Dissertations / Theses on the topic 'Apprentissage statistique profond'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 39 dissertations / theses for your research on the topic 'Apprentissage statistique profond.'
Next to every source in the list of references, there is an 'Add to bibliography' button. Press on it, and we will generate automatically the bibliographic reference to the chosen work in the citation style you need: APA, MLA, Harvard, Chicago, Vancouver, etc.
You can also download the full text of the academic publication as pdf and read online its abstract whenever available in the metadata.
Browse dissertations / theses on a wide variety of disciplines and organise your bibliography correctly.
Sors, Arnaud. "Apprentissage profond pour l'analyse de l'EEG continu." Thesis, Université Grenoble Alpes (ComUE), 2018. http://www.theses.fr/2018GREAS006/document.
Full textThe objective of this research is to explore and develop machine learning methods for the analysis of continuous electroencephalogram (EEG). Continuous EEG is an interesting modality for functional evaluation of cerebral state in the intensive care unit and beyond. Today its clinical use remains more limited that it could be because interpretation is still mostly performed visually by trained experts. In this work we develop automated analysis tools based on deep neural models.The subparts of this work hinge around post-anoxic coma prognostication, chosen as pilot application. A small number of long-duration records were performed and available existing data was gathered from CHU Grenoble. Different components of a semi-supervised architecture that addresses the application are imagined, developed, and validated on surrogate tasks.First, we validate the effectiveness of deep neural networks for EEG analysis from raw samples. For this we choose the supervised task of sleep stage classification from single-channel EEG. We use a convolutional neural network adapted for EEG and we train and evaluate the system on the SHHS (Sleep Heart Health Study) dataset. This constitutes the first neural sleep scoring system at this scale (5000 patients). Classification performance reaches or surpasses the state of the art.In real use for most clinical applications, the main challenge is the lack of (and difficulty of establishing) suitable annotations on patterns or short EEG segments. Available annotations are high-level (for example, clinical outcome) and therefore they are few. We search how to learn compact EEG representations in an unsupervised/semi-supervised manner. The field of unsupervised learning using deep neural networks is still young. To compare to existing work we start with image data and investigate the use of generative adversarial networks (GANs) for unsupervised adversarial representation learning. The quality and stability of different variants are evaluated. We then apply Gradient-penalized Wasserstein GANs on EEG sequences generation. The system is trained on single channel sequences from post-anoxic coma patients and is able to generate realistic synthetic sequences. We also explore and discuss original ideas for learning representations through matching distributions in the output space of representative networks.Finally, multichannel EEG signals have specificities that should be accounted for in characterization architectures. Each EEG sample is an instantaneous mixture of the activities of a number of sources. Based on this statement we propose an analysis system made of a spatial analysis subsystem followed by a temporal analysis subsystem. The spatial analysis subsystem is an extension of source separation methods built with a neural architecture with adaptive recombination weights, i.e. weights that are not learned but depend on features of the input. We show that this architecture learns to perform Independent Component Analysis if it is trained on a measure of non-gaussianity. For temporal analysis, standard (shared) convolutional neural networks applied on separate recomposed channels can be used
Moukari, Michel. "Estimation de profondeur à partir d'images monoculaires par apprentissage profond." Thesis, Normandie, 2019. http://www.theses.fr/2019NORMC211/document.
Full textComputer vision is a branch of artificial intelligence whose purpose is to enable a machine to analyze, process and understand the content of digital images. Scene understanding in particular is a major issue in computer vision. It goes through a semantic and structural characterization of the image, on one hand to describe its content and, on the other hand, to understand its geometry. However, while the real space is three-dimensional, the image representing it is two-dimensional. Part of the 3D information is thus lost during the process of image formation and it is therefore non trivial to describe the geometry of a scene from 2D images of it.There are several ways to retrieve the depth information lost in the image. In this thesis we are interested in estimating a depth map given a single image of the scene. In this case, the depth information corresponds, for each pixel, to the distance between the camera and the object represented in this pixel. The automatic estimation of a distance map of the scene from an image is indeed a critical algorithmic brick in a very large number of domains, in particular that of autonomous vehicles (obstacle detection, navigation aids).Although the problem of estimating depth from a single image is a difficult and inherently ill-posed problem, we know that humans can appreciate distances with one eye. This capacity is not innate but acquired and made possible mostly thanks to the identification of indices reflecting the prior knowledge of the surrounding objects. Moreover, we know that learning algorithms can extract these clues directly from images. We are particularly interested in statistical learning methods based on deep neural networks that have recently led to major breakthroughs in many fields and we are studying the case of the monocular depth estimation
Belilovsky, Eugene. "Apprentissage de graphes structuré et parcimonieux dans des données de haute dimension avec applications à l’imagerie cérébrale." Thesis, Université Paris-Saclay (ComUE), 2018. http://www.theses.fr/2018SACLC027.
Full textThis dissertation presents novel structured sparse learning methods on graphs that address commonly found problems in the analysis of neuroimaging data as well as other high dimensional data with few samples. The first part of the thesis proposes convex relaxations of discrete and combinatorial penalties involving sparsity and bounded total variation on a graph as well as bounded `2 norm. These are developed with the aim of learning an interpretable predictive linear model and we demonstrate their effectiveness on neuroimaging data as well as a sparse image recovery problem.The subsequent parts of the thesis considers structure discovery of undirected graphical models from few observational data. In particular we focus on invoking sparsity and other structured assumptions in Gaussian Graphical Models (GGMs). To this end we make two contributions. We show an approach to identify differences in Gaussian Graphical Models (GGMs) known to have similar structure. We derive the distribution of parameter differences under a joint penalty when parameters are known to be sparse in the difference. We then show how this approach can be used to obtain confidence intervals on edge differences in GGMs. We then introduce a novel learning based approach to the problem structure discovery of undirected graphical models from observational data. We demonstrate how neural networks can be used to learn effective estimators for this problem. This is empirically shown to be flexible and efficient alternatives to existing techniques
Delasalles, Edouard. "Inferring and Predicting Dynamic Representations for Structured Temporal Data." Electronic Thesis or Diss., Sorbonne université, 2020. http://www.theses.fr/2020SORUS296.
Full textTemporal data constitute a large part of data collected digitally. Predicting their next values is an important and challenging task in domains such as climatology, optimal control, or natural language processing. Standard statistical methods are based on linear models and are often limited to low dimensional data. We instead use deep learning methods capable of handling high dimensional structured data and leverage large quantities of examples. In this thesis, we are interested in latent variable models. Contrary to autoregressive models that directly use past data to perform prediction, latent models infer low dimensional vectorial representations of data on which prediction is performed. Latent vectorial spaces allow us to learn dynamic models that are able to generate high-dimensional and structured data. First, we propose a structured latent model for spatio-temporal data forecasting. Given a set of spatial locations where data such as weather or traffic are collected, we infer latent variables for each location and use spatial structure in the dynamic function. The model is also able to discover correlations between series without prior spatial information. Next, we focus on predicting data distributions, rather than point estimates. We propose a model that generates latent variables used to condition a generative model. Text data are used to evaluate the model on diachronic language modeling. Finally, we propose a stochastic prediction model. It uses the first values of sequences to generate several possible futures. Here, the generative model is not conditioned to an absolute epoch, but to a sequence. The model is applied to stochastic video prediction
Wolinski, Pierre. "Structural Learning of Neural Networks." Thesis, université Paris-Saclay, 2020. http://www.theses.fr/2020UPASS026.
Full textThe structure of a neural network determines to a large extent its cost of training and use, as well as its ability to learn. These two aspects are usually in competition: the larger a neural network is, the better it will perform the task assigned to it, but the more it will require memory and computing time resources for training. Automating the search of efficient network structures -of reasonable size and performing well- is then a very studied question in this area. Within this context, neural networks with various structures are trained, which requires a new set of training hyperparameters for each new structure tested. The aim of the thesis is to address different aspects of this problem. The first contribution is a training method that operates within a large perimeter of network structures and tasks, without needing to adjust the learning rate. The second contribution is a network training and pruning technique, designed to be insensitive to the initial width of the network. The last contribution is mainly a theorem that makes possible to translate an empirical training penalty into a Bayesian prior, theoretically well founded. This work results from a search for properties that theoretically must be verified by training and pruning algorithms to be valid over a wide range of neural networks and objectives
Malfante, Marielle. "Automatic classification of natural signals for environmental monitoring." Thesis, Université Grenoble Alpes (ComUE), 2018. http://www.theses.fr/2018GREAU025/document.
