7.1.1 SwiftHohenbergPseudoSpectral

• Keywords:
  Partial differential equation, Numerical solver, Self-organization
• Functional Description:

  This software solves the dimensionless Swift-Hohenberg Equation with a cubic-quadratic nonlinear term. The Swift-Hohenberg Equation is a widely studied model for pattern formation, derived using symmetry arguments. As such, it serves as a maximally symmetric model for systems exhibiting spontaneous pattern formation, making it broadly applicable to various physical contexts.

  The solver employs pseudospectral methods based on Fourier representations to ensure efficient and accurate computation. Importantly, the solver is physically constrained, using a Lyapunov functional to validate solutions, ensuring that the computed results remain consistent with the physical behavior expected from the system.

  Designed with versatility and automation in mind, the solver is capable of generating a large number of solutions with minimal supervision. This feature makes it particularly well-suited for creating synthetic datasets to be used in Physics Guided Machine Learning. Specifically, it addresses problems related to the femtosecond laser-induced self-organization of matter, where data augmentation is crucial for improving the performance and generalization of machine learning models.

• Publications:
  ujm-04157829, ujm-03823722, tel-04515418v1
• Contact:
  Eduardo Brandao

7.1.2 GUAP

• Name:
  Generalized Universal Adversarial Perturbations
• Keyword:
  Machine learning
• Functional Description:
  The software extends universal attacks by jointly learning a set of perturbations to choose from, aiming to maximize the success rate of attacks against deep neural network models.
• URL:
  https://­github.­com/­JordanFrecon/­guap
• Contact:
  Jordan Frecon Patracone
• Partner:
  LITIS

7.1.3 GraSPy

• Name:
  Graph Signal Processing for Python
• Keywords:
  Graph, Signal processing, Machine learning
• Functional Description:
  Implements classic components of such a toolbox, including (weighted) graph generalization and transformation and of their signals (a.k.a. vertex attributes), specific data visualization using Plotly, signal transformation with spectral methods (using IAGFTs), and graph learning.
• URL:
  https://­gitlab.­inria.­fr/­begiraul/­joint-graph-learning
• Publication:
  hal-04025968
• Contact:
  Benjamin Girault

7.1.4 Scikit-SpLearn

• Name:
  Toolbox for the spectral learning of weighted automata
• Keywords:
  Machine learning, Weighted automata
• Scientific Description:
  The core idea of the spectral learning of weighted automata is to use a rank factorization of a complete sub-block of the Hankel matrix of a target series to induce a weighted automaton.
• Functional Description:
  The toolkit contains an efficient implementation of weighted automata, a library to transformed any dataset containing strings of different length into the scitkit-learn format for data, 4 variants of the spectral learning algorithm.
• URL:
  https://­remieyraud.­github.­io/­scikit-splearn/
• Publication:
  hal-01777169
• Contact:
  Rémi Eyraud
• Partner:
  Laboratoire d'Informatique et des Systèmes (LIS) Université Aix-Marseille

7.1.5 SoundScapeExplorer

• Keywords:
  Acoustics, Data analysis, Data visualization, Soundscape
• Functional Description:

  SSE stands for SoundScapeExplorer and works with passive acoustic monitoring campaigns made of several microphones recording during hours, days or week.

  It analyzes these data by extracting various features and indicators for each signal window (typically 1 second) then aggregating them by chunks (e.g. from 5 seconds to 1 or more minutes). Using a selection of dimensionality reduction techniques, it allows to visualize the whole campaign as a point cloud where each point corresponds to a temporal chunk on one microphone. The visualization is interactive, allowing coloring by criteria, filtering, listening to audio data, cluster analysis and beyond.

• URL:
  https://­sound-scape-explorer.­github.­io/
• Contact:
  Rémi Emonet
• Partner:
  Laboratoire ENES

TAYSIR Benchmark

• Contributors:
  Rémi Eyraud
• Description:
  The Transformers+rnn: Algorithms to Yield Simple and Interpretable Representations (TAYSIR, the Arabic word for ‘simple’) Benchmark was created for an on-line challenge on extracting simpler models from already trained neural networks held in Spring 2023. These neural nets were trained on sequential categorial/symbolic data. Some of these data were artificial, some came from real world problems (such as Natural Language Processing, Bioinformatics, and Software Engineering). The trained models covered a wide spectrum of architectures, from Simple Recurrent neural Network (SRN) to Transformers, including Gated Recurrent Unit (GRU) and Long Short-Term memory (LSTM). The benchmark contains Neural Nets trained on two types of tasks: neural networks trained on Binary Classification tasks, and on Language Modeling tasks.
• Project link:
  https://remieyraud.github.io/TAYSIR/
• Publication:
  https://proceedings.mlr.press/v217/eyraud23a.html
• Contact:
  remi.eyraud@univ-st-etienne.fr

8.1.1 Approximation Error in Physics-informed Neural Networks

Participants: Benjamin Girault, Rémi Emonet, Amaury Habrard, Jordan Patracone, Marc Sebban.

Considering the key role played by derivatives in Partial Differential Equations (PDEs), using the tanh activation function in Physics-Informed Neural Networks(PINNs) yields useful smoothness properties to derive theoretical guarantees in Sobolev norm. In 2, we conduct an extensive functional analysis, unveiling tighter approximation bounds compared to prior works, especially for higher order PDEs. These better guarantees translate into smaller PINN architectures and improved generalization error with arbitrarily small Sobolev norms of the PDE residuals.

8.1.2 PAC-Bayes Theory for Generalization Bounds with Complexity Measures

Participants: Rémi Emonet, Amaury Habrard.

