Tesis sobre el tema "Neural network adaptation"

The PreventIT project is an EU Horizon 2020 project aimed at preventing early functional decline at younger old age. The analysis of causal links between risk factors and functional decline has been made possible by the cooperation of several research institutes' studies. However, since each research institute collects and delivers different kinds of data in different formats, so far the analysis has been assisted by expert geriatricians whose role is to detect the best candidates among hundreds of fields and offer a semantic interpretation of the values. This manual data harmonization approach is very common in both scientific and industrial environments. In this thesis project an alternative method for parsing heterogeneous data is proposed. Since all the datasets represent semantically related data, being all made from longitudinal studies on aging-related metrics, it is possible to train an artificial neural network to perform an automatic domain adaptation. To achieve this goal, a Stacked Denoising Autoencoder has been implemented and trained to extract a domain-invariant representation of the data. Then, from this high-level representation, multiple classifiers have been trained to validate the model and ultimately to predict the probability of functional decline of the patient. This innovative approach to the domain adaptation process can provide an easy and fast solution to many research fields that now rely on human interaction to analyze the semantic data model and perform cross-dataset analysis. Functional decline classifiers show a great improvement in their performance when trained on the domain-invariant features extracted by the Stacked Denoising Autoencoder. Furthermore, this project applies multiple deep neural network classifiers on top of the Stacked Denoising Autoencoder representation, achieving excellent results for the prediction of functional decline in a real case study that involves two different datasets.

As computers begin to pervade aspects of our everyday lives, so the problem of communication from man-to-machine becomes increasingly evident. In recent years, there has been a concerted interest in speech recognition offering a user to communicate freely with a machine. However, this deceptively simple means for exchanging information is in fact extremely complex. A single utterance can contain a wealth of varied information concerning the speaker's gender, age, dialect and mood. Numerous subtle differences such as intonation, rhythm and stress further add to the complexity, increasing the variability between inter- and intra-speaker utterances. These differences pose an enormous problem, especially for a multi-user system since it is impractical to train for every variation of every utterance from every speaker. Consequently adaptation is of great importance, allowing a system with limited knowledge to dynamically adapt towards a new speakers characteristics. A new modified artificial neural network (ANN) was proposed incorporating One-Class-OneNetwork (OCON) subnet architectures connected via a common front-end adaptation layer. Using vowel phonemes from the TIMIT speech database, the adaptation was concentrated on neurons within the front-end layer, resulting in only information common to all classes, primarily speaker characteristics, being adapted. In addition, this prevented new utterances from interfering with phoneme unique information in the corresponding OCON subnets. Hence a more efficient adaptation procedure was created which, after adaptation towards a single class, also aided in the recognition of the remaining classes within the network. Compared with a conventional multi-layer perceptron network, results for inter- and intraspeaker adaptation showed an equally marked improvement for the recognition of adapted phonemes during both full neuron and front-layer neuron adaptation within the new modified architecture. When testing the effects of adaptation on the remaining unadapted vowel phonemes, the modified architecture (allowing only the neurons in the front-end layer to adapt) yielded better results than the modified architecture allowing full neuron adaptation. These results highlighted the storing of speaker information, common to all classes, in the front-end layer allowing efficient inter- and intra-speaker dynamic adaptation.

Language is the principal medium for ideas, while dialogue is the most natural and effective way for humans to interact with and access information from machines. Natural language generation (NLG) is a critical component of spoken dialogue and it has a significant impact on usability and perceived quality. Many commonly used NLG systems employ rules and heuristics, which tend to generate inflexible and stylised responses without the natural variation of human language. However, the frequent repetition of identical output forms can quickly make dialogue become tedious for most real-world users. Additionally, these rules and heuristics are not scalable and hence not trivially extensible to other domains or languages. A statistical approach to language generation can learn language decisions directly from data without relying on hand-coded rules or heuristics, which brings scalability and flexibility to NLG. Statistical models also provide an opportunity to learn in-domain human colloquialisms and cross-domain model adaptations. A robust and quasi-supervised NLG model is proposed in this thesis. The model leverages a Recurrent Neural Network (RNN)-based surface realiser and a gating mechanism applied to input semantics. The model is motivated by the Long-Short Term Memory (LSTM) network. The RNN-based surface realiser and gating mechanism use a neural network to learn end-to-end language generation decisions from input dialogue act and sentence pairs; it also integrates sentence planning and surface realisation into a single optimisation problem. The single optimisation not only bypasses the costly intermediate linguistic annotations but also generates more natural and human-like responses. Furthermore, a domain adaptation study shows that the proposed model can be readily adapted and extended to new dialogue domains via a proposed recipe. Continuing the success of end-to-end learning, the second part of the thesis speculates on building an end-to-end dialogue system by framing it as a conditional generation problem. The proposed model encapsulates a belief tracker with a minimal state representation and a generator that takes the dialogue context to produce responses. These features suggest comprehension and fast learning. The proposed model is capable of understanding requests and accomplishing tasks after training on only a few hundred human-human dialogues. A complementary Wizard-of-Oz data collection method is also introduced to facilitate the collection of human-human conversations from online workers. The results demonstrate that the proposed model can talk to human judges naturally, without any difficulty, for a sample application domain. In addition, the results also suggest that the introduction of a stochastic latent variable can help the system model intrinsic variation in communicative intention much better.

The goal of this thesis is to advance the use of recurrent neural network language models (RNNLMs) for large vocabulary continuous speech recognition (LVCSR). RNNLMs are currently state-of-the-art and shown to consistently reduce the word error rates (WERs) of LVCSR tasks when compared to other language models. In this thesis we propose various advances to RNNLMs. The advances are: improved learning procedures for RNNLMs, enhancing the context, and adaptation of RNNLMs. We learned better parameters by a novel pre-training approach and enhanced the context using prosody and syntactic features. We present a pre-training method for RNNLMs, in which the output weights of a feed-forward neural network language model (NNLM) are shared with the RNNLM. This is accomplished by first fine-tuning the weights of the NNLM, which are then used to initialise the output weights of an RNNLM with the same number of hidden units. To investigate the effectiveness of the proposed pre-training method, we have carried out text-based experiments on the Penn Treebank Wall Street Journal data, and ASR experiments on the TED lectures data. Across the experiments, we observe small but significant improvements in perplexity (PPL) and ASR WER. Next, we present unsupervised adaptation of RNNLMs. We adapted the RNNLMs to a target domain (topic or genre or television programme (show)) at test time using ASR transcripts from first pass recognition. We investigated two approaches to adapt the RNNLMs. In the first approach the forward propagating hidden activations are scaled - learning hidden unit contributions (LHUC). In the second approach we adapt all parameters of RNNLM.We evaluated the adapted RNNLMs by showing the WERs on multi genre broadcast speech data. We observe small (on an average 0.1% absolute) but significant improvements in WER compared to a strong unadapted RNNLM model. Finally, we present the context-enhancement of RNNLMs using prosody and syntactic features. The prosody features were computed from the acoustics of the context words and the syntactic features were from the surface form of the words in the context. We trained the RNNLMs with word duration, pause duration, final phone duration, syllable duration, syllable F0, part-of-speech tag and Combinatory Categorial Grammar (CCG) supertag features. The proposed context-enhanced RNNLMs were evaluated by reporting PPL and WER on two speech recognition tasks, Switchboard and TED lectures. We observed substantial improvements in PPL (5% to 15% relative) and small but significant improvements in WER (0.1% to 0.5% absolute).

