Search

The determination of the number of mixture components (the order) of a finite mixture model has been an enduring problem in statistical inference. We prove that the closed testing principle leads to a sequential testing procedure (STP) that allows for confidence statements to be made regarding the order of a finite mixture model. We construct finite sample tests, via data splitting and data swapping, for use in the STP, and we prove that such tests are consistent against fixed alternatives. Simulation studies and real data examples are used to demonstrate the performance of the finite sample tests-based STP, yielding practical recommendations of their use as confidence estimators in combination with point estimates such as the Akaike information or Bayesian information criteria. In addition, we demonstrate that a modification of the STP yields a method that consistently selects the order of a finite mixture model, in the asymptotic sense. Our STP is not only applicable for order selection of finite mixture models, but is also useful for making confidence statements regarding any sequence of nested models.

Convolutional neural networks (CNN) have demonstrated their ability to segment 2D cardiac ultrasound images. However, despite recent successes according to which the intra-observer variability on end-diastole and end-systole images has been reached, CNNs still struggle to leverage temporal information to provide accurate and temporally consistent segmentation maps across the whole cycle. Such consistency is required to accurately describe the cardiac function, a necessary step in diagnosing many cardiovascular diseases. In this paper, we propose a framework to learn the 2D+time apical long-axis cardiac shape such that the segmented sequences can benefit from temporal and anatomical consistency constraints. Our method is a post-processing that takes as input segmented echocardiographic sequences produced by any state-of-the-art method and processes it in two steps to (i) identify spatio-temporal inconsistencies according to the overall dynamics of the cardiac sequence and (ii) correct the inconsistencies. The identification and correction of cardiac inconsistencies relies on a constrained autoencoder trained to learn a physiologically interpretable embedding of cardiac shapes, where we can both detect and fix anomalies. We tested our framework on 98 full-cycle sequences from the CAMUS dataset, which are available alongside this paper. Our temporal regularization method not only improves the accuracy of the segmentation across the whole sequences, but also enforces temporal and anatomical consistency., 12 pages, accepted for publication in IEEE TMI

The BumpHunter algorithm is widely used in the search for new particles in High Energy Physics analysis. This algorithm offers the advantage of evaluating the local and global p-values of a localized deviation in the observed data without making any hypothesis on the supposed signal. The increasing popularity of the Python programming language motivated the development of a new public implementation of this algorithm in Python, called pyBumpHunter, together with several improvements and additional features. It is the first public implementation of the BumpHunter algorithm to be added to Scikit-HEP. This paper presents in detail the BumpHunter algorithm as well as all the features proposed in this implementation. All these features have been tested in order to demonstrate their behaviour and performance., 14 pages, 9 figures

International audience; In the Global Navigation Satellite System (GNSS) context, the growing number of available satellites has lead to many challenges when it comes to choosing the most accurate pseudorange contributions, given the strong impact of biased measurements on positioning accuracy, particularly in singleepoch scenarios. This work leverages the potential of machine learning in predicting link-wise measurement quality factors and, hence, optimize measurement weighting. For this purpose, we use a customized matrix composed of heterogeneous features such as conditional pseudorange residuals and per-link satellite metrics (e.g., carrier-to-noise power density ratio and its empirical statistics, satellite elevation, carrier phase lock time). This matrix is then fed as an input to a recurrent neural network (RNN) (i.e., a long-short term memory (LSTM) network). Our experimental results on real data, obtained from extensive field measurements, demonstrate the high potential of our proposed solution being able to outperform traditional measurements weighting and selection strategies from state-of-the-art.

International audience; Reinforcement learning (RL) and trajectory optimization (TO) present strong complementary advantages. On one hand, RL approaches are able to learn global control policies directly from data, but generally require large sample sizes to properly converge towards feasible policies. On the other hand, TO methods are able to exploit gradient-based information extracted from simulators to quickly converge towards a locally optimal control trajectory which is only valid within the vicinity of the solution. Over the past decade, several approaches have aimed to adequately combine the two classes of methods in order to obtain the best of both worlds. Following on from this line of research, we propose several improvements on top of these approaches to learn global control policies quicker, notably by leveraging sensitivity information stemming from TO methods via Sobolev learning, and augmented Lagrangian techniques to enforce the consensus between TO and policy learning. We evaluate the benefits of these improvements on various classical tasks in robotics through comparison with existing approaches in the literature.

Understanding and controlling latent representations in deep generative models is a challenging yet important problem for analyzing, transforming and generating various types of data. In speech processing, inspiring from the anatomical mechanisms of phonation, the source-filter model considers that speech signals are produced from a few independent and physically meaningful continuous latent factors, among which the fundamental frequency $f_0$ and the formants are of primary importance. In this work, we start from a variational autoencoder (VAE) trained in an unsupervised manner on a large dataset of unlabeled natural speech signals, and we show that the source-filter model of speech production naturally arises as orthogonal subspaces of the VAE latent space. Using only a few seconds of labeled speech signals generated with an artificial speech synthesizer, we propose a method to identify the latent subspaces encoding $f_0$ and the first three formant frequencies, we show that these subspaces are orthogonal, and based on this orthogonality, we develop a method to accurately and independently control the source-filter speech factors within the latent subspaces. Without requiring additional information such as text or human-labeled data, this results in a deep generative model of speech spectrograms that is conditioned on $f_0$ and the formant frequencies, and which is applied to the transformation speech signals. Finally, we also propose a robust $f_0$ estimation method that exploits the projection of a speech signal onto the learned latent subspace associated with $f_0$., 23 pages, 7 figures, companion website: https://samsad35.github.io/site-sfvae/

International audience; Co-clustering aims at simultaneously partitioning both dimensions of a data matrix. It has demonstrated better performances than one-sided clustering for high-dimensional data. The Latent Block Model (LBM) is a probabilistic model for co-clustering based on mixture models that has proven useful for a broad class of data. In this paper, we propose to leverage prior knowledge in the form of pairwise semi-supervision in both row and column space to improve the clustering performances of the algorithms derived from this model. We present a general probabilistic framework for incorporating Must Link and Cannot Link relationships in the LBM based on Hidden Markov Random Fields (HMRF). We instantiate this framework on a model for count data and present two inference algorithms based on Variational and Classification EM. Extensive experiments on simulated data and on real-world attributed networks confirm the interest of our approach and demonstrate the effectiveness of our algorithms.

International audience; Stochastic Block Model (SBM) provides a statistical tool for modeling and clustering network data. In this paper, we propose an extension of this model for discrete-time dynamic networks that takes into account the variability in node degrees, allowing us to model a broader class of networks. We develop a probabilistic model that generates temporal graphs with a dynamic cluster structure and time-dependent degree corrections for each node. Thanks to these degree corrections, the nodes can have variable in-and out-degrees, allowing us to model complex cluster structures as well as interactions that decrease or increase over time. We compare the proposed model to a model without degree correction and highlight its advantages in the case of inhomogenous degree distributions in the clusters and in the recovery of unstable cluster dynamics. We propose an inference procedure based on Variational Expectation-Maximization (VEM) that also provides the means to estimate the time-dependent degree corrections. Extensive experiments on simulated and real datasets confirm the benefits of our approach and show the effectiveness of the proposed algorithm.

