- Talks by scientific committee members
- Scientific Highlights talks by LINCS members, Phd students, postdocs or engineers
- Survey talks by LINCS members
- Elevator pitch & poster session by Phd students & postdocs
Program
(click for printable version)
Thursday, July 10th, 2025 |
|
|
Coffee reception |
9:00/9:30 |
|
Opening by Sébastien Tixeuil (SU) |
9:30/9:35 |
|
Director speech by Daniel Kofman (IMT) |
9:35/9:45 |
|
1st LINCS Scientific Highlights |
9:45/10:00 |
|
1st Invited talk from the Scientific Committee Modeling the earth atmosphere is both a scientific and computational grand challenge. Much progress has been made in developing and implementing models that track the planet-scale evolution of the earth atmosphere and predict its response to various perturbations. Those models are unsurprisingly computationally onerous and running them at scale is feasible only through extensive parallelization. In this work, we target GEOS-Chem, an open-source software for simulating the earth atmosphere used by hundreds of researchers worldwide and capable of large-scale parallelism. GEOS-Chem inherent parallelism notwithstanding, leveraging it to realize effective speed-ups gives rise to interesting load-balancing and scheduling challenges. Of particular interest are the fact that computational workloads exhibit spatial and temporal variations and that the effectiveness of solutions to accommodate them vary based on the characteristics of the computational and communication platforms, e.g., whether data movement is realized through a combination of Infiniband and RDMA or Ethernet and (AWS) EFA affects the underlying trade-off between computations and communication. This talks explores those questions and the different approaches they gave rise to, and attempts to offer an illustration of the interplay between engineering and scientific investigations that effective solutions often require. * This is joint work with Daisy Wang, Jordan Sun, and Kunal Agrawal |
10:00/10:30 |
|
Coffee break |
10:30/11:00 |
|
PhD Elevator Pitch: Ahmad Nasser (NBL) / Amal Sakr (IMT) / Jules Sintes (Inria) / Alessa Mayer (IMT) / Baptiste Corban (Inria) / Thomas Le Corre (Inria) / (Inria) / Emanuele Mengoli (IMT) / Mohammed Amine Legheraba (Sorbonne University) / Ludmila Courtillat-Piazza (IMT) |
11:45/11:45 |
|
Lunch at the cantine + coffee in the inner garden |
11:45/13:30 |
|
1st LINCS Survey The 1st quantum revolution was about manipulating groups of quantum particles such as photons, electrons and atoms, and brought us technologies such as transistors and lasers. The 2nd quantum revolution, or “Quantum 2.0”, is about manipulating individual particles, and promises its own set of game-changing innovations. These include quantum computing, communications, networks, security, sensing and more. These technologies are not fully mature, however, and each one provides numerous research opportunities. In this talk, we give a quick overview of all things Quantum-2.0-related, and outline the past, current and future activities at LINCS in this field. |
13:30/14:00 |
|
2nd LINCS Scientific Highlights There is a long tradition of network management methods based on a precise model of the network and of the load. However, in practical situations it is impossible to build such a model, mainly because the load is uncertain and not known in advance. AI methods can overcome this model/reality gap, by continuously adjusting decisions based on streams of monitoring observations. In this talk, I will show how we applied AI to manage the cloud-to-edge continuum, focusing on the following decisions: pricing, placement of multiple “versions” of machine learning models, resource allocation. The methods we applied are Hidden Parameter Markov Decision Processes, Model-Based QLearning, Online Learning. I will show that, despite the uncertainty on the input load, we are able to provide analytic guarantees on the worst-case performance or on the average performance. Such guarantees are important to foster the applicability of AI algorithms, which is often hindered by their black-box nature. |
14:00/14:15 |
|
3rd LINCS Scientific Highlights Network Calculus is a theory that has been developed to compute performance guarantees, such as end-to-end delays, in networks. In this talk, I will discuss some improvement in analysis of multiclass network obtained in the past years (while working at Huawei), and present some perspective about the worst-case simulation. |
14:15/14:30 |
|
4th LINCS Scientific Highlights The rise of large reasoning language models (LRLMs) has unlocked new potential for solving complex tasks. These models operate with a thinking budget, that is, a predefined number of reasoning tokens used to arrive at a solution. We propose a novel approach, inspired by the generator/discriminator framework in generative adversarial networks, in which a critic model (potentially a specialized model) periodically probes to assess whether it has reasoned enough to reach a confident conclusion. If not, it continues reasoning until a target certainty threshold is met. We explore how model certainty can be quantified and integrated into the reasoning process, and discuss its practical implications. Through experiments and analysis, we show that certainty-guided reasoning improves accuracy while reducing unnecessary token usage. |
14:30/14:45 |
|
Poster session / Refreshment Break |
14:45/15:30 |
|