Les différences entre conditions d'apprentissage et conditions de test peuvent considérablement dégrader la qualité des transcriptions produites par un système de reconnaissance automatique de la parole (RAP). L'adaptation est un moyen efficace pour réduire l'inadéquation entre les modèles du système et les données liées à un locuteur ou un canal acoustique particulier. Il existe deux types dominants de modèles acoustiques utilisés en RAP : les modèles de mélanges gaussiens (GMM) et les réseaux de neurones profonds (DNN). L'approche par modèles de Markov cachés (HMM) combinés à des GMM (GMM-HMM) a été l'une des techniques les plus utilisées dans les systèmes de RAP pendant de nombreuses décennies. Plusieurs techniques d'adaptation ont été développées pour ce type de modèles. Les modèles acoustiques combinant HMM et DNN (DNN-HMM) ont récemment permis de grandes avancées et surpassé les modèles GMM-HMM pour diverses tâches de RAP, mais l'adaptation au locuteur reste très difficile pour les modèles DNN-HMM. L'objectif principal de cette thèse est de développer une méthode de transfert efficace des algorithmes d'adaptation des modèles GMM aux modèles DNN. Une nouvelle approche pour l'adaptation au locuteur des modèles acoustiques de type DNN est proposée et étudiée : elle s'appuie sur l'utilisation de fonctions dérivées de GMM comme entrée d'un DNN. La technique proposée fournit un cadre général pour le transfert des algorithmes d'adaptation développés pour les GMM à l'adaptation des DNN. Elle est étudiée pour différents systèmes de RAP à l'état de l'art et s'avère efficace par rapport à d'autres techniques d'adaptation au locuteur, ainsi que complémentaire<br>Differences between training and testing conditions may significantly degrade recognition accuracy in automatic speech recognition (ASR) systems. Adaptation is an efficient way to reduce the mismatch between models and data from a particular speaker or channel. There are two dominant types of acoustic models (AMs) used in ASR: Gaussian mixture models (GMMs) and deep neural networks (DNNs). The GMM hidden Markov model (GMM-HMM) approach has been one of the most common technique in ASR systems for many decades. Speaker adaptation is very effective for these AMs and various adaptation techniques have been developed for them. On the other hand, DNN-HMM AMs have recently achieved big advances and outperformed GMM-HMM models for various ASR tasks. However, speaker adaptation is still very challenging for these AMs. Many adaptation algorithms that work well for GMMs systems cannot be easily applied to DNNs because of the different nature of these models. The main purpose of this thesis is to develop a method for efficient transfer of adaptation algorithms from the GMM framework to DNN models. A novel approach for speaker adaptation of DNN AMs is proposed and investigated. The idea of this approach is based on using so-called GMM-derived features as input to a DNN. The proposed technique provides a general framework for transferring adaptation algorithms, developed for GMMs, to DNN adaptation. It is explored for various state-of-the-art ASR systems and is shown to be effective in comparison with other speaker adaptation techniques and complementary to them

Self-adaptive systems have limited time to adjust their configurations whenever their adaptation goals, i.e., quality requirements, are violated due to some runtime uncertainties. Within the available time, they need to analyze their adaptation space, i.e., a set of configurations, to find the best adaptation option, i.e., configuration, that can achieve their adaptation goals. Existing formal analysis approaches find the best adaptation option by analyzing the entire adaptation space. However, exhaustive analysis requires time and resources and is therefore only efficient when the adaptation space is small. The size of the adaptation space is often in hundreds or thousands, which makes formal analysis approaches inefficient in large-scale self-adaptive systems. In this thesis, we tackle this problem by presenting an online learning approach that enables formal analysis approaches to analyze large adaptation spaces efficiently. The approach integrates with the standard feedback loop and reduces the adaptation space to a subset of adaptation options that are relevant to the current runtime uncertainties. The subset is then analyzed by the formal analysis approaches, which allows them to complete the analysis faster and efficiently within the available time. We evaluate our approach on two different instances of an Internet of Things application. The evaluation shows that our approach dramatically reduces the adaptation space and analysis time without compromising the adaptation goals.

The application area of Fifth Generation New Radio (5G-NR) called Ultra-Reliable and Low-Latency Communication (URLLC) requires a reliability, the probability of receiving and decoding a data packet correctly, of 1 - 10^5. For this requirement to be fulfilled in a resource-efficient manner, it is necessary to have a good estimation of extremely low quan- tiles of the channel quality distribution, so that appropriate resources can be distributed to users of the network system.  This study proposes and evaluates two methods for estimating extreme quantiles of the downlink channel quality distribution, linear quantile regression and Quantile Regression Neural Network (QRNN). The models were trained on data from Ericsson’s system-level radio network simulator, and evaluated on goodness of fit and resourcefulness. The focus of this study was to estimate the quantiles 10^2, 10^3 and 10^4 of the distribution.  The results show that QRNN generally performs better than linear quantile regression in terms of pseudoR2, which indicates goodness of fit, when the sample size is larger. How- ever, linear quantile regression was more effective for smaller sample sizes. Both models showed difficulty estimating the most extreme quantiles. The less extreme quantile to esti- mate, the better was the resulting pseudoR2-score. For the largest sample size, the resulting pseudoR2-scores of the QRNN was 0.20, 0.12 and 0.07, and the scores of linear quantile regression was 0.16, 0.10 and 0.07 for the respective quantiles 10^2, 10^3 and 10^4.  It was shown that both evaluated models were significantly more resourceful than us- ing the average of the 50 last measures of channel quality subtracted with a fixed back-off value as a predictor. QRNN had the most optimistic predictions. If using the QRNN, theo- retically, on average 43% more data could be transmitted while fulfilling the same reliability requirement than by using the fixed back-off value.

This Master’s thesis deals with adaptation of fuzzy system parameters with main aim on artificial neural network. Current knowledge of methods connecting fuzzy systems and artificial neural networks is discussed in the search part of this work. The search in Student’s works is discussed either. Chapter focused on methods application deals with classifying ability verification of the chosen fuzzy-neural network with Kohonen learning algorithm. Later the model of fuzzy system with parameters adaptation based on fuzzyneural network with Kohonen learning algorithm is shown.

La détection des arcs électriques se produisant dans un réseau électrique par des approches d’apprentissage représente le cœur des travaux exposés dans cette thèse. Le problème a d’abord été vu comme la classification de séries temporelles à taille fixe en deux classes: normal et défaut. Cette première partie s’appuie sur les travaux de la littérature où les algorithmes de détection sont organisés principalement sur une étape de transformation des signaux acquis sur le réseau, suivie d’une étape d’extraction de caractéristiques descriptives et enfin d’une étape de décision. L’approche multicritères adoptée ici a pour objectif de répondre aux imprécisions systématiquement constatées. Une méthodologie de sélection des meilleures combinaisons et de transformation et de descripteurs a été proposée en exploitant des solutions d’apprentissage. La mise au point de descripteurs pertinents étant toujours difficile, l'évaluation des possibilités offertes par l'apprentissage profond a également été étudiée. Dans une seconde phase l’étude a porté sur les aspects variables dans le temps de la détection de défaut. Deux voies statistiques de décision ont été explorées l’une basée sur le test de probabilités séquentiel (SPRT) l’autre basée sur les réseaux de neurones artificiels LSTM (Long Short Time Memory Network) chacune de ces deux méthodes exploite à sa manière la durée d’une première étape de classification comprise entre 0 et 1 (normal, défaut). La décision par SPRT utilise une intégration de la classification initiale. LSTM apprend à classer des données à temps variable. Les résultats du réseau LSTM sont très prometteurs, il reste néanmoins quelques points à approfondir. L’ensemble de ces travaux s’appuie sur des expérimentations avec des données les plus complètes et les plus large possible sur le domaine des réseaux alternatifs 230V dans un contexte domestique et industriel. La précision obtenue approche les 100% dans la majorité des situations<br>The detection of electric arcs occurring in an electrical network by machine learning approaches represents the heart of the work presented in this thesis. The problem was first considered as a classification of fixed-size time series with two classes: normal and default. This first part is based on the work of the literature where the detection algorithms are organized mainly on a step of the transformation of the signals acquired on the network, followed by a step of extraction of descriptive characteristics and finally a step of decision. The multi-criteria approach adopted here aims to respond to systematic classification errors. A methodology for selecting the best combinations, transformation, and descriptors has been proposed by using learning solutions. As the development of relevant descriptors is always difficult, differents solutions offered by deep learning has also been studied. In a second phase, the study focused on the variable aspects in time of the fault detection. Two statistical decision paths have been explored, one based on the sequential probabilistic test (SPRT) and the other based on artificial neural networks LSTM (Long Short Time Memory Network). Each of these two methods exploits in its way the duration a first classification step between 0 and 1 (normal, default). The decision by SPRT uses an integration of the initial classification. LSTM learns to classify data with variable time. The results of the LSTM network are very promising, but there are a few things to explore. All of this work is based on experiments with the most complete and broadest possible data on the field of 230V alternative networks in a domestic and industrial context. The accuracy obtained is close to 100% in the majority of situations