International audience; Background: Speech disorders are amongst the first symptoms to appear in Parkinson's disease (PD). Objectives: We aimed to characterize PD voice signature from the prodromal stage (isolated rapid eye movement sleep behavior disorder, iRBD) to early PD using an automated acoustic analysis and compare male and female patients. We carried out supervised learning classifications to automatically detect patients using voice only. Methods: Speech samples were acquired in 256 French speakers (117 participants with early PD, 41 with iRBD, and 98 healthy controls), with a professional quality microphone, a computer microphone and their own telephone. High-level features related to prosody, phonation, speech fluency and rhythm abilities were extracted. Group analyses were performed to determine the most discriminant features, as well as the impact of sex, vocal tasks, and microphone type. These speech features were used as inputs of a support vector machine and were combined with classifiers using low-level features. Results: PD related impairments were found in prosody, pause durations and rhythmic abilities, from the prodromal stage. These alterations were more pronounced in men than in women. Early PD detection was achieved with a balanced accuracy of 89% in males and 70% in females. Participants with iRBD were detected with a balanced accuracy of 63% (reaching 70% in the subgroup with mild motor symptoms). Conclusion: This study provides new insight in the characterization of sex-dependent early PD speech impairments, and demonstrates the valuable benefit of including automated voice analysis in future diagnostic procedures of prodromal PD.

International audience; Tensor data analysis is the evolutionary step of data analysis to more than two dimensions. Dealing with tensor data is often based on tensor decomposition methods. The present paper focuses on unsupervised learning and provides a python package referred to as TensorClus including novel co-clustering algorithms of three-way data. All proposed algorithms are based on the latent block models and suitable to different types of data, sparse or not. They are successfully evaluated on challenges in text mining, recommender systems, and hyperspectral image clustering. TensorClus is an open-source Python package that allows easy interaction with other python packages such as NumPy and TensorFlow; it also offers an interface with some tensor decomposition packages namely Tensorly and TensorD on the one hand, and on the other, the co-clustering package Coclust. Finally, it provides CPU and GPU compatibility. The TensorClus library is available at https://pypi.org/project/TensorClus/ 1 .

Nowadays, the data used for decision-making come from a wide variety of sources which are difficult to manage using relational databases. To address this problem, many researchers have turned to Not only SQL (NoSQL) databases to provide scalability and flexibility for On-Line Analytical Processing (OLAP) systems. In this paper, we propose a set of formal rules to convert a multidimensional data model into a graph data model (MDM2G). These rules allow conventional star and snowflake schemas to fit into NoSQL graph databases. We apply the proposed rules to implement star-like and snowflake-like graph data warehouses. We compare their performances to similar relational ones focusing on the data model, dimensionality, and size. The experimental results show large differences between relational and graph implementations of a data warehouse. A relational implementation performs better for queries on a couple of tables, but conversely, a graph implementation is better when queries involve many tables. Surprisingly the performances of a star-like and snowflake-like graph data warehouses are very close. Hence a snowflake schema could be used in order to easily consider new sub-dimensions in a graph data warehouse.

In recent years, Empirical Bayesian (EB) inference has become an attractive approach for estimation in parametric models arising in a variety of real-life problems, especially in complex and high-dimensional scientific applications. However, compared to the relative abundance of available general methods for computing point estimators in the EB framework, the construction of confidence sets and hypothesis tests with good theoretical properties remains difficult and problem specific. Motivated by the universal inference framework of Wasserman et al. [2020], we propose a general and universal method, based on holdout likelihood ratios, and utilizing the hierarchical structure of the specified Bayesian model for constructing confidence sets and hypothesis tests that are finite sample valid. We illustrate our method through a range of numerical studies and real data applications, which demonstrate that the approach is able to generate useful and meaningful inferential statements in the relevant contexts.

International audience; Nowadays, power optimization has become a major interest for most digital hardware designers. Some, traditionally, might stick to offline power estimation especially in early design phases; some others resort to the modern and very promising runtime power management. Therefore, the Online Power Monitoring (OPM) is considered as an important feature serving real-time power optimization. OPM favors both in-situ power estimation and subsequent prediction, and can be exploited by the Dynamic Voltage and Frequency Scaling (DVFS) mechanism. DVFS, a modern technique for digital circuits power optimization, provides a realtime and adaptive voltage and/or frequency tuning while securing the systems' performance and integrity. In this paper, we present and evaluate an accurate online power monitoring methodology of FPGA IPs. We estimate power consumption using machine learning techniques, based on the IP's most power-influential operating modes and its inputs activity characteristics. The proposed online monitoring mechanism drastically reduces the communication-derived latency between the monitor and the DVFS. Experimental results show a mean absolute percentage error below 1.5% for the estimated power consumption.

Distributions of high-dimensional data can be learnt with unsupervised architectures, such as restricted Boltzmann machines (RBM). However, the resulting models are often uneasy to sample when the data distributions include multiple modes. We here consider deep tempering, a paralleltempering-like Monte Carlo sampling algorithm based on a chain of several restricted Boltzmann machines (RBM), where hidden configuration of a machine can be exchanged with the visible configurations of the next one along the chain. Replica exchanges between the different RBM is facilitated by the increasingly clustered representations learnt by deeper RBMs along the chain, allowing for fast transitions between the different modes of the data distribution. We explain why deep tempering works on hierarchical data, and introduce a theoretical framework to understand how hyperparameters, such as the aspect ratios of the RBMs and the weight regularization should be chosen. Our findings are illustrated on two datasets: MNIST and in silico Lattice Proteins.

The problem of learning a communication policy is investigated in this paper for the design of an event-triggered observer for discrete-time LTI systems. Firstly, the event-triggered observer problem is formulated as an optimisation problem. The existence of a solution to this problem (communication policy) is investigated and it is verified if this solution still preserves the stability of the estimation error dynamics. Secondly, an algorithm is provided to approximate this optimal solution using neural networks and deep learning. Simulation examples are provided to illustrate the effectiveness of the learned communication policies.

International audience; Deep learning methods have become very popular for the processing of natural images, and were then successfully adapted to the neuroimaging field. As these methods are non-transparent, interpretability methods are needed to validate them and ensure their reliability. Indeed, it has been shown that deep learning models may obtain high performance even when using irrelevant features, by exploiting biases in the training set. Such undesirable situations can potentially be detected by using interpretability methods. Recently, many methods have been proposed to interpret neural networks. However, this domain is not mature yet. Machine learning users face two major issues when aiming to interpret their models: which method to choose, and how to assess its reliability? Here, we aim at providing answers to these questions by presenting the most common interpretability methods and metrics developed to assess their reliability, as well as their applications and benchmarks in the neuroimaging context. Note that this is not an exhaustive survey: we aimed to focus on the studies which we found to be the most representative and relevant.

