32 Postdoctoral Fellowship at Inria, France

Atlantis is  a joint project-team  between Inria and  the Jean-Alexandre Dieudonné Mathematics Laboratory at  Université Côte d’Azur. The team  gathers applied mathematicians and  computational scientists who are collaboratively undertaking  research activities aiming at the design, analysis, development and  application of advanced numerical methods for solving systems of  partial differential equations (PDEs) modelling nanoscale light-matter interaction problems. In this context, the team is  developing  the   DIOGENeS  [https://diogenes.inria.fr/]  software suite,  which  implements  several Discontinuous  Galerkin  (DG)  type methods tailored to the systems  of time- and frequency-domain Maxwell equations  possibly coupled  to  differential  equations modeling  the behaviour of propagation  media at optical frequencies. DIOGENeS also includes a component dedicated to the optimization of geometrical characteristics of nanostructures driven by some performance objective in the contex of inverse design strategies of nanophotonic setups.   DIOGENeS is a unique  numerical   framework  leveraging   the  capabilities   of  DG techniques  for  the simulation  of  multiscale  problems relevant  to nanophotonics and nanoplasmonics.

One important line of research of the team during the last years has been dedicated to improve the capabilities of these numerical tools to produce novel inverse design methodologies for optical metasurfcaes. In the last decade metasurfaces, i.e. 2D arrays of optical nanoantennas with subwavelength size and separation [1] have revolutionized the field of linear optics with the promise to replace bulky and difficult-to-align optical components with ultrathin and flat devices like metagratings, metalenses and metaholograms, which can also implement new functionalities in terms of aberrations correction and arbitrary wavefront shaping. In the recent years, by combining a high-fidelity DG-based fullwave solver in the time-domain [2] with a statistical learning-based global optimization method [3], we have introduced innovative inverse design methodologies for mono-objective optimization of metadeflectors [4], multi-objective optimization of RGB metalenses [5] and robust optimization of metadeflectors [6].

This project is based on a collaboration between Scool and a medical team at Inserm/CHU de Lille that have been active for 5 years now.

The scientific goal of this collaboration is to investigate the exploitation of data to improve patient follow-up after surgery. We model this problem as a sequential decision making under uncertainty problem.

This collaboration has already produced interesting results, in the form of a website, presentations in conference, and publications in top machine learning conferences, and top medicine journals.

This collaboration involves about 10 people: data engineer, PhD student, post-doc, researchers, professors.

This post-doctoral position will be supported by the  HE Trumpet project, the HE Flute project and/or the PEPR IA Redeem project.      While this position will be in the MAGNET team in Lille, we will collaborate with the several European project partners.

While AI techniques are becoming ever more powerful, there is a growing concern about potential risks and abuses. As a result, there has been an increasing interest in research directions such as privacy-preserving machine learning, explainable machine learning, fairness and data protection legislation.
Privacy-preserving machine learning aims at learning (and publishing or applying) a model from data while the data is not revealed. Notions such as (local) differential privacy and its generalizations allow to bound the amount of information revealed.

 The MAGNET team is involved inthe related TRUMPET, FLUTE and REDEEM projects, and is looking for team members who can in close collaboration with other team members and national & international partners contribute to one or more of these projects.  All of these projects aim at researching and prototyping algoirhtms for secure, privacy-preserving federated learning in settings with potentially malicious participants.  The TRUMPET and FLUTE projects focus on applications in the field of oncology, while the REDEEM project has no a priori fixed application domain.

The research project will explore quantum protocols based on the concept of quantum nonlocality and quantum networks (see arXiv:2104.10700). A non-exhaustive list of potential projects is:

• Methods for characterizing quantum correlations beyond the Bell scenario:
  • mathematical foundation of these methods (C* algebras, noncommutative polynomial optimization): see e.g. arXiv:2210.09065, arXiv:2212.11299, arXiv:2301.12513
  • improve/find new algorithms for characterizing these correlations
  • numerical developpement of these algorithms, see e.g. arXiv:2211.04483
• Understanding the foundational implications of quantum correlations in networks, see e.g. arXiv:2101.10873 and arXiv:2105.09381
• Develop the applications of network nonlocality to certification protocols, such as
  • randomness generation: arXiv:2209.09921 
  • self testing of measurements and states: arXiv:1807.04956, arXiv:2201.05032
• Adapt existing protocols for their experimental implementation
• Develop practical benchmarks of the concept of ‘Genuine Multipartite Nonlocality’ introduced in arXiv:2105.09381
• Develop SDP relaxations for condensed matter problems, see e.g. arXiv:2212.03014, arXiv:2310.05844, arXiv:2311.18707, arXiv:2311.18706 
• Explore the limits of quantum distributed computing, see e.g. arXiv:1810.10838, arXiv:0903.113

