Search

In increasingly different contexts, it happens that a human player has to interact with artificial players who make decisions following decision-making algorithms. How should the human player play against these algorithms to maximize his utility? Does anything change if he faces one or more artificial players? The main goal of the paper is to answer these two questions. Consider n-player games in normal form repeated over time, where we call the human player optimizer, and the (n -- 1) artificial players, learners. We assume that learners play no-regret algorithms, a class of algorithms widely used in online learning and decision-making. In these games, we consider the concept of Stackelberg equilibrium. In a recent paper, Deng, Schneider, and Sivan have shown that in a 2-player game the optimizer can always guarantee an expected cumulative utility of at least the Stackelberg value per round. In our first result, we show, with counterexamples, that this result is no longer true if the optimizer has to face more than one player. Therefore, we generalize the definition of Stackelberg equilibrium introducing the concept of correlated Stackelberg equilibrium. Finally, in the main result, we prove that the optimizer can guarantee at least the correlated Stackelberg value per round. Moreover, using a version of the strong law of large numbers, we show that our result is also true almost surely for the optimizer utility instead of the optimizer's expected utility.

Item does not contain fulltext Intellectual disability with or without manifestations of autism and/or epilepsy affects 1-2% of the population, and it is estimated that more than 30-50% of these cases have a single genetic cause. More than 1000 genes and recurrent chromosomal abnormalities are involved in these genetic forms of neurodevelopmental disorders, which often remain insufficiently described in terms of clinical spectrum, associated medical problems, etc., due to their rarity and the often-limited number of patients' phenotypes reported. GenIDA is an international online participatory database that aims to better characterise the clinical manifestations and natural histories of these rare diseases. Clinical information is reported by parents of affected individuals using a structured questionnaire exploring physical parameters, cognitive and behavioural aspects, the presence or absence of neurological disorders or problems affecting major physiological functions, as well as autonomy and quality of life. This strengthens the implication in research of the concerned families. GenIDA aims to construct international cohorts of significant size of individuals affected by a given condition. As of July 2022, GenIDA counts some 1545 documented patient records from over 60 nationalities and collaborates with clinicians and researchers around the world who have access to the anonymized data collected to generate new, medically meaningful information to improve patient care. We present the GenIDA database here, together with an overview of the possibilities it offers to affected individuals, their families, and professionals in charge of the management of genetic forms of neurodevelopmental disorders. Finally, case studies of cohorts will illustrate the usefulness of GenIDA.

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

Internet traffic recognition is an essential tool for access providers since recognizing traffic categories related to different data packets transmitted on a network help them define adapted priorities. That means, for instance, high priority requirements for an audio conference and low ones for a file transfer, to enhance user experience. As internet traffic becomes increasingly encrypted, the mainstream classic traffic recognition technique, payload inspection, is rendered ineffective. This paper uses machine learning techniques for encrypted traffic classification, looking only at packet size and time of arrival. Spiking neural networks (SNN), largely inspired by how biological neurons operate, were used for two reasons. Firstly, they are able to recognize time-related data packet features. Secondly, they can be implemented efficiently on neuromorphic hardware with a low energy footprint. Here we used a very simple feedforward SNN, with only one fully-connected hidden layer, and trained in a supervised manner using the newly introduced method known as Surrogate Gradient Learning. Surprisingly, such a simple SNN reached an accuracy of 95.9% on ISCX datasets, outperforming previous approaches. Besides better accuracy, there is also a very significant improvement on simplicity: input size, number of neurons, trainable parameters are all reduced by one to four orders of magnitude. Next, we analyzed the reasons for this good accuracy. It turns out that, beyond spatial (i.e. packet size) features, the SNN also exploits temporal ones, mostly the nearly synchronous (within a 200ms range) arrival times of packets with certain sizes. Taken together, these results show that SNNs are an excellent fit for encrypted internet traffic classification: they can be more accurate than conventional artificial neural networks (ANN), and they could be implemented efficiently on low power embedded systems., Comment: 22 pages, 8 figures. Neurocomputing (2022)

Random forests have long been considered as powerful model ensembles in machine learning. By training multiple decision trees, whose diversity is fostered through data and feature subsampling, the resulting random forest can lead to more stable and reliable predictions than a single decision tree. This however comes at the cost of decreased interpretability: while decision trees are often easily interpretable, the predictions made by random forests are much more difficult to understand, as they involve a majority vote over hundreds of decision trees. In this paper, we examine different types of reasons that explain "why" an input instance is classified as positive or negative by a Boolean random forest. Notably, as an alternative to sufficient reasons taking the form of prime implicants of the random forest, we introduce majoritary reasons which are prime implicants of a strict majority of decision trees. For these different abductive explanations, the tractability of the generation problem (finding one reason) and the minimization problem (finding one shortest reason) are investigated. Experiments conducted on various datasets reveal the existence of a trade-off between runtime complexity and sparsity. Sufficient reasons - for which the identification problem is DP-complete - are slightly larger than majoritary reasons that can be generated using a simple linear- time greedy algorithm, and significantly larger than minimal majoritary reasons that can be approached using an anytime P ARTIAL M AX SAT algorithm., 21 pages

