Département d'Informatique

While logistic sigmoid neurons are more biologically plausible than hyperbolic tangent neurons, the latter work better for training multi-layer neural networks. This paper shows that rectifying neurons are an even better model of biological neurons and yield equal or better performance than hyperbolic tangent networks in spite of the hard non-linearity and non-differentiability at zero, creating sparse representations with true zeros, which seem remarkably suitable for naturally sparse data. Even though they can take advantage of semi-supervised setups with extra-unlabeled data, deep rectifier networks can reach their best performance without requiring any unsupervised pre-training on purely supervised tasks with large labeled datasets. Hence, these results can be seen as a new milestone in the attempts at understanding the difficulty in training deep but purely supervised neural networks, and closing the performance gap between neural networks learnt with and without unsupervised pre-training. 1

The interrogation of genetic markers in environmental meta-barcoding studies is currently seriously hindered by the lack of taxonomically curated reference data sets for the targeted genes. The Protist Ribosomal Reference database (PR(2), http://ssu-rrna.org/) provides a unique access to eukaryotic small sub-unit (SSU) ribosomal RNA and DNA sequences, with curated taxonomy. The database mainly consists of nuclear-encoded protistan sequences. However, metazoans, land plants, macrosporic fungi and eukaryotic organelles (mitochondrion, plastid and others) are also included because they are useful for the analysis of high-troughput sequencing data sets. Introns and putative chimeric sequences have been also carefully checked. Taxonomic assignation of sequences consists of eight unique taxonomic fields. In total, 136 866 sequences are nuclear encoded, 45 708 (36 501 mitochondrial and 9657 chloroplastic) are from organelles, the remaining being putative chimeric sequences. The website allows the users to download sequences from the entire and partial databases (including representative sequences after clustering at a given level of similarity). Different web tools also allow searches by sequence similarity. The presence of both rRNA and rDNA sequences, taking into account introns (crucial for eukaryotic sequences), a normalized eight terms ranked-taxonomy and updates of new GenBank releases were made possible by a long-term collaboration between experts in taxonomy and computer scientists.

Delineation of the left ventricular cavity, myocardium, and right ventricle from cardiac magnetic resonance images (multi-slice 2-D cine MRI) is a common clinical task to establish diagnosis. The automation of the corresponding tasks has thus been the subject of intense research over the past decades. In this paper, we introduce the "Automatic Cardiac Diagnosis Challenge" dataset (ACDC), the largest publicly available and fully annotated dataset for the purpose of cardiac MRI (CMR) assessment. The dataset contains data from 150 multi-equipments CMRI recordings with reference measurements and classification from two medical experts. The overarching objective of this paper is to measure how far state-of-the-art deep learning methods can go at assessing CMRI, i.e., segmenting the myocardium and the two ventricles as well as classifying pathologies. In the wake of the 2017 MICCAI-ACDC challenge, we report results from deep learning methods provided by nine research groups for the segmentation task and four groups for the classification task. Results show that the best methods faithfully reproduce the expert analysis, leading to a mean value of 0.97 correlation score for the automatic extraction of clinical indices and an accuracy of 0.96 for automatic diagnosis. These results clearly open the door to highly accurate and fully automatic analysis of cardiac CMRI. We also identify scenarios for which deep learning methods are still failing. Both the dataset and detailed results are publicly available online, while the platform will remain open for new submissions.

International audience

We tackle the problem of large scale visual place recognition, where the task is to quickly and accurately recognize the location of a given query photograph. We present the following three principal contributions. First, we develop a convolutional neural network (CNN) architecture that is trainable in an end-to-end manner directly for the place recognition task. The main component of this architecture, NetVLAD, is a new generalized VLAD layer, inspired by the "Vector of Locally Aggregated Descriptors" image representation commonly used in image retrieval. The layer is readily pluggable into any CNN architecture and amenable to training via backpropagation. Second, we develop a training procedure, based on a new weakly supervised ranking loss, to learn parameters of the architecture in an end-to-end manner from images depicting the same places over time downloaded from Google Street View Time Machine. Finally, we show that the proposed architecture significantly outperforms non-learnt image representations and off-the-shelf CNN descriptors on two challenging place recognition benchmarks, and improves over current state-of-the-art compact image representations on standard image retrieval benchmarks.