As part of a public collaboration with the Institute of Radiology of the University of São Paulo School of Medicine, the goal is to develop a software based on generative AI to generate 7 Tesla (T) magnetic resonance imaging (MRI) images from corresponding images acquired at standard clinical magnetic field strengths (e.g., 1.5 T or 3 T), to improve spatial resolution of MR images, their quality (contrast and signal-to-noise ratio), and ultimately their diagnostic capability.

Several stays in São Paulo will be planned for this position to collaborate with the partner team (Dr. Fabiola Macruz) at the Radiology Institute from the Clinics Hospital of the University of São Paulo Medical School. Travel expenses will be covered within the applicable scale.

Il est de développer une méthode d’estimation du connectome à partir des enregistrements simultanés EEG-fMRI.  Des différences dans les matrices de connectivité mesurées par EEG et IRMf sont attendues en raison des différences de résolution spatiale et temporelle mais aussi parce qu’elles capturent différents mécanismes de l’activité cérébrale. Une forte corrélation intermodale a été trouvée dans la bande de fréquence EEG-β.

La personne recrutée devra estimer les informations de connectivité communes et complémentaires et la relation entre l’organisation de la connectivité EEG et IRMf dans différentes bandes de fréquences, en utilisant la connectivité dynamique [5]. Cela sera fait en collaboration avec Jonathan Wirsich, Université de Genève, Genève, qui fournira une expertise dans l’estimation de la connectivité dynamique EEG. Le postdoc devra développer une approche

Exploiter les caractéristiques complémentaires et similaires avec une approche de fusion appropriée est crucial pour améliorer l’estimation du connectome et identifier de nouveaux biomarqueurs des maladies. Dans ce contexte, le postdoctorant devra construire des graphes multicouches contenant des matrices de connectivité IRMf et EEG.

This research engineer position takes place within the context of the ANR-SNF MetaboLinkAI project, which aspires to revolutionize the analysis and interpretation of metabolomics data through a multidisciplinary approach that combines a comprehensive knowledge hub (MetaKH) with cutting-edge artificial intelligence (AI) and machine learning (ML) techniques. The project’s main goals are to enhance the querying and ease of use of metabolomics data, improve research efficiency, and stimulate creativity in the field. These objectives are set to surpass current standards by creating an encyclopedic and expandable knowledge base, integrating advanced AI to handle the uncertainties of experimental data, and enabling a broader range of hypothesis testing and evaluation.

Within this context, this position will focus on the construction and querying of MetaKH, a decentralized, machine-readable knowledge hub federating and linking (1) pre-existing public knowledge and resources relevant for the use cases of the project (e.g. chemical entities description, biochemical pathways, metabolites information, relevant literature), (2) possibly newly created resources or the semantic lifting of existing resources not available in Semantic Web standards, and (3) and mass spectrometry datasets.

Crop protection often remains dependent on chemical pesticides, which are both harmful for the environment and human health. Resistant crops are an agroecological alternative to pesticides, but their extensive use may lead to the emergence/selection of virulent pathogens and resistance breakdown. Devising deployment strategies of resistant crops that are both efficient, i.e. that reduce crop damages, and durable, i.e. that limit the virulent pathogen populations, is hence a major issue.

The postdoctoral fellow will tackle this issue by means of a demo-genetic model, tailored for a specific pathosystem, the phoma stem canker of oilseed rape caused by fungus Leptosphaeria maculans. The emergence and development of virulent pathogens may vary according to the genetic determinisms of virulence (molecular mechanisms responsible for the transition to virulence, epistatic interactions, fitness costs), which are studied by other partners of the ENDURANCE project.

The work will be based on:

• time-series data of (i) phoma populations and resistance breakdowns, as well as (ii) resistance deployment in oilseed rape crops;
• a stochastic, discrete-time epidemiological model of an haploid monocyclic fungal pathogen, which includes features of the oilseed rape stem canker, such as interactions between resistance and avirulence genes;
• the corresponding C++ code.

