IMT Atlantique

Novel Class Discovery (NCD) is the problem of trying to discover novel classes in an unlabeled set, given a labeled set of different but related classes. The majority of NCD methods proposed so far only deal with image data, despite tabular data being among the most widely used type of data in practical applications. To interpret the results of clustering or NCD algorithms, data scientists need to understand the domain- and application-specific attributes of tabular data. This task is difficult and can often only be performed by a domain expert. Therefore, this interface allows a domain expert to easily run state-of-the-art algorithms for NCD in tabular data. With minimal knowledge in data science, interpretable results can be generated.

Diabetic Retinopathy (DR), a prevalent and severe complication of diabetes, affects millions of individuals globally, underscoring the need for accurate and timely diagnosis. Recent advancements in imaging technologies, such as Ultra-WideField Color Fundus Photography (UWF-CFP) imaging and Optical Coherence Tomography Angiography (OCTA), provide opportunities for the early detection of DR but also pose significant challenges given the disparate nature of the data they produce. This study introduces a novel multimodal approach that leverages these imaging modalities to notably enhance DR classification. Our approach integrates 2D UWF-CFP images and 3D high-resolution 6\(\,\times \,\)6 mm\(^3\) OCTA (both structure and flow) images using a fusion of ResNet50 and 3D-ResNet50 models, with Squeeze-and-Excitation (SE) blocks to amplify relevant features. Additionally, to increase the model’s generalization capabilities, a multimodal extension of Manifold Mixup, applied to concatenated multimodal features, is implemented. Experimental results demonstrate a remarkable enhancement in DR classification performance with the proposed multimodal approach compared to methods relying on a single modality only. The methodology laid out in this work holds substantial promise for facilitating more accurate, early detection of DR, potentially improving clinical outcomes for patients.

Recently, Deep Learning (DL)-based unmixing techniques have gained popularity owing to the robust learning of Deep Neural Networks (DNNs). In particular, the Autoencoder (AE) model, as a baseline network for unmixing, performs well in Hyperspectral Unmixing (HU) by automatically learning a new representation and recovering original data. However, patch-wise AE based architecture, which incorporates both spectral and spatial information through convolutional filters may blur the abundance maps due to the fixed kernel shape of the used window size. To cope with the above issue, we propose in this paper a novel methodology based on graph DL called DNGAE. Unlike the pixel-wise or patch-wise Convolutional AE (CAE), our proposed method incorporates the complementary spatial information based on graph spectral similarity. A neighborhood graph based on band correlations is firstly constructed. Then, our method attempts to aggregate similar spectra from the neighboring pixels of a target pixel. Consequently, this leads to better quality of both extracted endmembers and abundances. Extensive experiments performed on two real HSI benchmarks confirm the effectiveness of our proposed method compared to other DL models.

In the decision-making systems, sensors data are often affected by diversity types of imperfections which significantly reduce system performance. Thus, data quality assessment has become a primordial step in data learning process. In this work, feature quality assessment approach is used to select only useful and reliable information from those depth data in order to improve the performances of the staircase recognition system for the visually impaired. Possibility theory is utilized as a tool to deal with imperfect data and to represent knowledge. The developed feature quality assessment method is composed of two main steps. In the first step, the extracted features are classified in different quality levels. In the second step, an optimal feature subset is determined to discriminate between ascending and descending stairs. The proposed approach has experimentally shown to be very valuable with depth data, acquired from an ultrasonic sensor and a LiDAR sensor. The performance of the proposed approach has evaluated based on the classification accuracy which reached \(94.7\%\) using the ultrasonic sensor dataset and \(92.42\%\) using the laser rangefinder dataset.

Hyperspectral unmixing allows to represent mixed pixels as a set of pure materials weighted by their abundances. Spectral features alone are often insufficient, so it is common to rely on other features of the scene. Matrix models become insufficient when the hyperspectral image is represented as a high-order tensor with additional features in a multimodal, multi-feature framework. Tensor models such as Canonical polyadic decomposition allow for this kind of unmixing, but lack a general framework and interpretability of the results. In this paper, we propose an interpretable methodological framework for low-rank Multi-feature hyperspectral unmixing based on tensor decomposition (MultiHU-TD) which incorporates the abundance sum-to-one constraint in the Alternating optimization ADMM algorithm, and provide in-depth mathematical, physical and graphical interpretation and connections with the extended linear mixing model. As additional features, we propose to incorporate mathematical morphology and reframe a previous work on neighborhood patches within MultiHU-TD. Experiments on real hyperspectral images showcase the interpretability of the model and the analysis of the results.

