Search

Adopting Convolutional Neural Networks (CNNs) in the daily routine of primary diagnosis requires not only near-perfect precision, but also a sufficient degree of generalization to data acquisition shifts and transparency. Existing CNN models act as black boxes, not ensuring to the physicians that important diagnostic features are used by the model. Building on top of successfully existing techniques such as multi-task learning, domain adversarial training and concept-based interpretability, this paper addresses the challenge of introducing diagnostic factors in the training objectives. Here we show that our architecture, by learning end-to-end an uncertainty-based weighting combination of multi-task and adversarial losses, is encouraged to focus on pathology features such as density and pleomorphism of nuclei, e.g. variations in size and appearance, while discarding misleading features such as staining differences. Our results on breast lymph node tissue show significantly improved generalization in the detection of tumorous tissue, with best average AUC 0.89 (0.01) against the baseline AUC 0.86 (0.005). By applying the interpretability technique of linearly probing intermediate representations, we also demonstrate that interpretable pathology features such as nuclei density are learned by the proposed CNN architecture, confirming the increased transparency of this model. This result is a starting point towards building interpretable multi-task architectures that are robust to data heterogeneity. Our code is available at https://bit.ly/356yQ2u.

Despite the evolution of deep-learning-based visual-textual processing systems, precise multi-modal matching remains a challenging task. In this work, we tackle the task of cross-modal retrieval through image-sentence matching based on word-region alignments, using supervision only at the global image-sentence level. Specifically, we present a novel approach called Transformer Encoder Reasoning and Alignment Network (TERAN). TERAN enforces a fine-grained match between the underlying components of images and sentences, i.e., image regions and words, respectively, in order to preserve the informative richness of both modalities. TERAN obtains state-of-the-art results on the image retrieval task on both MS-COCO and Flickr30k datasets. Moreover, on MS-COCO, it also outperforms current approaches on the sentence retrieval task. Focusing on scalable cross-modal information retrieval, TERAN is designed to keep the visual and textual data pipelines well separated. Cross-attention links invalidate any chance to separately extract visual and textual features needed for the online search and the offline indexing steps in large-scale retrieval systems. In this respect, TERAN merges the information from the two domains only during the final alignment phase, immediately before the loss computation. We argue that the fine-grained alignments produced by TERAN pave the way towards the research for effective and efficient methods for large-scale cross-modal information retrieval. We compare the effectiveness of our approach against relevant state-of-the-art methods. On the MS-COCO 1K test set, we obtain an improvement of 5.7% and 3.5% respectively on the image and the sentence retrieval tasks on the Recall@1 metric. The code used for the experiments is publicly available on GitHub at https://github.com/mesnico/TERAN., Comment: Accepted in ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM). arXiv admin note: text overlap with arXiv:2004.09144

Adopting Convolutional Neural Networks (CNNs) in daily routine of primary diagnosis requires not only near-perfect precision, but also a sufficient degree of transparency and explainability of the decision making. With physicians being accountable for the diagnosis, it is fundamental that CNNs provide a clear interpretation of their learning paradigm, ensuring that relevant pathology features are being considered. Building on top of successfully existing techniques such as multi-task learning, domain adversarial training and concept-based interpretability, this paper addresses the challenge of introducing diagnostic factors in the training objectives. Here we show that our architecture, by learning end-to-end an uncertainty-based weighting combination of multi-task and adversarial losses, is encouraged to focus on pathology features such as density and pleomorphism of nuclei, e.g. variations in size and appearance, while discarding misleading features such as staining differences. Our results on the classification of tumor in breast lymph node tissue scans show significantly improved generalization, with best average AUC 0.89 (0.01) against the baseline AUC 0.86 (0.005). This result is a starting point towards building interpretable multi-task architectures that are robust to data heterogeneity. Our code is available at https://bit.ly/356yQ2u.

