Дисертації з теми "Apprentissage profond – Recherche de l'information"
Оформте джерело за APA, MLA, Chicago, Harvard та іншими стилями
Ознайомтеся з топ-50 дисертацій для дослідження на тему "Apprentissage profond – Recherche de l'information".
Біля кожної праці в переліку літератури доступна кнопка «Додати до бібліографії». Скористайтеся нею – і ми автоматично оформимо бібліографічне посилання на обрану працю в потрібному вам стилі цитування: APA, MLA, «Гарвард», «Чикаго», «Ванкувер» тощо.
Також ви можете завантажити повний текст наукової публікації у форматі «.pdf» та прочитати онлайн анотацію до роботи, якщо відповідні параметри наявні в метаданих.
Переглядайте дисертації для різних дисциплін та оформлюйте правильно вашу бібліографію.
Ayoub, Oussama. "Enrichissement sémantique non supervisé de longs documents spécialisés pour la recherche d’information." Electronic Thesis or Diss., Paris, HESAM, 2023. http://www.theses.fr/2023HESAC039.
Faced with the incessant growth of textual data that needs processing, Information Retrieval (IR) systems are confronted with the urgent need to adopt effective mechanisms for efficiently selecting document sets that are best suited to specific queries. A predominant difficulty lies in the terminological divergence between the terms used in queries and those present in relevant documents. This semantic disparity, particularly pronounced for terms with similar meanings in large-scale documents from specialized domains, poses a significant challenge for IR systems.In addressing these challenges, many studies have been limited to query enrichment via supervised models, an approach that proves inadequate for industrial application and lacks flexibility. This thesis proposes LoGE an innovative alternative with an unsupervised search system based on advanced Deep Learning methods. This system uses a masked language model to extrapolate associated terms, thereby enriching the textual representation of documents. The Deep Learning models used, pre-trained on extensive textual corpora, incorporate general or domain-specific knowledge, thus optimizing the document representation.The analysis of the generated extensions revealed an imbalance between the signal (relevant terms added) and the noise (irrelevant terms). To address this issue, we developed SummVD, an innovative extractive automatic summarization approach, using singular value decomposition to synthesize the information contained in documents and identify the most pertinent phrases. This method has been adapted to filter the terms of the extensions based on the local context of each document, thereby maintaining the relevance of the information while minimizing noise
Belkacem, Thiziri. "Neural models for information retrieval : towards asymmetry sensitive approaches based on attention models." Thesis, Toulouse 3, 2019. http://www.theses.fr/2019TOU30167.
This work is situated in the context of information retrieval (IR) using machine learning (ML) and deep learning (DL) techniques. It concerns different tasks requiring text matching, such as ad-hoc research, question answering and paraphrase identification. The objective of this thesis is to propose new approaches, using DL methods, to construct semantic-based models for text matching, and to overcome the problems of vocabulary mismatch related to the classical bag of word (BoW) representations used in traditional IR models. Indeed, traditional text matching methods are based on the BoW representation, which considers a given text as a set of independent words. The process of matching two sequences of text is based on the exact matching between words. The main limitation of this approach is related to the vocabulary mismatch. This problem occurs when the text sequences to be matched do not use the same vocabulary, even if their subjects are related. For example, the query may contain several words that are not necessarily used in the documents of the collection, including relevant documents. BoW representations ignore several aspects about a text sequence, such as the structure the context of words. These characteristics are important and make it possible to differentiate between two texts that use the same words but expressing different information. Another problem in text matching is related to the length of documents. The relevant parts can be distributed in different ways in the documents of a collection. This is especially true in large documents that tend to cover a large number of topics and include variable vocabulary. A long document could thus contain several relevant passages that a matching model must capture. Unlike long documents, short documents are likely to be relevant to a specific subject and tend to contain a more restricted vocabulary. Assessing their relevance is in principle simpler than assessing the one of longer documents. In this thesis, we have proposed different contributions, each addressing one of the above-mentioned issues. First, in order to solve the problem of vocabulary mismatch, we used distributed representations of words (word embedding) to allow a semantic matching between the different words. These representations have been used in IR applications where document/query similarity is computed by comparing all the term vectors of the query with all the term vectors of the document, regardless. Unlike the models proposed in the state-of-the-art, we studied the impact of query terms regarding their presence/absence in a document. We have adopted different document/query matching strategies. The intuition is that the absence of the query terms in the relevant documents is in itself a useful aspect to be taken into account in the matching process. Indeed, these terms do not appear in documents of the collection for two possible reasons: either their synonyms have been used or they are not part of the context of the considered documents. The methods we have proposed make it possible, on the one hand, to perform an inaccurate matching between the document and the query, and on the other hand, to evaluate the impact of the different terms of a query in the matching process. Although the use of word embedding allows semantic-based matching between different text sequences, these representations combined with classical matching models still consider the text as a list of independent elements (bag of vectors instead of bag of words). However, the structure of the text as well as the order of the words is important. Any change in the structure of the text and/or the order of words alters the information expressed. In order to solve this problem, neural models were used in text matching
Nguyen, Kim-Anh Laura. "Document Understanding with Deep Learning Techniques." Electronic Thesis or Diss., Sorbonne université, 2024. http://www.theses.fr/2024SORUS077.
The field of Document Understanding, which addresses the problem of solving an array of Natural Language Processing tasks for visually-rich documents, faces challenges due to the complex structures and diverse formats of documents. Real-world documents rarely follow a strictly sequential structure. The visual presentation of a document, especially its layout, conveys rich semantic information, highlighting the crucial need for document understanding systems to include multimodal information. Despite notable advancements attributed to the emergence of Deep Learning, the field still grapples with various challenges in real-world applications. This thesis addresses two key challenges: 1) developing efficient and effective methods to encode the multimodal nature of documents, and 2) formulating strategies for efficient and effective processing of long and complex documents, considering their visual appearance. Our strategy to address the first research question involves designing approaches that rely only on layout to build meaningful representations. Multimodal pre-trained models for Document Understanding often neglect efficiency and fail to fully capitalize on the strong correlation between text and layout. We address these issues by introducing an attention mechanism based exclusively on layout information, enabling performance improvement and attention sparsification. Furthermore, we introduce a strategy based solely on layout to address reading order issues. While layout inherently captures the correct reading order of documents, existing pre-training methods for Document Understanding rely solely on OCR or PDF parsing to establish the reading order of documents, potentially introducing inaccuracies that can impact the entire text processing pipeline. Therefore, we discard sequential position information and propose a model that strategically leverages layout information as an alternative means to determine the reading order of documents. In addressing the second research axis, we explore the potential of leveraging layout to enhance the performance of models for tasks related to long and complex documents. The importance of document structure in information processing, particularly in the context of long documents, underscores the need for efficient modeling of layout information. To fill a notable void in resources and approaches for multimodal long document modeling, we introduce a dataset collection for summarization of long documents with consideration for their visual appearance, and present novel baselines that can handle long documents with awareness of their layout
Chafik, Sanaa. "Machine learning techniques for content-based information retrieval." Thesis, Université Paris-Saclay (ComUE), 2017. http://www.theses.fr/2017SACLL008/document.
The amount of media data is growing at high speed with the fast growth of Internet and media resources. Performing an efficient similarity (nearest neighbor) search in such a large collection of data is a very challenging problem that the scientific community has been attempting to tackle. One of the most promising solutions to this fundamental problem is Content-Based Media Retrieval (CBMR) systems. The latter are search systems that perform the retrieval task in large media databases based on the content of the data. CBMR systems consist essentially of three major units, a Data Representation unit for feature representation learning, a Multidimensional Indexing unit for structuring the resulting feature space, and a Nearest Neighbor Search unit to perform efficient search. Media data (i.e. image, text, audio, video, etc.) can be represented by meaningful numeric information (i.e. multidimensional vector), called Feature Description, describing the overall content of the input data. The task of the second unit is to structure the resulting feature descriptor space into an index structure, where the third unit, effective nearest neighbor search, is performed.In this work, we address the problem of nearest neighbor search by proposing three Content-Based Media Retrieval approaches. Our three approaches are unsupervised, and thus can adapt to both labeled and unlabeled real-world datasets. They are based on a hashing indexing scheme to perform effective high dimensional nearest neighbor search. Unlike most recent existing hashing approaches, which favor indexing in Hamming space, our proposed methods provide index structures adapted to a real-space mapping. Although Hamming-based hashing methods achieve good accuracy-speed tradeoff, their accuracy drops owing to information loss during the binarization process. By contrast, real-space hashing approaches provide a more accurate approximation in the mapped real-space as they avoid the hard binary approximations.Our proposed approaches can be classified into shallow and deep approaches. In the former category, we propose two shallow hashing-based approaches namely, "Symmetries of the Cube Locality Sensitive Hashing" (SC-LSH) and "Cluster-based Data Oriented Hashing" (CDOH), based respectively on randomized-hashing and shallow learning-to-hash schemes. The SC-LSH method provides a solution to the space storage problem faced by most randomized-based hashing approaches. It consists of a semi-random scheme reducing partially the randomness effect of randomized hashing approaches, and thus the memory storage problem, while maintaining their efficiency in structuring heterogeneous spaces. The CDOH approach proposes to eliminate the randomness effect by combining machine learning techniques with the hashing concept. The CDOH outperforms the randomized hashing approaches in terms of computation time, memory space and search accuracy.The third approach is a deep learning-based hashing scheme, named "Unsupervised Deep Neuron-per-Neuron Hashing" (UDN2H). The UDN2H approach proposes to index individually the output of each neuron of the top layer of a deep unsupervised model, namely a Deep Autoencoder, with the aim of capturing the high level individual structure of each neuron output.Our three approaches, SC-LSH, CDOH and UDN2H, were proposed sequentially as the thesis was progressing, with an increasing level of complexity in terms of the developed models, and in terms of the effectiveness and the performances obtained on large real-world datasets
Tuo, Aboubacar. "Extraction d'événements à partir de peu d'exemples par méta-apprentissage." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG098.