Collaboration with Paul Viallard (Inria Rennes), Emilie Morvant (LabHC) and Valentina Zantedeschi (ServiceNow Research )

In statistical learning theory, a generalization bound usually involves a complexity measure imposed by the considered theoretical framework. This limits the scope of such bounds, as other forms of capacity measures or regularizations are used in algorithms. In 6, we leverage the framework of disintegrated PAC-Bayes bounds to derive a general generalization bound instantiable with arbitrary complexity measures. One trick to prove such a result involves considering a commonly used family of distributions: the Gibbs distributions. Our bound stands in probability jointly over the hypothesis and the learning sample, which allows the complexity to be adapted to the generalization gap as it can be customized to fit both the hypothesis class and the task.

8.1.3 A General Framework for the Practical Disintegration of PAC-Bayesian Bounds

Participants: Amaury Habrard.

Collaboration with Paul Viallard (Inria Rennes), Emilie Morvant (LabHC) and Pascal Germain (Dep. CS and Soft. Eng. [Québec])

PAC-Bayesian bounds are known to be tight and informative when studying the generalization ability of randomized classifiers. However, they require a loose and costly derandomization step when applied to some families of deterministic models such as neural networks. As an alternative to this step, we introduce in 7 new PAC-Bayesian generalization bounds that have the originality to provide disintegrated bounds, i.e., they give guarantees over one single hypothesis instead of the usual averaged analysis. Our bounds are easily optimizable and can be used to design learning algorithms. We illustrate this behavior on neural networks, and we show a significant practical improvement over the state-of-the-art framework.

8.1.4 Length Independent PAC-Bayes Bounds for Simple RNNs

Participants: Volodimir Mitarchuk, Rémi Eyraud, Rémi Emonet, Amaury Habrard.

Collaboration with Clara Lacroce from McGill (Canada) and Guillaume Rabusseau from MILA (Canada)

While the practical interest of Recurrent Neural Networks (RNNs) is attested, much remains to be done to develop a thorough theoretical understanding of their abilities, particularly in what concerns their learning capacities. A powerful framework to tackle this question is the one of PAC-Bayes theory, which allows one to derive bounds providing guarantees on the expected performance of learning models on unseen data. In 3, we provide an extensive study on the conditions leading to PAC-Bayes bounds for non-linear RNNs that are independent of the length of the data. The derivation of our results relies on a perturbation analysis on the weights of the network. We prove bounds that hold for $\beta$-saturated and DS $\beta$-saturated SRNs, classes of RNNs we introduce to formalize saturation regimes of RNNs. The first regime corresponds to the case where the values of the hidden state of the SRN are always close to the boundaries of the activation functions. The second one, closely related to practical observations, only requires that it happens at least once in each component of the hidden state on a sliding window of a given size.

8.1.5 A Theoretically Grounded Extension of Universal Attacks from the Attacker's Viewpoint

Participants: Jordan Patracone, Amaury Habrard.

Collaboration with Paul Viallard (Inria Rennes), Emilie Morvant (LabHC), Gilles Gasso (LITIS) and Stéphane Canu (LITIS)

In this work 5, we extend universal attacks by jointly learning a set of perturbations to choose from to maximize the chance of attacking deep neural network models. Specifically, we embrace the attacker’s perspective and introduce a theoretical bound quantifying how much the universal perturbations are able to fool a given model on unseen examples. An extension to assert the transferability of universal attacks is also provided. To learn such perturbations, we devise an algorithmic solution with convergence guarantees under Lipschitz continuity assumptions. Moreover, we demonstrate how it can improve the performance of state-of-the-art gradient-based universal perturbation. As evidenced by our experiments, these novel universal perturbations result in more interpretable, diverse, and transferable attacks.

8.1.6 Unsupervised Learning and Effective Complexity

Participants: Erick Gomez, Rémi Emonet, Marc Sebban.

Measuring the complexity of arbitrary data has been of interest to many scientific domains, including machine learning and particularly unsupervised learning. In 17, we cover relevant concepts including Kolmogorov complexity, entropy and minimum description length. We argue that these measures alone are failing to distinguish noise from meaningful complexity. We push for the concept sophistication which measures the complexity of the structured part of the data, ignoring unstructured noise. This concept is reified in two manners: using image compression algorithms and using autoencoders. We aim at leveraging this work for taking into account the complexity of physical dynamics into PIML algorithms.

8.2.1 Bregman Proximal Viewpoint on Neural Operators

Participants: Abdel-Rahim Mezidi, Jordan Patracone, Amaury Habrard, Rémi Emonet, Marc Sebban.

Collaboration with Saverio Salzo (DIAG) and Massimiliano Pontil (IIT, UCL)

In this work 19, we present several advances on neural operators by viewing the action of operator layers as the minimizers of Bregman regularized optimization problems over Banach function spaces. The proposed framework allows interpreting the activation operators as Bregman proximity operators from dual to primal space. This novel viewpoint is general enough to recover classical neural operators as well as a new variant, coined Bregman neural operators, which includes a skip-like connection and features the same expressivity of standard neural operators. Numerical experiments support the added benefits of the Bregman variant of Fourier neural operators for training deeper and more accurate models.

8.2.2 PIML for Better Understanding Laser-Matter Interaction

Participants: Fayad Ali Banna, Rémi Emonet, Marc Sebban.