We extend the study of learning and generalization in feed forward Boolean networks to random Boolean networks (RBNs). We explore the relationship between the learning capability and the network topology, the system size, the training sample size, and the complexity of the computational tasks. We show experimentally that there exists a critical connectivity Kc that improves the generalization and adaptation in networks. In addition, we show that in finite size networks, the critical K is a power-law function of the system size N and the fraction of inputs used during the training. We explain why adaptation improves at this critical connectivity by showing that the network ensemble manifests maximal topological diversity near Kc. Our work is partly motivated by self-assembled molecular and nanoscale electronics. Our findings allow to determine an automata network topology class for efficient and robust information processing.

The availability of annotated image datasets and recent advances in supervised deep learning methods are enabling the end-to-end derivation of representative image features that can impact a variety of image analysis problems. These supervised methods use prior knowledge derived from labelled training data and approaches, for example, convolutional neural networks (CNNs) have produced impressive results in natural (photographic) image classification. CNNs learn image features in a hierarchical fashion. Each deeper layer of the network learns a representation of the image data that is higher level and semantically more meaningful. However, the accuracy and robustness of image features with supervised CNNs are dependent on the availability of large-scale labelled training data. In medical imaging, these large labelled datasets are scarce mainly due to the complexity of manual annotation and inter- and intra-observer variability in label assignment. The concept of ‘transfer learning’ – the adoption of image features from different domains, e.g., image features learned from natural photographic images – was introduced to address the lack of large amounts of labelled medical image data. These image features, however, are often generic and do not perform well in specific medical image analysis problems. An alternative approach was to optimise these features by retraining the generic features using a relatively small set of labelled medical images. This ‘fine-tuning’ approach, however, is not able to match the overall accuracy of learning image features directly from large collections of data that are specifically related to the problem at hand. An alternative approach is to use unsupervised feature learning algorithms to build features from unlabelled data, which then allows unannotated image archives to be used. Many unsupervised feature learning algorithms such as sparse coding (SC), auto-encoder (AE) and Restricted Boltzmann Machines (RBMs), however, have often been limited to learning low-level features such as lines and edges. In an attempt to address these limitations, in this thesis, we present several new unsupervised deep learning methods to learn semantic high-level features from unlabelled medical images to address the challenge of learning representative visual features in medical image analysis. We present two methods to derive non-linear and non-parametric models, which are crucial to unsupervised feature learning algorithms; one method embeds a kernel learning within CNNs while the other couples clustering with CNNs. We then further improved the quality of image features using domain adaptation methods (DAs) that learn representations that are invariant to domains with different data distributions. We present a deep unsupervised feature extractor to transform the feature maps from the pre-trained CNN on natural images to a set of non-redundant and relevant medical image features. Our feature extractor preserves meaningful generic features from the pre-trained domain and learns specific local features that are more representative of the medical image data. We conducted extensive experiments on 4 public datasets which have diverse visual characteristics of medical images including X-ray, dermoscopic and CT images. Our results show that our methods had better accuracy when compared to other conventional unsupervised methods and competitive accuracy to methods that used state-of-the-art supervised CNNs. Our findings suggest that our methods could scale to many different transfer learning or domain adaptation approaches where they have none or small sets of labelled data.

We tackle the task of semi-supervised video object segmentation, i.e, pixel-level object classification of the images in video sequences using very limited ground truth training data of its corresponding video. Recently introduced online adaptation of convolutional neural networks for video object segmentation (OnAVOS) has achieved good results by pretraining the network, fine-tuning on the first frame and training the network at test time using its approximate prediction as newly obtained ground truth. We propose Flow Adaptive Video Object Segmentation (FAVOS) that refines the generated adaptive ground truth for online updates and utilizes temporal consistency between video frames with the help of optical flow. We validate our approach on the DAVIS Challenge and achieve rank 1 results on the DAVIS 2016 Challenge (single-object segmentation) and competitive scores on both DAVIS 2018 Semi-supervised Challenge and Interactive Challenge (multi-object segmentation). While most models tend to have increasing complexity for the challenging task of video object segmentation, FAVOS provides a simple and efficient pipeline that produces accurate predictions.

Pour avoir une idée sur la qualité du réseau, la majorité des acteurs concernés (opérateurs réseau, fournisseurs de service) se basent sur la Qualité de Service (Quality of Service). Cette mesure a montré des limites et beaucoup d'efforts ont été déployés pour mettre en place une nouvelle métrique qui reflète, de façon plus précise, la qualité du service offert. Cette mesure s'appelle la qualité d'expérience (Quality of Experience). La qualité d'expérience reflète la satisfaction de l'utilisateur par rapport au service qu'il utilise. Aujourd'hui, évaluer la qualité d'expérience est devenu primordiale pour les fournisseurs de services et les fournisseurs de contenus. Cette nécessité nous a poussés à innover et concevoir des nouvelles méthodes pour estimer la QoE. Dans cette thèse, nous travaillons sur l'estimation de la QoE (1) dans le cas des communications Voix sur IP et (2) dans le cas des services de diffusion Vidéo sur IP. Nous étudions les performances et la qualité des codecs iLBC, Speex et Silk pour la VoIP et les codecs MPEG-2 et H.264/SVC pour la vidéo sur IP. Nous étudions l'impact que peut avoir la majorité des paramètres réseaux, des paramètres sources (au niveau du codage) et destinations (au niveau du décodage) sur la qualité finale. Afin de mettre en place des outils précis d'estimation de la QoE en temps réel, nous nous basons sur la méthodologie Pseudo-Subjective Quality Assessment. La méthodologie PSQA est basée sur un modèle mathématique appelé les réseaux de neurones artificiels. En plus des réseaux de neurones, nous utilisons la régression polynomiale pour l'estimation de la QoE dans le cas de la VoIP.

In the recent years, remote sensing has faced a huge evolution. The constantly growing availability of remote sensing data has opened up new opportunities and laid the foundations for many new challenges. The continuous space missions and new constellations of satellites allow in fact more and more frequent acquisitions, at increasingly higher spatial resolutions, and at an almost total coverage of the globe. The availability of such an huge amount data has highlighted the need for automatic techniques capable of processing the data and exploiting all the available information. Meanwhile, the almost unlimited potential of machine learning has changed the world we live in. Artificial neural Networks have break trough everyday life, with applications that include computer vision, speech processing, autonomous driving but which are also the basis of commonly used tools such as online search engines. However, the vast majority of such models are of the supervised type and therefore their applicability rely on the availability of an enormous quantity of labeled data available to train the models themselves. Unfortunately, this is not the case with remote sensing, in which the enormous amounts of data are opposed to the almost total absence of ground truth. The purpose of this thesis is to find the way to exploit the most recent deep learning techniques, defining a common thread between two worlds, those of remote sensing and deep learning, which is often missing. In particular, this thesis proposes three novel contributions which face current issues in remote sensing. The first one is related to multisensor image registration and combines generative adversarial networks and non-linear optimization of crosscorrelation-like functionals to deal with the complexity of the setting. The proposed method was proved able to outperform state of the art approaches. The second novel contribution faces one of the main issues in deep learning for remote sensing: the scarcity of ground truth data for semantic segmentation. The proposed solution combines convolutional neural networks and probabilistic graphical models, two very active areas in machine learning for remote sensing, and approximate a fully connected conditional random field. The proposed method is capable of filling part of the gap which separate a densely trained model from a weakly trained one. Then, the third approach is aimed at the classification of high resolution satellite images for climate change purposes. It consist of a specific formulation of an energy minimization which allows to fuse multisensor information and the application a markov random field in a fast and efficient way for global scale applications. The results obtained in this thesis shows how deep learning methods based on artificial neural networks can be combined with statistical analysis to overcome their limitations, going beyond the classic benchmark environments and addressing practical, real and large-scale application cases.