Information Extraction (IE) is a research field with the objective of automatically identifying and extracting structured information within a given domain from unstructured or minimally structured text data. The implementation of such extractions often requires significant human efforts, either in the form of rule development or the creation of annotated data for systems based on machine learning. One of the current challenges in information extraction is to develop methods that minimize the costs and development time of these systems whenever possible. This thesis focuses on few-shot event extraction through a meta-learning approach that aims to train IE models from only few data. We have redefined the task of event extraction from this perspective, aiming to develop systems capable of quickly adapting to new contexts with a small volume of training data. First, we propose methods to enhance event trigger detection by developing more robust representations for this task. Then, we tackle the specific challenge raised by the "NULL" class (absence of events) within this framework. Finally, we evaluate the effectiveness of our proposals within the broader context of event extraction by extending their application to the extraction of event arguments
Chafik, Sanaa. "Machine learning techniques for content-based information retrieval." Electronic Thesis or Diss., Université Paris-Saclay (ComUE), 2017. http://www.theses.fr/2017SACLL008.
The amount of media data is growing at high speed with the fast growth of Internet and media resources. Performing an efficient similarity (nearest neighbor) search in such a large collection of data is a very challenging problem that the scientific community has been attempting to tackle. One of the most promising solutions to this fundamental problem is Content-Based Media Retrieval (CBMR) systems. The latter are search systems that perform the retrieval task in large media databases based on the content of the data. CBMR systems consist essentially of three major units, a Data Representation unit for feature representation learning, a Multidimensional Indexing unit for structuring the resulting feature space, and a Nearest Neighbor Search unit to perform efficient search. Media data (i.e. image, text, audio, video, etc.) can be represented by meaningful numeric information (i.e. multidimensional vector), called Feature Description, describing the overall content of the input data. The task of the second unit is to structure the resulting feature descriptor space into an index structure, where the third unit, effective nearest neighbor search, is performed.In this work, we address the problem of nearest neighbor search by proposing three Content-Based Media Retrieval approaches. Our three approaches are unsupervised, and thus can adapt to both labeled and unlabeled real-world datasets. They are based on a hashing indexing scheme to perform effective high dimensional nearest neighbor search. Unlike most recent existing hashing approaches, which favor indexing in Hamming space, our proposed methods provide index structures adapted to a real-space mapping. Although Hamming-based hashing methods achieve good accuracy-speed tradeoff, their accuracy drops owing to information loss during the binarization process. By contrast, real-space hashing approaches provide a more accurate approximation in the mapped real-space as they avoid the hard binary approximations.Our proposed approaches can be classified into shallow and deep approaches. In the former category, we propose two shallow hashing-based approaches namely, "Symmetries of the Cube Locality Sensitive Hashing" (SC-LSH) and "Cluster-based Data Oriented Hashing" (CDOH), based respectively on randomized-hashing and shallow learning-to-hash schemes. The SC-LSH method provides a solution to the space storage problem faced by most randomized-based hashing approaches. It consists of a semi-random scheme reducing partially the randomness effect of randomized hashing approaches, and thus the memory storage problem, while maintaining their efficiency in structuring heterogeneous spaces. The CDOH approach proposes to eliminate the randomness effect by combining machine learning techniques with the hashing concept. The CDOH outperforms the randomized hashing approaches in terms of computation time, memory space and search accuracy.The third approach is a deep learning-based hashing scheme, named "Unsupervised Deep Neuron-per-Neuron Hashing" (UDN2H). The UDN2H approach proposes to index individually the output of each neuron of the top layer of a deep unsupervised model, namely a Deep Autoencoder, with the aim of capturing the high level individual structure of each neuron output.Our three approaches, SC-LSH, CDOH and UDN2H, were proposed sequentially as the thesis was progressing, with an increasing level of complexity in terms of the developed models, and in terms of the effectiveness and the performances obtained on large real-world datasets
Tang, Anfu. "Leveraging linguistic and semantic information for relation extraction from domain-specific texts." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG081.
This thesis aims to extract relations from scientific documents in the biomedical domain, i.e. transform unstructured texts into structured data that is machine-readable. As a task in the domain of Natural Language Processing (NLP), the extraction of semantic relations between textual entities makes explicit and formalizes the underlying structures. Current state-of-the-art methods rely on supervised learning, more specifically the fine-tuning of pre-trained language models such as BERT. Supervised learning requires a large amount of examples that are expensive to produce, especially in specific domains such as the biomedical domain. BERT variants such as PubMedBERT have been successful on NLP tasks involving biomedical texts. We hypothesize that injecting external information such as syntactic information or factual knowledge into such BERT variants can compensate for the reduced number of annotated training data. To this end, this thesis consists of proposing several neural architectures based on PubMedBERT that exploit linguistic information obtained by syntactic parsers or domain knowledge from knowledge bases
Paumard, Marie-Morgane. "Résolution automatique de puzzles par apprentissage profond." Thesis, CY Cergy Paris Université, 2020. http://www.theses.fr/2020CYUN1067.
The objective of this thesis is to develop semantic methods of reassembly in the complicated framework of heritage collections, where some blocks are eroded or missing.The reassembly of archaeological remains is an important task for heritage sciences: it allows to improve the understanding and conservation of ancient vestiges and artifacts. However, some sets of fragments cannot be reassembled with techniques using contour information or visual continuities. It is then necessary to extract semantic information from the fragments and to interpret them. These tasks can be performed automatically thanks to deep learning techniques coupled with a solver, i.e., a constrained decision making algorithm.This thesis proposes two semantic reassembly methods for 2D fragments with erosion and a new dataset and evaluation metrics.The first method, Deepzzle, proposes a neural network followed by a solver. The neural network is composed of two Siamese convolutional networks trained to predict the relative position of two fragments: it is a 9-class classification. The solver uses Dijkstra's algorithm to maximize the joint probability. Deepzzle can address the case of missing and supernumerary fragments, is capable of processing about 15 fragments per puzzle, and has a performance that is 25% better than the state of the art.The second method, Alphazzle, is based on AlphaZero and single-player Monte Carlo Tree Search (MCTS). It is an iterative method that uses deep reinforcement learning: at each step, a fragment is placed on the current reassembly. Two neural networks guide MCTS: an action predictor, which uses the fragment and the current reassembly to propose a strategy, and an evaluator, which is trained to predict the quality of the future result from the current reassembly. Alphazzle takes into account the relationships between all fragments and adapts to puzzles larger than those solved by Deepzzle. Moreover, Alphazzle is compatible with constraints imposed by a heritage framework: at the end of reassembly, MCTS does not access the reward, unlike AlphaZero. Indeed, the reward, which indicates if a puzzle is well solved or not, can only be estimated by the algorithm, because only a conservator can be sure of the quality of a reassembly
Grivolla, Jens. "Apprentissage et décision automatique en recherche documentaire : prédiction de difficulté de requêtes et sélection de modèle de recherche." Avignon, 2006. http://www.theses.fr/2006AVIG0142.
This thesis is centered around the subject of information retrieval, with a focus on those queries that are particularly difficult to handle for current retrieval systems. In the application and evaluation settings we were concerned with, a user expresses his information need as a natural language query. There are different approaches for treating those queries, but current systems typically use a single approach for all queries, without taking into account the specific properties of each query. However, it has been shown that the performance of one strategy relative to another can vary greatly depending on the query. We have approached this problem by proposing methods that will permit to automatically identify those queries that will pose particular difficulties to the retrieval system, in order to allow for a specific treatment. This research topic was very new and barely starting to be explored at the beginning of my work, but has received much attention these last years. We have developed a certain number of quality predictor functions that obtain results comparable to those published recently by other research teams. However, the ability of individual predictors to accurately classify queries by their level of difficulty remains rather limited. The major particularity and originality of our work lies in the combination of those different measures. Using methods of automatic classification with corpus-based training, we have been able to obtain quite reliable predictions, on the basis of measures that individually are far less discriminant. We have also adapted our approach to other application settings, with very encouraging results. We have thus developed a method for the selective application of query expansion techniques, as well as the selection of the most appropriate retrieval model for each query
Oita, Marilena. "Inférer des objets sémantiques du Web structuré." Thesis, Paris, ENST, 2012. http://www.theses.fr/2012ENST0060/document.