International audience

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/

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

In this article, we propose an approach based on Gaussian Processes (GP) for large scale land cover pixel-basedclassification with Sentinel-2 satellite image time-series (SITS). We used a sparse approximation of the posterior combined with variational inference to learn the GP’s parameters. We applied stochastic gradient descent and GPU computing to optimize our GP models on massive data sets. The proposed GP model can be trained with hundreds of thousands of samples, compared to few thousands for traditional GP methods. Moreover, we included the spatial information by adding the geographic coordinates into the GP’s covariance function to efficiently exploit the spatio-spectro-temporal structure of the SITS. We ran experiments with Sentinel-2 SITS of the full year 2018 over an area of 200 000 km 2 (about 2 billion pixels) in the south of France, which is representative of an operational setting. Adding the spatial information significantly improved the results in terms of classification accuracy. With spatial information, GP models have an overall accuracy of 79.8. They are more than three points above Random Forest (the method used for current operational systems) and more than one point above a multi-layer perceptron. Compared to a Transformer-based model (which provides state ofthe art results in the literature, but are not applied in operational systems), GP models are only one point below.

International audience; Unsupervised anomaly detection using deep learning models is a popular computer-aided diagnosis approach because it does not need annotated data and is not restricted to the diagnosis of a disease seen during training. Such approach consists in first learning the distribution of anomaly free images. Images presenting anomalies are then detected as outliers of this distribution. These approaches have been widely applied in neuroimaging to detect sharp and localized anomalies such as tumors or white matter hyper-intensities from structural MRI. In this work, we aim to detect anomalies from FDG PET images of patients with Alzheimer's disease. In this context, the anomalies can be subtle and difficult to delineate, making the task more difficult and meaning that no ground truth exists to evaluate the approaches. We thus propose a framework to evaluate unsupervised anomaly detection approaches that consists in simulating realistic anomalies from images of healthy subjects. We demonstrate the use of this framework by evaluating an approach based on a 3D variational autoencoder.

Recent large and rapid growth in the healthcare sector has contributed to an increase in the elderly population and an increase in life expectancy. One of the important study topics in this field is the automatic fall detection system. Camera-video has been extensively employed recently for applications in surveillance, the home, and healthcare. Therefore a smart fall detection system is focusing on image and video analysis techniques. For that, our scientific work studied an actual vision-based fall detection system. It produces satisfactory outcomes, but there is still room for improvement. The system has a very high recall rate and can detect all falls, but it lacks precision and frequently reports false positives (more than 99 per-cent). In fact, due to the optimum camera quality, several ordinary activities with specific movements, such as wheelchair mobility, or the light changing in an empty room, can be mistaken for falls. To address this problem and increase precision, we propose a post-process approach, hybridizing a CNN model and a Haar Cascade Classifier to determine whether to confirm or reject an alert that has been identified as a fall. The system’s effectiveness will increase while the false positives are decreased.

Nous présentons une méthode d’apprentissage profond pour la classification supervisée d’images, intégrant des connaissances sous forme de graphe. A ces fins, nous introduisons une fonction de coût combinant à la fois une me- sure traditionnellement utilisée en apprentissage profond (entropie croisée) et une mesure originale qui prend en compte la représentation des nœuds après plongement du graphe de connaissances. Les connaissances ne sont utilisées que pendant la phase d’apprentissage et ne sont pas nécessaires pour l’évaluation d’un exemple en mode test. Les expérimentations sur plusieurs bases d’images démontrent l’amélioration en performances de notre méthode par rapport à l’état de l’art : d’une part en comparaison avec des algorithmes classiques d’apprentissage profond et d’autre part avec un algorithme très récent basé aussi sur la connaissance issue d’un graphe.

The U-Net model, introduced in 2015, is established as the state-of-the-art architecture for medical image segmentation, along with its variants UNet++, nnU-Net, V-Net, etc. Vision transformers made a breakthrough in the computer vision world in 2021. Since then, many transformer based architectures or hybrid architectures (combining convolutional blocks and transformer blocks) have been proposed for image segmentation, that are challenging the predominance of U-Net. In this paper, we ask the question whether transformers could overtake U-Net for medical image segmentation. We compare SegFormer, one of the most popular transformer architectures for segmentation, to U-Net using three publicly available medical image datasets that include various modalities and organs: segmentation of cardiac structures in ultrasound images from the CAMUS challenge, segmentation of polyp in endoscopy images and segmentation of instrument in colonoscopy images from the MedAI challenge. We compare them in the light of various metrics (segmentation performance, training time) and show that SegFormer can be a true competitor to U-Net and should be carefully considered for future tasks in medical image segmentation.