The objectives of this position are threefold:

1. adapt the model to take into account migration, mutation, and pathotype-dependent virulence costs, based on recent advances in the genetic determinisms of virulence;
2. develop a method to estimate model parameters from historical data, in order to gain deeper insights into the observed dynamics of resistance breakdown;
3. devise durable strategies for the deployment of multiple resistances.

(18) Postdoctoral Fellowship Position

Postdoc Fellowship Position summary/title: Post-Doctoral Research Visit F/M Neural Architecture Growth

Deadline : Open Until Filled

View details & Apply

A central challenge in distributed systems is how to efficiently design key-value (KV) stores that leverage network, CPU, and memory resources. A preliminary study of existing distributed KVs shows that they often use these resources suboptimally. For instance, while modern networks are capable of handling hundreds of millions of I/Os per second, existing KVs typically sustain only a few million requests per second.

The causes are manyfold. Current systems rely heavily on atomic operations over RDMA, which are limited in throughput. They also make extensive use of caching, which puts additional pressure on the already constrained memory resources of clients. Moreover, the exchanges between clients and servers often impose significant CPU overheads.

In this project, we aim to design a streamlined KV architecture that better exploits the capabilities of today’s extremely fast networks. Several open questions must be addressed:

• How can RPC mechanisms be designed to efficiently exploit RDMA and modern NICs? KVs are not limited by network speed but by CPU overheads and protocol inefficiencies. What mechanisms can be used to reduce the number of I/Os per request while avoiding expensive operations such as atomics?

• Where should most of the processing occur? One possibility is a client-centric, offloaded design where clients handle index lookups, replication, and sometimes even garbage collection. Another option is a server-centric model where memory nodes take responsibility for these tasks. How do these approaches compare in terms of performance, reliability, and implementation complexity, especially when memory nodes have modest but usable CPUs?

• What is the impact of client-side caching on memory efficiency and garbage collection? Client caching is widely used to hide network latency but increases client memory consumption and complicates consistency and garbage collection. Is it possible to design a KV store that minimizes or eliminates client-side caching without significantly sacrificing performance?

• How should concurrency control be managed in KV operations? Current designs often depend on RDMA atomics, which quickly become a bottleneck. Alternatives include managing concurrency entirely in software at the memory node or combining different strategies depending on workload characteristics. What are the trade-offs between these approaches in terms of throughput and tail latency?

• How do workload characteristics influence the optimal KV store design? Read-heavy, write-heavy, and mixed workloads stress different system components. Similarly, workloads with skewed access distributions or large object sizes may benefit from different design choices. How can KV systems adapt dynamically to such variations?

As part of the Inria OS Challenge, we have designed and developed the USM platform, which enables the implementation of specialized memory management policies for applications with specific memory access patterns, such as key–value stores (KVS). Under USM, policies developed by system administrators execute in user space, which greatly facilitates their deployment and maintenance. A functional prototype of USM already exists and has been used to specialize the memory management of several applications, including virtual machines.

Running memory management policies in user space opens up new opportunities, particularly the integration of AI models to predict memory accesses and anticipate page faults. At present, it is almost impossible to execute AI models in kernel space, since floating-point operations—required by AI workloads—are not supported there. USM makes it possible to replace traditional heuristics with AI-based techniques for memory management.

The objective of this project is to explore the use of AI within USM and to investigate the hardware and software conditions under which this approach is both relevant and practical.

The recruited postdoctoral researcher may be required to travel to France for meetings or to present the results of this work internationally. All travel expenses will be fully covered by the project.

A direct consequence of hosting resources in a distributed way at the Edge is their exposure and sensitivity to the heterogeneity, massiveness, and uncertainty in mobility and demands of smart devices, leading to non-optimal edge usage in the long run. We aim to deal with such impacting factors in device behaviors by bringing (i) perceptive and aware mobility/demand quasi-in-time anticipation, (ii) uncertainty handling, and (iii) self-adaptation to device-edge resource management.

In particular, the focus will be on smart devices, where perceptiveness and awareness of needs and behaviors (where, when, and for what resources are required) of users and applications dictate decision, reaction/action, and allocation/management at the edge. Previous knowledge of TRiBE and AGORA on modeling, uncertainty profiling, interpretative predictability, and personalized anticipation of mobility behaviors [1],[2],[3],[4] as well as of resource demands [5],[6][7] of networking users, will be leveraged. The first goal will be to design a framework for quasi-in-time anticipation of spatial-temporal resource demands.