The exponential increase in the availability of online reviews and recommendations makes sentiment classification an interesting topic in academic and industrial research. Reviews can span so many different domains that it is difficult to gather annotated training data for all of them. Hence, this paper studies the problem of domain adaptation for sentiment classifiers, hereby a system is trained on labeled reviews from one source domain but is meant to be deployed on another. We propose a deep learning approach which learns to extract a meaningful representation for each review in an unsupervised fashion. Sentiment classifiers trained with this high-level feature representation clearly outperform state-of-the-art methods on a benchmark composed of reviews of 4 types of Amazon products. Furthermore, this method scales well and allowed us to successfully perform domain adaptation on a larger industrial-strength dataset of 22 domains. 1.

The paper is devoted to a gnerai finite element approximation ofthe solution of the Stokes quations for an incompressible viscous fluid, Both conforming and nonconforming finite element methods are studied and various examples of simplicial lments well suitedfor the numerical treatment of the incompressibility condition are given. Optimal error estimtes are derived in the energy norm and in the L 2 -norm.

Abstract Motivation: A variety of probabilistic models describing the evolution of DNA or protein sequences have been proposed for phylogenetic reconstruction or for molecular dating. However, there still lacks a common implementation allowing one to freely combine these independent features, so as to test their ability to jointly improve phylogenetic and dating accuracy. Results: We propose a software package, PhyloBayes 3, which can be used for conducting Bayesian phylogenetic reconstruction and molecular dating analyses, using a large variety of amino acid replacement and nucleotide substitution models, including empirical mixtures or non-parametric models, as well as alternative clock relaxation processes. Availability: PhyloBayes is freely available from our web site http://www.phylobayes.org. It works under Linux, Mac OsX and Windows operating systems. Contact: nicolas.lartillot@umontreal.ca Supplementary information: Supplementary data are available at Bioinformatics online.

International audience

The human ageing process is universal, ubiquitous and inevitable. Every physiological function is being continuously diminished. There is a range between two distinct phenotypes of ageing, shaped by patterns of living - experiences and behaviours, and in particular by the presence or absence of physical activity (PA) and structured exercise (i.e., a sedentary lifestyle). Ageing and a sedentary lifestyle are associated with declines in muscle function and cardiorespiratory fitness, resulting in an impaired capacity to perform daily activities and maintain independent functioning. However, in the presence of adequate exercise/PA these changes in muscular and aerobic capacity with age are substantially attenuated. Additionally, both structured exercise and overall PA play important roles as preventive strategies for many chronic diseases, including cardiovascular disease, stroke, diabetes, osteoporosis, and obesity; improvement of mobility, mental health, and quality of life; and reduction in mortality, among other benefits. Notably, exercise intervention programmes improve the hallmarks of frailty (low body mass, strength, mobility, PA level, energy) and cognition, thus optimising functional capacity during ageing. In these pathological conditions exercise is used as a therapeutic agent and follows the precepts of identifying the cause of a disease and then using an agent in an evidence-based dose to eliminate or moderate the disease. Prescription of PA/structured exercise should therefore be based on the intended outcome (e.g., primary prevention, improvement in fitness or functional status or disease treatment), and individualised, adjusted and controlled like any other medical treatment. In addition, in line with other therapeutic agents, exercise shows a dose-response effect and can be individualised using different modalities, volumes and/or intensities as appropriate to the health state or medical condition. Importantly, exercise therapy is often directed at several physiological systems simultaneously, rather than targeted to a single outcome as is generally the case with pharmacological approaches to disease management. There are diseases for which exercise is an alternative to pharmacological treatment (such as depression), thus contributing to the goal of deprescribing of potentially inappropriate medications (PIMS). There are other conditions where no effective drug therapy is currently available (such as sarcopenia or dementia), where it may serve a primary role in prevention and treatment. Therefore, this consensus statement provides an evidence-based rationale for using exercise and PA for health promotion and disease prevention and treatment in older adults. Exercise prescription is discussed in terms of the specific modalities and doses that have been studied in randomised controlled trials for their effectiveness in attenuating physiological changes of ageing, disease prevention, and/or improvement of older adults with chronic disease and disability. Recommendations are proposed to bridge gaps in the current literature and to optimise the use of exercise/PA both as a preventative medicine and as a therapeutic agent.