Domain-wide recognized by their high value in human presence and activity studies, cellular network datasets (i.e., Charging Data Records, named CdRs), however, present accessibility, usability, and privacy issues, restricting their exploitation and research reproducibility. This paper tackles such challenges by modeling Cdrs that fulfill real-world data attributes. Our designed framework, named Zen follows a fourfold methodology related to (i) the LTSM-based modeling of users' traffic behavior, (ii) the realistic and flexible emulation of spatiotemporal mobility behavior, (iii) the structure of lifelike cellular network infrastructure and social interactions, and (iv) the combination of the three previous modules into realistic Cdrs traces with an individual basis, realistically. Results show that Zen's first and third models accurately capture individual and global distributions of a fully anonymized real-world Cdrs dataset, while the second model is consistent with the literature's revealed features in human mobility. Finally, we validate Zen Cdrs ability of reproducing daily cellular behaviors of the urban population and its usefulness in practical networking applications such as dynamic population tracing, Radio Access Network's power savings, and anomaly detection as compared to real-world CdRs.

With the increasing presence of robots in our every-day environments, improving their social skills is of utmost importance. Nonetheless, social robotics still faces many challenges. One bottleneck is that robotic behaviors need to be often adapted as social norms depend strongly on the environment. For example, a robot should navigate more carefully around patients in a hospital compared to workers in an office. In this work, we investigate meta-reinforcement learning (meta-RL) as a potential solution. Here, robot behaviors are learned via reinforcement learning where a reward function needs to be chosen so that the robot learns an appropriate behavior for a given environment. We propose to use a variational meta-RL procedure that quickly adapts the robots' behavior to new reward functions. As a result, given a new environment different reward functions can be quickly evaluated and an appropriate one selected. The procedure learns a vectorized representation for reward functions and a meta-policy that can be conditioned on such a representation. Given observations from a new reward function, the procedure identifies its representation and conditions the meta-policy to it. While investigating the procedures' capabilities, we realized that it suffers from posterior collapse where only a subset of the dimensions in the representation encode useful information resulting in a reduced performance. Our second contribution, a radial basis function (RBF) layer, partially mitigates this negative effect. The RBF layer lifts the representation to a higher dimensional space, which is more easily exploitable for the meta-policy. We demonstrate the interest of the RBF layer and the usage of meta-RL for social robotics on four robotic simulation tasks., 16 pages, 14 figures

International audience; La décomposition tensorielle a récemment fait l'objet d'une attention croissante dans la communauté de l'apprentissage automatique en raison de sa polyvalence dans le traitement des données à grande échelle. Cependant, cette tâche devient plus difficile lorsqu'il s'agit de prendre en comte la dimension temporelle. Dans cet article, nous étendons la décomposition tensorielle à l'extraction de phénotypes temporels, décrits comme un combinaison de caractéristiques sur une fenêtre de temps. Nous proposons un nouveau modèle de décomposition intégrant plusieurs régularisations pour améliorer l'interprétabilité des phénotypes extraits. Nous validons ce dernier sur des données synthétiques et réelles provenant de l'Assistance Publique-Hôpitaux de Paris (AP-HP). Les résultats montrent qu'il est plus performant que les modèles les plus récents de décomposition et qu'il découvre des phénotypes intéressants pour les cliniciens.

The increasingly widespread adoption of large language models has highlighted the need for improving their explainability. We present context length probing, a novel explanation technique for causal language models, based on tracking the predictions of a model as a function of the length of available context, and allowing to assign differential importance scores to different contexts. The technique is model-agnostic and does not rely on access to model internals beyond computing token-level probabilities. We apply context length probing to large pre-trained language models and offer some initial analyses and insights, including the potential for studying long-range dependencies. The source code and an interactive demo of the method are available., 11 pages, 9 figures. ACL 2023 short paper camera-ready. Demos at https://cifkao.github.io/context-probing/ and https://huggingface.co/spaces/cifkao/context-probing ; code at https://github.com/cifkao/context-probing/

Bandits are one of the most basic examples of decision-making with uncertainty. A Markovian restless bandit can be seen as the following sequential allocation problem: At each decision epoch, one or several arms are activated (pulled); all arms generate an instantaneous reward that depend on their state and their activation; the state of each arm then changes in a Markovian fashion, based on an underlying transition matrix. Both the rewards and the probability matrices are known, and the new state is revealed to the decision maker for its next decision. The word restless serves to emphasize the fact that arms that are not activated can also change states, hence generalizes the simpler rested bandits. In principle, the above problem can be solved by dynamic programming, since it is a Markov decision process. The challenge that we face is the curse of dimension, since the size of possible states and actions grows exponentially with the number of arms of the bandit. Consequently, the focus is to design policies that solve the dilemma of computational efficiency and close-to-optimal performance.In this thesis, we construct computationally efficient policies with provable performance bounds, that may differ depending on certain properties of the problem. In Part I , we first investigate the classical Whittle index policy (WIP) on infinite horizon problems, and prove that if it is asymptotically optimal under the global attractor assumption, then almost always it converges to the optimal value exponentially fast. The application of WIP has the additional technical assumption of indexability as a prerequisite, to get around this, we next study the LP-index policy, that is well-defined for any problem, and shares the same exponential speed of convergence as WIP under similar assumptions.In infinite horizon, we always need the global attractor assumption for asymptotic optimality. In Part II of the thesis, we study the problem under finite horizon, so that this assumption is no-longer a concern. Instead, the LP-compatibility and the non-degeneracy are required for the asymptotic optimality and a faster convergence rate. We construct the finite horizon LP-index policy, as well as the LP-update policy, that amounts to solving new LP-index policies during the evolution of the process. This latter LP-update policy is then generalized to the broader framework of weakly coupled MDPs, together with the generalization of the non-degenerate condition. This condition also allows a more efficient implementation of the LP-update policy, as well as a faster convergence rate, if it is satisfied on the weakly coupled MDPs., Les bandits sont l'un des exemples les plus élémentaires de prise de décision avec incertitude. Un bandit agité markovien peut être vu comme le problème d'allocation séquentielle suivant : à chaque époque de décision, un ou plusieurs bras sont activés (tirés) ; tous les bras génèrent une récompense instantanée qui dépend de leur état et de leur activation ; l'état de chaque bras change ensuite de façon markovienne, en fonction d'une matrice de transition sous-jacente. Les récompenses et les matrices de probabilité sont toutes deux connues, et le nouvel état est révélé au décideur pour sa prochaine décision. Le terme "agité" sert à souligner le fait que les bras qui ne sont pas activés peuvent également changer d'état, ce qui généralise les bandits reposés plus simples. En principe, le problème ci-dessus peut être résolu par programmation dynamique, puisqu'il s'agit d'un processus de décision de Markov. Le défi auquel nous sommes confrontés est la malédiction de la dimension, puisque la taille des états et des actions possibles croît de manière exponentielle avec le nombre de bras du bandit. Par conséquent, l'objectif est de concevoir des politiques qui résolvent ce dilemme de l'efficacité de calcul et de la performance proche de l'optimum.Dans cette thèse, nous construisons des politiques efficaces en termes de calcul avec des limites de performance prouvables, qui peuvent varier en fonction de certaines propriétés du problème. Dans la première partie, nous étudions d'abord la politique classique de l'indice de Whittle (WIP) sur des problèmes à horizon infini, et nous prouvons que si elle est asymptotiquement optimale sous l'hypothèse de l'attracteur global, alors elle converge presque toujours vers la valeur optimale de façon exponentielle. L'application de WIP requière l'hypothèse technique supplémentaire de l'indexabilité comme prérequis. Pour contourner cela, nous étudions ensuite la politique LP-index, qui est bien définie pour tout problème, et partage la même vitesse exponentielle de convergence que WIP sous des hypothèses similaires.En horizon infini, nous avons toujours besoin de l'hypothèse d'attracteur global pour l'optimalité asymptotique. Dans la partie II de la thèse, nous étudions le problème à horizon fini, de sorte que cette hypothèse n'est plus un problème. Au lieu de cela, la compatibilité LP et la non-dégénérescence sont requises pour l'optimalité asymptotique et un taux de convergence plus rapide. Nous construisons la politique LP-index à horizon fini, ainsi que la politique LP-mise à jour, qui revient à résoudre de nouvelles politiques LP-index au cours de l'évolution du processus. Cette dernière politique LP-update est ensuite généralisée au cadre plus large des MDPs faiblement couplés, ainsi que la généralisation de la condition de non-dégénérescence. Cette condition permet également une mise en œuvre plus efficace de la politique de mise à jour LP, ainsi qu'un taux de convergence plus rapide, si l'on tient compte de l'évolution du processus.

