Search

In an effort to reduce annotation costs in action recognition, unsupervised video domain adaptation methods have been proposed that aim to adapt a predictive model from a labelled dataset (i.e., source domain) to an unlabelled dataset (i.e., target domain). In this work we address a more realistic scenario, called open-set video domain adaptation (OUVDA), where the target dataset contains "unknown" semantic categories that are not shared with the source. The challenge lies in aligning the shared classes of the two domains while separating the shared classes from the unknown ones. In this work we propose to address OUVDA with an unified contrastive learning framework that learns discriminative and well-clustered features. We also propose a video-oriented temporal contrastive loss that enables our method to better cluster the feature space by exploiting the freely available temporal information in video data. We show that discriminative feature space facilitates better separation of the unknown classes, and thereby allows us to use a simple similarity based score to identify them. We conduct thorough experimental evaluation on multiple OUVDA benchmarks and show the effectiveness of our proposed method against the prior art., Currently under review at Computer Vision and Image Understanding (CVIU) journal

Foundational image-language models have generated considerable interest due to their efficient adaptation to downstream tasks by prompt learning. Prompt learning treats part of the language model input as trainable while freezing the rest, and optimizes an Empirical Risk Minimization objective. However, Empirical Risk Minimization is known to suffer from distributional shifts which hurt generalizability to prompts unseen during training. By leveraging the regularization ability of Bayesian methods, we frame prompt learning from the Bayesian perspective and formulate it as a variational inference problem. Our approach regularizes the prompt space, reduces overfitting to the seen prompts and improves the prompt generalization on unseen prompts. Our framework is implemented by modeling the input prompt space in a probabilistic manner, as an a priori distribution which makes our proposal compatible with prompt learning approaches that are unconditional or conditional on the image. We demonstrate empirically on 15 benchmarks that Bayesian prompt learning provides an appropriate coverage of the prompt space, prevents learning spurious features, and exploits transferable invariant features. This results in better generalization of unseen prompts, even across different datasets and domains.

This work focuses on the hydrodenitrogenation reaction modelisation. The removal of nitrogen-containing molecules is part of the hydrotreating process which is required before the catalytic conversion process. Impurities in the oil fraction are eliminated by mixing the feedstock with hydrogen and passing it through a fixed bed catalytic reactor at high temperature and pressure. Hydrotreating is a common operation in every oil refinery andallows satisfying the environmental regulations for the final products.Reactions take place in presence of catalyst and when supplying a catalyst, a vendor must guarantee its performance. The aim is to predict the nitrogen content (N) after the hydrotreating stage based on information on the feedstock and operating conditions (x). In order to model the nitrogen content evolution during the reaction, kinetic models (1) are used. The structure of f is fixed and it depends on parameters θ to be optimized.dNdt= fθ(N|x). (1)The predictive model fitting is based on experimental data and experiments are very expensive. New catalysts are constantly being developed so that each new generation of a catalyst requires a new model that is until now built from scratch from new experiments. The aim of this work is to build the best predictive model for a new catalyst from fewer observations and using the observations of previous generation catalysts. This task is known as transfer learning (Pan and Yang 2010 and Tsung et al. 2018).In order to adapt the past knowledge to the new catalyst, a Bayesian approach is considered. The Bayes Theorem gives the posterior distribution of the model parameters θ (2), where N is the vector of the nitrogen content for the different observations, X the matrix of new observations, π(θ) the prior distribution of parameters, L(N|θ,X) the likelihood and L(N|X) the marginal likelihood.π(θ|N,X) =π(θ)L(N|θ,X)L(N|X). (2)The likelihood represents the knowledge about the new observations, thus the posterior distribution will be modified when adding observations. The idea of the approach is to take as prior π(θ) a distribution centered on the previous model parameters, with variance large enough to allow parameter change and small enough to retain the information. Previous work has shown the effectiveness of this Bayesian transfer approach for modelling using linear models and kriging models (Iapteff et al. 2021), and this current work aims to adapt it on kinetic models. Results are good and show a reduction in the number of observations required to achieve good results.