The second goal will be the design of perceptive mobility-aware offloading policies and adaptive allocation strategy according to the quasi-in-time anticipation of spatial-temporal resource demands. The quasi-in-time anticipation will limit service interruptions due to networking uncertainties or overload. The third goal concerns the evaluation of the designed framework, policies, and strategies. Besides, the benefits and tradeoffs of decisions that are taken based on quasi-in-time spatiotemporal anticipation of demands will also be analyzed (e.g., energy or resource loss/gain).

Les  services  d’urgence  sont  régulièrement  saturés.   Ils doivent répondre à différents objectifs,  dont ceux de détecter et prendre en charge sans délai  la  fraction  d’urgences  médicales  graves  et  que  les  autres  cas  n’empirent  pas.   A cette  fin,  un  degré  de  priorité  est  donné  à  l’entrée,  qui  entraine  des  parcours-patients différenciés avec des moyens dédiés.  De plus le service doit être organisé pour répondre à des périodes de fortes affluences.

Les organisations sont très variables. Une étude mathématique permet de simuler des scénarios sans risque pour les patients afin d’en prédire le comportement. Mais, vu la complexité du service, les simulations ne sont pas suffisantes et des modèles s’imposent. Ils se décrivent en termes de réseaux stochastiques pour prendre en compte l’aléa présent dans le type de patients qui arrivent, les temps inter-arrivées, les durées de consultations ou d’examens médicaux, etc. L’analyse de ces modèles aléatoires utilise les outils probabilistes tels que les processus de Markov, leurs équations d’évolution et théorèmes limites, ainsi que les problèmes de Skorokhod et des techniques de couplage.

L’idée intuitive est que, selon la ressource qui manque (médecins, places dans une salle, etc.), différents régimes vont s’établir. L’objectif est de trouver les conditions des différents régimes du système par une analyse asymptotique. Notons que les différents régimes et limites peuvent être abordés heuristiquement. Vu la complexité et la spécificité des modèles de services d’urgence, on vise à automatiser l’obtention de ces régimes. Ensuite, il conviendrait de formaliser et de prouver les résultats de convergence qui ont pu être ainsi conjecturés.

The postdoctoral position will be in the Wimmics research team (Inria, Université Côted’Azur, CNRS, I3S) specialized in knowledge graphs, artificial intelligence and Web technologies. It is part of the MetaboLinkAI ANR-SNF project, which aims to revolutionize the analysis and interpretation of metabolomics data through a multidisciplinary approach. The project combines a comprehensive knowledge graph hub (MetaKH) with advanced artificial intelligence (AI) and machine learning (ML) techniques.

The main objectives of the project are to enhance the accessibility and querying of metabolomics data, improve research efficiency, and foster innovation in the field. The project aspires to go beyond current standards by developing an evolving encyclopedic knowledge base, integrating advanced AI approaches to handle experimental data uncertainties, and facilitating the exploration and evaluation of a broader range of hypotheses.

Within this framework, we will focus on developing innovative methodologies and tools, such as graph exploration methods, to improve data interaction, analytical capabilities, and uncertainty representation. A key challenge of metabolomics data (and thus MetaKH) lies in its incompleteness, variable reliability, and inherent uncertainty. AI will be leveraged to enhance the completeness and reliability of the knowledge graph while effectively addressing these uncertainties.

This postdoctoral position is specifically part of WP3.4, which aims to develop an AI-powered research assistant for MetaKH. It builds on recent advances in generative AI, natural language understanding, and knowledge graph integration. An initial version of this assistant has been designed and developed as an intuitive chatbot, facilitating researchers’ interaction with metabolomics data and the MetaKH knowledge graph. This chatbot enables users to query the graph in natural language and refine their searches incrementally.

The goal of this postdoc project is to investigate how to optimize resource and task allocation when deploying data streaming processing applications in the Fog. In particular, we want to investigate new optimization metrics and objectives when deploying streaming processing applications in the Fog, including latency, throughput, and maximum sustainable throughput. Accordingly, we will develop a new scheduling framework that relies, among others, on Machine Learning/Deep Learning models to decide on resource allocation and operator placement at runtime (based on the collected data and given the cost model of redeployment and process migration). The proposed framework will be integrated in one of state-of the art data stream engines such as Flick [6], Storm [7] or Spark [8] and evaluated at large-scale using syntactic applications and real-world stream data application.