Full textThis manuscript summarizes a three years work addressing the use of machine learning for the automatic analysis of natural signals. The main goal of this PhD is to produce efficient and operative frameworks for the analysis of environmental signals, in order to gather knowledge and better understand the considered environment. Particularly, we focus on the automatic tasks of detection and classification of natural events.This thesis proposes two tools based on supervised machine learning (Support Vector Machine, Random Forest) for (i) the automatic classification of events and (ii) the automatic detection and classification of events. The success of the proposed approaches lies in the feature space used to represent the signals. This relies on a detailed description of the raw acquisitions in various domains: temporal, spectral and cepstral. A comparison with features extracted using convolutional neural networks (deep learning) is also made, and favours the physical features to the use of deep learning methods to represent transient signals.The proposed tools are tested and validated on real world acquisitions from different environments: (i) underwater and (ii) volcanic areas. The first application considered in this thesis is devoted to the monitoring of coastal underwater areas using acoustic signals: continuous recordings are analysed to automatically detect and classify fish sounds. A day to day pattern in the fish behaviour is revealed. The second application targets volcanoes monitoring: the proposed system classifies seismic events into categories, which can be associated to different phases of the internal activity of volcanoes. The study is conducted on six years of volcano-seismic data recorded on Ubinas volcano (Peru). In particular, the outcomes of the proposed automatic classification system helped in the discovery of misclassifications in the manual annotation of the recordings. In addition, the proposed automatic classification framework of volcano-seismic signals has been deployed and tested in Indonesia for the monitoring of Mount Merapi. The software implementation of the framework developed in this thesis has been collected in the Automatic Analysis Architecture (AAA) package and is freely available
Baudry, Maximilien. "Quelques problèmes d’apprentissage statistique en présence de données incomplètes." Thesis, Lyon, 2020. http://www.theses.fr/2020LYSE1002.
Full textMost statistical methods are not designed to directly work with incomplete data. The study of data incompleteness is not new and strong methods have been established to handle it prior to a statistical analysis. On the other hand, deep learning literature mainly works with unstructured data such as images, text or raw audio, but very few has been done on tabular data. Hence, modern machine learning literature tackling data incompleteness on tabular data is scarce. This thesis focuses on the use of machine learning models applied to incomplete tabular data, in an insurance context. We propose through our contributions some ways to model complex phenomena in presence of incompleteness schemes, and show that our approaches outperform the state-of-the-art models
Chen, Mickaël. "Learning with weak supervision using deep generative networks." Electronic Thesis or Diss., Sorbonne université, 2020. http://www.theses.fr/2020SORUS024.
Full textMany successes of deep learning rely on the availability of massive annotated datasets that can be exploited by supervised algorithms. Obtaining those labels at a large scale, however, can be difficult, or even impossible in many situations. Designing methods that are less dependent on annotations is therefore a major research topic, and many semi-supervised and weakly supervised methods have been proposed. Meanwhile, the recent introduction of deep generative networks provided deep learning methods with the ability to manipulate complex distributions, allowing for breakthroughs in tasks such as image edition and domain adaptation. In this thesis, we explore how these new tools can be useful to further alleviate the need for annotations. Firstly, we tackle the task of performing stochastic predictions. It consists in designing systems for structured prediction that take into account the variability in possible outputs. We propose, in this context, two models. The first one performs predictions on multi-view data with missing views, and the second one predicts possible futures of a video sequence. Then, we study adversarial methods to learn a factorized latent space, in a setting with two explanatory factors but only one of them is annotated. We propose models that aim to uncover semantically consistent latent representations for those factors. One model is applied to the conditional generation of motion capture data, and another one to multi-view data. Finally, we focus on the task of image segmentation, which is of crucial importance in computer vision. Building on previously explored ideas, we propose a model for object segmentation that is entirely unsupervised
Novello, Paul. "Combining supervised deep learning and scientific computing : some contributions and application to computational fluid dynamics." Thesis, Institut polytechnique de Paris, 2022. http://www.theses.fr/2022IPPAX005.
Full textRecent innovations in mathematics, computer science, and engineering have enabled more and more sophisticated numerical simulations. However, some simulations remain computationally unaffordable, even for the most powerful supercomputers. Lately, machine learning has proven its ability to improve the state-of-the-art in many fields, notoriously computer vision, language understanding, or robotics. This thesis settles in the high-stakes emerging field of Scientific Machine Learning which studies the application of machine learning to scientific computing. More specifically, we consider the use of deep learning to accelerate numerical simulations.We focus on approximating some components of Partial Differential Equation (PDE) based simulation software by a neural network. This idea boils down to constructing a data set, selecting and training a neural network, and embedding it into the original code, resulting in a hybrid numerical simulation. Although this approach may seem trivial at first glance, the context of numerical simulations comes with several challenges. Since we aim at accelerating codes, the first challenge is to find a trade-off between neural networks’ accuracy and execution time. The second challenge stems from the data-driven process of the training, and more specifically, its lack of mathematical guarantees. Hence, we have to ensure that the hybrid simulation software still yields reliable predictions. To tackle these challenges, we thoroughly study each step of the deep learning methodology while considering the aforementioned constraints. By doing so, we emphasize interplays between numerical simulations and machine learning that can benefit each of these fields.We identify the main steps of the deep learning methodology as the construction of the training data set, the choice of the hyperparameters of the neural network, and its training. For the first step, we leverage the ability to sample training data with the original software to characterize a more efficient training distribution based on the local variation of the function to approximate. We generalize this approach to general machine learning problems by deriving a data weighting methodology called Variance Based Sample Weighting. For the second step, we introduce the use of sensitivity analysis, an approach widely used in scientific computing, to tackle neural network hyperparameter optimization. This approach is based on qualitatively assessing the effect of hyperparameters on the performances of a neural network using Hilbert-Schmidt Independence Criterion. We adapt it to the hyperparameter optimization context and build an interpretable methodology that yields competitive and cost-effective networks. For the third step, we formally define an analogy between the stochastic resolution of PDEs and the optimization process at play when training a neural network. This analogy leads to a PDE-based framework for training neural networks that opens up many possibilities for improving existing optimization algorithms. Finally, we apply these contributions to a computational fluid dynamics simulation coupled with a multi-species chemical equilibrium code. We demonstrate that we can achieve a time factor acceleration of 21 with controlled to no degradation from the initial prediction
Rossi, Simone. "Improving Scalability and Inference in Probabilistic Deep Models." Electronic Thesis or Diss., Sorbonne université, 2022. http://www.theses.fr/2022SORUS042.
Full textThroughout the last decade, deep learning has reached a sufficient level of maturity to become the preferred choice to solve machine learning-related problems or to aid decision making processes.At the same time, deep learning is generally not equipped with the ability to accurately quantify the uncertainty of its predictions, thus making these models less suitable for risk-critical applications.A possible solution to address this problem is to employ a Bayesian formulation; however, while this offers an elegant treatment, it is analytically intractable and it requires approximations.Despite the huge advancements in the last few years, there is still a long way to make these approaches widely applicable.In this thesis, we address some of the challenges for modern Bayesian deep learning, by proposing and studying solutions to improve scalability and inference of these models.The first part of the thesis is dedicated to deep models where inference is carried out using variational inference (VI).Specifically, we study the role of initialization of the variational parameters and we show how careful initialization strategies can make VI deliver good performance even in large scale models.In this part of the thesis we also study the over-regularization effect of the variational objective on over-parametrized models.To tackle this problem, we propose an novel parameterization based on the Walsh-Hadamard transform; not only this solves the over-regularization effect of VI but it also allows us to model non-factorized posteriors while keeping time and space complexity under control.The second part of the thesis is dedicated to a study on the role of priors.While being an essential building block of Bayes' rule, picking good priors for deep learning models is generally hard.For this reason, we propose two different strategies based (i) on the functional interpretation of neural networks and (ii) on a scalable procedure to perform model selection on the prior hyper-parameters, akin to maximization of the marginal likelihood.To conclude this part, we analyze a different kind of Bayesian model (Gaussian process) and we study the effect of placing a prior on all the hyper-parameters of these models, including the additional variables required by the inducing-point approximations.We also show how it is possible to infer free-form posteriors on these variables, which conventionally would have been otherwise point-estimated
Caucheteux, Charlotte. "Language representations in deep learning algorithms and the brain." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG031.