The increasing size of available data has led machine learning specialists to consider more complex models in order to achieve better performance. From a theoretical point of view, statistical learning under resource constraints has known a growing interest in the machine learning community. Many settings were developed to formalize budgeted limitations. In this thesis, we are motivated by these challenges, where we consider classical learning problems under the ``frugal lens". First, we tackle support recovery in a sparse linear regression problem, with one pass over data. We develop an online greedy algorithm named "online orthogonal matching pursuit" that actively selects covariates in a sequential way, with guarantees on its computational complexity that is adaptive to the unknown magnitude of the regression coefficients. Second, we consider the problem of model selection aggregation of experts. We present procedures that achieve fast rates under various budgeted settings and discuss the attainability of fast rates in different settings. Third, we tackle the problem of online prediction of individual sequences, where no distributional assumption is made in the process of generating data. We consider some natural budgeted constraints on the number of experts used for prediction and the number of observed feedbacks. We develop new strategies for each setting and discuss the attainability of constant regrets. Finally, we consider the problem of fixed confidence best arm identification. Given a confidence level, the learner wants to identify the arm with the largest mean using the least number of queries possible. We suppose that simultaneous queries are possible and prove that significant improvement can be made with respect to the BAI standard algorithms by taking the unknown covariance of the arms into consideration., Depuis le début du développement de la théorie de l'apprentissage statistique, un intérêt particulier a été porté aux méthodes efficaces en temps de calcul ainsi qu'en espace de stockage nécessaire, afin qu'elles soient utilisables en pratique. Ceci a motivé plusieurs théoriciens à formaliser différents problèmes d'apprentissage statistique sous contrainte d'accès aux données et aux ressources computationnelles. Dans cette thèse, nous avons considéré plusieurs problèmes d'apprentissage statistiques et d'apprentissage séquentiel, sous différents types de contraintes. Le premier problème traité concerne la régression parcimonieuse sous une contrainte de nature computationnelle. Nous développons un algorithme effectuant un seul passage sur les données (celles-ci sont supposées arriver en temps réel) avec une limitation sur l'espace mémoire disponible. Le deuxième problème traité concerne l'agrégation d'experts. Nous revisitons ce problème dans le cas où l'accès aux données est limité et développons des méthodes permettant d'atteindre des taux rapides pour l'excès de risques. Le problème suivant concerne l'agrégation d'experts pour la prédiction des suites individuelles fixes. Nous introduisant un formalisme similaire à celui utilisé dans le problème précédent: nous supposons que pour chaque tour, le joueur a une contrainte sur le nombre d'experts à utiliser pour la prédiction et une contrainte sur le nombre de pertes d'experts individuels observées après avoir fait une prédiction. Nous présentant des procédures pour chaque cas et développons des garanties théoriques sur le regret cumulé des stratégies présentées. Le dernier problème considéré est une instance du problème de l'identification du meilleur bras dans le cadre de la théorie des bandits stochastiques. Nous présentons une extension du formalisme standard en permettant le tirage de plusieurs bras simultanément. Dans ce cadre nous montrons que de nouvelles bornes, potentiellement meilleurs que les bornes classiques, sont possibles, et nous présentons des procédures permettant de les atteindre.

International audience; In federated learning, clients such as mobile devices or data silos (e.g. hospitals and banks) collaboratively improve a shared model, while maintaining their data locally. Multiple recent works show that client's private information can still be disclosed to an adversary who just eavesdrops the messages exchanged between the targeted client and the server. In this paper, we propose a novel model-based attribute inference attack in federated learning which overcomes the limits of gradient-based ones. Furthermore, we provide an analytical lower-bound for the success of this attack. Empirical results using real world datasets confirm that our attribute inference attack works well for both regression and classification tasks. Moreover, we benchmark our novel attribute inference attack against the state-ofthe-art attacks in federated learning. Our attack results in higher reconstruction accuracy especially when the clients' datasets are heterogeneous (as it is common in federated learning). Most importantly, our model-based fashion of designing powerful and explainable attacks enables an effective quantification of the privacy risk in FL.