The ability to generalize to unseen data is one of the fundamental, desired properties in a learning system. This thesis reports dierent research eorts in improving the generalization properties of machine learning systems at dierent levels, focusing on neural networks for computer vision tasks. First, a novel regularization method is presented, Curriculum Dropout. It combines Curriculum Learning and Dropout, and shows better regularization eects than the original algorithm in a variety of tasks, without requiring substantially any additional implementation eorts. While regularization methods are extremely powerful to better generalize to unseen data from the same distribution as the training one, they are not very successful in mitigating the dataset bias issue. This problem constitutes in models learning the peculiarities of the training set, and poorly generalizing to unseen domains. Unsupervised domain adaptation has been one of the main solutions to this problem. Two novel adaptation approaches are presented in this thesis. First, we introduce the DIFA algorithm, which combines domain invariance and feature augmentation to better adapt models to new domains by relying on adversarial training. Next, we propose an original procedure that exploits the \mode collapse" behavior of Generative Adversarial Networks. Finally, the general applicability of domain adaptation algorithms is questioned (due to the assumptions of knowing the target distribution a priori and being able to sample from it). A novel framework is presented to overcome its liabilities, where the goal is to generalize to unseen domains by relying only on data from a single source distribution. We face this problem through the lens of robust statistics, dening a worst-case formulation where the model parameters are optimized with respect to populations which are -distant from the source domain on a semantic space.

The vehicles of tomorrow will be more sophisticated, intelligent and safe than the vehicles of today. The future is leaning towards fully autonomous vehicles. This degree project provides a data driven solution for a speed adaptation system that can be used to compute a vehicle speed for curves, suitable for the underlying driving style of the driver, road properties and weather conditions. A speed adaptation system for curves aims to compute a vehicle speed suitable for curves that can be used in Advanced Driver Assistance Systems (ADAS) or in Autonomous Driving (AD) applications. This degree project was carried out at Volvo Car Corporation. Literature in the field of speed adaptation systems and factors affecting the vehicle speed in curves was reviewed. Naturalistic driving data was both collected by driving and extracted from Volvo's data base and further processed. A novel speed adaptation system for curves was invented, implemented and evaluated. This speed adaptation system is able to compute a vehicle speed suitable for the underlying driving style of the driver, road properties and weather conditions. Two different artificial neural networks and two mathematical models were used to compute the desired vehicle speed in curves. These methods were compared and evaluated.<br>Morgondagens fordon kommer att vara mer sofistikerade, intelligenta och säkra än dagens fordon. Framtiden lutar mot fullständigt autonoma fordon. Detta examensarbete tillhandahåller en datadriven lösning för ett hastighetsanpassningssystem som kan beräkna ett fordons hastighet i kurvor som är lämpligt för förarens körstil, vägens egenskaper och rådande väder. Ett hastighetsanpassningssystem för kurvor har som mål att beräkna en fordonshastighet för kurvor som kan användas i Advanced Driver Assistance Systems (ADAS) eller Autonomous Driving (AD) applikationer. Detta examensarbete utfördes på Volvo Car Corporation. Litteratur kring hastighetsanpassningssystem samt faktorer som påverkar ett fordons hastighet i kurvor studerades. Naturalistisk bilkörningsdata samlades genom att köra bil samt extraherades från Volvos databas och bearbetades. Ett nytt hastighetsanpassningssystem uppfanns, implementerades samt utvärderades. Hastighetsanpassningssystemet visade sig vara kapabelt till att beräkna en lämplig fordonshastighet för förarens körstil under rådande väderförhållanden och vägens egenskaper. Två olika artificiella neuronnätverk samt två matematiska modeller användes för att beräkna fordonets hastighet. Dessa metoder jämfördes och utvärderades.

Os equipamentos robóticos foram inicialmente criados para atuarem em ambientes industriais fechados. Com o passar do tempo, melhorias foram conquistadas. Atualmente, não se limitam mais à realização de tarefas simples e repetitivas em locais especialmente preparados. Novos equipamentos, capazes de atuarem em ambientes abertos e de realizarem as mais diversas atividades, estão sendo desenvolvidos. Para tanto, é necessário que seus sistemas de controle realizem uma efetiva interação com o mundo onde estão inseridos. Fazem-se necessários, portanto, novos sistemas controladores com capacidade de uma contínua adaptação ao ambiente dinâmico onde operam. As redes neurais artificiais, devido a sua capacidade de tratamento de problemas não lineares – matematicamente difíceis de serem resolvidos, estão sendo empregadas no controle destes processos. O gerenciamento da trajetória de um veículo móvel em ambientes abertos ou fechados é um procedimento altamente não-linear, logo, a aplicação das redes neurais artificiais é bastante promissora. Apesar de sua grande versatilidade, as redes neurais artificiais têm sido utilizadas apenas como sistemas de mapeamento. A grande maioria delas necessita de uma fase de treinamento para que possam armazenar a diversidade de estados possíveis do sistema. Quando atuam, elas simplesmente mapeiam os seus valores de entrada (estado atual) nas soluções previamente armazenadas. Contudo, esta não é a melhor abordagem para os sistemas abertos, ou seja, para os processos cujas situações e possibilidades não podem ser totalmente enumeradas e que podem ser mutáveis no decorrer do tempo. Este trabalho apresenta uma metodologia de controle neural adaptativo para guiar um veículo móvel até o seu destino em ambientes contendo obstáculos fixos ou móveis. Diferentemente das abordagens tradicionais, não existe a necessidade de um treinamento prévio da rede. A rede neural artificial escolhida promove uma contínua adaptação do sistema enquanto atua. Neste processo, são utilizados sensores que fornecem subsídios para que a rede possa gerar, adaptativamente, soluções parciais que façam com que o veículo autônomo se aproxime cada vez mais do seu objetivo, até, finalmente, atingi-lo.<br>The robotic equipments were created initially to actuate in closed industrial environments. Improvements have been acquieved in this area. Nowadays, they are no longer limited to perform simple and repetitive tasks in controlled places. New equipments, capable of acting in open environments and doing the most several activities, are being developed. For so much, it is necessary that its control systems accomplish an effective interaction with the world where they are inserted. Therefore, new systems controllers with capacity of a continuous adaptation to the dynamic environments are essential. Artificial neural networks, due to their capacity of dealing wit non-linear problems – mathematically difficult to be solved – are being used to control these kind of processes. Guide a mobile vehicle through an open or controlled environments is a highly non-linear procedure; therefore, the use of an artificial neural nets is quite promising. In spite of its great versatility, they have just been used as mapping systems. Most of them need a training phase so that they can store the diversity of system’s possible states. When they actuate, they simply map their input values (current state) to the solutions previously stored. However, this is not the best approach for open systems, i.e. systems whose situations and possibilities cannot be totally enumerated and that can change in time. This work presents an adaptive neural control methodology to guide a mobile vehicle to its target in environments with fixed or mobile obstacles. Differently from the traditional approaches, the need of a previous training phase of the neural network doesn't exist. The chosen model of artificial neural net promotes a continuous adaptation of the system while it actuates. Sensors are used to provide informations to the net. This way it generates partial solutions that makes the autonomous vehicle gets closer of its goal, until, finally, reach it.