Full textRecent deep language models -- like GPT-3 and ChatGPT -- are capable to produce text that closely resembles that of humans. Such similarity raises questions about how the brain and deep models process language, the mechanisms they use, and the internal representations they construct. In this thesis, I compare the internal representations of the brain and deep language models, with the goal of identifying their similarities and differences. To this aim, I analyze functional resonance imaging (fMRI) and magnetoencephalography (MEG) recordings of participants listening to and reading sentences, and compare them to the activations of thousands of language algorithms corresponding to these same sentences.Our results first highlight high-level similarities between the internal representations of the brain and deep language models. We find that deep nets' activations significantly predict brain activity across subjects for different cohorts (>500 participants), recording modalities (MEG and fMRI), stimulus types (isolated words, sentences, and natural stories), stimulus modalities (auditory and visual presentation), languages (Dutch, English and French), and deep language models. This alignment is maximal in brain regions repeatedly associated with language, for the best-performing algorithms and for participants who best understand the stories. Critically, we evidence a similar processing hierarchy between the two systems. The first layers of the algorithms align with low-level processing regions in the brain, such as auditory areas and the temporal lobe, while the deep layers align with regions associated with higher-level processing, such fronto-parietal areas.We then show how such similarities can be leveraged to build better predictive models of brain activity and better decompose several linguistic processes in the brain, such as syntax and semantics. Finally, we explore the differences between deep language models and the brain's activations. We find that the brain predicts distant and hierarchical representations, unlike current language models that are mostly trained to make short-term and word-level predictions. Overall, modern algorithms are still far from processing language in the same way that humans do. However, the direct links between their inner workings and that of the brain provide an promising platform for better understanding both systems, and pave the way for building better algorithms inspired by the human brain
Mohammad, Noshine. "Exploration des modèles d’apprentissage statistique profonds couplés à la spectrométrie de masse pour améliorer la surveillance épidémiologique des maladies infectieuses." Electronic Thesis or Diss., Sorbonne université, 2023. http://www.theses.fr/2023SORUS617.
Full textMALDI-TOF (matrix assisted laser desorption and ionisation time of flight) mass spectrometry is a rapid and robust diagnostic method for microbiology, enabling microorganism species to be identified on the basis of their protein fingerprint in the mass spectrum. However, the clinical and epidemiological applications of this technology remain limited by the bioinformatics tools available. This thesis focuses on the application of deep statistical learning models to MALDI-TOF mass spectrometry data for the purpose of epidemiological surveillance of infectious diseases. This includes the monitoring of fungal and mycobacterial epidemics in hospitals, as well as the characterisation of Anopheles vectors of malaria.We examined the impact of sample preparation methods and computer analysis of mass spectra on improving learning, in order to identify epidemic fungal clones in hospitals and prevent their spread. Our study showed that the convolution neural network (CNN) has a high potential for identifying the spectra of specific Candida parapsilosis clones, achieving 94% accuracy by optimising essential parameters (culture media, growth time, and the spectra acquisition machine). To detect epidemic Aspergillus flavus clones in multicentre hospital cohorts, the CNN was also able to classify most isolates correctly, achieving accuracy of over 93% for two of the three instruments used. We have also shown that by using optimised deep learning models, such as a CNN and a temporal convolution neural network (TCN), we can predict the age of mosquitoes with an average accuracy of two days (best mean absolute error: 1.74 days). This approach will enable us to effectively monitor the age structure of wild Anopheles mosquito populations and target them more effectively with control measures. Finally, we demonstrated the performance of various neural network architectures and mass spectra representation methods, using different cohorts covering various epidemiological issues such as age prediction, identification of closely related species of Anopheles mosquitoes, distinction between closely related subspecies, and detection of resistance in Mycobacterium abscessus. The study showed that of the different models evaluated, the best performing models, such as TCNs and a recurrent neural network, were able to achieve notable results, reaching an identification accuracy of 93% for closely related Anopheles species and 95% for Mycobacterium abscessus subspecies. In addition, the use of CNN and TCN enabled the detection of resistant strains in Mycobacterium abscessus with an accuracy of over 97%. This thesis highlights the use of deep learning in conjunction with MALDI-TOF, a hitherto little explored approach. With the widespread availability of MALDI-TOF instruments and the possibility of coupling analyses to online applications using deep learning, this approach looks promising, opening the way to other epidemiological applications beyond simple species identification, such as detecting epidemiological clusters of drug-resistant microorganisms, monitoring the transmission of bacterial and fungal diseases, and evaluating the effectiveness of targeted vector control interventions
Cutajar, Kurt. "Broadening the scope of gaussian processes for large-scale learning." Electronic Thesis or Diss., Sorbonne université, 2019. http://www.theses.fr/2019SORUS063.
Full textThe renewed importance of decision making under uncertainty calls for a re-evaluation of Bayesian inference techniques targeting this goal in the big data regime. Gaussian processes (GPs) are a fundamental building block of many probabilistic kernel machines; however, the computational and storage complexity of GPs hinders their scaling to large modern datasets. The contributions presented in this thesis are two-fold. We first investigate the effectiveness of exact GP inference on a computational budget by proposing a novel scheme for accelerating regression and classification by way of preconditioning. In the spirit of probabilistic numerics, we also show how the numerical uncertainty introduced by approximate linear algebra should be adequately evaluated and incorporated. Bridging the gap between GPs and deep learning techniques remains a pertinent research goal, and the second broad contribution of this thesis is to establish and reinforce the role of GPs, and their deep counterparts (DGPs), in this setting. Whereas GPs and DGPs were once deemed unfit to compete with alternative state-of-the-art methods, we demonstrate how such models can also be adapted to the large-scale and complex tasks to which machine learning is now being applied
Pavão, Adrien. "Methodology for Design and Analysis of Machine Learning Competitions." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG088.
Full textWe develop and study a systematic and unified methodology to organize and use scientific challenges in research, particularly in the domain of machine learning (data-driven artificial intelligence). As of today, challenges are becoming more and more popular as a pedagogic tool and as a means of pushing the state-of-the-art by engaging scientists of all ages, within or outside academia. This can be thought of as a form of citizen science. There is the promise that this form of community involvement in science might contribute to reproducible research and democratize artificial intelligence. However, while the distinction between organizers and participants may mitigate certain biases, there exists a risk that biases in data selection, scoring metrics, and other experimental design elements could compromise the integrity of the outcomes and amplify the influence of randomness. In extreme cases, the results could range from being useless to detrimental for the scientific community and, ultimately, society at large. Our objective is to structure challenge organization within a rigorous framework and offer the community insightful guidelines. In conjunction with the tools of challenge organization that we are developing as part of the CodaLab project, we aim to provide a valuable contribution to the community. This thesis includes theoretical fundamental contributions drawing on experimental design, statistics and game theory, and practical empirical findings resulting from the analysis of data from previous challenges
Barreau, Baptiste. "Machine Learning for Financial Products Recommendation." Thesis, université Paris-Saclay, 2020. http://www.theses.fr/2020UPAST010.