We introduce FENNIX (Force-Field-Enhanced Neural Network InteraXions), a hybrid approach between machine-learning and force-fields. We leverage state-of-the-art equivariant neural networks to predict local energy contributions and multiple atom-in-molecule properties that are then used as geometry-dependent parameters for physically-motivated energy terms which account for long-range electrostatics and dispersion. Using high-accuracy ab initio data (small organic molecules/dimers), we trained a first version of the model. Exhibiting accurate gas-phase energy predictions, FENNIX is transferable to the condensed phase. It is able to produce stable Molecular Dynamics simulations, including nuclear quantum effects, for water predicting accurate liquid properties. The extrapolating power of the hybrid physically-driven machine learning FENNIX approach is exemplified by computing: i) the solvated alanine dipeptide free energy landscape; ii) the reactive dissociation of small molecules.

In Constraint Programming, constraints are usually represented as predicates allowing or forbidding combinations of values. However, some algorithms can exploit a finer representation: error functions. By associating a function to each constraint type to evaluate the quality of an assignment, it extends the expressiveness of regular Constraint Satisfaction Problem/Constrained Optimization Problem formalisms. Their usage comes with a price though: it makes problem modeling significantly harder, since users must provide a set of error functions that are not always easy to define. Here, we propose a method to automatically learn an error function corresponding to a constraint, given its predicate version only. This is, to the best of our knowledge, the first attempt to automatically learn error functions for hard constraints. In this paper, we also give for the first time a formal definition of combinatorial problems with hard constraints represented by error functions. Our method aims to learn error functions in a supervised fashion, trying to reproduce either the Hamming or the Manhattan distance, by using a graph model we named Interpretable Compositional Networks. This model allows us to get interpretable results. We run experiments on 7 different constraints to show its versatility. Experiments show that our system can learn functions that scale to high dimensions, and can learn fairly good functions over incomplete spaces. We also show that learned error functions can be used efficiently to represent constraints in different classic problems.

While the ever-increasing emergence of online services has led to a growing interest in the development of recommender systems, the algorithms underpinning such systems have begun to be criticized for their role in limiting the variety of content exposed to users. In this context, the notion of diversity has been proposed as a way of mitigating the side effects resulting from the specialization of recommender systems. In this paper, using a well-known recommender system that makes use of collaborative filtering in the context of musical content, we analyze the diversity of recommendations generated through the lens of the recently proposed information network diversity measure. The results of our study offer significant insights into the effect of algorithmic recommendations. On the one hand, we show that the musical selections of a large proportion of users are diversified as a result of the recommendations. On the other hand, however, such improvements do not benefit all users. They are in fact mainly restricted to users with a low level of activity or whose past musical listening selections are very narrow. Through more in-depth investigations, we also discovered that while recommendations generally increase the variety of the songs recommended to users, they nonetheless fail to provide a balanced exposure to the different related categories.

In addition to being intrinsically interesting, calculus of variations constitutes an area that has been extensively considered in both theoretic and applied areas, underlying a large number of structures and phenomena, including the stationary action principle in physics. Basically, given a functional expressing some measurement to be optimized (minimized or maximized), as well as respective boundary conditions, calculus of variations allows the determination of the respective optimal solution, be it a function, a curve, surface, as well as a large range of mathematical structures. The present work aims at developing an introduction to variational calculus which could be hopefully accessible to readers with some background in multivariate calculus. After defining the problem, and presenting some brief historic remarks, we proceed to present a motivational example of calculus of variations respective to a specific type of variation function, namely second-order polynomials adhering to the boundary conditions. In addition to illustrating several concepts related to calculus of variations, this approach also allowed the visualization of the aimed functional behavior as a function of the variation parameter. Next, the Euler-Lagrange equation for functions of one variable was informally derived, and then illustrates its application respectively to the minimum arc-length and soap surface minimization. The interesting possibility of numerical approaches to calculus of variations is also briefly described and illustrated respectively to Euler's direct method. Appendices providing a brief presentation of the chain rule, integration by parts, multivariate derivatives, as well as the multivariate, multiconstraint Lagrange multipliers method are also provided for quick reference.