L’accélération promise par le calcul quantique s’appuie sur des travaux de recherche théoriques à propos de la complexité asymptotique des problèmes. La recherche actuelle semble néanmoins montrer que les algorithmes résultants pourraient permettre (au moins sur le papier) de réaliser des calculs inabordables par des méthodes conventionnelles. On passe donc de la recherche en algorithmique à la recherche en codage et optimisation, avec des études de cas.

On s’intéresse ici en particulier à la résolution de systèmes d’équations linéaires et différentielles avec l’algorithme HHL. L’équipe Inria/QuaCS est partenaire au sein du projet QLoop, avec comme tâche la compilation et l’optimisation de circuits résultant de cet algorithme. Le poste que nous proposons s’intègre dans ce projet.

Data are usually hosted in multiple geo-distributed locations (private and public clouds, distributed caches and across the edge-to-cloud continuum). Collectively processing these data is a must (e.g. distributed data analysis and queries), but this presents several challenges to existing data-intensive distributed workflow frameworks (e.g. MapReduce [1], Spark [2], TensorFlow [3] and Dataflow [4]). This is due to the low capacity of wide area network (WAN) links, as well as the heterogeneity of networks, computation power and monetary cost in geo-distributed environments [5].

Much effort has devoted on optimizing the performance of geo-distributed workflows [5, 6, 7, 8, 9]. These efforts mainly focus on reducing cross-data center data transfer and optimal task placement according to performance heterogeneity of the data centers [5, 6, 7]. However, these efforts do not consider the storage services heterogeneity [10] (input and intermediate data are stored on different devices), monetary cost heterogeneity in terms of computing, storage and network, or multi-tenancy when multiple workflows run concurrently.

This postdoc position will be in the context of IPCEI-CIS (Important Project of Common European Interest – Next Generation Cloud Infrastructure and Services) DXP (Data Exchange Platform) project involving Amadeus and three Inria research teams (COAST, CEDAR and MAGELLAN). This project aims to design and develop an open-source management solution for a federated and distributed data exchange platform (DXP), operating in an open, scalable, and massively distributed environment (cloud-edge continuum).

The postdoc will be located at The Inria Center of the University of Lorraine in the COAST team.

The Inria Center of the University of Lorraine is one of Inria’s nine centers and has twenty project teams, located in Nancy, Strasbourg and Saarbrücken. Its activities occupy over 400 people, scientists and research and innovation support staff, including 45 different nationalities. The Inria Center is a major and recognized player in the field of digital sciences. It is at the heart of a rich R&D and innovation ecosystem: highly innovative PMEs, large industrial groups, competitiveness clusters, research and higher education players, laboratories of excellence, technological research institutes, etc.

The Internet was conceived to enable computer resources’ time-sharing, but soon its main function became to deliver content to end users, but it is now called to play a new key role: to pervasively support machine learning (ML) operation both for model training and prediction serving.

There are two aspects calling for Internet-wide deployment of ML systems. First, data—one key ingredient of ML success—is often generated by users and devices at the edge of the network. The classic ML operation in the cloud requires such data to be collected at a single computing facility where training occurs. Data aggregation can be very costly, or simply impossible because of capacity constraints, privacy issues, or ownership ones. These scenarios call for distributed learning systems, where computation moves, at least in part, to the data. For example, Google’s federated learning [mcmahan17,kairouz21] enables mobile phones, or other devices with limited computing capabilities, to collaboratively learn an ML model while keeping all training data locally. Distributed ML training is already a difficult task in a cluster setting. Indeed, optimization techniques, distributed systems, and ML models are a triad difficult to untangle: e.g., relaxed state consistency across computing nodes increases system throughput but may jeopardize convergence of the optimization algorithm or affect the final solution selected, leading to models with very different generalization capabilities [chen16]. Additional challenges arise when training moves to the Internet. First, the system potentially scales up to billions of devices, against at most thousands of GPUs to break ML training records in a cluster. Second, local datasets are highly heterogeneous with very different sizes and feature/label distributions. Third, devices may have very different hardware and connectivity. Fourth, communications are often unreliable (devices can be switched off at any time), slow (latencies are 2 orders of magnitude larger), and expensive for battery-constrained devices. Fifth, privacy concerns are often important and limit the operations that can be performed during training to avoid inadvertently disclosing sensible information. Finally, training is more vulnerable to malicious attacks. For all these reasons, federated learning (as ML training over the Internet is now usually called) has emerged in the last years as a specific research topic—well distinct for example from high-performance computing or cloud computing—at the intersection of machine learning, optimization, distributed systems, and networking.