This thesis focuses on the extraction and analysis of Web data objects, investigated from different points of view: temporal, structural, semantic. We first survey different strategies and best practices for deriving temporal aspects of Web pages, together with a more in-depth study on Web feeds for this particular purpose, and other statistics. Next, in the context of dynamically-generated Web pages by content management systems, we present two keyword-based techniques that perform article extraction from such pages. Keywords, automatically acquired, guide the process of object identification, either at the level of a single Web page (SIGFEED), or across different pages sharing the same template (FOREST). We finally present, in the context of the deep Web, a generic framework that aims at discovering the semantic model of a Web object (here, data record) by, first, using FOREST for the extraction of objects, and second, representing the implicit rdf:type similarities between the object attributes and the entity of the form as relationships that, together with the instances extracted from the objects, form a labeled graph. This graph is further aligned to an ontology like YAGO for the discovery of the unknown types and relations
Oita, Marilena. "Inférer des objets sémantiques du Web structuré." Electronic Thesis or Diss., Paris, ENST, 2012. http://www.theses.fr/2012ENST0060.
This thesis focuses on the extraction and analysis of Web data objects, investigated from different points of view: temporal, structural, semantic. We first survey different strategies and best practices for deriving temporal aspects of Web pages, together with a more in-depth study on Web feeds for this particular purpose, and other statistics. Next, in the context of dynamically-generated Web pages by content management systems, we present two keyword-based techniques that perform article extraction from such pages. Keywords, automatically acquired, guide the process of object identification, either at the level of a single Web page (SIGFEED), or across different pages sharing the same template (FOREST). We finally present, in the context of the deep Web, a generic framework that aims at discovering the semantic model of a Web object (here, data record) by, first, using FOREST for the extraction of objects, and second, representing the implicit rdf:type similarities between the object attributes and the entity of the form as relationships that, together with the instances extracted from the objects, form a labeled graph. This graph is further aligned to an ontology like YAGO for the discovery of the unknown types and relations
Halin, Gilles. "Apprentissage pour la recherche interactive et progressive d'images : processus EXPRIM et prototype RIVAGE." Nancy 1, 1989. http://www.theses.fr/1989NAN10412.
Amini, Massih-Reza. "Apprentissage automatique et recherche de l'information : application a l'extraction d'information de surface et au resume de texte." Paris 6, 2001. http://www.theses.fr/2001PA066005.
Feutry, Clément. "Two sides of relevant information : anonymized representation through deep learning and predictor monitoring." Thesis, Université Paris-Saclay (ComUE), 2019. http://www.theses.fr/2019SACLS479.
The work presented here is for a first part at the cross section of deep learning and anonymization. A full framework was developed in order to identify and remove to a certain extant, in an automated manner, the features linked to an identity in the context of image data. Two different kinds of processing data were explored. They both share the same Y-shaped network architecture despite components of this network varying according to the final purpose. The first one was about building from the ground an anonymized representation that allowed a trade-off between keeping relevant features and tampering private features. This framework has led to a new loss. The second kind of data processing specified no relevant information about the data, only private information, meaning that everything that was not related to private features is assumed relevant. Therefore the anonymized representation shares the same nature as the initial data (e.g. an image is transformed into an anonymized image). This task led to another type of architecture (still in a Y-shape) and provided results strongly dependent on the type of data. The second part of the work is relative to another kind of relevant information: it focuses on the monitoring of predictor behavior. In the context of black box analysis, we only have access to the probabilities outputted by the predictor (without any knowledge of the type of structure/architecture producing these probabilities). This monitoring is done in order to detect abnormal behavior that is an indicator of a potential mismatch between the data statistics and the model statistics. Two methods are presented using different tools. The first one is based on comparing the empirical cumulative distribution of known data and to be tested data. The second one introduces two tools: one relying on the classifier uncertainty and the other relying on the confusion matrix. These methods produce concluding results
Angrisani, Armando. "The disparate impact of noise on quantum learning algorithms." Electronic Thesis or Diss., Sorbonne université, 2023. http://www.theses.fr/2023SORUS626.
Quantum computing, one of the most exciting scientific journeys of our time, holds remarkable potential by promising to rapidly solve computational problems. However, the practical implementation of these algorithms poses an immense challenge, with a universal and error-tolerant quantum computer remaining an elusive goal. Currently, short-term quantum devices are emerging, but they face significant limitations, including high levels of noise and limited entanglement capacity. The practical effectiveness of these devices, particularly due to quantum noise, is a subject of debate. Motivated by this situation, this thesis explores the profound impact of noise on quantum learning algorithms in three key dimensions. Firstly, it focuses on the influence of noise on variational quantum algorithms, especially quantum kernel methods. Our results reveal significant disparities between unital and non-unital noise, challenging previous conclusions on these noisy algorithms. Next, it addresses learning quantum dynamics with noisy binary measurements of the Choi-Jamiolkowski state, using quantum statistical queries. The Goldreich-Levin algorithm can be implemented in this way, and we demonstrate the efficiency of learning in our model. Finally, the thesis contributes to quantum differential privacy, demonstrating how quantum noise can enhance statistical security. A new definition of neighboring quantum states captures the structure of quantum encodings, providing stricter privacy guarantees. In the local model, we establish an equivalence between quantum statistical queries and local quantum differential privacy, with applications to tasks like asymmetric hypothesis testing. The results are illustrated by the efficient learning of parity functions in this model, compared to a classically demanding task
Cleuziou, Guillaume. "Une méthode de classification non-supervisée pour l'apprentissage de règles et la recherche d'information." Phd thesis, Université d'Orléans, 2004. http://tel.archives-ouvertes.fr/tel-00084828.
Nous proposons, dans cette étude, l'algorithme de clustering PoBOC permettant de structurer un ensemble d'objets en classes non-disjointes. Nous utilisons cette méthode de clustering comme outil de traitement dans deux applications très différentes.
- En apprentissage supervisé, l'organisation préalable des instances apporte une connaissance utile pour la tâche d'induction de règles propositionnelles et logiques.
- En Recherche d'Information, les ambiguïtés et subtilités de la langue naturelle induisent naturellement des recouvrements entre thématiques.
Dans ces deux domaines de recherche, l'intérêt d'organiser les objets en classes non-disjointes est confirmé par les études expérimentales adaptées.
De, Groc Clément. "Collecte orientée sur le Web pour la recherche d'information spécialisée." Phd thesis, Université Paris Sud - Paris XI, 2013. http://tel.archives-ouvertes.fr/tel-00853250.
Pouy, Léo. "OpenNas : un cadre adaptable de recherche automatique d'architecture neuronale." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG089.
When creating a neural network, the "fine-tuning" stage is essential. During this fine-tuning, the neural network developer must adjust the hyperparameters and the architecture of the network so that it meets the targets. This is a time-consuming and tedious phase, and requires experience on the part of the developer. So, to make it easier to create neural networks, there is a discipline called Automatic Machine Learning (Auto-ML), which seeks to automate the creation of Machine Learning. This thesis is part of this Auto-ML approach and proposes a method for creating and optimizing neural network architectures (Neural Architecture Search, NAS). To this end, a new search space based on block imbrication has been formalized. This space makes it possible to create a neural network from elementary blocks connected in series or in parallel to form compound blocks which can themselves be connected to form an even more complex network. The advantage of this search space is that it can be easily customized to influence the NAS for specific architectures (VGG, Inception, ResNet, etc.) and control the optimization time. Moreover, it is not constrained to any particular optimization algorithm. In this thesis, the formalization of the search space is first described, along with encoding techniques to represent a network from the search space by a natural number (or a list of natural numbers). Optimization strategies applicable to this search space are then proposed. Finally, neural architecture search experiments on different datasets and with different objectives using the developed tool (named OpenNas) are presented
Laporte, Léa. "La sélection de variables en apprentissage d'ordonnancement pour la recherche d'information : vers une approche contextuelle." Toulouse 3, 2013. http://thesesups.ups-tlse.fr/2170/.
