36 Postdoctoral Fellowship at Inria, France

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.

Currently, there is a significant transformation underway in car manufacturing, driven by the concept of Software Defined Vehicles. Software Defined Vehicles (SDVs) allow cars to be upgraded throughout their lifetimes via a centralized architecture, enabling updates and integration of new applications to enhance their functions. This represents the next revolution in the automotive industry, for which many manufacturers are preparing.

A primary objective of Software Defined Vehicles is to establish a flexible and scalable architecture, facilitating faster development and integration of new features over the vehicle’s lifespan.

In this context, HAL4SDV is an innovative collaborative European project that is starting, bringing together major industrial and academic partners to establish a set of standards and new development practices aimed at making SDV development more agile.

The Kairos team, a collaboration between Inria Sophia Antipolis and the I3S laboratory, is seeking a researcher to contribute to this ambitious European project.

Numerical evaluation of novel methods, a.k.a. benchmarking, is a pillar of the scientific method in machine learning. However, due to practical and statistical obstacles, the reproducibility of published results is currently insufficient: many details can invalidate numerical comparisons, from insufficient uncertainty quantification to improper methodology. In 2022, the benchopt initiative provided an open source Python package together with a framework to seamlessly run, reuse, share, and publish benchmarks in numerical optimization. In this project, we aim at bringing benchopt to the whole machine learning community, making it a new standard in benchmarking by empowering researchers and practitioners with efficient and valid benchmarking methods. Our goal is to ensure reproducibility and

Tout d’abord, nous proposons de revisiter le cadre riemannien afin d’incorporer plus systématiquement l’information préalable et la régularisation dans l’estimation des modèles de covariance. Deuxièmement, nous proposons d’aborder explicitement l’estimation des modèles de covariance pour les données structurées en
– Améliorer l’extraction d’informations à partir de données longitudinales, en distinguant les aspects spécifiques au sujet et les aspects longitudinaux du bruit d’échantillonnage.
– en renforçant l’invariance de l’origine dans les études multicentriques (voir par exemple [7]), mais le contexte de l’IRMf est très différent : dans l’IRMf, il n’y a pas d’incertitude quant à l’identification de la source. En ce qui concerne l’invariance par rapport à l’origine des données, nous proposons d’adopter une approche différente de celle de [10], en nous concentrant davantage sur la représentation riemannienne. En outre, nous aimerions réévaluer le cadre riemannien le mieux adapté : le cadre affine-invariant ou le cadre de Bures. L’avantage du cadre de Bures a été bien illustré dans [11]. bien illustré dans [11].

The Pl@ntNet team is looking for a post-doc with strong skills in deep learninn, Python development and, ideally, in GeoSpatial data. Pl@ntNet is a citizen observatory of plant biodiversity and a research platform at the crossroads of data science, ecology and artificial intelligence. The Pl@ntNet mobile application is used by tens of millions of users worldwide. 

You will work mainly in the Montpellier Computer Science Laboratory (LIRMM) as a starting researcher at Inria. Inria is a major player in computer science research worldwide; it is the managing organization of the Pl@ntNet consortium.
You will work in a quite unique environment at the forefront of digital technologies for biodiversity management. In particular, you will work in the context of the GUARDEN EU project aimed at mobilizing biodiversity data cubes in conjunction with other environmental data and scenarios, as the basis for models and indicators of past, current and future biodiversity.

This postdoc position will be located in the Inria-Inserm team COMPO (COMputational Pharmacology in Oncology), located in the La Timone health campus. The team is composed of mathematicians, data scientists, pharmacists and clinicians and is a unique multidisciplinary environment focused on developing novel computational tools for decision-making in clinical oncology.

 

The postdoc will join the INCa- (French National Cancer Institute) funded project COPYCAT: Combining Organoid technology with Mathematics to develop innovative models mimicking tumor cellular heterogeneity and plasticity for pediatric oncology. It is a collaboration between

He / she will also take part to the larger, INCa-funded, SouthRock consortium for pediatric cancers research between Lyon and Marseille.