High yield HIV testing strategies are critical to reach epidemic control in high prevalence and low-resource settings such as East and Southern Africa. In this study, we aimed to predict the HIV status of individuals living in Angola, Burundi, Ethiopia, Lesotho, Malawi, Mozambique, Namibia, Rwanda, Zambia and Zimbabwe with the highest precision and sensitivity for different policy targets and constraints based on a minimal set of socio-behavioural characteristics.We analysed the most recent Demographic and Health Survey from these 10 countries to predict individual's HIV status using four different algorithms (a penalized logistic regression, a generalized additive model, a support vector machine, and a gradient boosting trees). The algorithms were trained and validated on 80% of the data, and tested on the remaining 20%. We compared the predictions based on the F1 score, the harmonic mean of sensitivity and positive predictive value (PPV), and we assessed the generalization of our models by testing them against an independent left-out country. The best performing algorithm was trained on a minimal subset of variables which were identified as the most predictive, and used to 1) identify 95% of people living with HIV (PLHIV) while maximising precision and 2) identify groups of individuals by adjusting the probability threshold of being HIV positive (90% in our scenario) for achieving specific testing strategies.Overall 55,151 males and 69,626 females were included in the analysis. The gradient boosting trees algorithm performed best in predicting HIV status with a mean F1 score of 76.8% [95% confidence interval (CI) 76.0%-77.6%] for males (vs [CI 67.8%-70.6%] for SVM) and 78.8% [CI 78.2%-79.4%] for females (vs [CI 73.4%-75.8%] for SVM). Among the ten most predictive variables for each sex, nine were identical: longitude, latitude and, altitude of place of residence, current age, age of most recent partner, total lifetime number of sexual partners, years lived in current place of residence, condom use during last intercourse and, wealth index. Only age at first sex for male (ranked 10th) and Rohrer's index for female (ranked 6th) were not similar for both sexes. Our large-scale scenario, which consisted in identifying 95% of all PLHIV, would have required testing 49.4% of males and 48.1% of females while achieving a precision of 15.4% for males and 22.7% for females. For the second scenario, only 4.6% of males and 6.0% of females would have had to be tested to find 55.7% of all males and 50.5% of all females living with HIV.We trained a gradient boosting trees algorithm to find 95% of PLHIV with a precision twice higher than with general population testing by using only a limited number of socio-behavioural characteristics. We also successfully identified people at high risk of infection who may be offered pre-exposure prophylaxis or voluntary medical male circumcision. These findings can inform the implementation of new high-yield HIV tests and help develop very precise strategies based on low-resource settings constraints.

Twitter has significant potential as a source of Volunteered Geographic Information (VGI), as its content is updated at high frequency, with high availability thanks to dedicated interfaces. However, the diversity of content types and the low average accuracy of geographic information attached to individual tweets remain obstacles in this context. The contributions in this paper relate to the general goal of extracting actionable information regarding the impact of natural hazards on a specific region from social platforms, such as Twitter. Specifically, our contributions describe the construction of a model classifying whether given spatio-temporal coordinates, materialized by raster cells in a remote sensing context, lie in a flooded area. For training, remotely sensed data are used as the target variable, and the input covariates are built on the sole basis of textual and spatial data extracted from a Twitter corpus. Our contributions enable the use of trained models for arbitrary new Twitter corpora collected for the same region, but at different times, allowing for the construction of a flooded area measurement proxy available at a higher temporal frequency. Experimental validation uses true data that were collected during Hurricane Harvey, which caused significant flooding in the Houston urban area between mid-August and mid-September 2017. Our experimental section compares several spatial information extraction methods, as well as various textual representation and aggregation techniques, which were applied to the collected Twitter data. The best configuration yields a F1 score of 0.425, boosted to 0.834 if restricted to the 10% most confident predictions.

Computational pathology is a domain that aims to develop algorithms to automatically analyze large digitized histopathology images, called whole slide images (WSI). WSIs are produced scanning thin tissue samples that are stained to make specific structures visible. They show stain colour heterogeneity due to different preparation and scanning settings applied across medical centers. Stain colour heterogeneity is a problem to train convolutional neural networks (CNN), the state-of-the-art algorithms for most computational pathology tasks, since CNNs usually underperform when tested on images including different stain variations than those within data used to train the CNN. Despite several methods that were developed, stain colour heterogeneity is still an unsolved challenge that limits the development of CNNs that can generalize on data from several medical centers. This paper aims to present a novel method to train CNNs that better generalize on data including several colour variations. The method, called H&E-adversarial CNN, exploits H&E matrix information to learn stain-invariant features during the training. The method is evaluated on the classification of colon and prostate histopathology images, involving eleven heterogeneous datasets, and compared with five other techniques used to handle stain colour heterogeneity. H&E-adversarial CNNs show an improvement in performance compared to the other algorithms, demonstrating that it can help to better deal with stain colour heterogeneous images., Errata corrige Proceedings of the IEEE/CVF International Conference on Computer Vision 2021