Learning-to-rank aims at automatically optimizing a ranking function learned on training data by a machine learning algorithm. Existing approaches have two major drawbacks. Firstly, the ranking functions can use several thousands of features, which is an issue since algorithms have to deal with large scale data. This can also have a negative impact on the ranking quality. Secondly, algorithms learn an unique fonction for all queries. Then, nor the kind of user need neither the context of the query are taken into account in the ranking process. Our works focus on solving the large-scale issue and the context-aware issue by using feature selection methods dedicated to learning-to-rank. We propose five feature selection algorithms based on sparse Support Vector Machines (SVM). Three proceed to feature selection by reweighting the L2-norm, one solves a L1-regularized problem whereas the last algorithm consider nonconvex regularizations. Our methods are faster and sparser than state-of-the-art algorithms on benchmark datasets, while providing similar performances in terms of RI measures. We also evaluate our approches on a commercial dataset. Experimentations confirm the previous results. We propose in this context a relevance model based on users clicks, in the special case of multi-clickable documents. Finally, we propose an adaptative and query-dependent ranking system based on feature selection. This system considers several clusters of queries, each group defines a context. For each cluster, the system selects a group of features to learn a context-aware ranking function
Brodin, Elisabeth. "Interactions entre innovation, technologies de l'information et de la communication et apprentissage institutionnel des langues : l'exemple d'une recherche-action dans des lycées." Le Mans, 2002. http://cyberdoc.univ-lemans.fr/theses/2002/2002LEMA3003.pdf.
Veniat, Tom. "Neural Architecture Search under Budget Constraints." Electronic Thesis or Diss., Sorbonne université, 2021. http://www.theses.fr/2021SORUS443.
The recent increase in computation power and the ever-growing amount of data available ignited the rise in popularity of deep learning. However, the expertise, the amount of data, and the computing power necessary to build such algorithms as well as the memory footprint and the inference latency of the resulting system are all obstacles preventing the widespread use of these methods. In this thesis, we propose several methods allowing to make a step towards a more efficient and automated procedure to build deep learning models. First, we focus on learning an efficient architecture for image processing problems. We propose a new model in which we can guide the architecture learning procedure by specifying a fixed budget and cost function. Then, we consider the problem of sequence classification, where a model can be even more efficient by dynamically adapting its size to the complexity of the signal to come. We show that both approaches result in significant budget savings. Finally, we tackle the efficiency problem through the lens of transfer learning. Arguing that a learning procedure can be made even more efficient if, instead of starting tabula rasa, it builds on knowledge acquired during previous experiences. We explore modular architectures in the continual learning scenario and present a new benchmark allowing a fine-grained evaluation of different kinds of transfer
Heuillet, Alexandre. "Exploring deep neural network differentiable architecture design." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG069.
Artificial Intelligence (AI) has gained significant popularity in recent years, primarily due to its successful applications in various domains, including textual data analysis, computer vision, and audio processing. The resurgence of deep learning techniques has played a central role in this success. The groundbreaking paper by Krizhevsky et al., AlexNet, narrowed the gap between human and machine performance in image classification tasks. Subsequent papers such as Xception and ResNet have further solidified deep learning as a leading technique, opening new horizons for the AI community. The success of deep learning lies in its architecture, which is manually designed with expert knowledge and empirical validation. However, these architectures lack the certainty of an optimal solution. To address this issue, recent papers introduced the concept of Neural Architecture Search (NAS), enabling the learning of deep architectures. However, most initial approaches focused on large architectures with specific targets (e.g., supervised learning) and relied on computationally expensive optimization techniques such as reinforcement learning and evolutionary algorithms. In this thesis, we further investigate this idea by exploring automatic deep architecture design, with a particular emphasis on differentiable NAS (DNAS), which represents the current trend in NAS due to its computational efficiency. While our primary focus is on Convolutional Neural Networks (CNNs), we also explore Vision Transformers (ViTs) with the goal of designing cost-effective architectures suitable for real-time applications
Motta, Jesus Antonio. "VENCE : un modèle performant d'extraction de résumés basé sur une approche d'apprentissage automatique renforcée par de la connaissance ontologique." Doctoral thesis, Université Laval, 2014. http://hdl.handle.net/20.500.11794/26076.
Several methods and techniques of artificial intelligence for information extraction, pattern recognition and data mining are used for extraction of summaries. More particularly, new machine learning models with the introduction of ontological knowledge allow the extraction of the sentences containing the greatest amount of information from a corpus. This corpus is considered as a set of sentences on which different optimization methods are applied to identify the most important attributes. They will provide a training set from which a machine learning algorithm will can abduce a classification function able to discriminate the sentences of new corpus according their information content. Currently, even though the results are interesting, the effectiveness of models based on this approach is still low, especially in the discriminating power of classification functions. In this thesis, a new model based on this approach is proposed and its effectiveness is improved by inserting ontological knowledge to the training set. The originality of this model is described through three papers. The first paper aims to show how linear techniques could be applied in an original way to optimize workspace in the context of extractive summary. The second article explains how to insert ontological knowledge to significantly improve the performance of classification functions. This introduction is performed by inserting lexical chains of ontological knowledge based in the training set. The third article describes VENCE , the new machine learning model to extract sentences with the most information content in order to produce summaries. An assessment of the VENCE performance is achieved comparing the results with those produced by current commercial and public software as well as those published in very recent scientific articles. The use of usual metrics recall, precision and F_measure and the ROUGE toolkit showed the superiority of VENCE. This model could benefit other contexts of information extraction as for instance to define models for sentiment analysis.
Claveau, Vincent. "Acquisition automatique de lexiques sémantiques pour la recherche d'information." Phd thesis, Université Rennes 1, 2003. http://tel.archives-ouvertes.fr/tel-00524646.
Colombo, Pierre. "Learning to represent and generate text using information measures." Electronic Thesis or Diss., Institut polytechnique de Paris, 2021. http://www.theses.fr/2021IPPAT033.
Natural language processing (NLP) allows for the automatic understanding and generation of natural language. NLP has recently received growing interest from both industry and researchers as deep learning (DL) has leveraged the staggering amount of available text (e.g web, youtube, social media) and reached human-like performance in several tasks (e.g translation, text classification). Besides, Information theory (IT) and DL have developed a long-lasting partnership. Indeed, IT has fueled the adoption of deep neural networks with famous principles such as Minimum Description Length (MDL), Information Bottleneck (IB) or the celebrated InfoMax principle. In all these principles, different measures of information (e.g entropy, MI, divergences) are one of the core concepts. In this thesis, we address the interplay between NLP and measures of information. Our contributions focus on two types of NLP problems : natural language understanding (NLU) and natural language generation (NLG). NLU aims at automatically understand and extract semantic information from an input text where NLG aims at producing natural language that is both well-formed (i.e grammatically correct, coherent) and informative. Building spoken conversational agents is a challenging issue and dealing with spoken conversational data remains a difficult and overlooked problem. Thus, our first contributions, are turned towards NLU and we focus on learning transcript representations. Our contribution focuses on learning better transcript representations that include two important characteristics of spoken human conversations : namely the conversational and the multi-modal dimension. To do so, we rely on various measures of information and leverage the mutual information maximization principle. The second group of contributions addresses problems related to NLG. This thesis specifically focuses on two core problems. First, we propose a new upper bound on mutual information to tackle the problem of controlled generation via the learning of disentangled representation (i.e style transfer and conditional sentence generation). Secondly, we address the problem of automatic evaluation of generated texts by developing a new family of metrics using various measuresof information
Castagnos, Sylvain. "Modélisation de comportements et apprentissage stochastique non supervisé de stratégies d'interactions sociales au sein de systèmes temps réel de recherche et d'accès à l'information." Phd thesis, Université Nancy II, 2008. http://tel.archives-ouvertes.fr/tel-00341470.
Toutefois, les modèles utilisateurs souffrent d'un grand nombre de données manquantes. Notre approche consiste à exploiter collaborativement les données relatives à une population pour pallier le manque d'information inhérent à chaque utilisateur. L'emploi de techniques de filtrage collaboratif permet ainsi de bénéficier de l'expérience et des interactions au sein d'une population pour améliorer les services et prédire les futurs agissements d'un individu. Nous sommes partis du constat que, dans les approches centralisées, le nombre d'individus pris en compte dans la recherche des plus proches voisins ne peut excéder quelques milliers de candidats. Nos travaux nous ont donc conduit à distribuer le processus de filtrage sous plusieurs formes tant en terme de contenu que de calculs. L'objectif de cette thèse est de montrer comment il est possible d'assurer le passage à l'échelle, et faire face aux problèmes sous-jacents pouvant résulter de cette approche distribuée.
Moreno, José G. "Text-Based Ephemeral Clustering for Web Image Retrieval on Mobile Devices." Caen, 2014. http://www.theses.fr/2014CAEN2036.