Most cerebrovascular diseases (including strokes and aneurysms) are treated endovascularly with cathetersthat are navigated from the groin through the vessels to the brain. Many patients have complex anatomy of the aortic arch and supra-aortic vessels, which can make it difficult to select the best catheters for navigation, resulting in longer procedures and more complications or failures. To this end, we propose a framework dedicated to the analysis of the aortic arch and supra-aortic trunks. This framework can automatically compute anatomical and geometrical features from meshes segmented beforehand via CNN-based pipeline. These features such as arch type, tortuosity and angulations describe the navigational difficulties encountered during catheterization. Quantitative and qualitative validation was performed by experienced neuroradiologists, leading to reliable vessel characterization.

International audience; Despite great achievements of reinforcement learning based works, those methods are known for their poor sample eciency. This particular drawback usually means training agents in simulated environments is the only viable option, regarding time constraints. Furthermore, reinforcement learning agents have a strong tendency to overt on their environment, observing a drastic loss of performances at test time. As a result, tying the agent logic to its current body may very well make transfer unefcient. To tackle that issue, we propose the Universal Notice Network (UNN) method to enforce separation of the neural network layers holding information to solve the task from those related to robot properties, hence enabling easier transfer of knowledge between entities. We demonstrate the eciency of this method on a broad panel of applications, we consider dierent kinds of robots, with dierent morphological structures performing kinematic, dynamic single and multi-robot tasks. We prove that our method produces zero shot (without additionnal learning) transfers that may produce better performances than state-of-the art approaches and show that a fast tuning enhances those performances.

La originalidad es una de las exigencias que debe cumplir una creación para acceder a la protección del derecho de autor. A pesar de que no existe una definición jurídica en el derecho europeo o nacional de lo que es una obra original, los comentadores consideran que la originalidad se encuentra en la “expresión de la personalidad del autor”, es decir, en las intenciones y el estilo individual que expresa un determinado creador a través de la forma de la obra.En el marco de las creaciones realizadas con inteligencia artificial, el cumplimiento de esta exigencia ha sido objeto de discusión, debido a dos grandes problemas: el primero es que las creaciones realizadas con inteligencia artificial reproducen los datos de entrenamiento o de entrada utilizados por el desarrollador para la configuración del sistema. Este fenómeno, que denominamos “la huella de los datos”, puede configurar un problema para la caracterización de la originalidad, en la medida en que la reproducción de varios elementos de una obra protegida utilizada como dato de entrada puede poner en duda la protección del resultado del proceso de automatización. En segundo lugar, varios autores consideran que el hecho de que un sistema de inteligencia artificial ejecute automáticamente los actos de creación, se traduce en que el autor (persona natural) no expresa su personalidad.No obstante, veremos que estos dos obstáculos no impiden que una obra automatizada pueda ser original. Veremos que una obra que reproduce elementos preexistentes en otra creación puede ser original. También, que ciertos elementos como el estilo de un creador no son protegidos por el derecho de autor, y por ende la imitación de ciertos estilos no debería ser un gran obstáculo para acceder al derecho de autor. Finalmente, veremos que a pesar de que el sistema de inteligencia artificial automatiza el proceso de creación de una obra, esto no es impedimento para que ésta sea el resultado de la personalidad del autor.

Deploying Connected and Automated Vehicles (CAVs) on top of 5G and Beyond networks (5GB) makes them vulnerable to increasing vectors of security and privacy attacks. In this context, a wide range of advanced machine/deep learning-based solutions have been designed to accurately detect security attacks. Specifically, supervised learning techniques have been widely applied to train attack detection models. However, the main limitation of such solutions is their inability to detect attacks different from those seen during the training phase, or new attacks, also called zero-day attacks. Moreover, training the detection model requires significant data collection and labeling, which increases the communication overhead, and raises privacy concerns. To address the aforementioned limits, we propose in this paper a novel detection mechanism that leverages the ability of the deep auto-encoder method to detect attacks relying only on the benign network traffic pattern. Using federated learning, the proposed intrusion detection system can be trained with large and diverse benign network traffic, while preserving the CAVs’ privacy, and minimizing the communication overhead. The in-depth experiment on a recent network traffic dataset shows that the proposed system achieved a high detection rate while minimizing the false positive rate, and the detection delay., Le déploiement de véhicules connectés et automatisés (CAV) utilisant les réseaux 5G et 5G-B les rend vulnérables à de nouveaux vecteurs d'attaques. Afin de détecter et de mitiger ces attaques, plusieurs systèmes de détection d'intrusions (IDS) basés sur l'apprentissage automatique ont été proposés, dont la grande majorité utilise l'apprentissage profond supervisé. Cependant, la principale limite de ce type de solution est son incapacité à détecter des attaques différentes de celles vues lors de l'apprentissage de l'IDS, ou de nouvelles attaques, également appelées attaques Zero-day. De plus, l'apprentissage du modèle de détection nécessite une collecte et un étiquetage importants des données, ce qui augmente le coût de communication et soulève des problèmes de confidentialité. Afin de pallier ces problèmes, nous proposons dans cet article un nouveau IDS basé sur l'apprentissage fédéré non supervisé préservant la confidentialité, et dont l'apprentissage du système ne nécessite que le trafic réseau légitime. L'expérimentation de la solution proposée sur un jeu de données récent montre que le système proposé présente un taux de détection élevé tout en minimisant le taux de faux positifs et le délai de détection.

