such situations through a transfer learning solution based on physical models. Finally, we meticulously address the drones’ path planning problem to improve the air pollution mapping quality, and we provide a Multi-Agent Reinforcement Learning solution.

Keywords: Monitoring Dynamic Air Pollution, Spatio-temporal Forecasting, Deep Learning, Multi-Agent Reinforcement Learning, Drones.

Jury

NATALIZIO, Enrico Professeur des universités TII, Abu Dhabi Rapporteur
MITTON, Natalie Directrice de recherche INRIA Rapportrice
GARCIA, christophe Professeur des universités INSA-LYON Examinateur
CARNEIRO Viana, Aline Directrice de recherche INRIA Examinatrice
LABENTALl, Bérengère Directrice de recherche Université Gustave Eiffel Examinatrice
RIVANO, hervé Professeur des universités INSA-LYON Directeur de thèse
BECHKIT, Walid Maître de conférences INSA-LYON Co-directeur de thèse

CITI seminar – François Michaud (Université de Sherbrooke) – 26/05 at 11:00

Speaker: François is Prof. at University of Sherbrooke (Canada), and leading the IntroLab at the 3IT institute.

Date: 26/05/2023

Time: 11h00

Place: Amphi Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: Working Toward Human-Robot Symbiosis

Abstract: Human-robot symbiosis implies developing robotic systems that can collaborate with humans in open and ‘messy’ conditions, meaning unpredictable real-life settings, such as those found in assistive healthcare and work environments. Achieving human-robot symbiosis requires humanizing the sensing, perception, reasoning, and actuating capabilities based on evaluating human safety, well-being, acceptability, and usability. Researchers need to adopt a holistic approach enabling robots to seamlessly ‘see, hear and be’ in everyday settings, and design robots that are situationally balanced, in which complexity levels of sensory, motor, and artificial intelligence (AI)/cognitive capabilities are matched with the environment and people. This presentation addresses an overview of interactive robots and systems developed at IntRoLab, Université de Sherbrooke, involving compliant actuators, assistive robot platforms, telepresence robots, vision-based SLAM, drone intrusion, weed remoal robot, robot companion and robotic living labs.

Bio: François Michaud, Ph.D., is an engineer and full professor in the Department of Electrical and Computer Engineering at the Université de Sherbrooke, in Québec Canada. Holder of the Canada Research Chair in Mobile Robotics and Intelligent Autonomous Systems from 2001 to 2011, his research activities are aimed at integrating intelligent autonomous robotic systems into everyday operating conditions, to improve the well-being of people. His expertise is in human-robot interaction, assistive robotics, telepresence robotics, robot design and cognitive robotics. He has extensive experience in initiating and conducting interdisciplinary and intersectoral research projects involving collaborators in physiotherapy, occupational therapy, agriculture, child psychiatry, education, cognitive science, manufacturing, arts and automotive. He has published over 225 peer-reviewed papers in journals and international conferences (h-index 50), has been awarded 8 patents, has five significant distributed open source (software and hardware) contributions used by the robotics community, and has received funding over 50 M$ CAD supporting a broad range of research initiatives. He is the founding director of the Interdisciplinary Institute for Technological Innovation (3IT) (2008 – 2015), co-founder of Robotique FIRST Quebec (2010 – ), founder of Quebec Strategic Cluster INTER (Interactive Technologies in Rehabilitation Engineering) (2011 – ), and co-founder of a graduate training program CoRoM (COllaborative RObotics for Manufacturing). He is the Editor-in-Chief of Springer Nature Current Robotics Reports. He is also the founding director of the Bachelor of Robotics Engineering Program (2017 – ) at the Université de Sherbrooke, the first and only one in Canada.

CITI seminar – Frédéric Prost (CITI) – 04/05 at 12:15

Speaker: Frédéric Prost is an associate professor hosted by the CITI laboratory: https://lig-membres.imag.fr/prost/

Date: 04/05/2023

Time: 12h15

Place: Amphi Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: AI Risk: an Historical Perspective through the Game of Chess

Abstract: The game of chess as always been viewed as an iconic representation of intellectual prowess. Since the very beginning of computer science, the challenge of being able to program a computer capable of playing chess and beating humans has been alive and used both as a mark to measure hardware/software progresses and as an ongoing programming challenge leading to numerous discoveries.

Recent advances in AI (GPT-4, Midjourney etc.) have raised an important discussion on the societal risk of AI. Several articles, and a recent request for a moratorium of 6 months in AI research (signed by thousands of AI researchers and influential figures from politics, economics etc. https://futureoflife.org/open-letter/pause-giant-ai-experiments/) have been published in the last few weeks.