An approach to the time-accurate prediction of chaotic solutions is by learning temporal patterns from data. Echo State Networks (ESNs), which are a class of Reservoir Computing, can accurately predict the chaotic dynamics well beyond the predictability time. Existing studies, however, also showed that small changes in the hyperparameters may markedly affect the network's performance. The aim of this paper is to assess and improve the robustness of Echo State Networks for the time-accurate prediction of chaotic solutions. The goal is three-fold. First, we investigate the robustness of routinely used validation strategies. Second, we propose the Recycle Validation, and the chaotic versions of existing validation strategies, to specifically tackle the forecasting of chaotic systems. Third, we compare Bayesian optimization with the traditional Grid Search for optimal hyperparameter selection. Numerical tests are performed on two prototypical nonlinear systems that have both chaotic and quasiperiodic solutions. Both model-free and model-informed Echo State Networks are analysed. By comparing the network's robustness in learning chaotic versus quasiperiodic solutions, we highlight fundamental challenges in learning chaotic solutions. The proposed validation strategies, which are based on the dynamical systems properties of chaotic time series, are shown to outperform the state-of-the-art validation strategies. Because the strategies are principled-they are based on chaos theory such as the Lyapunov time-they can be applied to other Recurrent Neural Networks architectures with little modification. This work opens up new possibilities for the robust design and application of Echo State Networks, and Recurrent Neural Networks, to the time-accurate prediction of chaotic systems., Comment: 29 pages, 25 figures

In the context of sketching for compressive mixture modeling, we revisit existing proofs of the Restricted Isometry Property of sketching operators with respect to certain mixtures models. After examining the shortcomings of existing guarantees, we propose an alternative analysis that circumvents the need to assume importance sampling when drawing random Fourier features to build random sketching operators. Our analysis is based on new deterministic bounds on the restricted isometry constant that depend solely on the set of frequencies used to define the sketching operator; then we leverage these bounds to establish concentration inequalities for random sketching operators that lead to the desired RIP guarantees. Our analysis also opens the door to theoretical guarantees for structured sketching with frequencies associated to fast random linear operators.

Federated Learning (FL) is a novel approach enabling several clients holding sensitive data to collaboratively train machine learning models, without centralizing data. The cross-silo FL setting corresponds to the case of few ($2$--$50$) reliable clients, each holding medium to large datasets, and is typically found in applications such as healthcare, finance, or industry. While previous works have proposed representative datasets for cross-device FL, few realistic healthcare cross-silo FL datasets exist, thereby slowing algorithmic research in this critical application. In this work, we propose a novel cross-silo dataset suite focused on healthcare, FLamby (Federated Learning AMple Benchmark of Your cross-silo strategies), to bridge the gap between theory and practice of cross-silo FL. FLamby encompasses 7 healthcare datasets with natural splits, covering multiple tasks, modalities, and data volumes, each accompanied with baseline training code. As an illustration, we additionally benchmark standard FL algorithms on all datasets. Our flexible and modular suite allows researchers to easily download datasets, reproduce results and re-use the different components for their research. FLamby is available at~\url{www.github.com/owkin/flamby}., Accepted to NeurIPS, Datasets and Benchmarks Track, this version fixes typos in the datasets' table and the appendix

We introduce the information geometry module of the Python package Geomstats. The module first implements Fisher-Rao Riemannian manifolds of widely used parametric families of probability distributions, such as normal, gamma, beta, Dirichlet distributions, and more. The module further gives the Fisher-Rao Riemannian geometry of any parametric family of distributions of interest, given a parameterized probability density function as input. The implemented Riemannian geometry tools allow users to compare, average, interpolate between distributions inside a given family. Importantly, such capabilities open the door to statistics and machine learning on probability distributions. We present the object-oriented implementation of the module along with illustrative examples and show how it can be used to perform learning on manifolds of parametric probability distributions.

In this paper, we investigate the impact of neural networks (NNs) topology on adversarial robustness. Specifically, we study the graph produced when an input traverses all the layers of a NN, and show that such graphs are different for clean and adversarial inputs. We find that graphs from clean inputs are more centralized around highway edges, whereas those from adversaries are more diffuse, leveraging under-optimized edges. Through experiments on a variety of datasets and architectures, we show that these under-optimized edges are a source of adversarial vulnerability and that they can be used to detect adversarial inputs.

We propose a symbolic representation for support vector machines (SVMs) by means of abstract interpretation, a well-known and successful technique for designing and implementing static program analyses. We leverage this abstraction in two ways: (1) to enhance the interpretability of SVMs by deriving a novel feature importance measure, called abstract feature importance (AFI), that does not depend in any way on a given dataset of the accuracy of the SVM and is very fast to compute, and (2) for verifying stability, notably individual fairness, of SVMs and producing concrete counterexamples when the verification fails. We implemented our approach and we empirically demonstrated its effectiveness on SVMs based on linear and non-linear (polynomial and radial basis function) kernels. Our experimental results show that, independently of the accuracy of the SVM, our AFI measure correlates much more strongly with the stability of the SVM to feature perturbations than feature importance measures widely available in machine learning software such as permutation feature importance. It thus gives better insight into the trustworthiness of SVMs.

Normalizing Flows (NF) are powerful likelihood-based generative models that are able to trade off between expressivity and tractability to model complex densities. A now well established research avenue leverages optimal transport (OT) and looks for Monge maps, i.e. models with minimal effort between the source and target distributions. This paper introduces a method based on Brenier's polar factorization theorem to transform any trained NF into a more OT-efficient version without changing the final density. We do so by learning a rearrangement of the source (Gaussian) distribution that minimizes the OT cost between the source and the final density. The Gaussian preserving transformation is implemented with the construction of high dimensional divergence free functions and the path leading to the estimated Monge map is further constrain to lie on a geodesic in the space of volume-preserving diffeomorphisms thanks to Euler's equations. The proposed method leads to smooth flows with reduced OT costs for several existing models without affecting the model performance.

International audience; We investigate the role of inhibition in structuring new neural assemblies following a learning phase in the brain. Although many studies have been devoted to this subject, few of them have considered the specific role of inhibitory neurons. Therefore, this work aims at studying the impact of inhibition in adaptive neural networks. In particular we focus on the role of synchronized dynamics for the creation and maintenance of structural modules in the brain circuits. Specifically, we considered an excitatory-inhibitory neural network made of θ-neurons with plastic Hebbian synapses. The learning process is mimicked by stimulating different areas of the network with temporal alternating external stimuli. This stimulation process leads to the emergence of modular structures. Then, the network dynamics is analysed after the learning phase revealing the maintenance of co-active neural assemblies over the long term. We show that the presence of inhibitory neurons controls the emergence of the modular structures and their maintenance in the resting phase following the learning period. Furthermore, we demonstrate that the number of inhibitory neurons in the network is directly related to the maximal number of neural assemblies that can be maintained. These results support the idea that inhibition has a direct impact on the memory capacity of a neural network.

International audience

International audience; Due to abundant signal and user variability among others, BCIs remain difficult to control. To increase performance, adaptive methods are a necessary means to deal with such a vast spectrum of variable data. Typically, adaptive methods deal with the signal or classification corrections (adaptive spatial filters [1], co-adaptive calibration [2], adaptive classifiers [3]). As such, they do not necessarily account for the implicit alterations they perform on the feedback (in real-time), and in turn, on the user, creating yet another potential source of unpredictable variability. Namely, certain user's personality traits and states have shown to correlate with BCI performance, while feedback can impact user states [4]. For instance, altered (biased) feedback was distorting the participants' perception over their performance, influencing their feeling of control, and online performance [5]. Thus, one can assume that through feedback we might implicitly guide the user towards a desired state beneficial for BCI performance. We propose a novel, simple probabilistic, data-driven dynamic model to provide such feedback that will maximize performance.