La segmentation et regroupement en locuteurs (SRL) de collection cherche à répondre à la question « qui parle quand ? » dans une collection de documents multimédia. C’est un prérequis indispensable à l’indexation des contenus audiovisuels. La tâche de SRL consiste d’abord à segmenter chaque document en locuteurs, avant de les regrouper à l'échelle de la collection. Le but est de positionner des labels anonymes identifiant les locuteurs, y compris ceux apparaissant dans plusieurs documents, sans connaître à l'avance ni leur identité ni leur nombre. La difficulté posée par le regroupement en locuteurs à l'échelle d'une collection est le problème de la variabilité intra-locuteur/inter-document : selon les documents, un locuteur peut parler dans des environnements acoustiques variés (en studio, dans la rue...). Cette thèse propose deux méthodes pour pallier le problème. D'une part, une nouvelle méthode de compensation neuronale de variabilité est proposée, utilisant le paradigme de triplet-loss pour son apprentissage. D’autre part, un procédé itératif d'adaptation non supervisée au domaine est présenté, exploitant l'information, même imparfaite, que le système acquiert en traitant des données, pour améliorer ses performances sur le domaine acoustique cible. De plus, de nouvelles méthodes d'analyse en locuteurs des résultats de SRL sont étudiées, pour comprendre le fonctionnement réel des systèmes, au-delà du classique taux d'erreur de SRL (Diarization Error Rate ou DER). Les systèmes et méthodes sont évalués sur deux émissions télévisées d'une quarantaine d'épisodes, pour les architectures de SRL globale ou incrémentale, à l'aide de la modélisation locuteur à l'état de l'art<br>The task of speaker diarization and linking aims at answering the question "who speaks and when?" in a collection of multimedia recordings. It is an essential step to index audiovisual contents. The task of speaker diarization and linking firstly consists in segmenting each recording in terms of speakers, before linking them across the collection. Aim is, to identify each speaker with a unique anonymous label, even for speakers appearing in multiple recordings, without any knowledge of their identity or number. The challenge of the cross-recording linking is the modeling of the within-speaker/across-recording variability: depending on the recording, a same speaker can appear in multiple acoustic conditions (in a studio, in the street...). The thesis proposes two methods to overcome this issue. Firstly, a novel neural variability compensation method is proposed, using the triplet-loss paradigm for training. Secondly, an iterative unsupervised domain adaptation process is presented, in which the system exploits the information (even inaccurate) about the data it processes, to enhance its performances on the target acoustic domain. Moreover, novel ways of analyzing the results in terms of speaker are explored, to understand the actual performance of a diarization and linking system, beyond the well-known Diarization Error Rate (DER). Systems and methods are evaluated on two TV shows of about 40 episodes, using either a global, or longitudinal linking architecture, and state of the art speaker modeling (i-vector)

L’objectif de ces travaux est de fournir un modèle permettant la prédiction de la survie et l’identification de biomarqueurs dans le contexte du myélome multiple (MM) à l’aide de l’imagerie TEP (Tomographie à émission de positons) et de données cliniques. Cette thèse fut divisée en deux parties : La première permet d’obtenir un modèle basé sur les forêts de survie aléatoires (RSF). La seconde est basée sur l’adaptation de l’apprentissage profond à la survie et à nos données. Les contributions principales sont les suivantes : 1) Production d’un modèle basé sur les RSF et les images TEP permettant la prédiction d’un groupe de risque pour les patients atteints de MM.2) Détermination de biomarqueurs grâce à ce modèle3) Démonstration de l’intérêt des radiomiques TEP 4) Extension de l’état de l’art des méthodes d’adaptation de l’apprentissage profond à une petite base de données et à de petitesimages 5) Étude des fonctions de coût utilisées en survie. De plus, nous sommes, à notre connaissance, les premiers à investiguer l’utilisation des RSF dans le contexte du MM et des images TEP, à utiliser du pré-entraînement auto-supervisé avec des images TEP et, avec une tâche de survie, à adapter la fonction de coût triplet à la survie et à adapter un réseau de neurones convolutionnels à la survie du MM à partir de lésions TEP<br>The aim of this work is to provide a model for survival prediction and biomarker identification in the context of multiple myeloma (MM) using PET (Positron Emission Tomography) imaging and clinical data. This PhD is divided into two parts: The first part provides a model based on Random Survival Forests (RSF). The second part is based on the adaptation of deep learning to survival and to our data. The main contributions are the following: 1) Production of a model based on RSF and PET images allowing the prediction of a risk group for multiple myeloma patients. 2) Determination of biomarkers using this model.3) Demonstration of the interest of PET radiomics.4) Extension of the state of the art of methods for the adaptation of deep learning to a small database and small images. 5) Study of the cost functions used in survival. In addition, we are, to our knowledge, the first to investigate the use of RSFs in the context of MM and PET images, to use self-supervised pre-training with PET images, and, with a survival task, to fit the triplet cost function to survival and to fit a convolutional neural network to MM survival from PET lesions

In the field of machine learning (ML) a very commonly encountered problem is the lack of generalizability of learnt classification functions when subjected to new samples that are not representative of the training distribution. The discrepancy between the training (a.k.a. source) and test (a.k.a.target) distributions are caused by several latent factors such as change in appearance, illumination, viewpoints and so on, which is also popularly known as domain-shift. In order to make a classifier cope with such domain-shifts, a sub-field in machine learning called domain adaptation (DA) has emerged that jointly uses the annotated data from the source domain together with the unlabelled data from the target domain of interest. For a classifier to be adapted to an unlabelled target data set is of tremendous practical significance because it has no associated labelling cost and allows for more accurate predictions in the environment of interest. A majority of the DA methods which address the single source and single target domain scenario are not easily extendable to many practical DA scenarios. As there has been as increasing focus to make ML models deployable, it calls for devising improved methods that can handle inherently complex practical DA scenarios in the real world. In this work we build towards this goal of addressing more practical DA settings and help realize novel methods for more real world applications: (i) We begin our work with analyzing and addressing the single source and single target setting by proposing whitening-based embedded normalization layers to align the marginal feature distributions between two domains. To better utilize the unlabelled target data we propose an unsupervised regularization loss that encourages both confident and consistent predictions. (ii) Next, we build on top of the proposed normalization layers and use them in a generative framework to address multi-source DA by posing it as an image translation problem. This proposed framework TriGAN allows a single generator to be learned by using all the source domain data into a single network, leading to better generation of target-like source data. (iii) We address multi-target DA by learning a single classifier for all of the target domains. Our proposed framework exploits feature aggregation with a graph convolutional network to align feature representations of similar samples across domains. Moreover, to counteract the noisy pseudo-labels we propose to use a co-teaching strategy with a dual classifier head. To enable smoother adaptation, we propose a domain curriculum learning ,when the domain labels are available, that adapts to one target domain at a time, with increasing domain gap. (iv) Finally, we address the challenging source-free DA where the only source of supervision is a source-trained model. We propose to use Laplace Approximation to build a probabilistic source model that can quantify the uncertainty in the source model predictions on the target data. The uncertainty is then used as importance weights during the target adaptation process, down-weighting target data that do not lie in the source manifold.