Finding inherent or processed links within a dataset allows to discover potential knowledge. The main contribution of this article is to define a global framework that enables optimal knowledge discovery by visually rendering co-occurences (i.e. groups of linked data instances attached to a metadata reference) – either inherently present or processed – from a dataset as facets. Hypergraphs are well suited for modeling co-occurences since they support multi-adicity whereas graphs only support pairwise relationships. This article introduces an efficient navigation between different facets of an information space based on hypergraph modelisation and visualisation.

BackgroundThere is a need for high yield HIV testing strategies to reach epidemic control. We aimed to predict the HIV status of individuals based on socio-behavioural characteristics.MethodsWe analysed over 3,200 variables from the most recent Demographic Health Survey from 10 countries in East and Southern Africa. We trained four machine-learning algorithms and selected the best based on the f1 score. Training and validation were done on 80% of the data. The model was tested on the remaining 20% and on a left-out country which was rotated around. The best algorithm was retrained on the variables which were most predictive. We studied two scenarios: one aiming to identify 95% of people living with HIV (PLHIV) and one aiming to identify individuals with 95% or higher probability of being HIV positive.FindingsOverall 55,151 males and 69,626 females were included. XGBoost performed best in predicting HIV with a mean f1 of 76·8% [95% confidence interval 76·0%-77·6%] for males and 78·8% [78·2%-79·4%] for females. Among the ten most predictive variables, nine were identical for both sexes: longitude, latitude and, altitude of place of residence, current age, age of most recent partner, total lifetime number of sexual partners, years lived in current place of residence, condom use during last intercourse and, wealth index. Model performance based on these variables decreased minimally. For the first scenario, 7 males and 5 females would need to be tested to identify one HIV positive person. For the second scenario, 4·2% of males and 6·2% of females would have been identified as high-risk population.InterpretationWe were able to identify PLHIV and those at high risk of infection who may be offered pre-exposure prophylaxis and/or voluntary medical male circumcision. These findings can inform the implementation of HIV prevention and testing strategies.FundingSwiss National Science Foundation.

We present a novel cooperative strategy based on active learning and crowdsourcing, dedicated to provide a solution to the cold start stage, i.e. initializing the classification of a large set of data with no attached labels. The strategy is moreover designed to handle an imbalanced context in which random selection is highly inefficient. In this purpose, our method is guided by an unsupervised clustering, and the computation of cluster quality and impurity indexes, updated at each active learning step. The strategy is explained on a case study of annotating Twitter content w.r.t. a real flood event. We also show that our technique can cope with multiple heterogeneous data representations.

The estimation of local intrinsic dimensionality has applications ranging from adversarial attack disclosure to clustering and outlier detection, indexing, and data fingerprinting. In this paper, we analyze measures of network centrality in the kNN graph and their relation to LID measures. Our method ranks the dataset by its centrality, measured as the number of reverse or mutual kNN of each object. The computation of these measures involves only kNN queries, allowing a speedup in its computation using probabilistic indexing. A property of independent interest is the rank being independent of k for a wide range of k values, leading to parameter-free density estimation and applications.

International audience

We address the problem of the statistical description of 3D surfaces with the purpose of automatic classification and retrieval of archaeological potsherds. These are particularly interesting problems in archaeology, as pottery comprises a great volume of findings in archaeological excavations. Indeed, the analysis of potsherds brings relevant cues for understanding the culture of ancient groups. In particular, we develop a new local shape descriptor for 3D surfaces, called the histogram of spherical orientations (HoSO), which we use in combination with a bag-of-words approach to compute visual similarity between 3D surfaces. Given a point of interest on a 3D surface, its local shape descriptor (HoSO) captures the distribution of the spherical orientations of its neighboring points. In turn, those spherical orientations are computed with respect to the point of interest itself, both in the azimuth and the zenith axis. The proposed HoSO is invariant to scale transformations and highly robust to rotation and noise. In addition, it is efficient, as it only exploits the information of the position of the 3D points and disregards other types of information like faces or normals. We performed experiments on a set of 3D surfaces representing potsherds from the Teotihuacan civilization and further validations on a set of 3D models of generic objects. Our results show that our methodology is effective for describing 3D models and that it improves classification performance with respect to previous local descriptors.