One of the challenges in the field of human activity recognition in smart homes based on IoT sensors is the variability in the recorded data. This variability arises from differences in home configurations, sensor network setups, and the number and habits of inhabitants, resulting in a lack of data that accurately represent the application environment. Although simulators have been proposed in the literature to generate data, they fail to bridge the gap between training and field data or produce diverse datasets. In this article, we propose a solution to address this issue by leveraging the concept of digital twins to reduce the disparity between training and real-world data and generate more varied datasets. We introduce the Virtual Smart Home, a simulator specifically designed for modeling daily life activities in smart homes, which is adapted from the Virtual Home simulator. To assess its realism, we compare a set of activity data recorded in a real-life smart apartment with its replication in the VirtualSmartHome simulator. Additionally, we demonstrate that an activity recognition algorithm trained on the data generated by the VirtualSmartHome simulator can be successfully validated using real-life field data.

Background: Electronic health records (EHRs) contain valuable information for clinical research; however, the sensitive nature of healthcare data presents security and confidentiality challenges. Deidentification is therefore essential to protect personal data in EHRs and comply with government regulations. Named entity recognition (NER) methods have been proposed to remove personal identifiers, with deep learning-based models achieving better performance. However, manual annotation of training data is time-consuming and expensive. The aim of this study was to develop an automatic deidentification pipeline for all kinds of clinical documents based on a distant supervised method to significantly reduce the cost of manual annotations and to facilitate the transfer of the deidentification pipeline to other clinical centers. Methods: We proposed an automated annotation process for French clinical deidentification, exploiting data from the eHOP clinical data warehouse(CDW) of the CHU de Rennes and national knowledge bases, as well as other features. In addition, this paper proposes an assisted data annotation solution using the Prodigy annotation tool. This approach aims to reduce the cost required to create a reference corpus for the evaluation of state-of-the-art NER models. Finally, we evaluated and compared the effectiveness of different NER methods. Results: A French deidentification dataset was developed in this work, based on EHRs provided by the eHOP CDW at Rennes University Hospital, France. The dataset was rich in terms of personal information, and the distribution of entities was quite similar in the training and test datasets. We evaluated a Bi-LSTM + CRF sequence labeling architecture, combined with Flair + FastText word embeddings, on a test set of manually annotated clinical reports. The model outperformed the other tested models with a significant F1 score of 96,96%, demonstrating the effectiveness of our automatic approach for deidentifying sensitive information. Conclusions: This study provides an automatic deidentification pipeline for clinical notes, which can facilitate the reuse of EHRs for secondary purposes such as clinical research. Our study highlights the importance of using advanced NLP techniques for effective de-identification, as well as the need for innovative solutions such as distant supervision to overcome the challenge of limited annotated data in the medical domain.

Network troubleshooting usually requires packet level traffic capturing and analyzing. Indeed, the observation of emission patterns sheds some light on the kind of degradation experienced by a connection. In the case of reliable transport traffic where congestion control is performed, such as TCP and QUIC traffic, these patterns are the fruit of decisions made by the congestion control algorithm (CCA), according to its own perception of network conditions. The CCA estimates the bottleneck’s capacity via an exponential probing, during the so-called “Slow-Start” (SS) state. The bottleneck is considered reached upon reception of congestion signs, typically lost packets or abnormal packet delays depending on the version of CCA used. The SS state duration is thus a key indicator for the diagnosis of faults; this indicator is estimated empirically by human experts today, which is time-consuming and a cumbersome task with large error margins. This paper proposes a method to automatically identify the slow-start state from actively and passively obtained bidirectional packet traces. It relies on an innovative timeless representation of the observed packets series. We implemented our method in our active and passive probes and tested it with CUBIC and BBR under different network conditions. We then picked a few real-life examples to illustrate the value of our representation for easy discrimination between typical faults and for identifying BBR among CCAs variants.

A bstract In [1], we performed the first complete computation of the back-to-back inclusive dijet cross-section in Deeply Inelastic Scattering (DIS) at small x Bj to next-to-leading order (NLO) in the Color Glass Condensate effective field theory (CGC EFT). We demonstrate here that for dijets with relative transverse momentum P ⊥ and transverse momentum imbalance q ⊥ , to leading power in q ⊥ / P ⊥ , the cross-section for longitudinally polarized photons can be fully factorized into the product of a perturbative impact factor and the non-perturbative Weizsäcker-Williams (WW) transverse momentum dependent (TMD) gluon distribution to NLO accuracy. The impact factor can further be expressed as the product of a universal soft factor which resums Sudakov double and single logs in P ⊥ / q ⊥ and a coefficient function given by a remarkably compact analytic expression. We show that in the CGC EFT the WW TMD satisfies a kinematically constrained JIMWLK renormalization group evolution in rapidity. This factorization formula is valid to all orders in Q s / q ⊥ for q ⊥ , Q s ≪ P ⊥ , where Q s is the semi-hard saturation scale that grows with rapidity.

Internet of Things (IoT) applications have invaded several domains (supply chain, healthcare, etc.). To enhance the quality of the provided service in terms of latency, response time, etc., service providers such as Amazon, Google, and Microsoft turned to running IoT tasks near the end user by invoking the fog computing concept. Fog computing extends cloud services to the edge of the network. It provides a variety of computing resources in the form of fog nodes, which offer multiple services known as fog services. These latter are used to store and process the data generated by IoT devices. Fog services are characterized by their high reusability. It enables the construction of a composite service to provide complicated IoT tasks. In this paper, we introduce an adaptive requirements-aware approach for configuring IoT applications in fog computing. The configuration is based on a Composite Fog Service (CFS) model and is restricted by a set of constraints. The proposed approach is implemented in a Smart Car Parking (SCP) scenario. Simulation results reveal the effectiveness of our approach.