Abstract Artificial Life is an interdisciplinary scientific and engineering enterprise investigating the fundamental properties of living systems through the simulation and synthesis of life-like processes in artificial media. One of the avenues of investigation is autonomous robots and agents. Mimicking of the growth and adaptation of a biological neural circuit in an artificial medium is a challenging task owing to our limited knowledge of the complex process taking place in a living organism. By combining several developmental mechanisms, including the chemical, mechanical, genetic, and electrical, researchers have succeeded in developing networks with interesting topology, morphology, and function within Artificial Computational Chemistry. However, most of these approaches still fail to create neural circuits able to solve real problems in perception and robot control. In this thesis a phenomenological developmental model called a Stochastic Evolutionary Neuron Migration Process (SENMP) is proposed. Employing a spatial encoding scheme with lateral interaction of neurons for artificial neural networks, which represent candidate solutions within a neural network ensemble, neurons of the ensemble form problem-specific spatial patterns with the desired dynamics as they migrate under the selective pressure. The approach is applied to gain new insights into development, adaptation and plasticity in neural networks and to evolve purposeful behaviors for mobile robots. In addition, the approach is used to study the relationship of spatial patterns, composed of interacting entities, and their dynamics. The feasibility and advantages of the approach are demonstrated by evolving neural controllers for solving a non-Markovian double pole balancing problem and by evolving controllers that exhibit navigation behavior for simulated and real mobile robots in complex environments. Preliminary results regarding the behavior of the adapting neural network ensemble are also shown and, particularly, a phenomenon exhibiting Hebbian-like dynamics. This thesis is a step toward a long range goal that aims to create an intelligent robot that is capable of learning complex skills and adapts rapidly to environmental changes.

Learning representations is a central challenge in machine learning. For speech recognition, we are interested in learning robust representations that are stable across different acoustic environments, recording equipment and irrelevant inter– and intra– speaker variabilities. This thesis is concerned with representation learning for acoustic model adaptation to speakers and environments, construction of acoustic models in low-resource settings, and learning representations from multiple acoustic channels. The investigations are primarily focused on the hybrid approach to acoustic modelling based on hidden Markov models and artificial neural networks (ANN). The first contribution concerns acoustic model adaptation. This comprises two new adaptation transforms operating in ANN parameters space. Both operate at the level of activation functions and treat a trained ANN acoustic model as a canonical set of fixed-basis functions, from which one can later derive variants tailored to the specific distribution present in adaptation data. The first technique, termed Learning Hidden Unit Contributions (LHUC), depends on learning distribution-dependent linear combination coefficients for hidden units. This technique is then extended to altering groups of hidden units with parametric and differentiable pooling operators. We found the proposed adaptation techniques pose many desirable properties: they are relatively low-dimensional, do not overfit and can work in both a supervised and an unsupervised manner. For LHUC we also present extensions to speaker adaptive training and environment factorisation. On average, depending on the characteristics of the test set, 5-25% relative word error rate (WERR) reductions are obtained in an unsupervised two-pass adaptation setting. The second contribution concerns building acoustic models in low-resource data scenarios. In particular, we are concerned with insufficient amounts of transcribed acoustic material for estimating acoustic models in the target language – thus assuming resources like lexicons or texts to estimate language models are available. First we proposed an ANN with a structured output layer which models both context–dependent and context–independent speech units, with the context-independent predictions used at runtime to aid the prediction of context-dependent states. We also propose to perform multi-task adaptation with a structured output layer. We obtain consistent WERR reductions up to 6.4% in low-resource speaker-independent acoustic modelling. Adapting those models in a multi-task manner with LHUC decreases WERRs by an additional 13.6%, compared to 12.7% for non multi-task LHUC. We then demonstrate that one can build better acoustic models with unsupervised multi– and cross– lingual initialisation and find that pre-training is a largely language-independent. Up to 14.4% WERR reductions are observed, depending on the amount of the available transcribed acoustic data in the target language. The third contribution concerns building acoustic models from multi-channel acoustic data. For this purpose we investigate various ways of integrating and learning multi-channel representations. In particular, we investigate channel concatenation and the applicability of convolutional layers for this purpose. We propose a multi-channel convolutional layer with cross-channel pooling, which can be seen as a data-driven non-parametric auditory attention mechanism. We find that for unconstrained microphone arrays, our approach is able to match the performance of the comparable models trained on beamform-enhanced signals.

Deep Learning has achieved tremendous success in recent years in several areas such as image classification, text translation, autonomous agents, to name a few. Deep Neural Networks are able to learn non-linear features in a data-driven fashion from complex, large scale datasets to solve tasks. However, some fundamental issues remain to be fixed: the kind of data that is provided to the neural network directly influences its capability to generalize. This is especially true when training and test data come from different distributions (the so called domain gap or domain shift problem): in this case, the neural network may learn a data representation that is representative for the training data but not for the test, thus performing poorly when deployed in actual scenarios. The domain gap problem is addressed by the so-called Domain Adaptation, for which a large literature was recently developed. In this thesis, we first present a novel method to perform Unsupervised Domain Adaptation. Starting from the typical scenario in which we dispose of labeled source distributions and an unlabeled target distribution, we pursue a pseudo-labeling approach to assign a label to the target data, and then, in an iterative way, we refine them using Generative Adversarial Networks. Subsequently, we faced the debiasing problem. Simply speaking, bias occurs when there are factors in the data which are spuriously correlated with the task label, e.g., the background, which might be a strong clue to guess what class is depicted in an image. When this happens, neural networks may erroneously learn such spurious correlations as predictive factors, and may therefore fail when deployed on different scenarios. Learning a debiased model can be done using supervision regarding the type of bias affecting the data, or can be done without any annotation about what are the spurious correlations. We tackled the problem of supervised debiasing -- where a ground truth annotation for the bias is given -- under the lens of information theory. We designed a neural network architecture that learns to solve the task while achieving at the same time, statistical independence of the data embedding with respect to the bias label. We finally addressed the unsupervised debiasing problem, in which there is no availability of bias annotation. we address this challenging problem by a two-stage approach: we first split coarsely the training dataset into two subsets, samples that exhibit spurious correlations and those that do not. Second, we learn a feature representation that can accommodate both subsets and an augmented version of them.

Varying missile configurations may create uncertainty for a missile control algorithm developed with linear control theory, for instance the control system performance requirements may not be satisfied anymore. Missile configuration may change during the missile design period due to variations in subsystem locations, subsystem weights and missile geometry. Likewise, burning propellant, deployment of aerodynamic surfaces and wings with varying sweep angle can be considered as in-flight missile configuration changes. This thesis study addresses development and analysis of an adaptive missile control algorithm to account for the uncertain effects caused by varying missile configuration. Control algorithms, designed using pole placement, are augmented with adaptive neural networks. The resulting controller is a type of model reference adaptive controller. Adaptation characteristics of the augmented control algorithms are investigated to changing center of pressure location and missile geometry. Analyses are performed for three different missile configurations using simulation.

The attempt to evolve complete embodied and situated artificial creatures in which both morphological and control characteristics are adapted during the evolutionary process has been and still represents a long term goal key for the artificial life and the evolutionary robotics community. Loosely inspired by ancient biological organisms which are not provided with a central nervous system and by simple organisms such as stick insects, this thesis proposes a new genotype encoding which allows development and evolution of mor- phology and neural controller in artificial agents provided with a distributed neural network. In order to understand if this kind of network is appropriate for the evolution of non trivial behaviours in artificial agents, two experiments (description and results will be shown in chapter 3) in which evolution was applied only to the controller’s parameters were performed. The results obtained in the first experiment demonstrated how distributed neural networks can achieve a good level of organization by synchronizing the output of oscillatory elements exploiting acceleration/deceleration mechanisms based on local interactions. In the second experiment few variants on the topology of neural architecture were introduced. Results showed how this new control system was able to coordinate the legs of a simulated hexapod robot on two different gaits on the basis of the external circumstances. After this preliminary and successful investigation, a new genotype encoding able to develop and evolve artificial agents with no fixed morphology and with a distributed neural controller was proposed. A second set of experiments was thus performed and the results obtained confirmed both the effectiveness of genotype encoding and the ability of distributed neural network to perform the given task. The results have also shown the strength of genotype both in generating a wide range of different morphological structures and in favouring a direct co-adaptation between neural controller and morphology during the evolutionary process. Furthermore the simplicity of the proposed model has showed the effective role of specific elements in evolutionary experiments. In particular it has demonstrated the importance of the environment and its complexity in evolving non-trivial behaviours and also how adding an independent component to the fitness function could help the evolutionary process exploring a larger space solutions avoiding a premature convergence towards suboptimal solutions.