In recent years, Internet of Things (IoT) networks have become new favorite targets for attackers. Their fundamental characteristic such as their energy and calculation constraints are open to new attack vectors. In this thesis, we focus on the study of vulnerabilities present in wireless networks in order to create frameworks allowing to launch several types of attacks. We also show how these frameworks can also be used as a defense system.At the same time, the threat landscape is changing dramatically and new attacks, commonly referred to as smart attacks, are emerging through the use of new processes like machine learning. Attackers are now able to create more autonomous, robust and efficient attacks that manage to advance current detection and countermeasure systems. This is why studying the security of wireless networks in the face of these new types of attacks to better understand them has become an important issue in research. In this thesis, we evaluate several vulnerabilities present in wireless networks allowing to create new intelligent attacks. First, we are developing HARPAGON, a framework based on the Markov chains theory and exploiting the vulnerabilities generated by the duty cycle mechanism. The main advantage of HARPAGON is to predict the optimal moment to carry out its attack in order to reduce its probability of being detected. At the same time this framework also allows the attacker to conserve energy. Then we propose another framework called FOLPETTI allowing to create several types of attacks thwarting a well-known countermeasure in wireless networks: channel hopping. We show that with the help of FOLPETTI, an attacker is able to predict the future transmission channel in order to increase its impact. In order to evaluate their effectiveness, we have developed a new module on the NS-3 simulator to simulate jamming attacks. Then, after validating their components on the simulator, we assigned them through experimentation on a real testbed. These two solutions; which increase the performance of several attacks, do not require prior knowledge on the part of the attacker and can be implemented on inexpensive components.Finally, strongly inspired by the FOLPETTI and HARPAGON frameworks, we have implemented a new jamming attack, called ICARO, which targets illicit drones. In this case, we show how a jamming attack can be diverted as a defense method to counter drones flying over illicit areas. The main advantage of this contribution is that this new type of attack makes it possible to cut off the communication of an illicit drone with its controller without disrupting communications in the surrounding area., Ces dernières années, les réseaux de l'Internet des Objets (IoT) sont devenus les nouvelles cibles privilégiées des attaquants. Leur caractéristique fondamentale telle que leurs contraintes énergétiques et de calculs ont ouvert de nouveaux vecteurs d'attaques. Dans cette thèse, nous nous concentrons sur l'étude de vulnérabilités présents dans les réseaux sans fils afin de créer des frameworks permettant de lancer plusieurs types d'attaques. Nous montrons également comment ces frameworks peuvent aussi être détournés en système de défense.En parallèle, le paysage des menaces est en train de changer considérablement, et de nouvelles attaques, communément appelées attaques intelligentes, sont en train d'émerger grâce à l'utilisation de nouveaux procédés comme l'apprentissage automatique. Les attaquants sont maintenant capables de créer des attaques plus autonomes, robustes et efficaces qui arrivent à contourner les systèmes de détections et de contre-mesures actuels. C'est pourquoi, étudier la sécurité des réseaux sans fils en face de ces nouveaux types d'attaque pour mieux les comprendre est devenu un enjeu important dans la recherche. Dans cette thèse, nous évaluons plusieurs vulnérabilités présentes dans les réseaux sans fils permettant de créer de nouvelles attaques intelligentes. Dans un premier temps, nous développons HARPAGON, un framework basé sur la théorie des chaînes de Markov et exploitant les vulnérabilités générées par le mécanisme du cycle d'utilisation. Le principal avantages d'HARPAGON est de prédire le moment optimal pour effectuer son attaque afin de réduire sa probabilité d'être détecté. Dans un même temps ce framework permet également a l'attaquant de conserver son énergie. Puis nous proposons un autre framework nommé FOLPETTI permettant de créer plusieurs types d attaques déjouant une contre-mesure bien connu dans les réseaux sans fils: le saut de canal. Nous montrons qu'avec l'aide de FOLPETTI, un attaquant est capable de prédire le futur canal de transmission afin d'augmenter son impact. Afin d'évaluer leur efficacité, nous avons développé un nouveau module sur le simulateur NS-3 permettant de simuler les attaques de brouillages. Puis, après avoir validé leurs comportements sur le simulateur, nous les avons évaluées grâce à des expérimentations sur un réel banc d'essais. Ces deux solutions, qui permettent d'augmenter les performances de plusieurs attaques, ne requièrent pas de connaissance au préalable de la part de l'attaquant et peuvent être implémenté sur des composants bons marchés.Enfin, fortement inspirés des frameworks FOLPETTI et HARPAGON, nous avons implémenté une nouvelles attaques de brouillages, nomme ICARO, qui vise les drones illicites. Dans ce cas, nous montrons comment une attaque de brouillage peut être détournée en méthode de défense pour contrer des drones survolants des zones illicites. Le principal avantage de cette contribution est que ce nouveau type d'attaque permet de couper la communication d'un drone illicite avec son contrôleur sans sans perturber les communications aux alentours qui communiquent dans les mêmes fréquences.

International audience; Cell response heterogeneity upon treatment is a main obstacle in preclinical development of ef-ficacious cancer drugs, due to the emergence of drug-tolerant cells. We have previously developed asingle-cell pipeline called Fate-seq to profile drug-tolerant persisters, based on predictions of their drugresponse. To automatize and increase the prediction throughput, we present a mathematical frame-work, composed of machine-learning classification models augmented by an ODE model.First, an ODE model of the extrinsic apoptosis initiation by death ligands is calibrated on timetrajectories of hundreds of treated clonal HeLa cells. The resulting deterministic systems are thenanalysed, based on drug response, to highlight mechanistic features with predictive values for celldecision. In a second step, we combine the predictions of the ODE system with machine-learningclassification models to determine the drug response of each cell before it commits to an irreversibledecision that would alter their states before profiling.Here we show that the ODE model analysis could detect the time of cell decision shortly after treat-ment, thanks to the emergence of an additional regulation at the receptor level in drug-sensitive cells.Moreover, the parameters distribution of the deterministic system provided a biological threshold thatallows the prediction of cell response. Our mechanistic-informed approach, combining our ODE sys-tem with machine learning classifiers, outperformed classic machine learning approaches and enabledthe accurate cell response prediction of otherwise unpredictable cells (Meyer et al., Cell Systems 2020).

Spectral clustering is a popular approach for clustering undirected graphs, but its extension to directed graphs (digraphs) is much more challenging. A typical workaround is to naively symmetrize the adjacency matrix of the directed graph, which can however lead to discarding valuable information carried by edge directionality. In this paper, we present a generalized spectral clustering framework that can address both directed and undirected graphs. Our approach is based on the spectral relaxation of a new functional that we introduce as the generalized Dirichlet energy of a graph function, with respect to an arbitrary positive regularizing measure on the graph edges. We also propose a practical parametrization of the regularizing measure constructed from the iterated powers of the natural random walk on the graph. We present theoretical arguments to explain the efficiency of our framework in the challenging setting of unbalanced classes. Experiments using directed K-NN graphs constructed from real datasets show that our graph partitioning method performs consistently well in all cases, while outperforming existing approaches in most of them.