An increasing number of applications rely on complex inference tasks based on machine learning (ML). Currently, two options exist to run such tasks: either served directly by the end device (e.g., smartphones, IoT equipment, smart vehicles) or offloaded to a remote cloud. Both options may be unsatisfactory for many applications: local models may have inadequate accuracy, while the cloud may fail to meet delay constraints. In [SSCN+24], researchers from the Inria NEO and Nokia AIRL teams presented the novel idea of inference delivery networks (IDNs), networks of computing nodes that coordinate to satisfy ML inference requests achieving the best trade-off between latency and accuracy. IDNs bridge the dichotomy between device and cloud execution by integrating inference delivery at the various tiers of the infrastructure continuum (access, edge, regional data center, cloud). Nodes with heterogeneous capabilities can store a set of monolithic machine-learning models with different computational/memory requirements and different accuracy and inference requests that can be forwarded to other nodes if the local answer is not considered accurate enough.

In this postdoc, we will study the problem of AI model placement in an IDN. This is a challenging optimization problem that involves a non-trivial tradeoff between model effectiveness, inference latency, and resource availability while also dealing with the natural dynamicity of the network, e.g., due to users’ request process or changes in available computing and communication resources. 

We will also consider other metrics, such as energy consumption, in the objective functions and networking constraints for systems where the network presents some inelasticity (see also [1]). We will leverage multi-objective optimization techniques (e.g., Pareto efficient solutions) and transfer learning techniques to adapt models across nodes with different levels of knowledge and resource availability. We will also rely on online learning approaches to achieve model placements with adversarial guarantees regarding regret.

A posteriori error estimation for wave propagation is very involved. A novel approach to a posteriori error estimation of finite element discretizations of time-dependent wave propagation problems has been recently introduced and will be the main line of investigation of the postdoctoral research program.

Currently, this approach is limited to static meshes, which can be adapted in space but are fixed in time. However, to deliver optimal performance, adaptive schemes require dynamic meshes that are allowed to change as the wave propagates through the computational domain. The goal of the postdoctoral project is therefore to develop a posteriori error estimators that are able to handle dynamic meshes.

This project intends to provide efficient statistical modeling tools for processing tokamak simulation outputs. The goal of these tools will be to help identify anomalies or errors in the data, which can indicate problems with the HPC system or the underlying algorithms. This information can then be used to adjust the computations and ensure that the HPC system is running smoothly and efficiently. Moreover, statistical processing can also be used to analyze the outputs and detect events of interest to select the right snapshots of the simulation to allow post-processing of the results. By properly selecting the information to save, we can ensure that the HPC system is able to handle large volumes of output without running into storage or bandwidth limitations.

The postdoc work will be dedicated to enriching a qAOP model developed by the hosting team and collaborators. The first task will consist in coupling a size-structured (PDE-based) mechanistic model of oogenesis (the developmental process leading to the release of a fertilizable femate gamete), considering all maturation stages and their control (Bonnet et al. 2020, Ballif et al. 2024), with a physiologically-based pharmacokinetic (compartmental ODE-based) model representing the hormonal dynamics within the reproductive (hypothalamo-pituitary-gonadal) axis (Pery et al. 2014, Ly et al. 2023). The coupling will be based on careful mapping of functions involved in the different models, addition of missing variables when needed, and timescale separation when possible to enforce model parcimony.  The second task will be to design appropriate numerical schemes to simulate the model in both physiological and toxicological  situations (exposure to EDs) and perform an extensive parameter estimation and sensitivity/identifiability analysis. The calibrated coupled models will enable a detailed assessment of reprotoxic effects, in particular long-term effects and rebound/compensation mechanisms affecting the oocyte population and hormonal feedback loops.

The required research includes in-depth analysis of selected software libraries. Specific metadata information associated to each software under investigation needs to be identified and registered, including cybersecurity risks related to the usage of each software library.

Based on relevant state- of- the- art, preliminary ideas how to support the mitigation of cognitive biases – included in a speech and data processing environment that is used intelligence analysis.

Some other directions that the research can contribute concerning speech processing are :

• Speaker verification using machine learning model
• Knowledge reasoning for semantic audio segmentation

 

(18) Postdoctoral Fellowship Position

Postdoc Fellowship Position summary/title: Post-Doctoral Research Visit F/M Exploring Variability in Medical Image Segmentation: New Metrics and Frameworks

Deadline : 2024-10-31

View details & Apply

 

The main goals are to:

1. Put in place a posteriori error estimators which allow to quantify the error between the numerical approximation (known) and the exact solution (unknown).
2. Develop adaptive balancing strategy for the regularization parameter, the iterative linearization algorithm (Newton), and the iterative linear algebraic solver. This should lead to robust solvers and enable to significantly reduce the usual number of linear and nonlinear iterations.
3. Develop adaptive steering of the choice of time step and of the local mesh refinement. This will lead to adaptive front tracking and automatic recognition of viscous and gravitational phenomena.