[ES] El problema conocido como de secuencia a secuencia consiste en transformar una secuencia de entrada en una secuencia de salida. Bajo esta perspectiva se puede atacar una amplia cantidad de problemas, entre los cuales destacan la traducción automática o la descripción automática de objetos multimedia. La aplicación de redes neuronales profundas ha revolucionado esta disciplina, y se han logrado avances notables. Pero los sistemas automáticos todavía producen predicciones que distan mucho de ser perfectas. Para obtener predicciones de gran calidad, los sistemas automáticos se utilizan bajo la supervisión de un humano, quien corrige los errores. Esta tesis se centra principalmente en el problema de la traducción del lenguaje natural, usando modelos enteramente neuronales. Nuestro objetivo es desarrollar sistemas de traducción neuronal más eficientes. asentándonos sobre dos pilares fundamentales: cómo utilizar el sistema de una forma más eficiente y cómo aprovechar datos generados durante la fase de explotación del mismo. En el primer caso, aplicamos el marco teórico conocido como predicción interactiva a la traducción automática neuronal. Este proceso consiste en integrar usuario y sistema en un proceso de corrección cooperativo, con el objetivo de reducir el esfuerzo humano empleado en obtener traducciones de alta calidad. Desarrollamos distintos protocolos de interacción para dicha tecnología, aplicando interacción basada en prefijos y en segmentos, implementados modificando el proceso de búsqueda del sistema. Además, ideamos mecanismos para obtener una interacción con el sistema más precisa, manteniendo la velocidad de generación del mismo. Llevamos a cabo una extensa experimentación, que muestra el potencial de estas técnicas: superamos el estado del arte anterior por un gran margen y observamos que nuestros sistemas reaccionan mejor a las interacciones humanas. A continuación, estudiamos cómo mejorar un sistema neuronal mediante los datos generados como subproducto de este proceso de corrección. Para ello, nos basamos en dos paradigmas del aprendizaje automático: el aprendizaje muestra a muestra y el aprendizaje activo. En el primer caso, el sistema se actualiza inmediatamente después de que el usuario corrige una frase, aprendiendo de una manera continua a partir de correcciones, evitando cometer errores previos y especializándose en un usuario o dominio concretos. Evaluamos estos sistemas en una gran cantidad de situaciones y dominios diferentes, que demuestran el potencial que tienen los sistemas adaptativos. También llevamos a cabo una evaluación humana, con traductores profesionales. Éstos quedaron muy satisfechos con el sistema adaptativo. Además, fueron más eficientes cuando lo usaron, comparados con un sistema estático. El segundo paradigma lo aplicamos en un escenario en el que se deban traducir grandes cantidades de frases, siendo inviable la supervisión de todas. El sistema selecciona aquellas muestras que vale la pena supervisar, traduciendo el resto automáticamente. Aplicando este protocolo, redujimos de aproximadamente un cuarto el esfuerzo humano necesario para llegar a cierta calidad de traducción. Finalmente, atacamos el complejo problema de la descripción de objetos multimedia. Este problema consiste en describir en lenguaje natural un objeto visual, una imagen o un vídeo. Comenzamos con la tarea de descripción de vídeos pertenecientes a un dominio general. A continuación, nos movemos a un caso más específico: la descripción de eventos a partir de imágenes egocéntricas, capturadas a lo largo de un día. Buscamos extraer relaciones entre eventos para generar descripciones más informadas, desarrollando un sistema capaz de analizar un mayor contexto. El modelo con contexto extendido genera descripciones de mayor calidad que un modelo básico. Por último, aplicamos la predicción interactiva a estas tareas multimedia, disminuyendo el esfuerzo necesa<br>[CAT] El problema conegut com a de seqüència a seqüència consisteix en transformar una seqüència d'entrada en una seqüència d'eixida. Seguint aquesta perspectiva, es pot atacar una àmplia quantitat de problemes, entre els quals destaquen la traducció automàtica, el reconeixement automàtic de la parla o la descripció automàtica d'objectes multimèdia. L'aplicació de xarxes neuronals profundes ha revolucionat aquesta disciplina, i s'han aconseguit progressos notables. Però els sistemes automàtics encara produeixen prediccions que disten molt de ser perfectes. Per a obtindre prediccions de gran qualitat, els sistemes automàtics són utilitzats amb la supervisió d'un humà, qui corregeix els errors. Aquesta tesi se centra principalment en el problema de la traducció de llenguatge natural, el qual s'ataca emprant models enterament neuronals. El nostre objectiu principal és desenvolupar sistemes més eficients. Per a aquesta tasca, les nostres contribucions s'assenten sobre dos pilars fonamentals: com utilitzar el sistema d'una manera més eficient i com aprofitar dades generades durant la fase d'explotació d'aquest. En el primer cas, apliquem el marc teòric conegut com a predicció interactiva a la traducció automàtica neuronal. Aquest procés consisteix en integrar usuari i sistema en un procés de correcció cooperatiu, amb l'objectiu de reduir l'esforç humà emprat per obtindre traduccions d'alta qualitat. Desenvolupem diferents protocols d'interacció per a aquesta tecnologia, aplicant interacció basada en prefixos i en segments, implementats modificant el procés de cerca del sistema. A més a més, busquem mecanismes per a obtindre una interacció amb el sistema més precisa, mantenint la velocitat de generació. Duem a terme una extensa experimentació, que mostra el potencial d'aquestes tècniques: superem l'estat de l'art anterior per un gran marge i observem que els nostres sistemes reaccionen millor a les interacciones humanes. A continuació, estudiem com millorar un sistema neuronal mitjançant les dades generades com a subproducte d'aquest procés de correcció. Per a això, ens basem en dos paradigmes de l'aprenentatge automàtic: l'aprenentatge mostra a mostra i l'aprenentatge actiu. En el primer cas, el sistema s'actualitza immediatament després que l'usuari corregeix una frase. Per tant, el sistema aprén d'una manera contínua a partir de correccions, evitant cometre errors previs i especialitzant-se en un usuari o domini concrets. Avaluem aquests sistemes en una gran quantitat de situacions i per a dominis diferents, que demostren el potencial que tenen els sistemes adaptatius. També duem a terme una avaluació amb traductors professionals, qui varen quedar molt satisfets amb el sistema adaptatiu. A més, van ser més eficients quan ho van usar, si ho comparem amb el sistema estàtic. Pel que fa al segon paradigma, l'apliquem per a l'escenari en el qual han de traduir-se grans quantitats de frases, i la supervisió de totes elles és inviable. En aquest cas, el sistema selecciona les mostres que paga la pena supervisar, traduint la resta automàticament. Aplicant aquest protocol, reduírem en aproximadament un quart l'esforç necessari per a arribar a certa qualitat de traducció. Finalment, ataquem el complex problema de la descripció d'objectes multimèdia. Aquest problema consisteix en descriure, en llenguatge natural, un objecte visual, una imatge o un vídeo. Comencem amb la tasca de descripció de vídeos d'un domini general. A continuació, ens movem a un cas més específic: la descripció d''esdeveniments a partir d'imatges egocèntriques, capturades al llarg d'un dia. Busquem extraure relacions entre ells per a generar descripcions més informades, desenvolupant un sistema capaç d'analitzar un major context. El model amb context estés genera descripcions de major qualitat que el model bàsic. Finalment, apliquem la predicció interactiva a aquestes tasques multimèdia, di<br>[EN] The sequence-to-sequence problem consists in transforming an input sequence into an output sequence. A variety of problems can be posed in these terms, including machine translation, speech recognition or multimedia captioning. In the last years, the application of deep neural networks has revolutionized these fields, achieving impressive advances. However and despite the improvements, the output of the automatic systems is still far to be perfect. For achieving high-quality predictions, fully-automatic systems require to be supervised by a human agent, who corrects the errors. This is a common procedure in the translation industry. This thesis is mainly framed into the machine translation problem, tackled using fully neural systems. Our main objective is to develop more efficient neural machine translation systems, that allow for a more productive usage and deployment of the technology. To this end, we base our contributions on two main cornerstones: how to better use of the system and how to better leverage the data generated along its usage. First, we apply the so-called interactive-predictive framework to neural machine translation. This embeds the human agent and the system into a cooperative correction process, that seeks to reduce the human effort spent for obtaining high-quality translations. We develop different interactive protocols for the neural machine translation technology, namely, a prefix-based and a segment-based protocols. They are implemented by modifying the search space of the model. Moreover, we introduce mechanisms for achieving a fine-grained interaction while maintaining the decoding speed of the system. We carried out a wide experimentation that shows the potential of our contributions. The previous state of the art is overcame by a large margin and the current systems are able to react better to the human interactions. Next, we study how to improve a neural system using the data generated as a byproduct of this correction process. To this end, we rely on two main learning paradigms: online and active learning. Under the first one, the system is updated on the fly, as soon as a sentence is corrected. Hence, the system is continuously learning from the corrections, avoiding previous errors and specializing towards a given user or domain. A large experimentation stressed the adaptive systems under different conditions and domains, demonstrating the capabilities of adaptive systems. Moreover, we also carried out a human evaluation of the system, involving professional users. They were very pleased with the adaptive system, and worked more efficiently using it. The second paradigm, active learning, is devised for the translation of huge amounts of data, that are infeasible to being completely supervised. In this scenario, the system selects samples that are worth to be supervised, and leaves the rest automatically translated. Applying this framework, we obtained reductions of approximately a quarter of the effort required for reaching a desired translation quality. The neural approach also obtained large improvements compared with previous translation technologies. Finally, we address another challenging problem: visual captioning. It consists in generating a description in natural language from a visual object, namely an image or a video. We follow the sequence-to-sequence framework, under a a multimodal perspective. We start by tackling the task of generating captions of videos from a general domain. Next, we move on to a more specific case: describing events from egocentric images, acquired along the day. Since these events are consecutive, we aim to extract inter-eventual relationships, for generating more informed captions. The context-aware model improved the generation quality with respect to a regular one. As final point, we apply the intractive-predictive protocol to these multimodal captioning systems, reducing the effort required for correcting the outputs.<br>Section 5.4 describes an user evaluation of an adaptive translation system. This was done in collaboration with Miguel Domingo and the company Pangeanic, with funding from the Spanish Center for Technological and Industrial Development (Centro para el Desarrollo Tecnológico Industrial). [...] Most of Chapter 6 is the result of a collaboration with Marc Bolaños, supervised by Prof. Petia Radeva, from Universitat de Barcelona/CVC. This collaboration was supported by the R-MIPRCV network, under grant TIN2014-54728-REDC.<br>Peris Abril, Á. (2019). Interactivity, Adaptation and Multimodality in Neural Sequence-to-sequence Learning [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/134058<br>TESIS