In this work we address the problem of finding reliable pixel-level correspondences under difficult imaging conditions. We propose an approach where a single convolutional neural network plays a dual role: It is simultaneously a dense feature descriptor and a feature detector. By postponing the detection to a later stage, the obtained keypoints are more stable than their traditional counterparts based on early detection of low-level structures. We show that this model can be trained using pixel correspondences extracted from readily available large-scale SfM reconstructions, without any further annotations. The proposed method obtains state-of-the-art performance on both the difficult Aachen Day-Night localization dataset and the InLoc indoor localization benchmark, as well as competitive performance on other benchmarks for image matching and 3D reconstruction.

Estimating human pose, shape, and motion from images and videos are fundamental challenges with many applications. Recent advances in 2D human pose estimation use large amounts of manually-labeled training data for learning convolutional neural networks (CNNs). Such data is time consuming to acquire and difficult to extend. Moreover, manual labeling of 3D pose, depth and motion is impractical. In this work we present SURREAL (Synthetic hUmans foR REAL tasks): a new large-scale dataset with synthetically-generated but realistic images of people rendered from 3D sequences of human motion capture data. We generate more than 6 million frames together with ground truth pose, depth maps, and segmentation masks. We show that CNNs trained on our synthetic dataset allow for accurate human depth estimation and human part segmentation in real RGB images. Our results and the new dataset open up new possibilities for advancing person analysis using cheap and large-scale synthetic data.

Abstract Marine plastic debris floating on the ocean surface is a major environmental problem. However, its distribution in the ocean is poorly mapped, and most of the plastic waste estimated to have entered the ocean from land is unaccounted for. Better understanding of how plastic debris is transported from coastal and marine sources is crucial to quantify and close the global inventory of marine plastics, which in turn represents critical information for mitigation or policy strategies. At the same time, plastic is a unique tracer that provides an opportunity to learn more about the physics and dynamics of our ocean across multiple scales, from the Ekman convergence in basin-scale gyres to individual waves in the surfzone. In this review, we comprehensively discuss what is known about the different processes that govern the transport of floating marine plastic debris in both the open ocean and the coastal zones, based on the published literature and referring to insights from neighbouring fields such as oil spill dispersion, marine safety recovery, plankton connectivity, and others. We discuss how measurements of marine plastics (both in situ and in the laboratory), remote sensing, and numerical simulations can elucidate these processes and their interactions across spatio-temporal scales.

Successful methods for visual object recognition typically rely on training datasets containing lots of richly annotated images. Detailed image annotation, e.g. by object bounding boxes, however, is both expensive and often subjective. We describe a weakly supervised convolutional neural network (CNN) for object classification that relies only on image-level labels, yet can learn from cluttered scenes containing multiple objects. We quantify its object classification and object location prediction performance on the Pascal VOC 2012 (20 object classes) and the much larger Microsoft COCO (80 object classes) datasets. We find that the network (i) outputs accurate image-level labels, (ii) predicts approximate locations (but not extents) of objects, and (iii) performs comparably to its fully-supervised counterparts using object bounding box annotation for training.