In this talk I will tackle the issue of AI risk from an historical perspective. In chess the AI are stronger than humans for more than a quarter of century (Kasparov loss to Deep Blue dates back to 1997). We can use this history as a proxy to discuss fears/hopes and to explore what happens when AI develops super human capabilities (for instance how the chess community has evolved). Of course the range of the chess is wolrd is limited in its scope with relation to LLMs. But it is interesting and justified because, when one is trying to study a complex phenomenon, isolating experiments in a lab allows the reduction of noise.

Bio: Frédéric Prost est MdC à l’université Grenoble Alpes et au laboratoire LIG, hébergé au CITI. Il a principalement travaillé dans la théorie des langages de programmation (réécriture de graphes, sémantique des langages d’interrogation des BD graphes) et les problématiques de confidentialité (analyse de non interférence, anonymisation de bases de données graphe).

CITI seminar – Khac-Hoang Ngo (Chalmers University of Technology) – 20/04 at 10:00

Speaker: Khac-Hoang Ngo (Chalmers University of Technology)

Date: 20/04/2023

Time: 10h00

Place: Amphi Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: Unsourced Multiple Access: An Information-Theoretic Analysis

Abstract: The drastic growth of the number of connected devices gives rise to the Internet of Things (IoT). Massive IoT connectivity targets a large number of low-cost, battery-limited, narrowband devices—meters, sensors, trackers, wearables—that transmit small data volumes in a sporadic and uncoordinated manner. These key features are captured by the unsourced multiple access (UMA) model proposed by Polyanskiy (2017), where all users transmit their messages using the same codebook and the decoder returns an unordered list of messages. In this talk, we introduce the UMA framework and Polyanskiy’s random-coding achievability bound for the Gaussian UMA channel. We then extend this bound to the case of random and unknown number of active users, thus fully account for the random user activity. Finally, we investigate a setting where, on top of the standard UMA messages, the users transmit a common alarm message that needs to be decoded with higher reliability; we thereby study the coexistence of massive and critical IoT.

Bio: Khac-Hoang Ngo (https://khachoang1412.github.io/) received the B.Eng. degree (Hons.) in electronics and telecommunications from Vietnam National University, Hanoi, Vietnam, in 2014; and the M.Sc. degree (Hons.) and Ph.D. degree in wireless communications from CentraleSupélec, Paris-Saclay University, France, in 2016 and 2020, respectively. His Ph.D. thesis was also realized at Paris Research Center, Huawei Technologies France. Since September 2020, he has been a postdoc at Chalmers University of Technology, Sweden under a project funded by the MSCA Individual Fellowship. His research interests include wireless communications and information theory, with an emphasis on massive random access, edge computing, MIMO, noncoherent communications, coded caching, and network coding. He received the Honda Award for Young Engineers and Scientists in Vietnam in 2013 and the “Signal, Image & Vision Ph.D. Thesis Prize” by Club EEA, GRETSI and GdR-ISIS, France, in 2021.

CITI seminar – Pablo Rauzy (Université Paris 8) – 10/02 at 12:15

Speaker: Pablo Rauzy (Université Paris 8)

Date: 10/02/2023

Time: 12h15

Place: Amphi Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: Promesses et (dés)illusions : une introduction technocritique aux blockchains

Abstract: Une blockchain est un registre distribué et immuable dans lequel sont écrites des informations qui font consensus. ». Dans cet article, nous commencerons par donner du sens à cette phrase et à l’ensemble des termes qui y sont employés, en nous efforçant quand c’est nécessaire de rendre accessibles les notions informatiques (comme la décentralisation, la distribution, l’immuabilité, ou le consensus) et le fonctionnement technique des outils cryptographiques sous-jacents (comme les condensats, les signatures, ou la preuve de travail ou d’enjeu). L’objectif de cette introduction sera d’atteindre une compréhension réelle de ce qu’est une blockchain.

Ainsi équipé·es, nous discuterons ensuite de ce que les blockchains permettent effectivement d’accomplir, et donc surtout ce qu’elles ne permettent pas. Nous questionnerons alors les utilisations qui en sont proposées en nous concentrant sur des cas d’usage typiques des blockchains que nous étudierons plus en détails : les « cryptomonnaies » bien sûr, la certification de documents (avec l’exemple des diplômes), et nous mentionnerons également le cas des NFT. Cela nous permettra en conclusion de questionner de manière générale le caractère d’« innovation de rupture » que l’on associe souvent à cette technologie.