Full textAnticipating clients’ needs is crucial to any business — this is particularly true for corporate and institutional banks such as BNP Paribas Corporate and Institutional Banking due to their role in the financial markets. This thesis addresses the problem of future interests prediction in the financial context and focuses on the development of ad hoc algorithms designed for solving specific financial challenges.This manuscript is composed of five chapters:- Chapter 1 introduces the problem of future interests prediction in the financial world. The goal of this chapter is to provide the reader with all the keys necessary to understand the remainder of this thesis. These keys are divided into three parts: a presentation of the datasets we have at our disposal to solve the future interests prediction problem and their characteristics, an overview of the candidate algorithms to solve this problem, and the development of metrics to monitor the performance of these algorithms on our datasets. This chapter finishes with some of the challenges that we face when designing algorithms to solve the future interests problem in finance, challenges that will be partly addressed in the following chapters;- Chapter 2 proposes a benchmark of some of the algorithms introduced in Chapter 1 on a real-word dataset from BNP Paribas CIB, along with a development on the difficulties encountered for comparing different algorithmic approaches on a same dataset and on ways to tackle them. This benchmark puts in practice classic recommendation algorithms that were considered on a theoretical point of view in the preceding chapter, and provides further intuition on the analysis of the metrics introduced in Chapter 1;- Chapter 3 introduces a new algorithm, called Experts Network, that is designed to solve the problem of behavioral heterogeneity of investors on a given financial market using a custom-built neural network architecture inspired from mixture-of-experts research. In this chapter, the introduced methodology is experimented on three datasets: a synthetic dataset, an open-source one and a real-world dataset from BNP Paribas CIB. The chapter provides further insights into the development of the methodology and ways to extend it;- Chapter 4 also introduces a new algorithm, called History-augmented Collaborative Filtering, that proposes to augment classic matrix factorization approaches with the information of users and items’ interaction histories. This chapter provides further experiments on the dataset used in Chapter 2, and extends the presented methodology with various ideas. Notably, this chapter exposes an adaptation of the methodology to solve the cold-start problem and applies it to a new dataset;- Chapter 5 brings to light a collection of ideas and algorithms, successful or not, that were experimented during the development of this thesis. This chapter finishes on a new algorithm that blends the methodologies introduced in Chapters 3 and 4
Darmet, Ludovic. "Vers une approche basée modèle-image flexible et adaptative en criminalistique des images." Thesis, Université Grenoble Alpes, 2020. https://tel.archives-ouvertes.fr/tel-03086427.
Full textImages are nowadays a standard and mature medium of communication.They appear in our day to day life and therefore they are subject to concernsabout security. In this work, we study different methods to assess theintegrity of images. Because of a context of high volume and versatilityof tampering techniques and image sources, our work is driven by the necessity to developflexible methods to adapt the diversity of images.We first focus on manipulations detection through statistical modeling ofthe images. Manipulations are elementary operations such as blurring,noise addition, or compression. In this context, we are more preciselyinterested in the effects of pre-processing. Because of storagelimitation or other reasons, images can be resized or compressed justafter their capture. Addition of a manipulation would then be applied on analready pre-processed image. We show that a pre-resizing of test datainduces a drop of performance for detectors trained on full-sized images.Based on these observations, we introduce two methods to counterbalancethis performance loss for a pipeline of classification based onGaussian Mixture Models. This pipeline models the local statistics, onpatches, of natural images. It allows us to propose adaptation of themodels driven by the changes in local statistics. Our first method ofadaptation is fully unsupervised while the second one, only requiring a fewlabels, is weakly supervised. Thus, our methods are flexible to adaptversatility of source of images.Then we move to falsification detection and more precisely to copy-moveidentification. Copy-move is one of the most common image tampering technique. Asource area is copied into a target area within the same image. The vastmajority of existing detectors identify indifferently the two zones(source and target). In an operational scenario, only the target arearepresents a tampering area and is thus an area of interest. Accordingly, wepropose a method to disentangle the two zones. Our method takesadvantage of local modeling of statistics in natural images withGaussian Mixture Model. The procedure is specific for each image toavoid the necessity of using a large training dataset and to increase flexibility.Results for all the techniques described above are illustrated on publicbenchmarks and compared to state of the art methods. We show that theclassical pipeline for manipulations detection with Gaussian MixtureModel and adaptation procedure can surpass results of fine-tuned andrecent deep-learning methods. Our method for source/target disentanglingin copy-move also matches or even surpasses performances of the latestdeep-learning methods. We explain the good results of these classicalmethods against deep-learning by their additional flexibility andadaptation abilities.Finally, this thesis has occurred in the special context of a contestjointly organized by the French National Research Agency and theGeneral Directorate of Armament. We describe in the Appendix thedifferent stages of the contest and the methods we have developed, as well asthe lessons we have learned from this experience to move the image forensics domain into the wild
Chéron, Guilhem. "Structured modeling and recognition of human actions in video." Thesis, Paris Sciences et Lettres (ComUE), 2018. http://www.theses.fr/2018PSLEE058.
Full textAutomatic video understanding is expected to impact our lives through many applications such as autonomous driving, domestic robots, content search and filtering, gaming, defense or security. Video content is growing faster each year, for example on platforms such as YouTube, Twitter or Facebook. Automatic analysis of this data is required to enable future applications. Video analysis, especially in uncontrolled environments, presents several difficulties such as intraclass variability (samples from the same concept appear very differently) or inter-class confusion (examples from two different activities look similar). While these problems can be addressed with the supervised learning algorithms, fully-supervised methods are often associated with high annotation cost. Depending on both the task and the level of required supervision, the annotation can be prohibitive. For example, in action localization, a fully-supervised approach demands person bounding boxes to be annotated at every frames where an activity is performed. The cost of getting such annotation prohibits scalability and limits the number of training samples. Another issue is finding a consensus between annotators, which leads to labeling ambiguities (where does the action start? where does it end? what should be included in the bounding box? etc.). This thesis addresses above problems in the context of two tasks, namely human action classification and localization. The former aims at recognizing the type of activity performed in a short video clip trimmed to the temporal extent of the action. The latter additionally extracts the space-time locations of potentially multiple activities in much longer videos. Our approach to action classification leverages information from human pose and integrates it with appearance and motion descriptors for improved performance. Our approach to action localization models the temporal evolution of actions in the video with a recurrent network trained on the level of person tracks. Finally, the third method in this thesis aims to avoid a prohibitive cost of video annotation and adopts discriminative clustering to analyze and combine different levels of supervision
Ghrissi, Amina. "Ablation par catheter de fibrillation atriale persistante guidée par dispersion spatiotemporelle d’électrogrammes : Identification automatique basée sur l’apprentissage statistique." Thesis, Université Côte d'Azur, 2021. http://www.theses.fr/2021COAZ4026.
Full textCatheter ablation is increasingly used to treat atrial fibrillation (AF), the most common sustained cardiac arrhythmia encountered in clinical practice. A recent patient-tailored AF ablation therapy, giving 95% of procedural success rate, is based on the use of a multipolar mapping catheter called PentaRay. It targets areas of spatiotemporal dispersion (STD) in the atria as potential AF drivers. STD stands for a delay of the cardiac activation observed in intracardiac electrograms (EGMs) across contiguous leads.In practice, interventional cardiologists localize STD sites visually using the PentaRay multipolar mapping catheter. This thesis aims to automatically characterize and identify ablation sites in STD-based ablation of persistent AF using machine learning (ML) including deep learning (DL) techniques. In the first part, EGM recordings are classified into STD vs. non-STD groups. However, highly imbalanced dataset ratio hampers the classification performance. We tackle this issue by using adapted data augmentation techniques that help achieve good classification. The overall performance is high with values of accuracy and AUC around 90%. First, two approaches are benchmarked, feature engineering and automatic feature extraction from a time series, called maximal voltage absolute values at any of the bipoles (VAVp). Statistical features are extracted and fed to ML classifiers but no important dissimilarity is obtained between STD and non-STD categories. Results show that the supervised classification of raw VAVp time series itself into the same categories is promising with values of accuracy, AUC, sensi-tivity and specificity around 90%. Second, the classification of raw multichannel EGM recordings is performed. Shallow convolutional arithmetic circuits are investigated for their promising theoretical interest but experimental results on synthetic data are unsuccessful. Then, we move forward to more conventional supervised ML tools. We design a selection of data representations adapted to different ML and DL models, and benchmark their performance in terms of classification and computational cost. Transfer learning is also assessed. The best performance is achieved with a convolutional neural network (CNN) model for classifying raw EGM matrices. The average performance over cross-validation reaches 94% of accuracy and AUC added to an F1-score of 60%. In the second part, EGM recordings acquired during mapping are labeled ablated vs. non-ablated according to their proximity to the ablation sites then classified into the same categories. STD labels, previously defined by interventional cardiologists at the ablation procedure, are also aggregated as a prior probability in the classification task.Classification results on the test set show that a shallow CNN gives the best performance with an F1-score of 76%. Aggregating STD label does not help improve the model’s performance. Overall, this work is among the first attempts at the application of statistical analysis and ML tools to automatically identify successful ablation areas in STD-based ablation. By providing interventional cardiologists with a real-time objective measure of STD, the proposed solution offers the potential to improve the efficiency and effectiveness of this fully patient-tailored catheter ablation approach for treating persistent AF
Chesneau, Nicolas. "Learning to Recognize Actions with Weak Supervision." Thesis, Université Grenoble Alpes (ComUE), 2018. http://www.theses.fr/2018GREAM007/document.