2nd Invited talk from the Scientific Committee GPUs are increasingly used to accelerate ML applications. Often, they are used to offload inference tasks, which are usually latency-critical. But the arrival patterns of inference requests are often bursty and include periods without any load. Furthermore, inference tasks may not be able to fully utilize the compute resources of a GPU, even with larger batch sizes. Consequently, the average utilization of a GPU that is exclusively used for an inference service is low. Industry has recognized the problem of underutilization and offers solutions to co-locate applications, improving utilization and cost-efficiency. We show, however, that these state-of-the-art solutions only maintain low inference latency when their compute resources are significantly overprovisioned. We propose DRACO, a system to co-locate latency-critical inference tasks with a batch job, e.g., training an ML model, without violating the inference latency requirements. DRACO autonomously estimates the resource requirements of inference tasks and detects periods with low load by periodically sampling the GPU’s performance monitoring unit (PMU). Furthermore, DRACO manages a pool of streaming multiprocessor (SM) partitions and dynamically assigns them to inference tasks to meet latency requirements. Leftover resources are granted to the co-located batch job. Depending on the workload, DRACO increases throughput of the batch job by up to 4x compared to industrial solutions while keeping inference latency low. * This is a joint work with Theo Radig |
15:30/16:00 |
|
5rd LINCS Scientific Highlights Models of strategic decision making by human agents, usually assume that the agents are rational. However, in practice, humans exhibit different kinds of irrational behaviour in their decision making. Prospect theory is a mathematical framework that models some aspects of this irrationality. The application of prospect theory in the study of games and equilibria is an emerging area; this provides insights on change in equilibrium behaviour in the presence of irrationality. However, the application of prospect theory throws open |
16:00/16:15 |
|
6th LINCS Scientific Highlights The WebRTC API enables real-time communication of text, video, and audio media streams through a web browser without requiring third-party extensions. However, it was not designed with privacy in mind. We conduct an experiment to analyse privacy leaks associated with WebRTC. Our findings show that despite recent updates to the WebRTC specification and its implementations, sensitive public IP addresses can still be leaked during audio/video communication, particularly in large non-NAT corporate networks, even when using a VPN, SOCKS or HTTP/S proxy. To address the observed leaks, we develop a simple, easily maintainable, cross-platform open-source solution that confines the Mozilla Firefox web browser in a docker container. We also take into account the possibility of a malicious adversary compromising the browser. Our tests have shown that our containerised solution is effective in all situations without restricting applications. |
16:15/16:30 |
|
Transfer to Paris for Reception Cruise |
17:00 |
|
Peniche TIVANO – Escales de Grenelle : Métro : Ligne 6 – stop Bir-Hakeim / RER C : stop Champ de Mars Tour Eiffel / Parking Centre commercial Beaugrenelle, 5, Quai André Citroën (70015) / Parking Kennedy / Radio France, 1, av. du Pdt Kennedy (70016) |
18:30 |
|
Friday, July 11th, 2025 |
|
|
Coffee reception |
9:00/9:30 |
|
7th LINCS Scientific Highlights |
9:30/9:45 |
|
8th LINCS Scientific Highlights In modern architecture of communication network using LEO and MEO satellite constellation, it is crucial to know how frequently an user is performing handover. As a key performance metric of the system, handover frequency essentially reflects onto the cost of quality service. In this talk we consider a far more simplified model of the same flavor of LEO and MEO satellite constellation using stochastic geometry. In this dynamic communication model on the Euclidean plane we consider an user located at origin and it is served by the mobile base stations with initial locations given by a homogeneous Poisson point process. The base stations are moving at an identical speed in a random direction. The user stays connected to the nearest base station at any given point of time. Since the base stations are moving, the user disconnects and connects with different base stations over time, which ever base station is the closest. We determine the handover frequency first in the single-speed setting and use it as a inspiration to the multi-speed scenario. The model explored in this work have several important variants which are linked to these motivations. These variants include the finite visibility case, the case when the initial locations of the base stations are given by Poisson or Manhattan line Cox point processes. Their motion is along the underlying lines. The final variant is of course the spherical case. We shall briefly discuss about the steps towards relaxing these simplifications from the planar to spherical geometry. *This is a joint work with François Baccelli, Inria Paris & Telecom Paris. |
9:45/10:00 |
|
3rd Invited talk from the Scientific Committee |
10:00/10:30 |
|
Coffee break |
10:30/ 11:00 |
|
PhD Elevator Pitch: Alex Pierron (IMT) / Tiphaine George (IMT) / Julien Cardinal (Inria) / Luis Muneca Tomas (NBL) / Aoyu Pang (NBL) / Iain Burges (IMT) / Hugo Rimlinger (Sorbonne University) / Shu Li (Inria) |
11:00/11:45 |
|
Lunch at the cantine + coffee in the inner garden |
11:45/13:30 |
|
2nd LINCS Survey This presentation explains how to make AI more eco-friendly. First, we assess the current environmental impact of AI and its future evolution, more specifically regarding to energy consumption. Secondly, we present tools estimating the energy footprint of an AI. Thirdly, we present techniques to mitigate this problem. Finally, we conclude this presentation with a list of best practices.