Collaboration with Jean-Philippe Colombier (LabHC)

Physics-informed machine learning typically assumes that the underlying physical laws are known and abundant training data is available. These assumptions do not hold in the context of self-organization of matter, a phenomenon that leads to the emergence of patterns when a surface is irradiated with an ultrafast laser beam. Indeed, due to the constraints of the electronic data acquisition devices, the creation of large datasets is made impossible. Moreover, modeling this dynamic process is challenging as it involves coupling between electromagnetism, thermodynamics and fluid mechanics under far-from-equilibrium conditions that are not yet fully understood. This work 12 aims at taking a step forward towards a better understanding of this complex phenomenon. We specifically focus on the laser energy absorption of the surface, which is governed by the distinctive characteristics of Maxwell's equations in an inhomogeneous lossy medium. This involves modelling physics at the nano scale and incurs high simulation costs that make any exploration impractical. To address this major issue, we investigate different physics-informed learning models. In this low data regime, our study reveals that learning a simple U-Net-based surrogate model surpasses (i) more sophisticated neural architectures and (ii) the FDTD-based solver in speed by several orders of magnitude. Interestingly, our study highlights a link between the formation of patterns and the magnitude of absorbed energy.

8.2.3 Self-Consuming Generative Models

Participants: Quentin Bertrand.

The rapid progress in generative models has resulted in impressive leaps in generation quality, blurring the lines between synthetic and real data. Web-scale datasets are now prone to the inevitable contamination by synthetic data, directly impacting the training of future generated models. Already, some theoretical results on self-consuming generative models (a.k.a., iterative retraining) have emerged in the literature, showcasing that either model collapse or stability could be possible depending on the fraction of generated data used at each retraining step. However, in practice, synthetic data is often subject to human feedback and curated by users before being used and uploaded online. For instance, many interfaces of popular text-to-image generative models, such as Stable Diffusion or Midjourney, produce several variations of an image for a given query which can eventually be curated by the users. In 1, we theoretically study the impact of data curation on iterated retraining of generative models and show that it can be seen as an implicit preference optimization mechanism. However, unlike standard preference optimization, the generative model does not have access to the reward function or negative samples needed for pairwise comparisons. Moreover, our study doesn't require access to the density function, only to samples. We prove that, if the data is curated according to a reward model, then the expected reward of the iterative retraining procedure is maximized. We further provide theoretical results on the stability of the retraining loop when using a positive fraction of real data at each step. Finally, we conduct illustrative experiments on both synthetic datasets and on CIFAR10 showing that such a procedure amplifies biases of the reward model. We aim at leveraging this work for better filtering synthetic examples; This might be of great help in data augmentation, especially in the low-data regimes observed in surface engineering-informed machine learning.

9.1.1 CIFRE Theses with Thalès (2024-2027)

Participants: Rémi Eyraud.

Partners: SESAM (LabHC-UJM), Thalès

During Falls 2024, a collaboration with the DIS (Digital Identity & Security) lab of the Thalès firm started via a CIFRE PhD (Wissal Ghamour). Rémi Eyraud took the academic direction of the PhD together with members of the SESAME team of the Hubert Curien Laboratory. The subject aims at leveraging the tools of Machine Learning for Side Channel Attacks of embedded microchips. Information from physics can be added to the process, since the leak mainly relies on an electro-magnetic sensor. This followed the Ph.D of Dorian Llavata that started in 2022 on the same subject in collaboration with the CESTI Leti of the CEA.

9.1.2 I-Démo Région "GREENAI"

Participants: Jordan Patracone.

Partners: Dracula Technologies, ASYGN.

The project I-Démo Région "GREENAI" involves three key actors working collaboratively towards sustainable Internet of Things (IoT) solutions. Dracula Technologies specializes in developing organic photovoltaic cells that harness ambient light to power IoT devices, aligning with environmental sustainability goals. ASYGN contributes by designing ultra-low-power hardware accelerators to enable advanced AI processing directly on IoT devices. The Laboratoire Hubert Curien (Jordan Patracone ) focuses on energy-efficient artificial intelligence for computer vision, optimizing neural network architectures for low-resource hardware through two scientific theses (among which, that of Ben Gao). This work is sponsored by a public grant overseen by the Auvergne-Rhône-Alpes region, Grenoble Alpes Métropole, and BPIFrance.

9.2.1 "Baby Cry" project - AXA Fundation (2024-2026)

Participants: Rémi Emonet.

Partners: ENES (UJM), SAINBIOSE-MoVE, CHU Saint-Étienne.

Lead by the ENES, the AXA Baby Cry 1000 project aims at recording 1000 babies, each for one day after they are born to help early diagnosis of cognitive development issues and compare with the development of premature babies. The project is a joint effort by three labs (ENES, Laboratoire Hubert Curien (Rémi Emonet), and SAINBIOSE-MoVE, CHU Saint-Étienne) and is funded by the AXA Fundation for a total of 1M€ and a duration of 3 years.

10.1.1 Visits of international scientists

Three professors visited our team in 2024.

Participants: Rémi Eyraud.

• Status
  Researcher
• Institution of origin:
  New York State University at Stony Brook
• Country:
  USA
• Dates:
  April 2024 (one month)
• Mobility program:
  Research stay funded by "Invited Professors" UJM grant

Participants: Jordan Patracone.

• Status
  Associate Professor
• Institution of origin:
  Sapienza Università di Roma
• Country:
  Italy
• Dates:
  September 2024 (two weeks)
• Mobility program:
  Research stay funded by "Invited Professors" UJM grant

Participants: Marc Sebban.

• Status
  Researcher
• Institution of origin:
  University of Alicante
• Country:
  Spain
• Dates:
  July 2024 (one month)
• Mobility program:
  Research stay funded by "Invited Professors" UJM grant

10.1.2 Visits to international teams

Participants: Quentin Bertrand.

• Visited institution:
  Mila and Université de Montréal
• Country:
  Canada
• Dates:
  December 2024 (one week)
• Mobility program/type of mobility:
  Research stay funded by Inria

Participants: Marc Sebban.