This article presents a biological neural network model driven by inhomogeneous Poisson processes accounting for the intrinsic randomness of synapses. The main novelty is the introduction of local interactions: each firing neuron triggers an instantaneous increase in electric potential to a fixed number of randomly chosen neurons. We prove that, as the number of neurons approaches infinity, the finite network converges to a nonlinear meanfield process characterised by a jump-type stochastic differential equation. We show that this process displays a phase transition: the activity of a typical neuron in the infinite network either rapidly dies out, or persists forever, depending on the global parameters describing the intensity of interconnection. This provides a way to understand the emergence of persistent activity triggered by weak input signals in large neural networks.

International audience; With the increase of component complexity, protection against single event effects becomes a critical point for the disponibility and reliability of space systems. In this paper, machine learning is investigated to improve the detection of radiation faults. An algorithm named DYD² that meets space application requirements is proposed. In addition, a study to improve the characterisation of single event effects through feature extraction is described. Finally, results of experimentation based on a heavy-ion campaign test are discussed.

Biclustering is an unsupervised machine learning technique that simultaneously clusters rows and columns in a data matrix. Biclustering has emerged as an important approach and plays an essential role in various applications such as bioinformatics, text mining, and pattern recognition. However, finding significant biclusters is an NP-hard problem that can be formulated as an optimization problem. Therefore, different metaheuristics have been applied to biclustering problems because of their exploratory capability of solving complex optimization problems in reasonable computation time. Although various surveys on biclustering have been proposed, there is a lack of a comprehensive survey on the biclustering problem using metaheuristics. This chapter will present a survey of metaheuristics approaches to address the biclustering problem. The review focuses on the underlying optimization methods and their main search components: representation, objective function, and variation operators. A specific discussion on single versus multi-objective approaches is presented. Finally, some emerging research directions are presented., Comment: 32 pages, 6 figures, 2 tables, chapter book

This document presents additional experimental results and details that could not be included in the original paper because of space limitations.

International audience; Reservoir Computing (RC) is a type of recurrent neural network (RNNs) where learning is restricted to the output weights. RCs are often considered as temporal Support Vector Machines (SVMs) for the way they project inputs onto dynamic non-linear high-dimensional representations. This paradigm, mainly represented by Echo State Networks (ESNs), has been successfully applied on a wide variety of tasks, from time series forecasting to sequence generation. They offer de facto a fast, simple yet efficient way to train RNNs. We present in this paper a library that facilitates the creation of RC architectures, from simplest to most complex, based on the Python scientific stack (NumPy, Scipy). This library offers memory and time efficient implementations for both online and offline training paradigms, such as FORCE learning or parallel ridge regression. The flexibility of the API allows to quickly design ESNs including re-usable and customizable components. It enables to build models such as DeepESNs as well as other advanced architectures with complex connectivity between multiple reservoirs with feedback loops. Extensive documentation and tutorials both for newcomers and experts are provided through GitHub and ReadTheDocs websites. The paper introduces the main concepts supporting the library, illustrated with code examples covering popular RC techniques from the literature. We argue that such flexible dedicated library will ease the creation of more advanced architectures while guarantying their correct implementation and reproducibility across the RC community.

Deep reinforcement learning uses simulators as abstract oracles to interact with the environment. In continuous domains of multi body robotic systems, differentiable simulators have recently been proposed but are yet under utilized, even though we have the knowledge to make them produce richer information. This problem when juxtaposed with the usually high computational cost of exploration-exploitation in high dimensional state space can quickly render reinforcement learning algorithms impractical. In this paper, we propose to combine learning and simulators such that the quality of both increases, while the need to exhaustively search the state space decreases. We propose to learn value function and state, control trajectories through the locally optimal runs of model based trajectory optimizer. The learned value function, along with an estimate of optimal state and control policies, is subsequently used in the trajectory optimizer: the value function estimate serves as a proxy for shortening the preview horizon, while the state and control approximations serve as a guide in policy search for our trajectory optimizer. The proposed approach demonstrates a better symbiotic relation, with super linear convergence, between learning and simulators, that we need for end-to-end learning of complex poly articulated systems.

Simulating trajectories of virtual crowds is a commonly encountered task in Computer Graphics. Several recent works have applied Reinforcement Learning methods to animate virtual agents, however they often make different design choices when it comes to the fundamental simulation setup. Each of these choices comes with a reasonable justification for its use, so it is not obvious what is their real impact, and how they affect the results. In this work, we analyze some of these arbitrary choices in terms of their impact on the learning performance, as well as the quality of the resulting simulation measured in terms of the energy efficiency. We perform a theoretical analysis of the properties of the reward function design, and empirically evaluate the impact of using certain observation and action spaces on a variety of scenarios, with the reward function and energy usage as metrics. We show that directly using the neighboring agents' information as observation generally outperforms the more widely used raycasting. Similarly, using nonholonomic controls with egocentric observations tends to produce more efficient behaviors than holonomic controls with absolute observations. Each of these choices has a significant, and potentially nontrivial impact on the results, and so researchers should be mindful about choosing and reporting them in their work., Comment: Accepted for publication at MIG 2022

International audience; Disease progression models are crucial to understanding degenerative diseases. Mixed-effects models have been consistently used to model clinical assessments or biomarkers extracted from medical images, allowing missing data imputation and prediction at any timepoint. However, such progression models have seldom been used for entire medical images. In this work, a Variational Auto Encoder is coupled with a temporal linear mixed-effect model to learn a latent representation of the data such that individual trajectories follow straight lines over time and are characterised by a few interpretable parameters. A Monte Carlo estimator is devised to iteratively optimize the networks and the statistical model. We apply this method on a synthetic data set to illustrate the disentanglement between time dependant changes and inter-subjects variability, as well as the predictive capabilities of the method. We then apply it to 3D MRI and FDG-PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) to recover well documented patterns of structural and metabolic alterations of the brain.