In this thesis, we present a study about Web image results visualization on mobile devices. Our main findings were inspired by the recent advances in two main research areas - Information Retrieval and Natural Language Processing. In the former, we considered different topics such as search results clustering, Web mobile interfaces, query intent mining, to name but a few. In the latter, we were more focused in collocation measures, high order similarity metrics, etc. Particularly in order to validate our hypothesis, we performed a great deal of different experiments with task specific datasets. Many characteristics are evaluated in the proposed solutions. First, the clustering quality in which classical and recent evaluation metrics are considered. Secondly, the labeling quality of each cluster is evaluated to make sure that all possible query intents are covered. Thirdly and finally, we evaluate the user's effort in exploring the images in a gallery-based interface. An entire chapter is dedicated to each of these three aspects in which the datasets - some of them built to evaluate specific characteristics - are presented. For the final results, we can take into account two developed algorithms, two datasets and a SRC evaluation tool. From the algorithms, Dual C-means is our main product. It can be seen as a generalization of our previously developed algorithm, the AGK-means. Both are based in text-based similarity metrics. A new dataset for a complete evaluation of SRC algorithms is developed and presented. Similarly, a new Web image dataset is developed and used together with a new metric to measure the users effort when a set of Web images is explored. Finally, we developed an evaluation tool for the SRC problem, in which we have implemented several classical and recent SRC metrics. Our conclusions are drawn considering the numerous factors that were discussed in this thesis. However, additional studies could be motivated based in our findings. Some of them are discussed in the end of this study and preliminary analysis suggest that they are directions that have potential
Ferré, Sébastien. "Systèmes d'information logiques : un paradigme logico-contextuel pour interroger, naviguer et apprendre." Rennes 1, 2002. http://www.theses.fr/2002REN10143.
Chifu, Adrian-Gabriel. "Adaptation des systèmes de recherche d'information aux contextes : le cas des requêtes difficiles." Thesis, Toulouse 3, 2015. http://www.theses.fr/2015TOU30061/document.
The field of information retrieval (IR) studies the mechanisms to find relevant information in one or more document collections, in order to satisfy an information need. For an Information Retrieval System (IRS) the information to find is represented by "documents" and the information need takes the form of a "query" formulated by the user. IRS performance depends on queries. Queries for which the IRS fails (little or no relevant documents retrieved) are called in the literature "difficult queries". This difficulty may be caused by term ambiguity, unclear query formulation, the lack of context for the information need, the nature and structure of the document collection, etc. This thesis aims at adapting IRS to contexts, particularly in the case of difficult queries. The manuscript is organized into five main chapters, besides acknowledgements, general introduction, conclusions and perspectives. The first chapter is an introduction to RI. We develop the concept of relevance, the retrieval models from the literature, the query expansion models and the evaluation framework that was employed to validate our proposals. Each of the following chapters presents one of our contributions. Every chapter raises the research problem, indicates the related work, our theoretical proposals and their validation on benchmark collections. In chapter two, we present our research on treating the ambiguous queries. The query term ambiguity can indeed lead to poor document retrieval of documents by the search engine. In the related work, the disambiguation methods that yield good performance are supervised, however such methods are not applicable in a real IR context, as they require the information which is normally unavailable. Moreover, in the literature, term disambiguation for IR is declared under optimal
Goswami, Parantapa. "Learning information retrieval functions and parameters on unlabeled collections." Thesis, Grenoble, 2014. http://www.theses.fr/2014GRENM089.
The present study focuses on (a) predicting parameters of already existing standard IR models and (b) learning new IR functions. We first explore various statistical methods to estimate the collection parameter of family of information based models (Chapter 2). This parameter determines the behavior of a term in the collection. In earlier studies, it was set to the average number of documents where the term appears, without full justification. We introduce here a fully formalized estimation method which leads to improved versions of these models over the original ones. But the method developed is applicable only to estimate the collection parameter under the information model framework. To alleviate this we propose a transfer learning approach which can predict values for any parameter for any IR model (Chapter 3). This approach uses relevance judgments on a past collection to learn a regression function which can infer parameter values for each single query on a new unlabeled target collection. The proposed method not only outperforms the standard IR models with their default parameter values, but also yields either better or at par performance with popular parameter tuning methods which use relevance judgments on target collection. We then investigate the application of transfer learning based techniques to directly transfer relevance information from a source collection to derive a "pseudo-relevance" judgment on an unlabeled target collection (Chapter 4). From this derived pseudo-relevance a ranking function is learned using any standard learning algorithm which can rank documents in the target collection. In various experiments the learned function outperformed standard IR models as well as other state-of-the-art transfer learning based algorithms. Though a ranking function learned through a learning algorithm is effective still it has a predefined form based on the learning algorithm used. We thus introduce an exhaustive discovery approach to search ranking functions from a space of simple functions (Chapter 5). Through experimentation we found that some of the discovered functions are highly competitive with respect to standard IR models
De, Groc Clément. "Collecte orientée sur le Web pour la recherche d’information spécialisée." Thesis, Paris 11, 2013. http://www.theses.fr/2013PA112073/document.
Vertical search engines, which focus on a specific segment of the Web, become more and more present in the Internet landscape. Topical search engines, notably, can obtain a significant performance boost by limiting their index on a specific topic. By doing so, language ambiguities are reduced, and both the algorithms and the user interface can take advantage of domain knowledge, such as domain objects or characteristics, to satisfy user information needs.In this thesis, we tackle the first inevitable step of a all topical search engine : focused document gathering from the Web. A thorough study of the state of art leads us to consider two strategies to gather topical documents from the Web: either relying on an existing search engine index (focused search) or directly crawling the Web (focused crawling).The first part of our research has been dedicated to focused search. In this context, a standard approach consists in combining domain-specific terms into queries, submitting those queries to a search engine and down- loading top ranked documents. After empirically evaluating this approach over 340 topics, we propose to enhance it in two different ways: Upstream of the search engine, we aim at formulating more relevant queries in or- der to increase the precision of the top retrieved documents. To do so, we define a metric based on a co-occurrence graph and a random walk algorithm, which aims at predicting the topical relevance of a query. Downstream of the search engine, we filter the retrieved documents in order to improve the document collection quality. We do so by modeling our gathering process as a tripartite graph and applying a random walk with restart algorithm so as to simultaneously order by relevance the documents and terms appearing in our corpus.In the second part of this thesis, we turn to focused crawling. We describe our focused crawler implementation that was designed to scale horizontally. Then, we consider the problem of crawl frontier ordering, which is at the very heart of a focused crawler. Such ordering strategy allows the crawler to prioritize its fetches, maximizing the number of in-domain documents retrieved while minimizing the non relevant ones. We propose to apply learning to rank algorithms to efficiently order the crawl frontier, and define a method to learn a ranking function from existing crawls
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.
This 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
Tmar, Mohamed. "Modèle auto-adaptatif de filtrage d'information : apprentissage incrémental du profil et de la fonction de décision." Toulouse 3, 2002. http://www.theses.fr/2002TOU30081.
Chen, Jianan. "Deep Learning Based Multimodal Retrieval." Electronic Thesis or Diss., Rennes, INSA, 2023. http://www.theses.fr/2023ISAR0019.
Multimodal tasks play a crucial role in the progression towards achieving general artificial intelligence (AI). The primary goal of multimodal retrieval is to employ machine learning algorithms to extract relevant semantic information, bridging the gap between different modalities such as visual images, linguistic text, and other data sources. It is worth noting that the information entropy associated with heterogeneous data for the same high-level semantics varies significantly, posing a significant challenge for multimodal models. Deep learning-based multimodal network models provide an effective solution to tackle the difficulties arising from substantial differences in information entropy. These models exhibit impressive accuracy and stability in large-scale cross-modal information matching tasks, such as image-text retrieval. Furthermore, they demonstrate strong transfer learning capabilities, enabling a well-trained model from one multimodal task to be fine-tuned and applied to a new multimodal task, even in scenarios involving few-shot or zero-shot learning. In our research, we develop a novel generative multimodal multi-view database specifically designed for the multimodal referential segmentation task. Additionally, we establish a state-of-the-art (SOTA) benchmark and multi-view metric for referring expression segmentation models in the multimodal domain. The results of our comparative experiments are presented visually, providing clear and comprehensive insights
Theobald, Claire. "Bayesian Deep Learning for Mining and Analyzing Astronomical Data." Electronic Thesis or Diss., Université de Lorraine, 2023. http://www.theses.fr/2023LORR0081.
In this thesis, we address the issue of trust in deep learning predictive systems in two complementary research directions. The first line of research focuses on the ability of AI to estimate its level of uncertainty in its decision-making as accurately as possible. The second line, on the other hand, focuses on the explainability of these systems, that is, their ability to convince human users of the soundness of their predictions.The problem of estimating the uncertainties is addressed from the perspective of Bayesian Deep Learning. Bayesian Neural Networks assume a probability distribution over their parameters, which allows them to estimate different types of uncertainties. First, aleatoric uncertainty which is related to the data, but also epistemic uncertainty which quantifies the lack of knowledge the model has on the data distribution. More specifically, this thesis proposes a Bayesian neural network can estimate these uncertainties in the context of a multivariate regression task. This model is applied to the regression of complex ellipticities on galaxy images as part of the ANR project "AstroDeep''. These images can be corrupted by different sources of perturbation and noise which can be reliably estimated by the different uncertainties. The exploitation of these uncertainties is then extended to galaxy mapping and then to "coaching'' the Bayesian neural network. This last technique consists of generating increasingly complex data during the model's training process to improve its performance.On the other hand, the problem of explainability is approached from the perspective of counterfactual explanations. These explanations consist of identifying what changes to the input parameters would have led to a different prediction. Our contribution in this field is based on the generation of counterfactual explanations relying on a variational autoencoder (VAE) and an ensemble of predictors trained on the latent space generated by the VAE. This method is particularly adapted to high-dimensional data, such as images. In this case, they are referred as counterfactual visual explanations. By exploiting both the latent space and the ensemble of classifiers, we can efficiently produce visual counterfactual explanations that reach a higher degree of realism than several state-of-the-art methods
Ramu, Jean-Philippe. "Efficience d'une documentation opérationnelle contextuelle sur la performance des pilotes de transport aérien." Toulouse, ISAE, 2008. http://www.theses.fr/2008ESAE0020.