Bio: Pablo Rauzy est maître de conférences en informatique à l’Université Paris 8 et membre de l’équipe PASTIS du LIASD. Avant sa prise de poste à Paris 8 en 2016, il a été doctorant dans l’équipe SEN de Télécom ParisTech puis post-doctorant dans l’équipe Inria Privatics au CITI. Ses travaux de recherche portent de manière générale sur la sécurité et touchent à différents aspects du domaine — privacy et contrôle, formalisation et modélisation, cryptographie et implémentation —, toujours avec la volonté d’une approche émancipatrice consciente du caractère non-neutre des sciences et technologies. C’est ce souci qui l’a conduit à devoir finalement s’intéresser aux blockchains dans le cadre de sa recherche, pour rendre accessible au plus grand nombre leur fonctionnement, leurs limites, et leurs dangers.

CITI seminar – Eleftherios Kofidis (University of Patras) – 24/01 at 14:00

Speaker: Eleftherios Kofidis (University of Patras)

Date: 24/01/2023

Time: 14h00

Place: Amphi Chappe/Lamarr, 6 avenue des arts, La Doua Campus

Title: Tensor methods and applications

Abstract: Tensor models have been well established as a natural and powerful way of representing systems and data that involve multiple aspects/dimensions. Assisted by their unique ability to unveil latent information through tensor decomposition methods, they have proved successful in numerous applications. In this talk, I will present some of our recent work and results on tensor methods, with example applications including wireless communications and remote sensing.

Bio: Eleftherios Kofidis received the Diploma (MEng) and Ph.D. degrees in 1990 and 1996, respectively, both from the Department of Computer Engineering and Informatics, University of Patras, Patras, Greece. From 1996 to 1998 he served in the Hellenic Army. In the period 1998 to 2000, he was a postdoctoral fellow at the Institut National des Télécommunications (INT), Évry, France (now Télécom SudParis). From 2001 to 2004, he was a research associate at the University of Athens, and adjunct professor at the Universities of Peloponnese and Piraeus, Greece. In 2004, he joined the Dept. of Statistics and Insurance Science, University of Piraeus, Greece, where he is now Associate Professor. He is also affiliated with the Computer Technology Institute & Press “Diophantus” (CTI), Greece. His research interests are in signal processing and machine learning, with applications including communications and medical imaging. He has served as technical program co-chair and in the organization and technical program committees of a number of conferences. Dr. Kofidis has served as Associate Editor in the IEEE Transactions on Signal Processing, the EURASIP Journal Advances in Signal Processing, and the IET Signal Processing journal.

CITI seminar – Prasenjit Mitra (Penn State) – 13/12 at 10:00

Speaker: Prasenjit Mitra (Penn State)

Date: 13/12/2022

Time: 10h00

Place: Amphi Chappe/Lamarr

Title: Secure Federated Learning: Lessons Learned, and Future Directions

Abstract: In this talk, I will introduce the topic of federated learning and discuss about its implications to computer security and machine learning. Federated learning has a wide range of applications in several areas where machine learning is gaining prominence but the need to preserve privacy arises. For example, hospitals prevented from sharing patient data can nevertheless collaborate to build a better model that utilizes data from multiple hospitals; next generation cars can share their data to enable dynamic, personalized, just-in-time preventive maintenance that can save customers money as well as make cars reliable, while preserving the privacy of individual drivers. The talk will survey the state-of-the-art for the technology, outline several open issues, and briefly mention our previous work on making federated learning solutions robust when adversaries attack. We will also briefly mention issues related to fairness and federated learning and how we can implement explainable and interpretable federated learning. The objective will be to raise questions that are of interest to the community that we can jointly address to improve the state-of-the-art in federated learning with respect to a wide range of attributes, e.g., computational complexity, security, privacy, fairness and equity, robustness, interpretability, etc.

Bio: Prasenjit Mitra is a Professor at The Pennsylvania State University and a visiting Professor at the L3S Center at the Leibniz University at Hannover, Germany. He obtained his Ph.D. from Stanford University in 2003 in Electrical Engineering and has been at Penn State since. His research interests are in artificial intelligence, applied machine learning, natural language processing, etc. His research has been supported by the NSF CAREER award, the DoE, DoD, Microsoft Research, Raytheon, Lockheed Martin, Dow Chemicals, McDonnell Foundation, etc. His has published over 200 peer-reviewed papers at top conferences and journals, supervised or co-supervised 15-20 Ph.D. dissertations; his work has been widely cited (h-index 60) and over 12,500 citations. Along with his co-authors, he has won the test of time award at the IEEE VIS and a best paper award at ISCRAM, etc.