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.

Ropeways – chairlifts, gondola lifts, bicable aerial ropeway, etc. – are subject to dynamic phenomena linked mainly to three factors: the movement of the cable required for installation, loads due to operation (acceleration, braking, load changes) and external loads to the system, mainly climatic (wind, etc.). These phenomena can cause vehicles (chairs or cabins) to oscillate, with consequences ranging from discomfort (motion sickness) to premature wear on installations (accelerated fatigue of material) and even incidents with an impact on safety (movements incompatible with the system, collisions with infrastructure).

A number of events have occurred in France, two of which are being investigated by the bureau d’enquêtes sur les accidents de transport terrestre (BEA-TT): the fall of 5 cabins (during maintenance) on the “Aulp-de-Véran” gondola lift on 13 October 2011 in the Flaine ski area [1] and the immobilisation and late evacuation of the “Panoramic Mont-Blanc” ropeway on 8 September 2016 in Chamonix [2].

If the static of cables [3-4] is well known, their dynamics are still a subject that needs to be studied in more detail. There is a large amount of work [5-6] on analytical modelling, while the field of finite element modelling needs to be explored in more depth [7-9].

This post-doc will be supervised by Vincent Acary (Inria) and M. Weiss (STRMTG). The postdoctoral student will be hosted in the Tripop project team, a research team focusing on the modelling, simulation and control of non-smooth dynamic systems. The STRMTG is the technical department of the Ministry of Transport responsible for the safety of mechanical lifts and guided transport.

The person will be recruited into the LACODAM team at the Inria Rennes research center. This research center is distinguished by its excellence in digital research and technological innovation, driving major advances in fields such as artificial intelligence, cybersecurity, and digital health, while working closely with industrial and academic partners to shape the technologies of tomorrow.

The recruited individual will be supervised by Élisa Fromont and Romaric Gaudel, and will collaborate with Paul Viallard. All three are researchers in Machine Learning with publications in leading conferences on the theoretical aspects of Machine Learning, particularly on multi-armed bandit problems.

The goal of the project is to solve a class of combinatorial optimization problems with multi-armed bandit feedback through approaches based on local search techniques. The envisioned approach falls within the framework of reinforcement learning and significantly reduces both the regret and computational cost associated with finding an optimal policy.

The specific application problem under consideration is the selection of a connected subgraph in order to maximize flow. More precisely, we are faced with a set of nodes that wish to communicate and a set of possible connections between these nodes. The objective is to maximize the communication flow while minimizing the number of edges between the nodes. Therefore, the task is to identify the edges to retain for this purpose.

This research project is funded as part of a partnership agreement between Inria and the Defense Innovation Agency of the French Ministry of the Armed Forces.

The project context revolves around the detection of cardiac arrhythmias, specifically ventricular fibrillation, which is responsible for the vast majority of the 350,000 sudden cardiac deaths occurring annually in Europe. One of the most promising recent approaches to detecting these cardiac rhythm disorders is electrocardiographic imaging (ECGi), currently under study at IHU-Liryc. It is a non-invasive imaging technique that reconstructs the heart’s electrical activity based on electrical measurements taken on the patient’s chest using a vest adorned with numerous electrodes, commonly referred to as the “vest.” However, current ECGi resolution techniques do not allow for sufficiently accurate reconstruction of cardiac electrical activity for complex clinical cases. In particular, it is widely accepted that uncertainties in organ and chest movement, as well as inter-individual variations in tissue conductivities, play a role in the loss of ECGi precision.

The position is in the QInfo team. The candidate will collaborate with Omar Fawzi and Alastair Abbott, and other members of the team.

Within the framework of a partnership: collaboration with the members of the PEPR DIQKD project including CEA Saclay, Université de Nice and CNRS Sorbonne Université

The Lyon centre is active in the fields of software, distributed and high-performance computing, embedded systems, quantum computing and privacy in the digital world, but also in digital health and computational biology.

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.