Full textWith the rapid growth of digital video content, automaticvideo understanding has become an increasingly important task. Video understanding spansseveral applications such as web-video content analysis, autonomous vehicles, human-machine interfaces (eg, Kinect). This thesismakes contributions addressing two major problems in video understanding:webly-supervised action detection and human action localization.Webly-supervised action recognition aims to learn actions from video content on the internet, with no additional supervision. We propose a novel approach in this context, which leverages thesynergy between visual video data and the associated textual metadata, to learnevent classifiers with no manual annotations. Specifically, we first collect avideo dataset with queries constructed automatically from textual descriptionof events, prune irrelevant videos with text and video data, and then learn thecorresponding event classifiers. We show the importance of both the main steps of our method, ie,query generation and data pruning, with quantitative results. We evaluate this approach in the challengingsetting where no manually annotated training set is available, i.e., EK0 in theTrecVid challenge, and show state-of-the-art results on MED 2011 and 2013datasets.In the second part of the thesis, we focus on human action localization, which involves recognizing actions that occur in a video, such as ``drinking'' or ``phoning'', as well as their spatial andtemporal extent. We propose a new person-centric framework for action localization that trackspeople in videos and extracts full-body human tubes, i.e., spatio-temporalregions localizing actions, even in the case of occlusions or truncations.The motivation is two-fold. First, it allows us to handle occlusions and camera viewpoint changes when localizing people, as it infers full-body localization. Second, it provides a better reference grid for extracting action information than standard human tubes, ie, tubes which frame visible parts only.This is achieved by training a novel human part detector that scores visibleparts while regressing full-body bounding boxes, even when they lie outside the frame. The core of our method is aconvolutional neural network which learns part proposals specific to certainbody parts. These are then combined to detect people robustly in each frame.Our tracking algorithm connects the image detections temporally to extractfull-body human tubes. We evaluate our new tube extraction method on a recentchallenging dataset, DALY, showing state-of-the-art results
Boonkongkird, Chotipan. "Deep learning for Lyman-alpha based cosmology." Electronic Thesis or Diss., Sorbonne université, 2023. https://accesdistant.sorbonne-universite.fr/login?url=https://theses-intra.sorbonne-universite.fr/2023SORUS733.pdf.
Full textAs cosmological surveys advance and become more sophisticated, they provide data with increasing resolution and volume. The Lyman-α forest has emerged as a powerful probe to study the intergalactic medium (IGM) properties up to a very high redshift. Analysing this extensive data requires advanced hydrodynamical simulations capable of resolving the observational data, which demands robust hardware and a considerable amount of computational time. Recent developments in machine learning, particularly neural networks, offer potential solutions. With their ability to function as universal fitting mechanisms, neural networks are gaining traction in various disciplines, including astrophysics and cosmology. In this doctoral thesis, we explore a machine learning framework, specifically, an artificial neural network to emulate hydrodynamical simulations from N-body simulations of dark matter. The core principle of this work is based on the fluctuating Gunn-Peterson approximation (FGPA), a framework commonly used to emulate the Lyman-α forest from dark matter. While useful for physical understanding, the FGPA misses to properly predict the absorption by neglecting non-locality in the construction of the IGM. Instead, our method includes the diversity of the IGM while being interpretable, which does not exclusively benefit the Lyman-α forest and extends to other applications. It also provides a more efficient solution to generate simulations, significantly reducing time compared to standard hydrodynamical simulations. We also test its resilience and explore the potential of using this framework to generalise to various astrophysical hypotheses of the IGM physics using a transfer learning method. We discuss how the results relate to other existing methods. Finally, the Lyman-α simulator typically constructs the observational volume using a single timestep of the cosmological simulations. This implies an identical astrophysical environment everywhere, which does not reflect the real universe. We explore and experiment to go beyond this limitation with our emulator, accounting for variable baryonic effects along the line of sight. While this is still preliminary, it could become a framework for constructing consistent light-cones. We apply neural networks to interpolate astrophysical feedback across different cells in simulations to provide mock observables more realistic to the real universe, which would allow us to understand the nature of IGM better and to constrain the ΛCDM model
Rohé, Marc-Michel. "Représentation réduite de la segmentation et du suivi des images cardiaques pour l’analyse longitudinale de groupe." Thesis, Université Côte d'Azur (ComUE), 2017. http://www.theses.fr/2017AZUR4051/document.
Full textThis thesis presents image-based methods for the analysis of cardiac motion to enable group-wise statistics, automatic diagnosis and longitudinal study. This is achieved by combining advanced medical image processing with machine learning methods and statistical modelling. The first axis of this work is to define an automatic method for the segmentation of the myocardium. We develop a very-fast registration method based on convolutional neural networks that is trained to learn inter-subject heart registration. Then, we embed this registration method into a multi-atlas segmentation pipeline. The second axis of this work is focused on the improvement of cardiac motion tracking methods in order to define relevant low-dimensional representations. Two different methods are developed, one relying on Barycentric Subspaces built on ref- erences frames of the sequence, and another based on a reduced order representation of the motion from polyaffine transformations. Finally, in the last axis, we apply the previously defined representation to the problem of diagnosis and longitudinal analysis. We show that these representations encode relevant features allowing the diagnosis of infarcted patients and Tetralogy of Fallot versus controls and the analysis of the evolution through time of the cardiac motion of patients with either cardiomyopathies or obesity. These three axes form an end to end framework for the study of cardiac motion starting from the acquisition of the medical images to their automatic analysis. Such a framework could be used for diagonis and therapy planning in order to improve the clinical decision making with a more personalised computer-aided medicine
Sanchez, Théophile. "Reconstructing our past ˸ deep learning for population genetics." Electronic Thesis or Diss., université Paris-Saclay, 2022. http://www.theses.fr/2022UPASG032.
Full textConstant improvement of DNA sequencing technology that produces large quantities of genetic data should greatly enhance our knowledge of evolution, particularly demographic history. However, the best way to extract information from this large-scale data is still an open problem. Neural networks are a strong candidate to attain this goal, considering their recent success in machine learning. These methods have the advantages of handling high-dimensional data, adapting to most applications and scaling efficiently to available computing resources. However, their performance dependents on their architecture, which should match the data properties to extract the maximum information. In this context, this thesis presents new approaches based on deep learning, as well as the principles for designing architectures adapted to the characteristics of genomic data. The use of convolution layers and attention mechanisms allows the presented networks to be invariant to the sampled haplotypes' permutations and to adapt to data of different dimensions (number of haplotypes and polymorphism sites). Experiments conducted on simulated data demonstrate the efficiency of these approaches by comparing them to more classical network architectures, as well as to state-of-the-art methods. Moreover, coupling neural networks with some methods already proven in population genetics, such as the approximate Bayesian computation, improves the results and combines their advantages. The practicality of neural networks for demographic inference is tested on whole genome sequence data from real populations of Bos taurus and Homo sapiens. Finally, the scenarios obtained are compared with current knowledge of the demographic history of these populations
Suzano, Massa Francisco Vitor. "Mise en relation d'images et de modèles 3D avec des réseaux de neurones convolutifs." Thesis, Paris Est, 2017. http://www.theses.fr/2017PESC1198/document.