• Visited institution:
  University of Alicante
• Country:
  Spain
• Dates:
  May 2024 (one week)
• Mobility program/type of mobility:
  Research stay funded by Univ. Alicante

10.2.1 ML4Health - Transform4Europe (2024-2026)

Participants: Amaury Habrard, Marc Sebban.

Partners: University of Alicante, Dpto.Tecnologia Informatica y Computacion (DTIC)

Type: Seed Funding T4EU (co-funded by the European Union)

Duration: 2 years (2024-2026)

The objective of ML4Health is two-fold: (i) from an training perspective, develop a double master degree in AI between the two T4EU partners, built from two existing complementary programmes (Machine Learning & Data Mining master track at UJM and Máster Universitario en Inteligencia Artificial at UA); (ii) from a scientific standpoint, benefit from the expertise in statistical learning and physics-informed Machine Learning (UJM) and in 3D perception and intelligent systems (DTIC) to focus on the morphological evolution of bodies of obese patients and the early detection of neurodegenerative diseases. In particular, the two partners study (data+theory) hybrid machine learning models combining both 3D+t images and biological/morphometric knowledge.

10.3.1 ANR MELISSA (2024-2029)

MELISSA: MEthodological contributions in statistical Learning InSpired by SurfAce engineering

Participants: Eduardo Brandao, Rémi Emonet, Benjamin Girault, Amaury Habrard, Jordan Patracone, Marc Sebban.

Partners: MAGNET (Inria Lille), MLIA (ISIR, PSL) and F. Garrelie and JP Colombier (LabHC-UJM)

Type: ANR PRC

The underlying dynamics of many physical problems are governed by parameterized partial differential equations (PDEs). Despite important scientific advances in numerical simulation, solving efficiently PDEs remains complex and often prohibitively expensive. Physics-informed Machine Learning (PiML) has recently emerged as a promising way to learn efficient surrogate solvers, and augment the physical laws by leveraging knowledge extracted from data. Despite indisputable advances, several open problems remain to be addressed in PIML: (i) Deriving generalization guarantees; (ii) Learning with a limited amount of data; (iii) Augmenting partially known physical laws; (v) Modeling uncertainty; (vi) Building foundation models for physics. MELISSA will deal with these problems from both theoretical and algorithmic perspectives. The objective is to design the next generation of provably accurate PIML algorithms in the challenging context of laser-matter interaction where data is scarce and the available physical laws only partially explain the observed dynamics.

10.3.2 ANR TAUDoS (2021-2026)

Participants: Rémi Emonet, Rémi Eyraud, Amaury Habrard, Marc Sebban.

TAUDoS: Theory and Algorithms for the Understanding of Deep learning On Sequential data

Partners: LIS (Aix-Marseille University), EURA NOVA (Firm), MILA (Canadian State)

Type: ANR PRCE

Website: TAUDoS

The ambition of this project is to provide a better understanding of the mechanisms that allow the amazing recent achievements of Machine Learning, and in particular of Deep Learning. This is achieved by providing elements that allow a better scientific comprehension of the models, strengthening our experimental results by theoretical guarantees, incorporating components dedicated to interpretability within the models, and allowing trustful quantitative comparison between learned models.

The originality and the specificities of TAUDoS are due to three major characteristics:

• The focus on models for sequential data, such as Recurrent Neural Networks (RNN), while most works concentrate on feed-forward networks
• The will to analyze these models in the light of formal language theory
• The goal to target both rigorous theoretical analyses and empirical evidence related to interpretability.

10.3.3 AI4OP (2021-2024)

Participants: Rémi Eyraud.

AI4OP: Artificial Intelligence for Onco-Plasma

Partners: LIS (Aix-Marseille University), AP-HM, INT (Aix-Marseille University)

Type: ITMO-Cancer Math-Info for Cancers

Website: AI4OP

The objective of this project was to create a universal and non-invasive method to help diagnose cancers. The proposed method is based on the existence of unique plasma denaturation profiles (signatures) for different cancers. The plasma denaturation profile represents the total denaturation curve (under the influence of temperature) of its constituent proteins. Due to homeostasis, the plasma denaturation profiles of healthy individuals do not vary significantly. However, due to disease, the composition of the blood or the thermal stability of circulating proteins may change, thus altering the plasma denaturation profile.

In the context of this project, we demonstrated that nanoDSF can unambiguously distinguish several cancers using developed machine learning methods from a simple plasma sample. To achieve these objectives, we worked closely with oncologists to select cohorts of blood plasma from patients different types of cancers, including melanoma, lung, colorectal cancer, gliobastoma, breast., as well as from healthy donors. We determined the denaturation profiles for all plasma samples using the Prometheus NT.Plex. The profiles generated by each tool were stored in a developed in this project database with web interface and automatically classified using machine learning methods design in the frame of the proposed project.

10.3.4 ANR SAFE (2022-2026)

SAFE: Controlling networks with safety bounded and interpretable machine learning

Participants: Amaury Habrard.