Given the ever increasing cost of designing de novo molecules to target causes of diseases, and the huge amount of currently available biological data, the development of systematic explorative pipelines for drug development has become of paramount importance. In my thesis, I focused on drug repurposing, which is a paradigm that aims at identifying new therapeutic indications for known chemical compounds. Due to the already large collection of transcriptomic data -that is, related to protein production through the transcription of gene DNA sequences- which is publicly available, I investigated how to process in a transparent and controllable way this information about gene activity to screen molecules. The current state of research in drug development indicates that such generic approaches might considerably fasten the discovery of promising therapies, especially for neglected or rare diseases research. First, noting that transcriptomic measurements are the product of a complex dynamical system of co- and inter-gene activity regulations, I worked on integrating in an automated fashion diverse types of biological information in order to build a model of these regulations. That is where gene regulatory networks, and more specifically, Boolean networks, intervene. Such models are useful for both explaining observed transcription levels, and for predicting the result of gene activity perturbations through molecules. Second, these models allow online in silico drug testing. While using the predictive features of Boolean networks can be costly, the core assumption of this thesis is that, combining them with sequential learning algorithms, such as multi-armed bandits, might mitigate that effect, and help control the error rate in recommended therapeutic candidates. This is the drug testing procedure suggested throughout my PhD. The question of the proper integration of known side information about the chemical compounds into multi-armed bandits is crucial, and has also been investigated further. Finally, I applied part of my work to ranking different treatment protocols for neurorepair in the case of encephalopathy in premature infants. On the theoretical side, I also contributed to the design of an algorithm which is able to extend the drug testing procedure in a distributed way, for instance across several tested populations, disease models, or research teams., À cause du coût toujours croissant de la conception de molécules de novo ciblant les causes d'une maladie, et la masse considérable de données biologiques disponibles actuellement, la création de méthodes d'exploration systématique pour le développement de thérapies est devenue un enjeu crucial. Lors de ma thèse, je me suis concentrée sur le paradigme du repositionnement de médicaments, qui vise l'identification de nouvelles indications thérapeutiques pour des molécules chimiques connues. Étant donné la quantité déjà importante de données transcriptionnelles (relative à la production de protéines à travers la transcription des séquences ADN géniques) qui est accessible, j'ai cherché à déterminer comment traiter ces données sur l'activité génique de manière transparente et contrôlable pour le criblage de molécules. Une revue de l'état de la recherche en développement de médicaments montre que de telles approches génériques peuvent permettre de considérablement accélérer la découverte de thérapies prometteuses, plus particulièrement contre les maladies rares ou tropicales négligées. Premièrement, en remarquant que les mesures d'activité transcriptionnelle résultent d'un réseau dynamique complexe d'interactions coordonnées de régulation génique, j'ai travaillé sur l'intégration de façon automatique d'information biologique de formes variées afin de construire un modèle de ces régulations géniques. C'est là que les réseaux de régulation génique, et, plus spécifiquement, les réseaux booléens, interviennent. Ces modèles permettent à la fois d'expliquer les mesures d'origine transcriptionnelle observées, et de prédire le résultat de perturbations de l'activité de certains gènes par des molécules. Ensuite, ces modèles permettent d'effectuer des essais in silico de médicaments. Tandis que l'utilisation des prédictions faites par des réseaux booléens peut s'avérer coûteuse, l'hypothèse centrale de ma thèse est que leur combinaison avec des algorithmes d'apprentissage statistique séquentiels, comme les bandits à bras multiples, peuvent non seulement réduire ce coût, mais également contrôler le taux d'erreur dans les recommandations de candidats thérapeutiques. Cette démarche est la procédure d'essai clinique in silico analysée tout au long de mon travail de thèse. Le problème, capital, d'intégration des caractéristiques connues des composants chimiques dans les bandits à bras multiples a également été étudié plus en profondeur. Enfin, j'ai appliqué une partie de mon travail de thèse au classement de différents protocoles de traitement pour de la neuroréparation dans le cas d'encéphalopathies chez des enfants prématurés. D'un point de vue théorique, j'ai également contribué à la conception d'un algorithme qui permet d'étendre la procédure d'essai médicamenteux in silico à un cadre distribué à travers plusieurs populations test, modèles de maladies, ou équipes de recherche.

Fairness in machine learning has attained significant focus due to the widespread application in high-stake decision-making tasks. Unregulated machine learning classifiers can exhibit bias towards certain demographic groups in data, thus the quantification and mitigation of classifier bias is a central concern in fairness in machine learning. In this paper, we aim to quantify the influence of different features in a dataset on the bias of a classifier. To do this, we introduce the Fairness Influence Function (FIF). This function breaks down bias into its components among individual features and the intersection of multiple features. The key idea is to represent existing group fairness metrics as the difference of the scaled conditional variances in the classifier's prediction and apply a decomposition of variance according to global sensitivity analysis. To estimate FIFs, we instantiate an algorithm FairXplainer that applies variance decomposition of classifier's prediction following local regression. Experiments demonstrate that FairXplainer captures FIFs of individual feature and intersectional features, provides a better approximation of bias based on FIFs, demonstrates higher correlation of FIFs with fairness interventions, and detects changes in bias due to fairness affirmative/punitive actions in the classifier. The code is available at https://github.com/ReAILe/bias-explainer., Proceedings of FAccT, 2023

We consider the problem of minimizing the sum of two convex functions. One of those functions has Lipschitz-continuous gradients, and can be accessed via stochastic oracles, whereas the other is "simple". We provide a Bregman-type algorithm with accelerated convergence in function values to a ball containing the minimum. The radius of this ball depends on problem-dependent constants, including the variance of the stochastic oracle. We further show that this algorithmic setup naturally leads to a variant of Frank-Wolfe achieving acceleration under parallelization. More precisely, when minimizing a smooth convex function on a bounded domain, we show that one can achieve an $\epsilon$ primal-dual gap (in expectation) in $\tilde{O}(1/ \sqrt{\epsilon})$ iterations, by only accessing gradients of the original function and a linear maximization oracle with $O(1/\sqrt{\epsilon})$ computing units in parallel. We illustrate this fast convergence on synthetic numerical experiments.

Diagnosing erythema migrans (EM) skin lesion, the most common early symptom of Lyme disease using deep learning techniques can be effective to prevent long-term complications. Existing works on deep learning based EM recognition only utilizes lesion image due to the lack of a dataset of Lyme disease related images with associated patient data. Physicians rely on patient information about the background of the skin lesion to confirm their diagnosis. In order to assist the deep learning model with a probability score calculated from patient data, this study elicited opinion from fifteen doctors. For the elicitation process, a questionnaire with questions and possible answers related to EM was prepared. Doctors provided relative weights to different answers to the questions. We converted doctors' evaluations to probability scores using Gaussian mixture based density estimation. For elicited probability model validation, we exploited formal concept analysis and decision tree. The elicited probability scores can be utilized to make image based deep learning Lyme disease pre-scanners robust.

International audience; ReservoirPy is a simple user-friendly library based on Python scientific modules. It provides a flexible interface to implement efficient Reservoir Computing (RC) [2] architectures with a particular focus on Echo State Networks (ESN) [1]. Advanced features of ReservoirPy allow to improve computation time efficiency on a simple laptop compared to basic Python implementation. Some of its features are: offline and online training, parallel implementation, sparse matrix computation, fast spectral initialization, advanced learning rules (e.g. Intrinsic Plasticity) etc. It also makes possible to easily create complex architectures with multiple reservoirs (e.g. deep reservoirs), readouts, and complex feedback loops. Moreover, graphical tools are included to easily explore hyperparameters with the help of the hyperopt library. It includes several tutorials exploring exotic architectures and examples of scientific papers reproduction.