A bstract Measurements of the production of electrons from heavy-flavour hadron decays in pp collisions at $$ \sqrt{s} $$ s = 13 TeV at midrapidity with the ALICE detector are presented down to a transverse momentum ( p T ) of 0.2 GeV/ c and up to p T = 35 GeV/ c , which is the largest momentum range probed for inclusive electron measurements in ALICE. In p-Pb collisions, the production cross section and the nuclear modification factor of electrons from heavy-flavour hadron decays are measured in the p T range 0 . 5 < p T < 26 GeV/ c at $$ \sqrt{s_{\textrm{NN}}} $$ s NN = 8 . 16 TeV. The nuclear modification factor is found to be consistent with unity within the statistical and systematic uncertainties. In both collision systems, first measurements of the yields of electrons from heavy-flavour hadron decays in different multiplicity intervals normalised to the multiplicity-integrated yield (self-normalised yield) at midrapidity are reported as a function of the self-normalised charged-particle multiplicity estimated at midrapidity. The self-normalised yields in pp and p-Pb collisions grow faster than linear with the self-normalised multiplicity. A strong p T dependence is observed in pp collisions, where the yield of high- p T electrons increases faster as a function of multiplicity than the one of low- p T electrons. The measurement in p-Pb collisions shows no p T dependence within uncertainties. The self-normalised yields in pp and p-Pb collisions are compared with measurements of other heavy-flavour, light-flavour, and strange particles, and with Monte Carlo simulations.

The release of CO2 into the atmosphere has become a primary issue nowadays. Recently, researchers found Metal-Organic Frameworks M-CPO-27 (M = Mg, Co, Ni, and Zn) to be revolutionary for CO2 adsorption due to the presence of open metal sites enhancing CO2 binding and leading to higher capacity. This study aims to select the best metal center for CPO-27 with the high performance of CO2 adsorption by screening metal centers using simulation as a preliminary selection method. Then, the different metal centers were synthesized using the solvothermal process for validation. The synthesis of MOFs is confirmed through PXRD and FTIR analysis. Subsequently, by using simulation and experimental methods, it is discovered that Ni-CPO-27 gives the best performance compared with magnesium, zinc, and cobalt metal centers. The CO2 adsorption capacity of synthesized Ni-CPO-27 is 5.6 mmol/g, which is almost 20% higher than other MOFs. In conclusion, the prospective outcome of changing the metal from Mg-CPO-27 to Ni-CPO-27 would be advantageous in this investigation owing to its excellent performance in capturing CO2.

The metaverse and Web 3.0 have created a new digital world with specific properties and behaviours replicating and influencing the behaviours and processes of physical entities. This study aims to advance our understanding of how the metaverse will impact supply chain and operations management (SCOM). Using elements of a structured literature search and building on the concepts of cyber-physical systems, digital supply chain twins, cloud supply chains, and Industry 4.0/Industry 5.0, we propose a framework for metaverse SCOM encompassing multiple socio-technological dimensions. We conclude that further metaverse developments could result in a co-existence of physical SCOM, metaverse SCOM, and SCOM for coordination of the physical and metaverse worlds. We offer a structured future research agenda pointing to new research questions and topics stemming from metaverse-driven visibility, computational power for data analytics, digital collaboration, and connectivity. New research areas can emerge for the novel metaverse SCOM processes and decision-making areas (e.g. joint demand forecasting for metaverse and physical products, digital inventory allocation in the metaverse, integrated production planning for the metaverse and physical worlds, and pricing and contracting for digital products), as well as new performance measures (e.g. virtual customer experience level, availability of digital products, and digital resilience and sustainability).

We report on the first search for nuclear recoils from dark matter in the form of weakly interacting massive particles (WIMPs) with the XENONnT experiment, which is based on a two-phase time projection chamber with a sensitive liquid xenon mass of 5.9 ton. During the (1.09±0.03) ton yr exposure used for this search, the intrinsic Kr85 and Rn222 concentrations in the liquid target are reduced to unprecedentedly low levels, giving an electronic recoil background rate of (15.8±1.3) events/ton yr keV in the region of interest. A blind analysis of nuclear recoil events with energies between 3.3 and 60.5 keV finds no significant excess. This leads to a minimum upper limit on the spin-independent WIMP-nucleon cross section of 2.58×1047 cm2 for a WIMP mass of 28 GeV/c2 at 90% confidence level. Limits for spin-dependent interactions are also provided. Both the limit and the sensitivity for the full range of WIMP masses analyzed here improve on previous results obtained with the XENON1T experiment for the same exposure.