Partners: XLIM / Univ. Poitiers, IRISA / Univ. Rennes 1, Huawei, QOS DESIGN and B. Jeudy & K. Singh (LabHC-UJM)

Type: ANR PRCE

When applied to communication networks, traditional approaches for control and decision-making require a comprehensive knowledge of system and user behaviours, which is unrealistic in practice when there is an increase in scale and complexity. Data-driven AI approaches do not have this drawback, but offer no safety bounds and are difficult to interpret. The SAFE project aims to design an innovative approach by combining the best of both worlds. In this new approach, intelligence is distributed in the network between a global AI (at the central level) and local AIs (at the edge level) collaborating with each other by integrating traditional models with graph neural networks and reinforcement learning. The approach, developed for partially or completely observable/controllable environments, will natively integrate safety bounds, interpretability and provide self-adaptive systems for routing, traffic engineering and scheduling. SAFE has following scientific objectives with an open source strategy: 1) Hierarchical architecture: Assuming modern network architectures, we will design a ML architecture based on global AI (running at central controller level) and local AI (running at edge device level) for decision-making in partially as well as fully observable and controllable environments. Global AI will be able to control, configure and install policies on local AI. 2) Algorithms for partially observable environments: We will design new safety bounded and interpretable algorithms for self-adaptive traffic engineering, automatic scheduling algorithms for partially observable and controllable environments. These methods find use cases in SD-WAN (Software-Defined Wide Area Networks), where edge devices present at customer premises need to collaboratively operate in overlay on top of partially observable core networks. 3) Algorithms for fully observable environments: We will investigate the application of the global and local AI architecture for fully observable and controllable environments. Specifically, we will design new safety bounded and interpretable algorithms for software-defined routing and traffic engineering, which find use cases in data centers as well as private WANs connecting multiple sites.

10.3.5 ANR FAMOUS (2023-2027)

Participants: Rémi Eyraud, Amaury Habrard.

Famous: Fair Multi-modal Learning

Partners: LIS Aix-Marseille, LITIS Rouen, INT Marseille, Euranova and A. Gouru, C. Largeron & E. Morvant (LabHC-UJM)

The aim of this project is to explore the first avenues of research into the contribution of multimodality in datasets to meet the requirements of fair learning. Fairness refers here to the biases (in the data and/or induced), while being interested in the interpretability of the models to help their certification. Each modality has its own statistical and topological characteristics, which requires upstream research on the adjustment of distributions when biased, adapted metrics, etc. Moreover, each one being itself a bias of observation of the data, this will be taken into account to establish a joint distribution (trans-modal) unbiased on all these modalities. With theoretical research in cross-modal statistical learning, we will study methods for reducing some types of identified biases (non iid, imbalances, sensitive variables) in the case of multimodal data. Two levels of treatment are privileged: (1) cross-modal pre-processing of biases in the data, by learning metrics, neural representations, and optimization constraints on kernel pre-images; (2) cross-modal algorithms for eliminating biases in model learning: cross-modal optimization algorithms, as well as optimal transfer and transport approaches between modalities to debias the concerned ones, based on the theoretical results previously obtained. Parsimony will be considered for scaling and explainability. Transversally, our work will be based on problems arising from real data sets in biology and health, multi-modal and presenting various types of bias, and on toy data sets to be generated. They have modalities where the data are structured in graphs: all our fundamental works will be declined to take into account this specificity impacting the treatment of the considered biases.

10.3.6 EUR SLEIGHT PIMALEA (2023-2024)

Participants: Eduardo Brandao, Rémi Emonet, Marc Sebban.

PIMALEA: Physics-Informed MAchine LEARning: From extraction to transfer of knowledge in surface engineering

Partners: JP Colombier (co-coordinator)

Type: EUR Manutech SLEIGHT

During the past few years, machine learning has been massively used to better understand multiscale physics, e.g. (i) by overcoming the limitations of costly numerical solvers of Partial Differential Equations (PDE), (ii) by learning from data the residuals of known physical models or more recently (iii) by discovering the latter explaining the underlying dynamics of observed data. In this context, a new generation of machine learning models emerged by integrating physical information to guide the learning process. This collection of techniques is generally grouped in under the "Physics-guided" or "Physics-informed" machine learning topic. In order to train the corresponding deep neural networks, these methods typically assume that they have access to a large enough amount of empirical data and/or they know the underlying physics allowing the generation of simulated examples. The ambitious goal of this project PIMALEA is to achieve scientific breakthroughs in the domain of self-organization of matter whose specificities constitute important pitfalls for a direct use of machine learning.

10.3.7 EUR SLEIGHT TREASURF (2024-2027)

TREASURF: TRansfer lEarning for Frugal and Accurate modeling of SURface Functionalization prediction –application to multicomponent alloys

Participants: Amaury Habrard, Rémi Emonet, Marc Sebban.

Partners: F. Garrelie and JP Colombier (LabHC-UJM)

Type: EUR Manutech SLEIGHT

TREASURF is an interdisciplinary project focusing on the development of novel machine learning approaches for the prediction of surface functionalization of different families of metals and metal alloys by (femto)laser irradiation. The ability to predict the micro- or nanopatterns induced by laser functionalization is a crucial challenge for an optimal use of surface properties. In this context, machine learning methods have been subject of a growing interest recently but they have to cope with of limited amounts of experimental data due to the very high acquisition costs. In the TREASURF project, we propose to address this problem by developing methods able to transfer the knowledge of a prediction model learned from a given metal or alloy to another, different but sharing certain properties. Re-training a new model is not a plausible hypothesis, mainly because of the difficulties involved in acquiring large quantities of data (laser irradiation + nanoscale imaging). This project is therefore situated in a difficult context of "frugal" learning. Our aim is to focus primarily on topographic predictions for two or more different alloy families. The project also envisages taking into account variability due to chemical changes to guide the transfer process. The advances made in this project will enable us to better characterize the impact of laser-matter interaction with the perspective of designing new surface functionalizations on various novel metal alloys, opening the door to new application prospects in numerous societal challenges related to health, energy, space, nuclear or defense.