Permutation-based indexing is one of the most popular techniques for the approximate nearest-neighbor search problem in high-dimensional spaces. Due to the exponential increase of multimedia data, the time required to index this data has become a serious constraint. One of the possible steps towards faster index construction is utilization of massively parallel platforms such as the GPGPU architectures. In this paper, we have analyzed the computational costs of individual steps of the permutation-based index construction in a high-dimensional feature space and summarized our hybrid CPU-GPU solution. Our experience gained from this research may be utilized in other individual problems that require computing Lp distances in high-dimensional spaces, parallel top-k selection, or partial sorting of multiple smaller sets. We also provide guidelines how to balance workload in hybrid CPU-GPU systems.

International audience; Sea Surface Salinity (SSS) acquired by the Soil Moisture and Ocean Salinity (SMOS) satellite mission are subject to systematic errors originating from various non-geophysical contaminations such as land contamination. These systematic errors reach more than 2 pss in some regions close to the land with very strong spatial gradients according to the coast orientation and the across-track position within the satellite swath.An empirical method to estimate and correct the time independent systematic errors from resampled quasi L2 SMOS SSS is presented. The method is based on self-consistency hypothesis of long term (July 2010–June 2014) variability of SMOS SSS among SMOS dwell-lines and orbit orientation (ascending and descending). The bias correction is performed by first adjusting SSS relative variations among dwell-lines and orbits orientation and then by determining a mean correction over four years. A reference time series of SSS and the associated relative systematic error for each dwell-line and orbit orientation is estimated using a least square approach. Then, the 4-year mean value of corrected SMOS SSS is adjusted relative to a 4-year mean in situ data climatology.Four years of salinity maps mitigated from systematic errors are presented. Independent validation using in situ thermosalinograph SSS from ships of opportunity is presented. The new SMOS bias corrected SSS over the global ocean shows an improvement of 32% of the RMSD.

Indexing exploits assumptions on the inner structures of a dataset to make the nearest neighbor queries cheaper to resolve. Datasets are generally indexed at once into a unique index for similarity search. By indexing a given dataset as a whole, one faces the parameters of its global structure, which may be adverse. A typical well-studied example is a high global dimensionality of the dataset, making any indexing strategy inefficient due to the curse of dimensionality. We conjecture that a dataset may be partitioned into subsets of variable indexability. The strategy is, therefore, to define a procedure to extract parts of the dataset with predictable indexability and to adapt the index structure to this parameter. In this paper, we define and discuss indexability related to the curse of dimensionality and propose a related heuristic to partition the dataset into low-dimensional parts. Each data object is ranked according to its degree centrality, under a connected sparse graph, the Half-Space Proximal Graph (HSP). We postulate centrality measures are good predictors of dimensionality and indexability. In view of validation, we conducted an experiment using the degree centrality of the HSP graph as unique dimensionality/indexability measure. We ranked the data objects by their respective centrality degree under the HSP graph, then extracted the lower dimensional subsets, recomputed the HSP and repeated. Subsets were then indexed with an exact method in increasing, decreasing and random order. We measured the complexity of a fixed set of queries for each of the three arrangements. For each set we used a fixed dataset with 250 queries. The above single experiment demonstrated that the heuristic can extract low dimensional subsets, and also that those subsets are easier to index. This initial results demonstrate the validity of our conjecture and motivate the need for exploring further the notion of indexability and related dataset partitioning strategies.

We propose a new method for input variable selection in nonlinear regression. The method is embedded into a kernel regression machine that can model general nonlinear functions, not being a priori limited to additive models. This is the first kernel-based variable selection method applicable to large datasets. It sidesteps the typical poor scaling properties of kernel methods by mapping the inputs into a relatively low-dimensional space of random features. The algorithm discovers the variables relevant for the regression task together with learning the prediction model through learning the appropriate nonlinear random feature maps. We demonstrate the outstanding performance of our method on a set of large-scale synthetic and real datasets. Code related to this paper is available at: https://bitbucket.org/dmmlgeneva/srff_pytorch.