Full textThe recent availability of large catalogs of 3D models enables new possibilities for a 3D reasoning on photographs. This thesis investigates the use of convolutional neural networks (CNNs) for relating 3D objects to 2D images.We first introduce two contributions that are used throughout this thesis: an automatic memory reduction library for deep CNNs, and a study of CNN features for cross-domain matching. In the first one, we develop a library built on top of Torch7 which automatically reduces up to 91% of the memory requirements for deploying a deep CNN. As a second point, we study the effectiveness of various CNN features extracted from a pre-trained network in the case of images from different modalities (real or synthetic images). We show that despite the large cross-domain difference between rendered views and photographs, it is possible to use some of these features for instance retrieval, with possible applications to image-based rendering.There has been a recent use of CNNs for the task of object viewpoint estimation, sometimes with very different design choices. We present these approaches in an unified framework and we analyse the key factors that affect performance. We propose a joint training method that combines both detection and viewpoint estimation, which performs better than considering the viewpoint estimation separately. We also study the impact of the formulation of viewpoint estimation either as a discrete or a continuous task, we quantify the benefits of deeper architectures and we demonstrate that using synthetic data is beneficial. With all these elements combined, we improve over previous state-of-the-art results on the Pascal3D+ dataset by a approximately 5% of mean average viewpoint precision.In the instance retrieval study, the image of the object is given and the goal is to identify among a number of 3D models which object it is. We extend this work to object detection, where instead we are given a 3D model (or a set of 3D models) and we are asked to locate and align the model in the image. We show that simply using CNN features are not enough for this task, and we propose to learn a transformation that brings the features from the real images close to the features from the rendered views. We evaluate our approach both qualitatively and quantitatively on two standard datasets: the IKEAobject dataset, and a subset of the Pascal VOC 2012 dataset of the chair category, and we show state-of-the-art results on both of them
Kozyrskiy, Bogdan. "Exploring the Intersection of Bayesian Deep Learning and Gaussian Processes." Electronic Thesis or Diss., Sorbonne université, 2023. https://accesdistant.sorbonne-universite.fr/login?url=https://theses-intra.sorbonne-universite.fr/2023SORUS064archi.pdf.
Full textDeep learning played a significant role in establishing machine learning as a must-have instrument in multiple areas. The use of deep learning poses several challenges. Deep learning requires a lot of computational power for training and applying models. Another problem with deep learning is its inability to estimate the uncertainty of the predictions, which creates obstacles in risk-sensitive applications. This thesis presents four projects to address these problems: We propose an approach making use of Optical Processing Units to reduce energy consumption and speed up the inference of deep models. We address the problem of uncertainty estimates for classification with Bayesian inference. We introduce techniques for deep models that decreases the cost of Bayesian inference. We developed a novel framework to accelerate Gaussian Process regression. We propose a technique to impose meaningful functional priors for deep models through Gaussian Processes
Maignant, Elodie. "Plongements barycentriques pour l'apprentissage géométrique de variétés : application aux formes et graphes." Electronic Thesis or Diss., Université Côte d'Azur, 2023. http://www.theses.fr/2023COAZ4096.
Full textAn MRI image has over 60,000 pixels. The largest known human protein consists of around 30,000 amino acids. We call such data high-dimensional. In practice, most high-dimensional data is high-dimensional only artificially. For example, of all the images that could be randomly generated by coloring 256 x 256 pixels, only a very small subset would resemble an MRI image of a human brain. This is known as the intrinsic dimension of such data. Therefore, learning high-dimensional data is often synonymous with dimensionality reduction. There are numerous methods for reducing the dimension of a dataset, the most recent of which can be classified according to two approaches.A first approach known as manifold learning or non-linear dimensionality reduction is based on the observation that some of the physical laws behind the data we observe are non-linear. In this case, trying to explain the intrinsic dimension of a dataset with a linear model is sometimes unrealistic. Instead, manifold learning methods assume a locally linear model.Moreover, with the emergence of statistical shape analysis, there has been a growing awareness that many types of data are naturally invariant to certain symmetries (rotations, reparametrizations, permutations...). Such properties are directly mirrored in the intrinsic dimension of such data. These invariances cannot be faithfully transcribed by Euclidean geometry. There is therefore a growing interest in modeling such data using finer structures such as Riemannian manifolds. A second recent approach to dimension reduction consists then in generalizing existing methods to non-Euclidean data. This is known as geometric learning.In order to combine both geometric learning and manifold learning, we investigated the method called locally linear embedding, which has the specificity of being based on the notion of barycenter, a notion a priori defined in Euclidean spaces but which generalizes to Riemannian manifolds. In fact, the method called barycentric subspace analysis, which is one of those generalizing principal component analysis to Riemannian manifolds, is based on this notion as well. Here we rephrase both methods under the new notion of barycentric embeddings. Essentially, barycentric embeddings inherit the structure of most linear and non-linear dimension reduction methods, but rely on a (locally) barycentric -- affine -- model rather than a linear one.The core of our work lies in the analysis of these methods, both on a theoretical and practical level. In particular, we address the application of barycentric embeddings to two important examples in geometric learning: shapes and graphs. In addition to practical implementation issues, each of these examples raises its own theoretical questions, mostly related to the geometry of quotient spaces. In particular, we highlight that compared to standard dimension reduction methods in graph analysis, barycentric embeddings stand out for their better interpretability. In parallel with these examples, we characterize the geometry of locally barycentric embeddings, which generalize the projection computed by locally linear embedding. Finally, algorithms for geometric manifold learning, novel in their approach, complete this work
Beghini, Federica. "À la recherche de la « pépite d'or » : Étude textométrique de l'œuvre de Milan Kundera." Electronic Thesis or Diss., Université Côte d'Azur, 2023. https://intranet-theses.unice.fr/2023COAZ2020.
Full textThis study consists of an integrated linguistic analysis of the work of Milan Kundera. By integrated analysis, we mean a linguistic study carried out through qualitative and quanti-tative methods. These methods belong to the field of textometry, a discipline whose objective is to analyse textual corpora through computer processing (Guiraud, 1960; Lebart, Salem, 1994; Pincemin, 2020). More generally, this work could therefore be included in the field of stylometry, since this textometric analysis is functional to the characterization of a style of writing (Magri, 2010). Indeed, the main objective of this research is to detect by contrast the elements that define Kundera's prose. To this end, two corpora were composed : a corpus of study and a reference corpus (Rastier, 2011). The first comprehends almost all the texts of Kundera's Œuvre I, II (Gallimard, Pléiade). The second is representative of the French literary landscape of the period in which Kundera published his texts (1968-2013).The corpora were first digitised and then examined using the textometry software Hyperbase (web and standard version), which employs both classical statistical methods and deep learning techniques (CNN, Convolutional neural network).This software allows various analyses on lexical, morphosyntactic and semantic levels. In particular, the following elements have been investigated : the vocabulary structure, morphological and syntactic aspects, morphosyntactic and multidimensional patterns, and finally the thematic structure.These elements were examined in an endogenous analysis of the corpus of study and in a series of exogenous analyses between the corpus of study and the reference corpus. Indeed, comparative studies between Kundera's work and the contrastive norm represented by the reference corpus aim to isolate the linguistic characteristics of the literary language of the time in novels, essays and short stories, in order to detect the distinguishing elements of Kundera's prose that differ from the linguistic model of his contemporaries' literary language. In addition, endogenous analyses of Kundera's work - made possible by the compilation of subcorpora - can account for linguistic constants that are independent of genre, period and/or language, as well as for linguistic variants determined by literary genre, diachronic and/or linguistic variability. In conclusion, this study employs an integrated methodology (linguistics, literature, statistics, deep learning) with the aim of defining the prototypical features of Kundera's idiolect, that is, the most significant elements that distinguish his writing from that of a representative sample of his contemporary French authors
Pajot, Arthur. "Incorporating physical knowledge into deep neural network." Electronic Thesis or Diss., Sorbonne université, 2019. http://www.theses.fr/2019SORUS290.
Full textA physical process is a sustained phenomenon marked by gradual changes through a series of states occurring in the physical world. Physicists and environmental scientists attempt to model these processes in a principled way through analytic descriptions of the scientist’s prior knowledge of the underlying processes. Despite the undeniable Deep Learning success, a fully data-driven approach is not yet ready to challenge the classical approach for modeling dynamical systems. We will try to demonstrate in this thesis that knowledge and techniques accumulated for modeling dynamical systems processes in well-developed fields such as maths or physics could be useful as a guideline to design efficient learning systems and conversely, that the ML paradigm could open new directions for modeling such complex phenomena. We describe three tasks that are relevant to the study and modeling of Deep Learning and Dynamical System : Forecasting, hidden state discovery and unsupervised signal recovery
Resmerita, Diana. "Compression pour l'apprentissage en profondeur." Thesis, Université Côte d'Azur, 2022. http://www.theses.fr/2022COAZ4043.