Zouambi, Meyssa. "Optimizing deep learning : navigating the field of neural architecture search from theory to practice." Electronic Thesis or Diss., Université de Lille (2022-....), 2023. http://www.theses.fr/2023ULILB054.
In the realm of deep learning, the design and optimization of neural architectures are crucial for achieving high-performance models. This process, based on trial and error, has been done manually for decades and is both time and resource-consuming. This thesis delves into the domain of Neural Architecture Search (NAS), a promising technique that seeks to automate this process. The research explores the complexities inherent in NAS, highlighting the challenges of navigating the vast search space of potential architectures. It investigates methods based on Local Search and proposes two efficient algorithms built around it, namely LS-Weight and LS-PON. Each method offers a distinctive approach to integrate knowledge during the search to offer efficient and more frugal strategies to NAS. Furthermore, as deep learning models are often governed by multiple competing objectives such as accuracy, complexity, and computational efficiency, this research also delves into multi-objective optimization within NAS. This ensures that the resulting architectures are not only performant for the task they are designed for but also aligned with multiple criteria essential for real-world applications. For this purpose, this research offers an alternative approach to multi-objective NAS that addresses certain issues found in strategies from the literature. On top of that, it also analyzes the complexity of moving from benchmarks to real data, offering a protocol that guides practitioners in their usage of NAS for their applications. Lastly, by recognizing the importance of domain applications, this work focuses on healthcare images to validate these contributions. It also presents a detailed survey on the use of NAS for healthcare, by analyzing more than 40 contributions in the literature and laying the ground for future works in the field
Zakaria, Ahmad. "Batch steganography and pooled steganalysis in JPEG images." Thesis, Montpellier, 2020. http://www.theses.fr/2020MONTS079.
ABSTRACT:Batch steganography consists of hiding a message by spreading it out in a set of images, while pooled steganalysis consists of analyzing a set of images to conclude whether or not a hidden message is present. There are many strategies for spreading a message and it is reasonable to assume that the steganalyst does not know which one is being used, but it can be assumed that the steganographer uses the same embedding algorithm for all images. In this case, it can be shown that the most appropriate solution for pooled steganalysis is to use a single quantitative detector (i.e. one that predicts the size of the hidden message), to evaluate for each image the size, the hidden message (which can be zero if there is none), and to average the sizes (which are finally considered as scores) obtained over all the images.What would be the optimal solution if now the steganalyst could discriminate the spreading strategy among a set of known strategies. Could the steganalyst use a pooled steganalysis algorithm that is better than averaging the scores? Could the steganalyst obtain results close to the so-called "clairvoyant" scenario where it is assumed that the steganalyst knows exactly the spreading strategy?In this thesis, we try to answer these questions by proposing a pooled steganalysis architecture based on a quantitative image detector and an optimized score pooling function. The first contribution is a study of quantitative steganalysis algorithms in order to decide which one is best suited for pooled steganalysis. For this purpose, we propose to extend this comparison to binary steganalysis algorithms and we propose a methodology to switch from binary steganalysis results to quantitative steganalysis and vice versa.The core of the thesis lies in the second contribution. We study the scenario where the steganalyst does not know the spreading strategy. We then propose an optimized pooling function of the results based on a set of spreading strategies which improves the accuracy of the pooled steganalysis compared to a simple average. This pooling function is computed using supervised learning techniques. Experimental results obtained with six different spreading strategies and a state-of-the-art quantitative detector confirm our hypothesis. Our pooling function gives results close to a clairvoyant steganalyst who is supposed to know the spreading strategy.Keywords: Multimedia Security, Batch Steganography, Pooled Steganalysis, Machine Learning
Erbacher, Pierre. "Proactive models for open-domain conversational search." Electronic Thesis or Diss., Sorbonne université, 2024. http://www.theses.fr/2024SORUS009.
Conversational systems are increasingly becoming important gateways to information in a wide range of application domains such as customer service, health, education, office work, online shopping, and web search.While existing language models are able to follow long conversations, answer questions, and summarize documents with impressive fluency, they cannot be considered as true conversational search systems.Beyond providing natural language answers, a key capability of conversational search systems is their (pro)active participation in the conversation with users. This allows conversational search systems to better capture users' needs but also guide, and assist them during search sessions. In particular, when users cannot browse the list of documents to assess the relevance, as in pure speech interactions, the system needs to take the initiative to ask for additional context, ask for confirmation, or suggest more information to help the user navigate virtually and reduce his cognitive load. Additionally, these models are expected not only to take the initiate in conversation with users but also to proactively interact with a diverse range of other systems or database, including various tools (calendar, calculator ), internet (search engines), and various other APIs (weather, maps, e-commerce, booking.. ). However, due to the high cost of collecting and annotating such data, available conversational datasets for information access are typically small, hand-crafted, and limited to domain-specific applications such as recommendation or conversational question-answering, which are typically user-initiated and contain simple or a series of contextualized questions. In addition, it is particularly challenging to properly evaluate conversational search systems because of the nature of the interactions.In this thesis, we aim to improve conversational search by enabling more complex and useful interactions with users. We propose multiple methods and approaches to achieve this goal.First, in chapter 1 and 2, we investigate how user simulations can be used to train and evaluate systems that perform query refinement through sequential interactions with the user. We focus on sequential click-based interaction with a user simulation for clarifying queries.Then, in chapter 3 and chapter 4, we explore how existing IR datasets can be enhanced with simulated interactions to improve IR capabilities in conversational search and how mixed-initiative interactions can serve document retrieval and query disambiguation. In chapter 4, we propose to augment the AmbigNQ dataset with clarifying questions to better train and evaluate systems to perform pro-active question-answering tasks, where systems are expected to disambiguate the initial user questions before answering. To our knowledge, PAQA is the first dataset providing both questions, answers, supporting documents, and clarifying questions covering multiple types of ambiguity (entity references, event references, properties, time-dependent…) with enough examples for fine-tuning models. Finally, in the last chapter, we focused on the interaction between systems and an external search engine. We introduced a new approach method to teach a language model to internally assess its ability to answer properly a given query, without using anything more than data comprised used for its training. The resulting model can directly identify its ability to answer a given question, with performances comparable -if not superior- to widely accepted hallucination detection baselines such as perplexity-based approaches which are strong exogenous baselines. It allows models to proactively query search API depending on its ability to answer the question
Nguyen, Gia Hung. "Modèles neuronaux pour la recherche d'information : approches dirigées par les ressources sémantiques." Thesis, Toulouse 3, 2018. http://www.theses.fr/2018TOU30233.
In this thesis, we focus on bridging the semantic gap between the documents and queries representations, hence improve the matching performance. We propose to combine relational semantics from knowledge resources and distributed semantics of the corpus inferred by neural models. Our contributions consist of two main aspects: (1) Improving distributed representations of text for IR tasks. We propose two models that integrate relational semantics into the distributed representations: a) an offline model that combines two types of pre-trained representations to obtain a hybrid representation of the document; b) an online model that jointly learns distributed representations of documents, concepts and words. To better integrate relational semantics from knowledge resources, we propose two approaches to inject these relational constraints, one based on the regularization of the objective function, the other based on instances in the training text. (2) Exploiting neural networks for semantic matching of documents}. We propose a neural model for document-query matching. Our neural model relies on: a) a representation of raw-data that models the relational semantics of text by jointly considering objects and relations expressed in a knowledge resource, and b) an end-to-end neural architecture that learns the query-document relevance by leveraging the distributional and relational semantics of documents and queries
Chenu, Alexandre. "Leveraging sequentiality in Robot Learning : Application of the Divide & Conquer paradigm to Neuro-Evolution and Deep Reinforcement Learning." Electronic Thesis or Diss., Sorbonne université, 2023. http://www.theses.fr/2023SORUS342.