11.1.1 Scientific events: organisation

11.1.2 Scientific events: selection

11.1.3 Journal

11.1.4 Invited talks

• Quentin Bertrand : "Some Challenges Around Retraining Generative Models on their Own Data", ECCV 2024 Workshop The Dark Side of Generative AIs and Beyond, October 2024.
• Rémi Emonet "Détection d’Anomalies dans les Images par Apprentissage Auto-supervisé", Journée FIL, Lyon, 13 June 2024.
• Benjamin Girault : "Learning a Graph and Importances of its Vertices", Machine Learning and Signal Processing Seminar, ENS Lyon, Jan 2024.
• Benjamin Girault : "Learning a Graph and Importances of its Vertices", Laboratoire Jean Kuntzman, Grenoble, May 2024.
• Amaury Habrard : "Domain Adaptation and Transfer Learning", invited speaker, Inter-PEPR Day (DIADEM, IA, NumPex), 18 November 2024.
• Amaury Habrard : "Choosing the right complexity measure: a data-driven artificial intelligence perspective", Congrès de l'IUF, 29 May 2024.
• Marc Sebban : "Physics-informed Machine Learning: From algorithms to theoretical questions", Journée scientifique "Simulation et IA" 3DS & Inria, Paris, 29 November 2024.
• Marc Sebban : "Is it possible to learn well from both data and physical knowledge?", invited speaker, Univ. Alicante, Spain, 29 May 2024.

11.1.5 Scientific expertise

• Rémi Emonet , scientific expert for the ANR AAPG2024 call.
• Amaury Habrard , scientific expert for the ANR AAPG2024 call, scientific expert for the ANRT (CIFRE), scientific expert for ERC project call (STG)
• Marc Sebban , scientific expert for the "Dispositif Doctorants" Normandy Region, 2024

11.1.6 Research administration

• Rémi Emonet is head of the Machine Learning project at laboratoire Hubert Curien.
• Amaury Habrard is head of the Data Intelligence team, member of the scientific board of the MIAI AI Cluster, co-head of the AI and ML axis of the Fédération des Laboratoires d'Informatique de Lyon (FIL), co-responsible of the working group on Physics-aware Machine Learning of the GdR IASIS
• Marc Sebban is deputy director of the Hubert Curien Lab (UMR CNRS 5516), Member of the Board of Directors of the Fédération des Laboratoires d'informatique de Lyon (FIL), Member of the executive committee of the EUR Manutech Sleight, Member of the COMEX Labex MILYON, Member of the COS Inria Lyon Centre.

11.2.1 Teaching

• Quentin Bertrand :
  • Computer Science Master, Numerical Optimal Transport for Machine and Deep Learning, 10h, M2, ENS Lyon.
• Eduardo Brandao :
  • Diplôme Universitaire Cycle initial en Technologie de l’Information de Saint-Étienne: Mathèmatiques pour l'ingénieur 1, 52.5h, L1, Télécom Saint-Étienne, UJM
  • Formation Ingénieur Télécom Saint-Étienne: Programmation Orientée Objet, 30h, 1AL3, Télécom Saint- Étienne, UJM
  • Formation Ingénieur Télécom Saint-Étienne: Mini-projet, 12h, 1AL3, Télécom Saint- Étienne, UJM
• Farah Cherfaoui :
  • Licence Mathématiques-Physique-Chimie: Outils informatiques, 91h, L1, Faculté des Sciences, UJM.
  • Licence informatique: programmation impérative, 56h, L1, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Advanced Machine Learning, 4h, M2, Faculté des Sciences, UJM.
• Rémi Emonet :
  • Master Machine Learning and Data Mining: Probablistic Graphical Models, 20h, M2, Faculté des Sciences, UJM.
  • Master Données et Systèmes Connectés: Programmation Web Avancée, 30h, M1, Faculté des Sciences, UJM.
• Rémi Eyraud :
  • Master Machine Learning and Data Mining: Research Methodology, 20h, 1A, Faculté des Sciences, UJM.
  • Master Données et Systèmes Connectés: Research Methodology, 20h, 2A, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Machine Learning Fundamentals and Algorithms, 30h, 1A, Faculté des Sciences, UJM.
  • Master Données et Systèmes Connectés: Machine Learning Fundamentals and Algorithms, 30h, 1A, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Advanced Machine Learning, 3h, M2, Faculté des Sciences, UJM.
  • GACO: Base de Données, 18h, 2A, IUT de Saint-Etienne, UJM.
  • GACO: Conception de Sites Web Dynamiques, 112h, 2A, IUT de Saint-Etienne, UJM.
  • GACO: Etablir le Diagnostique Marketing d'une Organisation (SAE), 8h, 2A, IUT de Saint-Etienne, UJM.
  • GACO: Traitement des Données, 20h, 3A, IUT de Saint-Etienne, UJM.
  • GACO: Environnement Informatique, 16h, 1A, IUT de Saint-Etienne, UJM.
• Benjamin Girault :
  • Master Machine Learning and Data Mining: Deep Learning I, 20h, M1, Faculté des Sciences, UJM.
• Amaury Habrard :
  • Master Machine Learning and Data Mining: Advanced Algorithms and Programming, 20h, M2, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Advanced Machine Learning, 20h, M2, Faculté des Sciences, UJM.
  • joint course Master Machine Learning and Data Mining and Master Données and Systèmes Connectés: Deep Learning II, 15h, M2, Faculté des Sciences, UJM.
• Jordan Patracone :
  • Formation Ingénieur Télécom Saint-Étienne: Langage C Algorithmie et structures de données, 32h, 1A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur Télécom Saint-Étienne: Algorithmique et structures de données, 13h, 1A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur Télécom Saint-Étienne: Projet recherche et innovation, 5h, 3A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur Télécom Saint-Étienne: Big Data Project, 10h, 3A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur Télécom Saint-Étienne: Projet d'ingéniérie, 14h, 2A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur en Apprentissage Data Engineering: Statistiques inférentielles, 33h, 2A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur en Apprentissage Data Engineering: Algorithms for data analysis, 28h, 2A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur en Apprentissage Data Engineering: Qualité de fonctionnement, 11h, 3A, Telecom Saint-Etienne, UJM.
  • Formation Ingénieur en Apprentissage Data Engineering: Innovation en data, 2h, 3A, Telecom Saint-Etienne, UJM.
  • Master Machine Learning and Data Mining: Advanced Machine Learning, 6h, M2, Faculté des Sciences, UJM.
• Marc Sebban :
  • Licence Informatique: Probabilités-Statistiques, 24h, L3, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Introduction to Machine Learning, 20h, M1, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Data Analysis, 24h, M1, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Advanced Machine Learning, 12h, M2, Faculté des Sciences, UJM.
  • Master Machine Learning and Data Mining: Machine Learning Project, 20h, M2, Faculté des Sciences, UJM.