Full textAutonomous cars are complex applications that need powerful hardware machines to be able to function properly. Tasks such as staying between the white lines, reading signs, or avoiding obstacles are solved by using convolutional neural networks (CNNs) to classify or detect objects. It is highly important that all the networks work in parallel in order to transmit all the necessary information and take a common decision. Nowadays, as the networks improve, they also have become bigger and more computational expensive. Deploying even one network becomes challenging. Compressing the networks can solve this issue. Therefore, the first objective of this thesis is to find deep compression methods in order to cope with the memory and computational power limitations present on embedded systems. The compression methods need to be adapted to a specific processor, Kalray's MPPA, for short term implementations. Our contributions mainly focus on compressing the network post-training for storage purposes, which means compressing the parameters of the network without retraining or changing the original architecture and the type of the computations. In the context of our work, we decided to focus on quantization. Our first contribution consists in comparing the performances of uniform quantization and non-uniform quantization, in order to identify which of the two has a better rate-distortion trade-off and could be quickly supported in the company. The company's interest is also directed towards finding new innovative methods for future MPPA generations. Therefore, our second contribution focuses on comparing standard floating-point representations (FP32, FP16) to recently proposed alternative arithmetical representations such as BFloat16, msfp8, Posit8. The results of this analysis were in favor for Posit8. This motivated the company Kalray to conceive a decompressor from FP16 to Posit8. Finally, since many compression methods already exist, we decided to move to an adjacent topic which aims to quantify theoretically the effects of quantization error on the network's accuracy. This is the second objective of the thesis. We notice that well-known distortion measures are not adapted to predict accuracy degradation in the case of inference for compressed neural networks. We define a new distortion measure with a closed form which looks like a signal-to-noise ratio. A set of experiments were done using simulated data and small networks, which show the potential of this distortion measure
Kodi, Ramanah Doogesh. "Bayesian statistical inference and deep learning for primordial cosmology and cosmic acceleration." Thesis, Sorbonne université, 2019. http://www.theses.fr/2019SORUS169.
Full textThe essence of this doctoral research constitutes the development and application of novel Bayesian statistical inference and deep learning techniques to meet statistical challenges of massive and complex data sets from next-generation cosmic microwave background (CMB) missions or galaxy surveys and optimize their scientific returns to ultimately improve our understanding of the Universe. The first theme deals with the extraction of the E and B modes of the CMB polarization signal from the data. We have developed a high-performance hierarchical method, known as the dual messenger algorithm, for spin field reconstruction on the sphere and demonstrated its capabilities in reconstructing pure E and B maps, while accounting for complex and realistic noise models. The second theme lies in the development of various aspects of Bayesian forward modelling machinery for optimal exploitation of state-of-the-art galaxy redshift surveys. We have developed a large-scale Bayesian inference framework to constrain cosmological parameters via a novel implementation of the Alcock-Paczyński test and showcased our cosmological constraints on the matter density and dark energy equation of state. With the control of systematic effects being a crucial limiting factor for modern galaxy redshift surveys, we also presented an augmented likelihood which is robust to unknown foreground and target contaminations. Finally, with a view to building fast complex dynamics emulators in our above Bayesian hierarchical model, we have designed a novel halo painting network that learns to map approximate 3D dark matter fields to realistic halo distributions
Schmitt, Thomas. "Appariements collaboratifs des offres et demandes d’emploi." Thesis, Université Paris-Saclay (ComUE), 2018. http://www.theses.fr/2018SACLS210/document.
Full textOur research focuses on the recommendation of new job offers that have just been posted and have no interaction history (cold start). To this objective, we adapt well-knowns recommendations systems in the field of e-commerce by exploiting the record of use of all job seekers on previous offers. One of the specificities of the work presented is to have considered real data, and to have tackled the challenges of heterogeneity and noise of textual documents. The presented contribution integrates the information of the collaborative data to learn a new representation of text documents, which is required to make the so-called cold start recommendation of a new offer. The new representation essentially aims to build a good metric. The search space considered is that of neural networks. Neural networks are trained by defining two loss functions. The first seeks to preserve the local structure of collaborative information, drawing on non-linear dimension reduction approaches. The second is inspired by Siamese networks to reproduce the similarities from the collaborative matrix. The scaling up of the approach and its performance are based on the sampling of pairs of offers considered similar. The interest of the proposed approach is demonstrated empirically on the real and proprietary data as well as on the CiteULike public benchmark. Finally, the interest of the approach followed is attested by our participation in a good rank in the international challenge RecSys 2017 (15/100, with millions of users and millions of offers)
Martineau, Maxime. "Deep learning onto graph space : application to image-based insect recognition." Thesis, Tours, 2019. http://www.theses.fr/2019TOUR4024.
Full textThe goal of this thesis is to investigate insect recognition as an image-based pattern recognition problem. Although this problem has been extensively studied along the previous three decades, an element is to the best of our knowledge still to be experimented as of 2017: deep approaches. Therefore, a contribution is about determining to what extent deep convolutional neural networks (CNNs) can be applied to image-based insect recognition. Graph-based representations and methods have also been tested. Two attempts are presented: The former consists in designing a graph-perceptron classifier and the latter graph-based work in this thesis is on defining convolution on graphs to build graph convolutional neural networks. The last chapter of the thesis deals with applying most of the aforementioned methods to insect image recognition problems. Two datasets are proposed. The first one consists of lab-based images with constant background. The second one is generated by taking a ImageNet subset. This set is composed of field-based images. CNNs with transfer learning are the most successful method applied on these datasets
Tran, Gia-Lac. "Advances in Deep Gaussian Processes : calibration and sparsification." Electronic Thesis or Diss., Sorbonne université, 2020. https://accesdistant.sorbonne-universite.fr/login?url=https://theses-intra.sorbonne-universite.fr/2020SORUS410.pdf.
Full textGaussian Processes (GPs) are an attractive specific way of doing non-parametric Bayesian modeling in a supervised learning problem. It is well-known that GPs are able to make inferences as well as predictive uncertainties with a firm mathematical background. However, GPs are often unfavorable by the practitioners due to their kernel's expressiveness and the computational requirements. Integration of (convolutional) neural networks and GPs are a promising solution to enhance the representational power. As our first contribution, we empirically show that these combinations are miscalibrated, which leads to over-confident predictions. We also propose a novel well-calibrated solution to merge neural structures and GPs by using random features and variational inference techniques. In addition, these frameworks can be intuitively extended to reduce the computational cost by using structural random features. In terms of computational cost, the exact Gaussian Processes require the cubic complexity to training size. Inducing point-based Gaussian Processes are a common choice to mitigate the bottleneck by selecting a small set of active points through a global distillation from available observations. However, the general case remains elusive and it is still possible that the required number of active points may exceed a certain computational budget. In our second study, we propose Sparse-within-Sparse Gaussian Processes which enable the approximation with a large number of inducing points without suffering a prohibitive computational cost
Desir, Chesner. "Classification Automatique d'Images, Application à l'Imagerie du Poumon Profond." Phd thesis, Université de Rouen, 2013. http://tel.archives-ouvertes.fr/tel-00879356.
Full textDesir, Chesner. "Classification automatique d'images, application à l'imagerie du poumon profond." Phd thesis, Rouen, 2013. http://www.theses.fr/2013ROUES053.
Full textThis thesis deals with automated image classification, applied to images acquired with alveoscopy, a new imaging technique of the distal lung. The aim is to propose and develop a computer aided-diagnosis system, so as to help the clinician analyze these images never seen before. Our contributions lie in the development of effective, robust and generic methods to classify images of healthy and pathological patients. Our first classification system is based on a rich and local characterization of the images, an ensemble of random trees approach for classification and a rejection mechanism, providing the medical expert with tools to enhance the reliability of the system. Due to the complexity of alveoscopy images and to the lack of expertize on the pathological cases (unlike healthy cases), we adopt the one-class learning paradigm which allows to learn a classifier from healthy data only. We propose a one-class approach taking advantage of combining and randomization mechanisms of ensemble methods to respond to common issues such as the curse of dimensionality. Our method is shown to be effective, robust to the dimension, competitive and even better than state-of-the-art methods on various public datasets. It has proved to be particularly relevant to our medical problem
Fissore, Giancarlo. "Generative modeling : statistical physics of Restricted Boltzmann Machines, learning with missing information and scalable training of Linear Flows." Electronic Thesis or Diss., université Paris-Saclay, 2022. http://www.theses.fr/2022UPASG028.