“To succeed, planning alone is insufficient. One must improvise as well.” This quote from Isaac Asimov, founding father of robotics and author of the Three Laws of Robotics, emphasizes the importance of being able to adapt and think on one’s feet to achieve success. Although robots can nowadays resolve highly complex tasks, they still need to gain those crucial adaptability skills to be deployed on a larger scale. Robot Learning uses learning algorithms to tackle this lack of adaptability and to enable robots to solve complex tasks autonomously. Two types of learning algorithms are particularly suitable for robots to learn controllers autonomously: Deep Reinforcement Learning and Neuro-Evolution. However, both classes of algorithms often cannot solve Hard Exploration Problems, that is problems with a long horizon and a sparse reward signal, unless they are guided in their learning process. One can consider different approaches to tackle those problems. An option is to search for a diversity of behaviors rather than a specific one. The idea is that among this diversity, some behaviors will be able to solve the task. We call these algorithms Diversity Search algorithms. A second option consists in guiding the learning process using demonstrations provided by an expert. This is called Learning from Demonstration. However, searching for diverse behaviors or learning from demonstration can be inefficient in some contexts. Indeed, finding diverse behaviors can be tedious if the environment is complex. On the other hand, learning from demonstration can be very difficult if only one demonstration is available. This thesis attempts to improve the effectiveness of Diversity Search and Learning from Demonstration when applied to Hard Exploration Problems. To do so, we assume that complex robotics behaviors can be decomposed into reaching simpler sub-goals. Based on this sequential bias, we try to improve the sample efficiency of Diversity Search and Learning from Demonstration algorithms by adopting Divide & Conquer strategies, which are well-known for their efficiency when the problem is composable. Throughout the thesis, we propose two main strategies. First, after identifying some limitations of Diversity Search algorithms based on Neuro-Evolution, we propose Novelty Search Skill Chaining. This algorithm combines Diversity Search with Skill- Chaining to efficiently navigate maze environments that are difficult to explore for state-of-the-art Diversity Search. In a second set of contributions, we propose the Divide & Conquer Imitation Learning algorithms. The key intuition behind those methods is to decompose the complex task of learning from a single demonstration into several simpler goal-reaching sub-tasks. DCIL-II, the most advanced variant, can learn walking behaviors for under-actuated humanoid robots with unprecedented efficiency. Beyond underlining the effectiveness of the Divide & Conquer paradigm in Robot Learning, this work also highlights the difficulties that can arise when composing behaviors, even in elementary environments. One will inevitably have to address these difficulties before applying these algorithms directly to real robots. It may be necessary for the success of the next generations of robots, as outlined by Asimov
Mustar, Agnès. "Modeling User-Machine Interactions During The Information Retrieval Process." Electronic Thesis or Diss., Sorbonne université, 2023. http://www.theses.fr/2023SORUS136.
While today’s search engines work well for simple queries, there are situations where search results are not satisfactory. To cope with such situations, user-machine interactions have increased significantly since the early days of retrieval systems. The exchanges between users and retrieval systems during search sessions may contain information that is critical to the success of the information search. Meanwhile, the Transformer-based architectures relying on the attention mechanism have led to great improvements in several NLP tasks, such as summarization or translation. The architecture was soon applied to other domains, including information retrieval. Several retrieval models have benefited from this architecture’s ability to focus on document and query terms to estimate their relationships. However, the majority of these works have focused on ad hoc retrieval. The goal of this thesis is to study user modeling and user-machine interactions with transformer-based models. The contributions of this thesis can be divided into two parts, those related to user modeling and those related to interactive systems. In the former, I analyze existing user models, and in particular the text generation process of transformer-based models for query suggestions (Mustar, Lamprier, and Piwowarski, 2020; Mustar, Lamprier, and Piwowarski, 2021). In the latter, I present a new user/machine interaction framework based on the studied user models (Mustar, Lamprier, and Piwowarski, 2022)
Alassaf, Yaqdhan. "Integration des TIC dans l'enseignement/apprentissage du FLE en Irak : enjeux institutionnels, organisationnels et pédagogiques." Thesis, Lille 3, 2016. http://www.theses.fr/2016LIL30062.
The teaching/learning of French as a Foreign Language (FLE) has been influenced by ICT in recent years. We note that the latest evolutions have not been observed in Iraq where pedagogical practice, in particular with the help of Information and Communication Technologies (ICT), is marked by specific constraints. IT equipment is not sufficient and computer aided practice do not quite meet innovation and global digital development. Reform of the educational system and the methods used in Iraq could also permit an evolution of practice. Thus, this research focuses on the changing practices of the teaching/learning of FLE that could result from the integration of ICT. An action research has been conducted at the University of Mosul in order to measure the level of practice and the representation of different kinds of users. Interviews and questionnaires have permitted to gather a certain amount of data that have been analyzed and verified statistically. This work aims to emphasize the usefulness of an innovative environment for the teaching/learning of FLE. To be in phase with this development, and jointly ensure, in the context of Iraq, the use of platforms, the environment had to consider "the social appropriation process" and consider three levels of intervention, macro, meso and micro to initiate change. Thanks to the results obtained during the experimental phase, the development of innovation through the integration of ICT for education is likely to become operational
Sayadi, Karim. "Classification du texte numérique et numérisé. Approche fondée sur les algorithmes d'apprentissage automatique." Thesis, Paris 6, 2017. http://www.theses.fr/2017PA066079/document.
Different disciplines in the humanities, such as philology or palaeography, face complex and time-consuming tasks whenever it comes to examining the data sources. The introduction of computational approaches in humanities makes it possible to address issues such as semantic analysis and systematic archiving. The conceptual models developed are based on algorithms that are later hard coded in order to automate these tedious tasks. In the first part of the thesis we propose a novel method to build a semantic space based on topics modeling. In the second part and in order to classify historical documents according to their script. We propose a novel representation learning method based on stacking convolutional auto-encoder. The goal is to automatically learn plot representations of the script or the written language
Panovski, Dancho. "Simulation, optimization and visualization of transportation data." Electronic Thesis or Diss., Institut polytechnique de Paris, 2020. http://www.theses.fr/2020IPPAS016.
Today all major metropolises in France, Europe and the rest of the world suffer from severe problems of congestion and saturation of infrastructures, which concern both individual and public transport. Current transportation systems are reaching capacity limits and appear unable to absorb increases in passenger flows in the future. The transport of the future is part of the various so-called Smart City initiatives and should be ”intelligent”, that is to say not only react to the demands but anticipate them, relying on the data exchanged between the end user and the information system of transportation operators.Within this context, one of the main challenges is the creation of appropriate methodologies for analysis of geo-localized transport data for instantaneous storage, analysis, management and dissemination of massive (typically thousands of instant geo-localizations , with a refresh rate of the order of a few seconds) data flows. The related algorithms must be capable of managing event lists of several tens of minutes to calculate real trajectories, instantaneous occupations, traffic lights changing cycles as well as vehicular traffic flow forecasts.In this thesis, we address two different issues related to this topic.A first contribution concerns the optimization of the traffic lights systems. The objective is to minimize the total journey time of the vehicles that are present in a certain part of a city. To this purpose, we propose a PSO (Particle Swarm Optimization) technique. The experimental results obtained show that such an approach makes it possible to obtain significant gains (5.37% - 21.53%) in terms of global average journey time.The second part of the thesis is dedicated to the issue of traffic flow prediction. In particular, we focus on prediction of the bus arrival time in the various bus stations existent over a given itinerary. Here, our contributions first concern the introduction of a novel data model, so-called TDM (Traffic Density Matrix), which captures dynamically the situation of the traffic along a given bus itinerary. Then, we show how different machine learning (ML) techniques can exploit such a structure in order to perform efficient prediction. To this purpose, we consider first traditional ML techniques, including linear regression and support vector regression with various kernels. The analysis of the results show that increasing the level of non-linearity can lead to superior results. Based on this observation, we propose various deep learning techniques with hand-crafted networks that we have specifically adapted to our objectives. The proposed approach include recurrent neural networks, LSTM (Long Short Time Memory) approaches, fully connected and convolutional networks. The analysis of the obtained experimental results confirm our intuition and demonstrate that such highly non-linear techniques outperform the traditional approaches and are able to deal with the singularities of the data that in this case correspond to localized traffic jams that globally affect the behavior of the system.Due to the lack of availability of such highly sensitive type of geo-localized information, all the data considered in our experiments has been produced with the help of the SUMO (Simulation of Urban Mobility) microscopic simulator. We notably show how SUMO can be exploited to construct realistic scenarios, close to real-life situations and exploitable for analysis purposes.Interpretation and understanding the data is of vital importance, nevertheless an adequate visualization platform is needed to present the results in a visually pleasing and understandable manner. To this purpose, we finally propose two different visualization application, a first one dedicated to the operators and the second one to clients. To ensure the deployment and compatibility of such applications on different devices (desktop PCs, Laptops, Smartphones, tablets…) a scalable solution is proposed
Blot, Michaël. "Étude de l'apprentissage et de la généralisation des réseaux profonds en classification d'images." Electronic Thesis or Diss., Sorbonne université, 2018. http://www.theses.fr/2018SORUS412.