An important contribution of satellite Sea Surface Salinity (SSS) is the spatio-temporal monitoring of rivers fresh water plumes at mesoscale. In this paper, we detail a new correction for systematic errors in the Soil Moisture and Ocean Salinity (SMOS) measurements that is implemented in the Centre Aval de Traitement des Donnees SMOS (CATDS). With this new mitigation, the SMOS and Soil Moisture Active Passive (SMAP) SSS monitor very consistent features in most areas close to continents. The rms-difference between bi-weekly SMOS and SMAP SSS over 20 months and in selected coastal regions is about 0.3pss (once outliers are filtered out), rather consistent with the rms-difference between satellite and in situ SSS (on the order of 0.2pss). The coefficient of determination (r2) between SMOS and SMAP SSS is above than 0.8 in very fresh areas (river plumes). Over the open ocean, the rms difference between SMOS and ship SSS is 0.2pss.

This article presents the information-theoretic based feature information interaction, a measure that can describe complex feature dependencies in multivariate settings. According to the theoretical development, feature interactions are more accurate than current, bivariate dependence measures due to their stable and unambiguous definition. In experiments with artificial and real data we compare first the empirical dependency estimates of correlation, mutual information and 3-way feature interaction. Then, we present feature selection and classification experiments that show superior performance of interactions over bivariate dependence measures for the artificial data, for real world data this goal is not achieved yet.

International audience; What role does the notion of ‘everyday life’ play in Sextus Empiricus’s skepticism? On the basis of an analysis of the concept of βιωτικὴ τήρησις, this paper purports to show (i) that everyday life, as opposed to ‘philosophical life’, is an empirical fact that allows the Pyrrhonist to act without holding beliefs, and (ii) that everyday life is a genuine value of the Pyrrhonian philosophy. Even though these two theses may seem contradictory, the aim of the present paper is to show that Sextus’s philosophical skepticism make them compatible.; Quel rôle joue le concept de vie quotidienne dans le scepticisme de Sextus Empiricus ? À partir d’une analyse du concept de βιωτικὴ τήρησις, il s’agit de faire apparaître, d’une part (i) que la vie quotidienne, par opposition à la vie philosophique, est un fait empirique qui permet au sceptique d’agir sans pour autant avoir d’opinions et d’autre part que (ii) la vie quotidienne est une valeur qui norme et donne sens à la philosophie sceptique. Bien que ces deux approches paraissent contradictoires, le but de cet article est de montrer que le scepticisme philosophique proposé par Sextus les rend compatibles.

The K-Nearest Neighbor (K-NN) search problem is the way to find the K closest and most similar objects to a given query. The K-NN is essential for many applications such as information retrieval and visualization, machine learning and data mining. The exponential growth of data imposes to find approximate approaches to this problem. Permutation-based indexing is one of the most recent techniques for approximate similarity search. Objects are represented by permutation lists ordering their distances to a set of selected reference objects, following the idea that two neighboring objects have the same surrounding. In this paper, we propose a novel quantized representation of permutation lists with its related data structure for effective retrieval on single and multicore architectures. Our novel permutation-based indexing strategy is built to be fast, memory efficient and scalable. This is experimentally demonstrated in comparison to existing proposals using several large-scale datasets of millions of documents and of different dimensions. HighlightsA Multi-core indexing and searching implementations of our data structure.Test our proposal on the full CoPhIR dataset 106-million features.Compare our proposal to all the available permutation based indexing technique with larger datasets (1-million and 10-million).Compare our proposal to other approximate similarity search techniques like LSH-Forest and AM-Tree.

This article presents an integrated framework for multimedia access and analysis of ancient Maya epigraphic resources, which is developed as an interdisciplinary effort involving epigraphers (someone who deciphers ancient inscriptions) and computer scientists. Our work includes several contributions: a definition of consistent conventions to generate high-quality representations of Maya hieroglyphs from the three most valuable ancient codices, which currently reside in European museums and institutions; a digital repository system for glyph annotation and management; as well as automatic glyph retrieval and classification methods. We study the combination of statistical Maya language models and shape representation within a hieroglyph retrieval system, the impact of applying language models extracted from different hieroglyphic resources on various data types, and the effect of shape representation choices for glyph classification. A novel Maya hieroglyph data set is given, which can be used for shape analysis benchmarks, and also to study the ancient Maya writing system.