In the context of mobile robotics, this work reports research addressing scene understanding using embedded multi-modal perception systems. Scene understanding remains a complex problem, essential for high society impact applications such as Intelligent Transportation Systems, involving in particular ADAS and Autonomous Vehicles. Results and outlook based on the conducted research are detailed and structured following two research perspectives., Dans le domaine de la robotique mobile, mes travaux de recherche portent sur l’analyse de scènes dynamiques par des systèmes de perception multimodale embarqués. Ceci demeure un verrou scientifique complexe, mais essentiel pour des applications avec un fort impact sociétal telles que les systèmes de transport intelligents où se situent les systèmes d’assistance à la conduite et les voitures à conduite automatisée. Une synthèse et des perspectives structurées sur deux axes thématiques sont présentés.

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.

Machine learning algorithms can produce biased outcome/prediction, typically, against minorities and under-represented sub-populations. Therefore, fairness is emerging as an important requirement for the large scale application of machine learning based technologies. The most commonly used fairness notions (e.g. statistical parity, equalized odds, predictive parity, etc.) are observational and rely on mere correlation between variables. These notions fail to identify bias in case of statistical anomalies such as Simpson's or Berkson's paradoxes. Causality-based fairness notions (e.g. counterfactual fairness, no-proxy discrimination, etc.) are immune to such anomalies and hence more reliable to assess fairness. The problem of causality-based fairness notions, however, is that they are defined in terms of quantities (e.g. causal, counterfactual, and path-specific effects) that are not always measurable. This is known as the identifiability problem and is the topic of a large body of work in the causal inference literature. This paper is a compilation of the major identifiability results which are of particular relevance for machine learning fairness. The results are illustrated using a large number of examples and causal graphs. The paper would be of particular interest to fairness researchers, practitioners, and policy makers who are considering the use of causality-based fairness notions as it summarizes and illustrates the major identifiability results, Comment: arXiv admin note: text overlap with arXiv:2010.09553

Setting up IoT monitoring solutions requires a significant workforce for deployment and maintenance. This becomes particularly costly when a massive number of sensors are deployed. In this paper, we propose a generic solution based on energyautonomous LPWAN sensors deployed on the fly. On one side, we propose an efficient management strategy that adapts the transmission period of sensors in order to receive a chosen amount of information on each sensor cluster that emits similar information. Moreover, we develop a method that generates realtime estimates based on the sensor message history. After showing the relevance of our estimation method, we show that our sensor management method is efficient by comparing it to a baseline. Additionally, our tools are robust to the hazards occurring in the IoT: sensor arrivals and departures, noisy measuring in a highly variable environment, misplacement of sensors. Overall, this paper proposes a way to scale up IoT, breaking the deployment cost barrier and paving the way for universal and versatile monitoring solutions.

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 report lifts the curtain on the Brazilian data workers involved in global AI production chains. They do 'micro-tasks' to generate and annotate data for machine learning, while also checking algorithmic outputs and occasionally, taking the place of failing automation. They use international digital labor platforms through which they execute on-demand tasks, mostly for AI producers located in the Global North. As platform workers, they are paid by piecework and have no long-term commitment to the buyers of their outputs. Some highlights:- Three out of five Brazilian data workers are women, while in most other previously-surveyed countries, women are a minority (one in three or less). - 9 reais (1.73 euros) per hour is the average amount earned on platforms.- There are at least 54 micro-working platforms operating in Brazil.- One third of Brazilian data workers have no other source of income, and depend on microworking platforms for subsistence.- Two out of five Brazilian data workers are (apart from this activity) unemployed, without professional activity, or in informality. In Brazil, platform microwork arises out of widespread unemployment and informalization of work.- Three out of five of data workers have completed undergraduate education, although they mostly do repetitive and unchallenging online data tasks, suggesting some form of skill mismatch.- The worst microtasks involve moderation of violent and pornographic contents on social media, as well as data training in tasks that workers may find uncomfortable or weird, such as taking pictures of dog poop in domestic environments to train data for "vacuuming robots".- Workers' main grievances are linked to uncertainty, lack of transparency, job insecurity, fatigue and lack of social interaction on platforms.