Artificial intelligence is experiencing a resurgence in recent years. This is due to the growing ability to collect and store a considerable amount of digitized data. These huge databases allow machine learning algorithms to respond to certain tasks through supervised learning. Among the digitized data, images remain predominant in the modern environment. Huge datasets have been created. moreover, the image classification has allowed the development of previously neglected models, deep neural networks or deep learning. This family of algorithms demonstrates a great facility to learn perfectly datasets, even very large. Their ability to generalize remains largely misunderstood, but the networks of convolutions are today the undisputed state of the art. From a research and application point of view of deep learning, the demands will be more and more demanding, requiring to make an effort to bring the performances of the neuron networks to the maximum of their capacities. This is the purpose of our research, whose contributions are presented in this thesis. We first looked at the issue of training and considered accelerating it through distributed methods. We then studied the architectures in order to improve them without increasing their complexity. Finally, we particularly study the regularization of network training. We studied a regularization criterion based on information theory that we deployed in two different ways
Ouyang, Wei. "Deep Learning for Advanced Microscopy." Thesis, Sorbonne Paris Cité, 2018. http://www.theses.fr/2018USPCC174/document.
Background: Microscopy plays an important role in biology since several centuries, but its resolution has long been limited to ~250nm due to diffraction, leaving many important biological structures (e.g. viruses, vesicles, nuclear pores, synapses) unresolved. Over the last decade, several super-resolution methods have been developed that break this limit. Among the most powerful and popular super-resolution techniques are those based on single molecular localization (single molecule localization microscopy, or SMLM) such as PALM and STORM. By precisely localizing positions of isolated fluorescent molecules in thousands or more sequentially acquired diffraction limited images, SMLM can achieve resolutions of 20-50 nm or better. However, SMLM is inherently slow due to the necessity to accumulate enough localizations to achieve high resolution sampling of the fluorescent structures. The drawback in acquisition speed (typically ~30 minutes per super-resolution image) makes it difficult to use SMLM in high-throughput and live cell imaging. Many methods have been proposed to address this issue, mostly by improving the localization algorithms to localize overlapping spots, but most of them compromise spatial resolution and cause artifacts.Methods and results: In this work, we applied deep learning based image-to-image translation framework for improving imaging speed and quality by restoring information from rapidly acquired low quality SMLM images. By utilizing recent advances in deep learning including the U-net and Generative Adversarial Networks, we developed our method Artificial Neural Network Accelerated PALM (ANNA-PALM) which is capable of learning structural information from training images and using the trained model to accelerate SMLM imaging by tens to hundreds folds. With experimentally acquired images of different cellular structures (microtubules, nuclear pores and mitochondria), we demonstrated that deep learning can efficiently capture the structural information from less than 10 training samples and reconstruct high quality super-resolution images from sparse, noisy SMLM images obtained with much shorter acquisitions than usual for SMLM. We also showed that ANNA-PALM is robust to possible variations between training and testing conditions, due either to changes in the biological structure or to changes in imaging parameters. Furthermore, we take advantage of the acceleration provided by ANNA-PALM to perform high throughput experiments, showing acquisition of ~1000 cells at high resolution in ~3 hours. Additionally, we designed a tool to estimate and reduce possible artifacts is designed by measuring the consistency between the reconstructed image and the experimental wide-field image. Our method enables faster and gentler imaging which can be applied to high-throughput, and provides a novel avenue towards live cell high resolution imaging. Deep learning methods rely on training data and their performance can be improved even further with more training data. One cheap way to obtain more training data is through data sharing within the microscopy community. However, it often difficult to exchange or share localization microscopy data, because localization tables alone are typically several gigabytes in size, and there is no dedicated platform for localization microscopy data which provide features such as rendering, visualization and filtering. To address these issues, we developed a file format that can losslessly compress localization tables into smaller files, alongside with a web platform called ShareLoc (https://shareloc.xyz) that allows to easily visualize and share 2D or 3D SMLM data. We believe that this platform can greatly improve the performance of deep learning models, accelerate tool development, facilitate data re-analysis and further promote reproducible research and open science
Carsault, Tristan. "Introduction of musical knowledge and qualitative analysis in chord extraction and prediction tasks with machine learning. : application to human-machine co-improvisation." Electronic Thesis or Diss., Sorbonne université, 2020. http://www.theses.fr/2020SORUS247.
This thesis investigates the impact of introducing musical properties in machine learning models for the extraction and inference of musical features. Furthermore, it discusses the use of musical knowledge to perform qualitative evaluations of the results. In this work, we focus on musical chords since these mid-level features are frequently used to describe harmonic progressions in Western music. Hence, amongs the variety of tasks encountered in the field of Music Information Retrieval (MIR), the two main tasks that we address are the Automatic Chord Extraction (ACE) and the inference of symbolic chord sequences. In the case of musical chords, there exists inherent strong hierarchical and functional relationships. Indeed, even if two chords do not belong to the same class, they can share the same harmonic function within a chord progression. Hence, we developed a specifically-tailored analyzer that focuses on the functional relations between chords to distinguish strong and weak errors. We define weak errors as a misclassification that still preserves the relevance in terms of harmonic function. This reflects the fact that, in contrast to strict transcription tasks, the extraction of high-level musical features is a rather subjective task. Moreover, many creative applications would benefit from a higher level of harmonic understanding rather than an increased accuracy of label classification. For instance, one of our application case is the development of a software that interacts with a musician in real-time by inferring expected chord progressions. In order to achieve this goal, we divided the project into two main tasks : a listening module and a symbolic generation module. The listening module extracts the musical structure played by the musician, where as the generative module predicts musical sequences based on the extracted features. In the first part of this thesis, we target the development of an ACE system that could emulate the process of musical structure discovery, as performed by musicians in improvisation contexts. Most ACE systems are built on the idea of extracting features from raw audio signals and, then, using these features to construct a chord classifier. This entail two major families of approaches, as either rule-based or statistical models. In this work, we identify drawbacks in the use of statistical models for ACE tasks. Then, we propose to introduce prior musical knowledge in order to account for the inherent relationships between chords directly inside the loss function of learning methods. In the second part of this thesis, we focus on learning higher-level relationships inside sequences of extracted chords in order to develop models with the ability to generate potential continuations of chord sequences. In order to introduce musical knowledge in these models, we propose both new architectures, multi-label training methods and novel data representations
Nguyen, Nhu Van. "Représentations visuelles de concepts textuels pour la recherche et l'annotation interactives d'images." Phd thesis, Université de La Rochelle, 2011. http://tel.archives-ouvertes.fr/tel-00730707.
Angulo, Mendoza Gustavo Adolfo. "Renforcer la présence en formation à la recherche dans le deuxième cycle universitaire par les communautés d’apprentissage : encourager la collaboration pour moduler la distance pédagogique." Doctoral thesis, Université Laval, 2020. http://hdl.handle.net/20.500.11794/66683.
This doctoral research aims to determine how an increased presence can help to modulate the educational distance in a context of learning the scientific research process in master's degree programs. In other words, the main goal is to determine if, and in what way, the social interactions taking place in a technology-mediated community can lessen the difficulties associated with educational distance and how these interactions would support the learning of scientific research process. The reference framework for this study consists of two essential concepts: educational distance (Jacquinot, 1993; Moore, 1993; Moore et Kearsley, 2011) and presence (Jézégou, 2012; Shin, 2002). From these concepts we developed an analytical framework that we called "global transactional presence in a graduate level community". This study aims to document a global and emerging portrait of these elements in a rarely studied context, namely research training in master's degree programs. From an interpretive and comprehensive perspective with an exploratory, but also descriptive and explanatory scope, this research is based on a case study that took place in a community of research and mutual aid in a North American Francophone university. This community aims to develop graduate students' scientific skills, support their research work and gradually integrate them into the professional community. For this research, the main source of empirical data is transcripts of semi-structured interviews with 15 students and 4 faculty members. The data collection was supplemented by observations of activities taking place in synchronous meetings (face-to-face or online) and discussions in asynchronous thematic forums. A mixed content analysis method was used including quantification of code co-occurrences and an interpretative analysis of participants' comments. The study shows that, in a context of research training in master's degree programs, increasing the global transactional presence through interactions within a technology-mediated community promotes studentresearchers' perceptions regarding the availability of peers and faculty and, in turn, reinforces the sense of connection between them. The study highlights the importance of peer interaction to support future researchers’ training in several dimensions: learning about the academic research process, scientific enculturation, socialization, psychological aspects, counselling and orientation needs. It identifies eight successful educational practices that can contribute to the development of student-researchers’ scientific skills: research clinics, presentations when milestones are achieved, training activities, writing workshops, closed symposia, forums, debates and reading clubs. Finally, it reports nine key conditions for a successful technology-mediated learning community for graduate research training: adherence to a socio-constructivist approach, sharing responsibility between faculty and students, definition of a disciplinary perimeter, planning of activities considering both present and online students, balance between individual and collective supervision, establishment of a structure promoting participation (frequency and duration of activities), developing skills to provide critical and constructive feedback, building a common knowledge base and promoting awareness of community activities.