IFP group develops catalysts and has to guarantee their performances. It is therefore crucial to have good predictive models for all new catalysts. These models are built upon very expensive experimental data. In order to minimize costs, we aim at reducing the number of new data points to measure to fit a model on the new catalyst.Previous work has shown that a Bayesian transfer approach can improve the quality of models when the number of observations is reduced. In this paper, experimental designs are studied in order to determine how to select this reduced number of observations to obtain the best models by Bayesian transfer. This article shows that the Kennard and Stone algorithm can, under certain conditions, offer better results than optimal designs., Le groupe IFP commercialise des catalyseurs et doit s’engager sur leur performance. Il est donc nécessaire de disposer de modèles prédictifs fiables pour chaque nouvelle génération de catalyseurs. Ces modèles sont construits à partir de données expérimentales très couteuses. Afin d’optimiser les coûts, notre ambition est de réduire le nombre d’expérimentations nécessaires pour estimer un modèle associé à un nouveautype de catalyseur. De précédents travaux ont montré qu’une approche de transfert Bayésien permettait d’améliorer la qualité des modèles lorsque le nombre d’observations est réduit. Dans cet article, les plans d’expériences sont étudiés afin de déterminer comment sélectionner ce nombre réduit d’observations permettant d’obtenir les meilleurs modèles par transfert Bayésien. Cet article montre que l’algorithme Kennard and Stonepeut, sous certaines conditions, offrir de meilleurs résultats que des plans optimaux.

It was discovered in 2010 that Croconic Acid, in its crystal form, has the highest polarization among organic ferroelectrics. In the context of eliminating toxic substances from electronic devices, Croconic Acid has a great potential as a sublimable lead-free ferroelectric. However, studies on ferroelectric properties of its thin films are only in their early stages and its capability to be incorporated in nanoscale devices is unknown. In this work, we demonstrate, upon ferroelectric switching at the nanoscale, stable and enduring room temperature polarization with no leakage current in Croconic Acid thin films. We thus show that it is a promising lead-free organic ferroelectric toward integration in nanoscale devices. The challenging switching current and polarization reversal characterization at the nanoscale was done using a unique combination of piezoresponse force microscopy, polarization switching current spectroscopy and the concurrent electromechanical strain response. Indeed, this combination can help to rationalize otherwise asymmetric polarization-voltage data and distorted hysteresis due to current jumps below the background noise, which are statistically washed away in macrojunctions but become prevalent at the nanoscale. These results are valid irrespective of the ferroelectrics' nature, organic or inorganic. Beyond the potential of Croconic Acid as an ecological ferroelectric material in devices, our detection of a clear nanoscopic polarization switching current thus paves the way for a fundamental understanding and technological applications of the polarization reversal mechanism at the nanoscale.

In severe cases of COVID-19, it is important to note that some laboratory signs may alert to the presence of underlying macrophage activation syndrome (MAS), and we ratify that the classic signs of primary MAS are often not present. Here we show a case report of COVID-19 complicated by MAS treated with high doses of methylprednisolone and intravenous Immunoglobulin, with excellent clinical outcome, avoiding orotracheal intubation indeed. The interpretation of laboratory signs leads to early diagnosis and the introduction of effective therapy.

As more data are produced each day, and faster, data stream mining is growing in importance, making clear the need for algorithms able to fast process these data. Data stream mining algorithms are meant to be solutions to extract knowledge online, specially tailored from continuous data problem. Many of the current algorithms for data stream mining have high processing and memory costs. Often, the higher the predictive performance, the higher these costs. To increase predictive performance without largely increasing memory and time costs, this paper introduces a novel algorithm, named Online Local Boosting (OLBoost), which can be combined into online decision tree algorithms to improve their predictive performance without modifying the structure of the induced decision trees. For such, OLBoost applies a boosting to small separate regions of the instances space. Experimental results presented in this paper show that by using OLBoost the online learning decision tree algorithms can significantly improve their predictive performance. Additionally, it can make smaller trees perform as good or better than larger trees., To appear on the 8th Brazilian Conference on Intelligent Systems (BRACIS)