Contrary to artificial systems, biological systems are able to perform navigation tasks over large distances and long duration. This ability is partly based on the robustness of the learning mechanisms involved in their cognition, on their ability to exploit couplings between different sensors and to adapt themselves continuously to changes in the environment. The modeling of their cognition could thus be an important vector of innovation in robotics, bringing to existing models new efficient and original mechanisms. In order to propose new robotic architectures, this thesis explores the development of models inspired by mammalian cognition for the autonomous vehicle. This work starts from the "PerAc navigation model", a neurocybernetic model designed to reproduce the observations made in neurobiology in small environments. The model has demonstrated good navigation capabilities on small mobile robots, notably linked to its specific way of representing the environment. However, it is not immediately suitable for large-scale operation because of its high computational and memory cost. In particular, recent work has shown the need to improve the performance of its system of visual localisation, the main bottleneck of the model. In order to overcome this problem, the first part of the thesis was dedicated to the integration of population coding mechanisms in the visual localisation system of the model. Thus, several models of visual information encoding have been proposed, based on a new method of sparse coding inspired by the functioning of the visual cortex. In particular, an effort was made to adopt a conventional method of model evaluation, in order to obtain a strong measure of their performance comparable with the state of the art. Finally, the results obtained demonstrated the efficiency of the proposed methods, with significant gains in computational cost and memory over the original model. The second part of the thesis was devoted to the development of a complete neurorobotic navigation model for autonomous vehicles. The original model has therefore been taken over in its entirety and significantly modified to suit the operation of an autonomous vehicle. This work led to the construction of the MpNav model, a new end-to-end model of visual navigation, ables to learn a trajectory in a few iterations. Thus, its use on an experimental vehicle of VEDECOM has permitted to perform several navigation trajectories in complete autonomy, with a record of 2km on a looping trajectory., Contrairement aux systèmes artificiels, les systèmes biologiques sont capables d’effectuer des tâches de navigation sur de grandes distances et de longues durées. Cette capacité reposerait en partie sur la robustesse des mécanismes d’apprentissage impliqués dans leur cognition, sur leur faculté à exploiter les couplages entre différents capteurs et à s’adapter en permanence aux changements survenant dans l’environnement. La modélisation de leur cognition pourrait ainsi être un vecteur important d’innovation en robotique, apportant aux modèles existants de nouveaux mécanismes efficace et originaux. Dans le but de proposer de nouvelles architectures de robotique, cette thèse explore le développement de modèles inspirés de la cognition des mammifères pour le véhicule autonome. Ces travaux sont basés sur le modèle de navigation "PerAc", un modèle de neurocybernétique conçu pour reproduire les observations faites en neurobiologie chez le rongeur naviguant dans de petits environnements. Le modèle a ainsi démontré de bonnes capacités de navigation sur de petits robots mobiles, liées notamment à sa façon spécifique de représenter l’environnement. Il n’est cependant pas adapté pour fonctionner sur de grandes échelles de fonctionnement à cause de ses coûts computationnel et mémoire trop élevés. Des travaux ont démontré le besoin d’améliorer les performances de son système de localisation visuelle, principal goulot d’étranglement du modèle. En vue de surmonter ce problème, la première partie de la thèse a été consacrée à l’intégration de mécanismes de codage en population dans le système de localisation visuelle du modèle. Plusieurs modèles d’encodage des informations visuelles ont ainsi été proposés, s’appuyant sur une nouvelle méthode de codage parcimonieux inspirée du fonctionnement du cortex visuel. Un effort a notamment été réalisé pour adopter une méthode conventionnelle d’évaluation des modèles, afin d’obtenir une mesure forte de ses performances comparable avec l’état de l’art. Les résultats obtenus ont ainsi démontré l’efficacité des méthodes proposées, avec des gains importants en coûts computationnel et mémoire sur le modèle d’origine. La deuxième partie de la thèse, dans la continuité des travaux entrepris, a été consacrée au développement d’un modèle complet de navigation neurorobotique pour un véhicule autonome. Le modèle d’origine a ainsi été repris dans son intégralité et a été fortement modifié pour convenir aux particularités, en termes de contrôle, d’un véhicule autonome. Ces travaux ont abouti à la construction du modèle MpNav, un nouveau modèle de navigation visuelle, qui, bien que proche du type "end-to-end" n’est pas basé sur un apprentissage profond et permet d’apprendre une trajectoire de navigation en peu d’itérations. Son utilisation sur un véhicule expérimental de VEDECOM a ainsi permis d’effectuer en totale autonomie plusieurs trajectoires de navigation, avec un record de 2km sans intervention de l’opérateur