|
13:30/14:00 |
|
9tjLINCS Scientific Highlights Networks of the future, including 5G and Beyond 5G, as well as IoT, are connecting more and more devices to the Internet, improving the connectivity and increasing the service offer, at the expense of their exposure to malicious actors. Although many countermeasures exist, it remains a daunting task to |
14:00/14:15 |
|
10th LINCS Scientific Highlights Accurate timing advance (TA) computation is critical in 5G non-terrestrial networks (NTN). It is necessary to compute it accurately to avoid inter user interference in the uplink at the satellite (BS) level. Estimating TA in low earth orbit (LEO) satellite networks is more challenging than in classical terrestrial deployments due to the larger path loss and high-speed movement of non-stationary LEO satellites. Capturing the doppler shift also becomes very pertinent in such scenarios. The problem becomes more challenging in the event of the UE being mobile itself. In this talk, we first showcase an extended Kalman filter (EKF) based recursive Bayesian framework to accurately estimate the TA and Doppler shift in the presence of LEO satellite-UE joint motion dynamics. The framework first accurately models the joint motion dynamics and then constructs a Jacobian to linearize the inherent non-linearities present in the motion process. Probabilistic insights are also provided. The proposed framework is also useful when the satellite and UE clocks are not in sync, with the corresponding clock drift a function of the measured time difference of arrivals. Our results showcase the efficacy and robustness of the proposed EKF framework to estimate the TA and Doppler shift, even at very high UE speeds. The work is expected to be extremely useful in realizing LEO satellite based non-terrestrial networks. Further, as a current work, we are working on statistical characterization of the Doppler shift experienced at the UE. We showcase the probability densities of the Doppler shift in a 2D isotropic scenario and the engineering insights through it. |
14:15/14:30 |
|
Poster session / Refreshment Break |
14:30/15:15 |
|
11th LINCS Scientific Highlights We explore the applicability to high-speed NFV systems of causal discovery, a framework of statistical and algorithmic techniques that aims to uncover causal relationships. This framework highlights the ‘true’ structure of the processes that lead to observed outcomes while transcending spurious correlations. Causal discovery is crucial in the NFV domain, where introducing new levels of abstraction in the execution of virtualized services may diminish an observer’s ability to understand configuration or runtime issues. As a drawback, however, strict assumptions must be established concerning the data collection and the underlying system behavior. Most causal discovery techniques have been exercised on synthetic data, which lack the complexity and subtlety of real-world data generation processes. In this paper, we instrument a testbed to allow the controlled deployment and perturbation of NFV topologies and evaluate the algorithms’ robustness, defined as their ability to successfully reconstruct a correct configuration from observational and interventional probings. We then consider some ramifications of discovery quality for anomaly detection. |
15:15/15:30 |
|
12h LINCS Scientific Highlights Previous studies have shown that Instant-Runoff Voting (IRV) is highly resistant to coalitional manipulation (CM), though the theoretical reasons for this remain unclear. To address this gap, we analyze the susceptibility to CM of three major voting rules—Plurality, Two-Round System, and IRV—within the Perturbed Culture model. Our findings reveal that each rule undergoes a phase transition at a critical value ?_? of the concentration of preferences: the probability of CM for large electorates converges exponentially fast to 1 below ?_? and to 0 above ?_?. We introduce the Super Condorcet Winner (SCW), showing that its presence is a key factor of IRV’s resistance to coalitional manipulation, both theoretically and empirically. Notably, we use this notion to prove that for IRV, ?_? = 0, making it resistant to CM with even minimal preference concentration. |
15:30/15:45 |
|
13h LINCS Scientific Highlights Low Earth orbit (LEO) and medium Earth orbit (MEO) satellite networks consist of multiple orbits which are populated with many satellites. A widely used spatial architecture for LEO or MEO satellites is the Walker constellation, where the longitudes of orbits are equally spaced and the satellites are equally spaced along the orbits. In this paper, we develop a stochastic geometry model for the Walker constellations. This proposed model enables an analysis based on dynamical system theory, which allows one to address essential structural properties such as periodicity and ergodicity. It also enables a stochastic geometry analysis under which we derive the performance of downlink communications of a typical user at a given latitude, as a function of the key constellation parameters. |
15:45/16:00 |
|
Public Comment by the LINCS Scientific Committee |
16:00/16:30 |
|
Workshop Closing |