Dealing with memory and time constraints are current challenges when learning from data streams with a massive amount of data. Many algorithms have been proposed to handle these difficulties, among them, the Very Fast Decision Tree (VFDT) algorithm. Although the VFDT has been widely used in data stream mining, in the last years, several authors have suggested modifications to increase its performance, putting aside memory concerns by proposing memory-costly solutions. Besides, most data stream mining solutions have been centred around ensembles, which combine the memory costs of their weak learners, usually VFDTs. To reduce the memory cost, keeping the predictive performance, this study proposes the Strict VFDT (SVFDT), a novel algorithm based on the VFDT. The SVFDT algorithm minimises unnecessary tree growth, substantially reducing memory usage and keeping competitive predictive performance. Moreover, since it creates much more shallow trees than VFDT, SVFDT can achieve a shorter processing time. Experiments were carried out comparing the SVFDT with the VFDT in 11 benchmark data stream datasets. This comparison assessed the trade-off between accuracy, memory, and processing time. Statistical analysis showed that the proposed algorithm obtained similar predictive performance and significantly reduced processing time and memory use. Thus, SVFDT is a suitable option for data stream mining with memory and time limitations, recommended as a weak learner in ensemble-based solutions., Comment: 7 pages, 26 figures, Under R1 revision in Pattern Recognition Letters

Laser reshaping of cartilage is an emerging technology aimed at replacing conventional techniques for aesthetic and reconstructive surgery. Little is known about the mechanisms of wound healing following the photothermal heating during laser reshaping and, ultimately, how collagen remodels in the irradiated tissue. Healthy hyaline and elastic cartilage as found in the ear, nose, larynx, and trachea does not express collagen type I which is characteristic of fibro-cartilage and scar tissue. The aim of the study was to determine if collagen I and II gene expression occurs within laser irradiated rabbit septal cartilage.Nasal septum harvested from freshly euthanized New Zealand White rabbits were irradiated with an Nd:YAG laser. After 2 weeks in culture, the laser spot and surrounding non-irradiated regions were imaged using immunofluorescence staining and evaluated using reverse transcription polymerase chain reaction (RT-PCR) to determine the presence of collagen I and II, and ascertain collagen I and II gene expression, respectively.All laser irradiated specimens showed a cessation in collagen II gene expression within the center of the laser spot. Collagen II was expressed in the surrounding region encircling the laser spot and within the non-irradiated periphery in all specimens. Immunohistochemistry identified only type II collagen. Neither collagen I gene expression nor immunoreactivity were identified in any specimens regardless or irradiation parameters.Laser irradiation of rabbit septal cartilage using dosimetry parameters similar to those used in laser reshaping does not result in the detection of either collagen I gene expression or immunoreactivity. Only collagen type II was noted after laser exposure in vitro following cell culture, which suggests that the cellular response to laser irradiation is distinct from that observed in conventional wound healing. Laser irradiation of cartilage can leave an intact collagen matrix which likely allows chondrocyte recovery on an intact scaffold.

As technology evolves and electronic devices become widespread, the amount of data produced in the form of stream increases in enormous proportions. Data streams are an online source of data, meaning that it keeps producing data continuously. This creates the need for fast and reliable methods to analyse and extract information from these sources. Stream mining algorithms exist for this purpose, but the use of supervised machine learning is extremely limited in the stream domain since it is unfeasible to label every data instance requested to be processed. Tackling this problem, our paper proposes the use of active learning techniques for stream mining algorithms, specifically incremental Hoeffding trees-based. It is important to mention that the active learning techniques were implemented to match the stream mining constraints regarding low computational cost. We took advantage of the incremental tree original structure to avoid overburdening the original computational cost when selecting a label. In other words, the statistical strategy to grow each incremental tree has supported the execution of active learning. Using techniques of uncertainty sampling, we were able to drastically reduce the number of labels required at the cost of a very small reduction in accuracy. Particularly with Budget Entropy there was an average negative impact of accuracy about \(4\%\) using only \(14\%\) of samples labelled.

Mobile botnets are a growing threat to the internet security field. These botnets target less secure devices with lower computational power, while sometimes taking advantage of features specific to them, e.g., SMS messages. We propose a host-based approach using machine learning techniques to detect mobile botnets with features derived from system calls. Patterns created tend to be shared among applications with similar actions. Therefore, different botnets are likely to share similar system call patterns. To measure the effectiveness of our approach, a dataset containing multiple botnets and legitimate applications was created. We carried out three experiments, namely finding out the best time-window, and performing feature selection and hyperparameter tuning. A high performance (over 84%) was achieved in multiple metrics across multiple machine learning algorithms. An in-depth analysis of the features is also presented to help future work with a solid discussion about system call-based features.