Background and objective: 3D rotational angiography (3DRA) provides high quality images of the cerebral arteriovenous malformation (AVM) nidus that can be reconstructed in 3D. However, these reconstructions are limited to only 3D visualization without possible interactive exploration of geometric characteristics of cerebral structures. Refined understanding of the AVM angioarchitecture prior to treatment is mandatory and vascular segmentation is an important preliminary step that allow physicians analyze the complex vascular networks and can help guide microcatheters navigation and embolization of AVM.Methods: A deep learning method was developed for the segmentation of 3DRA images of AVM patients. The method uses a fully convolutional neural network with a U-Net-like architecture and a DenseNet backbone. A compound loss function, combining Cross Entropy and Focal Tversky, is employed for robust segmentation. Binary masks automatically generated from region-growing segmentation have been used to train and validate our model.Results: The developed network was able to achieve the segmentation of the vessels and the malformationand significantly outperformed the region-growing algorithm. Our experiments were performed on 9 AVMpatients. The trained network achieved a Dice Similarity Coefficient (DSC) of 80.43%, surpassing other U-Net like architectures and the region-growing algorithm on the manually approved test set by physicians.Conclusions: This work demonstrates the potential of a learning-based segmentation method for characterizing very complex and tiny vascular structures even when the training phase is performed with the results of an automatic or a semi-automatic method. The proposed method can contribute to the planning and guidance of endovascular procedures.

Historiquement, l'intelligence artificielle (IA) s'est divisée en 2 courants selon les hypothèses faites pour modéliser l'intelligence humaine : l'IA symbolique, supposant que des symboles sont nécessaires, et l'IA statistique (plus particulièrement l'IA connexionniste) affirmant le contraire. Dernièrement, l'IA neuro-symbolique tente de réconcilier les 2. Les travaux présentés dans cet article sont en en collaboration avec la Police Scientifique française, dans le contexte du Plan National Stup. Nous présentons les problématiques métiers en lien avec le projet, puis en déduisons les problématiques scientifiques. Après un rappel des domaines de l'IA et leurs limites, nous présentons un état des lieux du domaine de l'IA neuro-symbolique. Nous discutons les défis applicatifs avant de finir par une présentation des premiers travaux qui ont été conduits. Des pistes d'approches basées sur l'état des lieux de l'IA neurosymbolique sont aussi présentées pour la suite des travaux.

Agent-based modelling has been used in many studies of urban expansion, land use and land cover change patterns. While representing a powerful tool for depicting and formulating predictions about the evolution of interconnected complex systems, this method also poses a series of challenges to the researcher community, most notably in terms of model calibration and validation, and output data visualisation. Based on these findings, we present an agent-based model developed in GAMA, coupled with a data exploration platform coded in python, for analysing commuting patterns in the canton of Geneva, Switzerland. Output datasets generated from a set of simple evolution rules for the agents, are distributed in open access together with the code for the associated data visualisation platform. This prototype is early work in developing a series of agent-based models for simulating urban expansion and land cover change dynamics, together with their own data exploration platforms for calibration, validation and output data analysis purposes.

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 .

International audience; Representation learning is a central problem of Attributed Networks data analysis in a variety of fields. Given an attributed graph, the objectives are to obtain a representation of nodes and a partition of the set of nodes. Usually these two objectives are pursued separately via two tasks that are performed sequentially, and any benefit that may be obtained by performing them simultaneously is lost. In this paper we propose a Power Attributed Graph Embedding and clustering (PAGEC for short) in which the two tasks, embedding and clustering, are considered together. To jointly encode data affinity between node links and attributes, we use a new powered proximity matrix. We formulate a new matrix decomposition model to obtain node representation and node clustering simultaneously. Theoretical analysis shows the close connections between the new proximity matrix and the random walk theory on a graph. Experimental results demonstrate that the PAGEC algorithm performs better, in terms of clustering and embedding, than stateof-the-art algorithms including deep learning methods designed for similar tasks in relation to attributed network datasets with different characteristics.

International audience; Image reconstruction from a sequence of a few linear measurements that are corrupted by signal-dependent normally distributed noise is an inverse problem with many biomedical imaging applications, such as computerized tomography and optical microscopy. In this study, we focus on single-pixel imaging, where the set-up acquires a down-sampled Hadamard transform of an image of the scene. Deep learning is a computationally efficient framework to solve inverse problems in imaging. Several neural-network architectures provide a link between deep and optimization-based image reconstruction methods. These deep-learning methods rely on a forward operator and lead to more interpretable networks. Here, we propose a novel network architecture obtained by unrolling an heuristic expectation-maximization algorithm. In particular, we compute the maximum a posteriori estimate of the unknown image given measurements corrupted by normally distributed signal-dependent noise. We show that the so-called expectation-maximization reconstruction network (EM-Net) applies to mixed Skellam-Gaussian noise models that are common in single-pixel imaging. We present reconstruction results from simulated and experimental single-pixel acquisitions. We show that EM-Net generalizes very well to noise levels not seen during training, despite having fewer learned parameters than alternative methods. The proposed EM-Net generally reconstructs images with fewer artifacts and higher signal-to-noise ratios, in particular in high-noise situations compared to other state of the art reconstruction algorithms that do not estimate the noise covariance.