In this paper, we address the problem of estimating and removing non-uniform motion blur from a single blurry image. We propose a deep learning approach to predicting the probabilistic distribution of motion blur at the patch level using a convolutional neural network (CNN). We further extend the candidate set of motion kernels predicted by the CNN using carefully designed image rotations. A Markov random field model is then used to infer a dense non-uniform motion blur field enforcing motion smoothness. Finally, motion blur is removed by a non-uniform deblurring model using patch-level image prior. Experimental evaluations show that our approach can effectively estimate and remove complex non-uniform motion blur that is not handled well by previous approaches.

The purpose of this paper is to show that logic provides a convenient formalism for studying classical database problems. There are two main parts to the paper, devoted respectively to conventional databases and deductive databases. In the first part, we focus on query languages, integrity modeling and maintenance, query optimization, and data

This paper studies automatic image classification by modeling soft assignment in the popular codebook model. The codebook model describes an image as a bag of discrete visual words selected from a vocabulary, where the frequency distributions of visual words in an image allow classification. One inherent component of the codebook model is the assignment of discrete visual words to continuous image features. Despite the clear mismatch of this hard assignment with the nature of continuous features, the approach has been successfully applied for some years. In this paper, we investigate four types of soft assignment of visual words to image features. We demonstrate that explicitly modeling visual word assignment ambiguity improves classification performance compared to the hard assignment of the traditional codebook model. The traditional codebook model is compared against our method for five well-known data sets: 15 natural scenes, Caltech-101, Caltech-256, and Pascal VOC 2007/2008. We demonstrate that large codebook vocabulary sizes completely deteriorate the performance of the traditional model, whereas the proposed model performs consistently. Moreover, we show that our method profits in high-dimensional feature spaces and reaps higher benefits when increasing the number of image categories.

Epidemiological, migration, intervention, and genetic studies in humans and animals provide very strong evidence of a causal link between high salt intake and high blood pressure. The mechanisms by which dietary salt increases arterial pressure are not fully understood, but they seem related to the inability of the kidneys to excrete large amounts of salt. From an evolutionary viewpoint, the human species is adapted to ingest and excrete <1 g of salt per day, at least 10 times less than the average values currently observed in industrialized and urbanized countries. Independent of the rise in blood pressure, dietary salt also increases cardiac left ventricular mass, arterial thickness and stiffness, the incidence of strokes, and the severity of cardiac failure. Thus chronic exposure to a high-salt diet appears to be a major factor involved in the frequent occurrence of hypertension and cardiovascular diseases in human populations.

This paper proposes a novel algorithm for calibrated multi-view stereopsis that outputs a (quasi) dense set of rectangular patches covering the surfaces visible in the input images. This algorithm does not require any initialization in the form of a bounding volume, and it detects and discards automatically outliers and obstacles. It does not perform any smoothing across nearby features, yet is currently the top performer in terms of both coverage and accuracy for four of the six benchmark datasets presented in [20]. The keys to its performance are effective techniques for enforcing local photometric consistency and global visibility constraints. Stereopsis is implemented as a match, expand, and filter procedure, starting from a sparse set of matched keypoints, and repeatedly expanding these to nearby pixel correspondences before using visibility constraints to filter away false matches. A simple but effective method for turning the resulting patch model into a mesh appropriate for image-based modeling is also presented. The proposed approach is demonstrated on various datasets including objects with fine surface details, deep concavities, and thin structures, outdoor scenes observed from a restricted set of viewpoints, and "crowded" scenes where moving obstacles appear in different places in multiple images of a static structure of interest.

A scattering transform defines a locally translation invariant representation which is stable to time-warping deformation. It extends MFCC representations by computing modulation spectrum coefficients of multiple orders, through cascades of wavelet convolutions and modulus operators. Second-order scattering coefficients characterize transient phenomena such as attacks and amplitude modulation. A frequency transposition invariant representation is obtained by applying a scattering transform along log-frequency. State-the-of-art classification results are obtained for musical genre and phone classification on GTZAN and TIMIT databases, respectively.