Full textNeural network models able to approximate and sample high-dimensional probability distributions are known as generative models. In recent years this class of models has received tremendous attention due to their potential in automatically learning meaningful representations of the vast amount of data that we produce and consume daily. This thesis presents theoretical and algorithmic results pertaining to generative models and it is divided in two parts. In the first part, we focus our attention on the Restricted Boltzmann Machine (RBM) and its statistical physics formulation. Historically, statistical physics has played a central role in studying the theoretical foundations and providing inspiration for neural network models. The first neural implementation of an associative memory (Hopfield, 1982) is a seminal work in this context. The RBM can be regarded to as a development of the Hopfield model, and it is of particular interest due to its role at the forefront of the deep learning revolution (Hinton et al. 2006).Exploiting its statistical physics formulation, we derive a mean-field theory of the RBM that let us characterize both its functioning as a generative model and the dynamics of its training procedure. This analysis proves useful in deriving a robust mean-field imputation strategy that makes it possible to use the RBM to learn empirical distributions in the challenging case in which the dataset to model is only partially observed and presents high percentages of missing information. In the second part we consider a class of generative models known as Normalizing Flows (NF), whose distinguishing feature is the ability to model complex high-dimensional distributions by employing invertible transformations of a simple tractable distribution. The invertibility of the transformation allows to express the probability density through a change of variables whose optimization by Maximum Likelihood (ML) is rather straightforward but computationally expensive. The common practice is to impose architectural constraints on the class of transformations used for NF, in order to make the ML optimization efficient. Proceeding from geometrical considerations, we propose a stochastic gradient descent optimization algorithm that exploits the matrix structure of fully connected neural networks without imposing any constraints on their structure other then the fixed dimensionality required by invertibility. This algorithm is computationally efficient and can scale to very high dimensional datasets. We demonstrate its effectiveness in training a multylayer nonlinear architecture employing fully connected layers
Dancette, Corentin. "Shortcut Learning in Visual Question Answering." Electronic Thesis or Diss., Sorbonne université, 2023. http://www.theses.fr/2023SORUS073.
Full textThis thesis is focused on the task of VQA: it consists in answering textual questions about images. We investigate Shortcut Learning in this task: the literature reports the tendency of models to learn superficial correlations leading them to correct answers in most cases, but which can fail when encountering unusual input data. We first propose two methods to reduce shortcut learning on VQA. The first, which we call RUBi, consists of an additional loss to encourage the model to learn from the most difficult and less biased examples -- those which cannot be answered solely from the question. We then propose SCN, a model for the more specific task of visual counting, which incorporates architectural priors designed to make it more robust to distribution shifts. We then study the existence of multimodal shortcuts in the VQA dataset. We show that shortcuts are not only based on correlations between the question and the answer but can also involve image information. We design an evaluation benchmark to measure the robustness of models to multimodal shortcuts. We show that existing models are vulnerable to multimodal shortcut learning. The learning of those shortcuts is particularly harmful when models are evaluated in an out-of-distribution context. Therefore, it is important to evaluate the reliability of VQA models, i.e. We propose a method to improve their ability to abstain from answering when their confidence is too low. It consists of training an external ``selector'' model to predict the confidence of the VQA model. This selector is trained using a cross-validation-like scheme in order to avoid overfitting on the training set
Rosar, Kós Lassance Carlos Eduardo. "Graphs for deep learning representations." Thesis, Ecole nationale supérieure Mines-Télécom Atlantique Bretagne Pays de la Loire, 2020. http://www.theses.fr/2020IMTA0204.
Full textIn recent years, Deep Learning methods have achieved state of the art performance in a vast range of machine learning tasks, including image classification and multilingual automatic text translation. These architectures are trained to solve machine learning tasks in an end-to-end fashion. In order to reach top-tier performance, these architectures often require a very large number of trainable parameters. There are multiple undesirable consequences, and in order to tackle these issues, it is desired to be able to open the black boxes of deep learning architectures. Problematically, doing so is difficult due to the high dimensionality of representations and the stochasticity of the training process. In this thesis, we investigate these architectures by introducing a graph formalism based on the recent advances in Graph Signal Processing (GSP). Namely, we use graphs to represent the latent spaces of deep neural networks. We showcase that this graph formalism allows us to answer various questions including: ensuring generalization abilities, reducing the amount of arbitrary choices in the design of the learning process, improving robustness to small perturbations added to the inputs, and reducing computational complexity
Martens, Corentin. "Patient-Derived Tumour Growth Modelling from Multi-Parametric Analysis of Combined Dynamic PET/MR Data." Doctoral thesis, Universite Libre de Bruxelles, 2021. https://dipot.ulb.ac.be/dspace/bitstream/2013/320127/5/contratCM.pdf.
Full textLes gliomes sont les tumeurs cérébrales primitives les plus communes et sont associés à un mauvais pronostic. Parmi ces derniers, les gliomes diffus – qui incluent la forme la plus agressive, le glioblastome (GBM) – sont connus pour être hautement infiltrants. Le diagnostic et le suivi des gliomes s'appuient sur la tomographie par émission de positons (TEP) ainsi que l'imagerie par résonance magnétique (IRM). Cependant, ces techniques d'imagerie ne permettent actuellement pas d'évaluer l'étendue totale de tumeurs aussi infiltrantes ni d'anticiper leurs schémas d'invasion préférentiels, conduisant à une planification sous-optimale du traitement. La modélisation mathématique de la croissance tumorale a été proposée pour répondre à ce problème. Les modèles de croissance tumorale de type réaction-diffusion, qui sont probablement les plus communément utilisés pour la modélisation de la croissance des gliomes diffus, proposent de capturer la prolifération et la migration des cellules tumorales au moyen d'une équation aux dérivées partielles. Bien que le potentiel de tels modèles ait été démontré dans de nombreux travaux pour le suivi des patients et la planification de thérapies, seules quelques applications cliniques restreintes semblent avoir émergé de ces derniers. Ce travail de thèse a pour but de revisiter les modèles de croissance tumorale de type réaction-diffusion en utilisant des technologies de pointe en imagerie médicale et traitement de données, avec pour objectif d'y intégrer des données TEP/IRM multi-paramétriques pour personnaliser davantage le modèle. Le problème de la segmentation des tissus cérébraux dans les images IRM est d'abord adressé, avec pour but de définir un domaine propre au patient pour la résolution du modèle. Une méthode proposée précédemment permettant de dériver un tenseur de diffusion tumoral à partir du tenseur de diffusion de l'eau évalué par imagerie DTI a ensuite été implémentée afin de guider la migration anisotrope des cellules tumorales le long des fibres de matière blanche. L'utilisation de l'imagerie TEP dynamique à la [S-méthyl-11C]méthionine ([11C]MET) est également investiguée pour la génération de cartes de potentiel prolifératif propre au patient afin de nourrir le modèle. Ces investigations ont mené au développement d'un modèle compartimental pour le transport des traceurs TEP dérivés des acides aminés dans les gliomes. Sur base des résultats du modèle compartimental, une nouvelle méthodologie est proposée utilisant l'analyse en composantes principales pour extraire des cartes paramétriques à partir de données TEP dynamiques à la [11C]MET. Le problème de l'estimation des conditions initiales du modèle à partir d'images IRM est ensuite adressé par le biais d'une étude translationelle combinant IRM et histologie menée sur un cas de GBM non-opéré. Différentes stratégies de résolution numérique basées sur les méthodes des différences et éléments finis sont finalement implémentées et comparées. Tous ces développements sont embarqués dans un framework commun permettant d'étudier in silico la croissance des gliomes et fournissant une base solide pour de futures recherches dans le domaine. Cependant, certaines hypothèses communément admises reliant les délimitations des anormalités visibles en IRM à des iso-contours de densité de cellules tumorales ont été invalidée par l'étude translationelle menée, laissant ouverte les questions de l'initialisation et de la validation du modèle. Par ailleurs, l'analyse de l'évolution temporelle de cas réels de gliomes multi-traités démontre les limitations du modèle. Ces dernières affirmations mettent en évidence les obstacles actuels à l'application clinique de tels modèles et ouvrent la voie à de nouvelles possibilités d'amélioration.
Doctorat en Sciences de l'ingénieur et technologie
info:eu-repo/semantics/nonPublished
Matcha, Wyao. "Identification des composants prioritaires pour les tests unitaires dans les systèmes OO : une approche basée sur l'apprentissage profond." Thèse, 2020. http://depot-e.uqtr.ca/id/eprint/9420/1/eprint9420.pdf.
Full text