Modern feedforward convolutional neural networks (CNNs) can now solve some computer vision tasks at super-human levels. However, these networks only roughly mimic human visual perception. One difference from human vision is that they do not appear to perceive illusory contours (e.g. Kanizsa squares) in the same way humans do. Physiological evidence from visual cortex suggests that the perception of illusory contours could involve feedback connections. Would recurrent feedback neural networks perceive illusory contours like humans? In this work we equip a deep feedforward convolutional network with brain-inspired recurrent dynamics. The network was first pretrained with an unsupervised reconstruction objective on a natural image dataset, to expose it to natural object contour statistics. Then, a classification decision layer was added and the model was finetuned on a form discrimination task: squares vs. randomly oriented inducer shapes (no illusory contour). Finally, the model was tested with the unfamiliar ''illusory contour'' configuration: inducer shapes oriented to form an illusory square. Compared with feedforward baselines, the iterative ''predictive coding'' feedback resulted in more illusory contours being classified as physical squares. The perception of the illusory contour was measurable in the luminance profile of the image reconstructions produced by the model, demonstrating that the model really ''sees'' the illusion. Ablation studies revealed that natural image pretraining and feedback error correction are both critical to the perception of the illusion. Finally we validated our conclusions in a deeper network (VGG): adding the same predictive coding feedback dynamics again leads to the perception of illusory contours., Manuscript under review

Purpose: DYRK1A syndrome is among the most frequent monogenic forms of intellectual disability (ID). We refined the molecular and clinical description of this disorder and developed tools to improve interpretation of missense variants, which remains a major challenge in human genetics. Methods: We reported clinical and molecular data for 50 individuals with ID harboring DYRK1A variants and developed (1) a specific DYRK1A clinical score; (2) amino acid conservation data generated from 100 DYRK1A sequences across different taxa; (3) in vitro overexpression assays to study level, cellular localization, and kinase activity of DYRK1A mutant proteins; and (4) a specific blood DNA methylation signature. Results: This integrative approach was successful to reclassify several variants as pathogenic. However, we questioned the involvement of some others, such as p.Thr588Asn, still reported as likely pathogenic, and showed it does not cause an obvious phenotype in mice. Conclusion: Our study demonstrated the need for caution when interpreting variants in DYRK1A, even those occurring de novo. The tools developed will be useful to interpret accurately the variants identified in the future in this gene. Graphic abstract: [Figure not available: see fulltext.]

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/

This paper focuses on multi-label learning from small number of labelled data. We demonstrate that the straightforward binary-relevance extension of the interpolated label propagation algorithm, the harmonic function, is a competitive learning method with respect to many widely-used evaluation measures. This is achieved mainly by a new transition matrix that better captures the underlying manifold structure. Furthermore, we show that when there exists label dependence, we can use the outputs of a competitive learning method as part of the input to the harmonic function to provide improved results over those of the original model. Finally, since we are using multiple metrics to thoroughly evaluate the performance of the algorithm, we propose to use the game-theory based method of Kalai and Smorodinsky to output a single compromise solution. This method can be applied to any learning model irrespective of the number of evaluation measures used.

Containing the medical data of millions of patients, clinical data warehouses (CDWs) represent a great opportunity to develop computational tools. Magnetic resonance images (MRIs) are particularly sensitive to patient movements during image acquisition, which will result in artefacts (blurring, ghosting and ringing) in the reconstructed image. As a result, a significant number of MRIs in CDWs are corrupted by these artefacts and may be unusable. Since their manual detection is impossible due to the large number of scans, it is necessary to develop tools to automatically exclude (or at least identify) images with motion in order to fully exploit CDWs. In this paper, we propose a novel transfer learning method for the automatic detection of motion in 3D T1-weighted brain MRI. The method consists of two steps: a pre-training on research data using synthetic motion, followed by a fine-tuning step to generalise our pre-trained model to clinical data, relying on the labelling of 4045 images. The objectives were both (1) to be able to exclude images with severe motion, (2) to detect mild motion artefacts. Our approach achieved excellent accuracy for the first objective with a balanced accuracy nearly similar to that of the annotators (balanced accuracy>80 %). However, for the second objective, the performance was weaker and substantially lower than that of human raters. Overall, our framework will be useful to take advantage of CDWs in medical imaging and highlight the importance of a clinical validation of models trained on research data.