The second driver to distribute ML processes over the Internet is real-time inference. In fact, ML models are often trained for inference’s purposes, i.e., to make predictions on new data. Model predictions need then to be served to the final users. ML training is a computationally expensive operation and is the object of much research effort. Inference does not involve complex iterative algorithms and is therefore generally assumed to be easy, but it also presents fundamental challenges that are likely to become dominant as ML adoption increases [stoica17]. AI systems will be ubiquitously deployed and will need to make timely and safe decisions in unpredictable environments. In this case, inference must run in real-time, and predictions may need to be served at a very high rate. The big cloud players—Amazon, Microsoft, and Google—have all started pushing their “machine learning as a service” (MLaaS) solutions. Running the models in the cloud guarantees high scalability, but may fail to meet delay constraints. As an example, already deployed applications, such as recommendation systems, voice assistants, and ad-targeting, need to serve predictions from ML models in less than 20 ms [simsek16]. Future wireless services, such as connected and autonomous cars, industrial robotics, mobile gaming, augmented/virtual reality, have even stricter latency requirements, often below 10 ms and below 1 ms for the so-called tactile internet. It is then imperative to run these services closer to the user at the network edge. 5G deployment can provide computing and storage capabilities at the edge, but those will still be very limited in comparison to the cloud and need to be wisely used. In conclusion, inference will require complex resource orchestration across users’ devices, edge computing servers, and the cloud.

The recruitee will work in close relation with both teams, Scool and the Inserm team.

With Scool, the goal is to explore how the problem at hand can be modeled, most likely as a contextual bandit problem, propose algorithms, study their theoretical properties.

With the Inserm team, the goal is to investigate the application of this work to the real practical case.

These two lines of work are intrinsically strongly tied to each others.

The position will be funded through a QuantERA project involving a large european consortium: see the project website https://project.inria.fr/compute/

It will be achieved by collaborating with Marc-Olivier Renou, other members of the group (Xiangling Xu, Lucas Tendick, Isadora Veeren, and also Titouan Carette and Filippo Vicentini), as well as the consortium members:

• Mariami Gachechiladze (TU Darmstadt, Germany) and David Gross (Cologne)
• Victor Magron (LAAS-CNRS, Toulouse, France)
• Igor Klep (University of Ljubljana, Slovenia)
• Antonio Acín (ICFO Barcelona, Spain)

The recruited person will work in close collaboration with Bruno Lévy, in Nancy, and will work in collaboration with the other members of the PARMA team in Orsay (mainly Quentin Mérigot and Hugo Leclerc)

We consider the context where control and protection functions for energy transportation systems have to be executed on a distributed heterogeneous Cloud-Edge computing infrastructure, according to the needs and specificities of these applications.

In this context, this deployment also has to be reconfigured in a self-adaptive way, in a feedback loop according to variations and events at the different levels of application (e.g., variations in the need for precision in the control functions), infrastructure (e.g., loss or degradation of computing nodes, or of communication links) environment (e.g., heat wave compromising use of fan-less or airco-less nodes in edge posts, wind, rain or flooding).

We pursue software architectures that support separation of concerns between self-adaptations at the levels of computing and communication infrastructure (hardware), management and allocation of computing tasks implementing control and protection (middleware), and modes of the applications (business or process).

We will consider different reconfiguration strategies, according to criteria such as security properties, non-functional properties (e.g., latency), minimization of computing resources. The work will be specifically targeted at the application to an energy infrastructure, e.g., the use case from project partner RTE.

We recently introduced the Trap space semantics for Logic Programs. One objective of this project is to generalize the trap space semantics
to disjunctive logic programs.
We also want to find efficient methods for computing ≤s-minimal stable trap spaces (equivalently regular models) of a logic program.
Finally, we would like to explore more new results on relating the recently introduced dependence graph and models of a logic program.

The objective of this research project is de develop a mathematical model and asociated numerical methods for the simulation of the hemodynamics in the four cavities of the heart. The main motivation of this work is driven by the developement of a simulation tool to simulate patohological scenarios, such as the Hypoplastic Left Heart Syndrome (HLHS), and numerically asses different surgicals options. This work will be performed in the framework of the MEDITWIN project, which involves a collaboration beetwen COMMEDIA projec-team, Dassault Systèmes and SIMBIOTX project-team.