11.2.2 Supervision

• PhD in progress: Fayad Ali Banna. Physics-guided Machine Learning, since Octo. 2022. Marc Sebban and Rémi Emonet .
• PhD in progress: Hind Atbir. Learning fair and robust kernel-based models with generalization guarantees, since Oct. 2024. Remi Eyraud and Farah Cherfaoui .
• Postdoc in progress: Antoine Caradot. On the estimation of integrals of functions: applications in PIML, since Oct. 2024. Rémi Emonet , Amaury Habrard and Marc Sebban .
• PhD in progress: Wissal Ghamour. Nouvelles approches de l’Intelligence Artificielle pour les attaques par canaux auxiliaires, since Dec. 2024. Rémi Eyraud
• PhD in progress: Dorian Llavata. Apprentissage Profond Diversement Supervisé Pour Les Attaques Par Canaux Auxiliaires, Rémi Eyraud , since 2022
• PhD in progress: Volodimir Mitarchuk, Theory and Algorithms for the Understanding of Deep Neural Network on Sequential Data, since 2021. Rémi Eyraud , Amaury Habrard , Rémi Emonet .
• PhD in progress: Robin Mermillod-blondin, Optimisation multicritère de couleurs plasmoniques pour les documents d’identité. Rémi Emonet .
• PhD in progress: Sayan Chaki, Unsupervised Analysis of Ornaments extracted from Ancient Books, Rémi Emonet .
• PhD in progress: Abdel-Rahim Mezidi, Unveiling and Incorporating Knowledge in Physics-Guided Machine Learning Models, Amaury Habrard , Jordan Patracone .
• PhD in progress: Ben Gao, Toward frugal machine learning with physics-aware models, Jordan Patracone .
• PhD in progress: Thibault Girardin, Prediction of multidimensional colors printed by laser on plasmonic metamaterials using deep learning and adaptive strategies, Amaury Habrard .
• PhD in progress: Erick Gomez, TRansfer lEarning for Frugal and Accurate modeling of SURface Functionalization prediction – application to multicomponent alloys, Amaury Habrard .
• PhD defended in 01/2024: Rehan Jhuboo. Morphometry-guided Super Resolution for Bone Microstructure CT lmaging, Marc Sebban

11.2.3 Juries

• Marc Sebban : Antonio Rios Vila (Univ. Alicante, President), Etienne Le Naour (PSL, Member), Eduardo Brandao (Univ. Saint-Etienne, Member), Rehan Jhuboo (Univ. Saint-Etienne, Director), Rémi Emonet (HDR, Univ. Saint-Etienne, Tutor).
• Rémi Emonet : Ashna Jose (PhD, Université Grenoble Alpes, Reviewer), Eduardo Brandao (Univ. Saint-Etienne, Supervisor).
• Rémi Eyraud : Reda Marzouk (Univ. Nantes, Member)
• Amaury Habrard : Steeven Janny (INSA Lyon, President), Emmanuel Menier (U.Paris-Saclay, Reviewer), Antoine de Mathelin (ENS-Paris-Saclay, Reviewer), Marin Scalbert (U.Paris-Saclay, Reviewer), Hoel Le Capitaine (HDR, Univ. Nantes, member), Eduardo Brandao (Univ. Saint-Etienne, invited)

11.3.1 Productions (articles, videos, podcasts, serious games, ...)

• Marc Sebban : interviews leading to several articles related to the creation of the MALICE team (IF Saint-Etienne, l’ESSOR, L’Usine Nouvelle, Dépêche AEF Info, Carnot TSN).
• Amaury Habrard : co-author of the white paper "Generative AI and Hypertrucages" from Minalogic for the Auvergne Rhone-Alpes Region
• Rémi Emonet, Quentin Bertrand, collaboration with Ockham, we wrote a friendly blog post on recent flow matching techniques.

11.3.2 Participation in Live events

• Rémi Eyraud : 2h of conference titled "de l’Algorithme à l’IA". Conférences pour Tous.tes – MJC Annonay (14/03/2024) ; Association Culturelle des Monts du Matin – St-Genis l'Argentière (10/12/02024)
• Rémi Eyraud : 2h conference titled "Démystifier l’IA". Saison culturelle, villes de Sorbiers et St-Jean-Bonnefond, 26/11/2024.

11.3.3 Others science outreach relevant activities

• Jordan Patracone : presentation to the executive committee of CARSAT Rhône-Alpes (Artificial intelligence at the service of public organizations)
• Rémi Eyraud : round table on "IA: au-delà des craintes", Comité Loire Connect, Département de la Loire.
• Amaury Habrard : round table on Artificial Intelligence, Club Eco - agence d’urbanisme EPURES (04/12/2024).
• Amaury Habrard : round table on the future of the Labex MILyon (14/10/2024)

International journals

• 8 articleK.Kacper Bilko, R. G.Rubén García Alía, M. S.Mario Sacristan Barbero, S.Sylvain Girard, Y.Ygor Aguiar, M.Matteo Cecchetto, C.Camille Belanger-Champagne, S.Salvatore Danzeca, W.Wojtek Hajdas, A.Alex Hands, P. M.Pedro Martín Holgado, Y. M.Yolanda Morilla Garcia, A. R.Amor Romero Maestre, D.Daniel Prelipcean, F.Federico Ravotti and M.Marc Sebban. Mixed-field Radiation Monitoring and Beam Characterisation Through Silicon Diode Detectors.IEEE Transactions on Nuclear Science714January 2024, 777-784HALDOI
• 9 articleK.Kacper Bilko, R. G.Rubén García Alía, A.Alessandra Constantino, A.Andrea Coronetti, S.Salvatore Danzeca, M.Marc Delrieux, N.Natalia Emriskova, M. A.Matthew Alexander Fraser, S.Sylvain Girard, E. P.Eliott Philippe Johnson, M.Marc Sebban, F.Federico Ravotti and A.Andreas Waets. CHARM High-energy Ions for Micro Electronics Reliability Assurance (CHIMERA).IEEE Transactions on Nuclear ScienceJanuary 2024, 1-1HALDOI
• 10 articleV.Volodimir Mitarchuk, R.Rémi Emonet, R.Rémi Eyraud and A.Amaury Habrard. On the theoretical limit of gradient descent for Simple Recurrent Neural Networks with finite precision.Transactions on Machine Learning Research Journal31242024HAL
• 11 articleD.Damien Robissout, L.Lilian Bossuet and A.Amaury Habrard. Scoring the predictions: a way to improve profiling side-channel attacks.Journal of Cryptographic EngineeringApril 2024HALDOI

International peer-reviewed conferences

• 12 inproceedingsF.Fayad Ali Banna, J.-P.Jean-Philippe Colombier, R.Rémi Emonet, M.Marc Sebban and R.Rémi Emonet. Physics-informed Machine Learning for Better Understanding Laser-Matter Interaction.IEEEThe 36th IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2024)11-7Herndon, VA, United StatesIEEE Trans2024, 7HALback to textback to text
• 13 inproceedingsJ.Jorge Azorin-Lopez, M.Marc Sebban, N.Nahuel Garcia-d'Urso, A.Amaury Habrard and A.Andres Fuster-Guillo. Generative shape deformation with optimal transport using learned transformations.IEEE 2024, ISBN 979-8-3503-5931-2International Joint Conference on Neural Networks (IJCNN)YOKOHAMA, JapanJune 2024HAL
• 14 inproceedingsS. K.Sayan Kumar Chaki, Z. S.Zeynep Sonat Baltaci, E.Elliot Vincent, R.Rémi Emonet, F.Fabienne Vial-Bonacci, C.Christelle Bahier-Porte, M.Mathieu Aubry and T.Thierry Fournel. Historical Printed Ornaments: Dataset and Tasks.ICDAR 2024 - International Conference on Document Analysis and Recognition14806Lecture Notes in Computer ScienceAthens, GreeceSpringer Nature SwitzerlandSeptember 2024, 251-270HALDOI
• 15 inproceedingsS.Sami Marouani, K.Kamal Singh, B.Baptiste Jeudy, A.Abbas Bradai and A.Amaury Habrard. Advanced Traffic Engineering in WAN Using Graph Attention Networks.Proceedings of the 20th International Conference on Wireless and Mobile Computing, Networking and Communications, Wimob 2024The 20th International Conference on Wireless and Mobile Computing, Networking and Communications, Wimob 2024Paris, FranceOctober 2024HAL
• 16 inproceedingsV.Volodimir Mitarchuk and R.Rémi Eyraud. A Theoretical Analysis of the Incremental Counting Ability of LSTM in Finite Precision.LearnAut workshop 2024Tallinn, Estonia2024HAL
• 17 inproceedingsE. G.Erick Gomez Soto, R.Rémi Emonet and M.Marc Sebban. Unsupervised Learning and Effective Complexity: introducing JPG and Neural Sophistication.International Conference on Tools with Artificial Intelligence (ICTAI)Herndon, United StatesOctober 2024HALback to textback to text

Conferences without proceedings

• 18 inproceedingsJ.-P.Jean-Philippe Colombier, E.Eduardo Brandao, A.Anthony Nakhoul, F.Fayad Ali Banna, R.Rémi Emonet, A.Amaury Habrard, F.François Jacquenet, F.Florence Garrelie and M.Marc Sebban. Deciphering the complexity behind laser-induced selforganized nanopatterns.17th International Conference on Laser Ablation (COLA 2024)Hersonissos, GreeceSeptember 2024HAL

Reports & preprints

• 19 miscA.-R.Abdel-Rahim Mezidi, J.Jordan Patracone, S.Saverio Salzo, A.Amaury Habrard, M.Massimiliano Pontil, R.Rémi Emonet and M.Marc Sebban. Bregman Proximal Viewpoint on Neural Operators.June 2024HALback to textback to text
• 20 miscM.Marianna de Santis, J.Jordan Patracone, F.Francesco Rinaldi, S.Saverio Salzo, M.Martin Schmidt and S.Sara Venturini. Relax and penalize: a new bilevel approach to mixed-binary hyperparameter optimization.January 2025HAL

Other scientific publications

• 21 inproceedingsF.Fayad Ali Banna, R.Rémi Emonet, A.Anton Rudenko, M.Marc Sebban and J.-P.Jean-Philippe Colombier. Predicting laser energy absorption on nanostructured surfaces with deep learning.Machine Learning in PhotonicsStrasbourg, FranceSPIEApril 2024, 74HALDOI