Artificial neural networks are learning paradigms which mimic the biological neural system. The temporal coding Spiking Neural Network, a relatively new artificial neural network paradigm, is considered to be computationally more powerful than the conventional neural network. Research on the network of spiking neurons is an emerging field and has potential for wider investigation. This research explores alternative learning models with temporal coding spiking neural networks for clustering and classification tasks. Neurons are known to be operating in two modes namely, as integrators and coincidence detectors. Previous temporal coding spiking neural networks, realising spiking neurons as integrators, were utilised for analytical studies. Temporal coding spiking neural networks applied successfully for clustering and classification tasks realised spiking neurons as coincidence detectors and encoded input in formation in the connection delays through a weight adaptation technique. These learning models select suitably delayed connections by enhancing the weights of those connections while weakening the others. This research investigates the learning in temporal coding spiking neural networks with spiking neurons as integrators and coincidence detectors. Focus is given to both supervised and unsupervised learning through weight as well as through delay adaptation. Three novel models for learning in temporal coding spiking neural networks are presented in this research. The first spiking neural network model, Self- Organising Weight Adaptation Spiking Neural Network (SOWA_SNN) realises the spiking neuron as integrator. This model adapts and encodes input information in its connection weights. The second learning model, Self-Organising Delay Adaptation Spiking Neural Network (SODA_SNN) and the third model, Super vised Delay Adaptation Spiking Neural Network (SDA_SNN) realise the spiking neuron as coincidence detector. These two models adapt the connection delays in order to detect temporal patterns through coincidence detection. The first two models were developed for clustering applications and the third for classification tasks. All three models employ Hebbian-based learning rules to update the network connection parameters by utilising the difference between the input and output spike times. The proposed temporal coding spiking neural network models were implemented as discrete models in software and their characteristics and capabilities were analysed through simulations on three bench mark data sets and a high dimensional data set. All three models were able to cluster or classify the analysed data sets efficiently with a high degree of accuracy. The performance of the proposed models, was found to be better than the existing spiking neural network models as well as conventional neural networks. The proposed learning paradigms could be applied to a wide range of applications including manufacturing, business and biomedical domains.

State-of-the-art biochemical systems for medical applications and chemical computing are application-specific and cannot be re-programmed or trained once fabricated. The implementation of adaptive biochemical systems that would offer flexibility through programmability and autonomous adaptation faces major challenges because of the large number of required chemical species as well as the timing-sensitive feedback loops required for learning. Currently, biochemistry lacks a systems vision on how the user-level programming interface and abstraction with a subsequent translation to chemistry should look like. By developing adaptation in chemistry, we could replace multiple hard-wired systems with a single programmable template that can be (re)trained to match a desired input-output profile benefiting smart drug delivery, pattern recognition, and chemical computing. I aimed to address these challenges by proposing several approaches to learning and adaptation in Chemical Reaction Networks (CRNs), a type of simulated chemistry, where species are unstructured, i.e., they are identified by symbols rather than molecular structure, and their dynamics or concentration evolution are driven by reactions and reaction rates that follow mass-action and Michaelis-Menten kinetics. Several CRN and experimental DNA-based models of neural networks exist. However, these models successfully implement only the forward-pass, i.e., the input-weight integration part of a perceptron model. Learning is delegated to a non-chemical system that computes the weights before converting them to molecular concentrations. Autonomous learning, i.e., learning implemented fully inside chemistry has been absent from both theoretical and experimental research. The research in this thesis offers the first constructive evidence that learning in CRNs is, in fact, possible. I have introduced the original concept of a chemical binary perceptron that can learn all 14 linearly-separable logic functions and is robust to the perturbation of rate constants. That shows learning is universal and substrate-free. To simplify the model I later proposed and applied the "asymmetric" chemical arithmetic providing a compact solution for representing negative numbers in chemistry. To tackle more difficult tasks and to serve more complicated biochemical applications, I introduced several key modular building blocks, each addressing certain aspects of chemical information processing and learning. These parts organically combined into gradually more complex systems. First, instead of simple static Boolean functions, I tackled analog time-series learning and signal processing by modeling an analog chemical perceptron. To store past input concentrations as a sliding window I implemented a chemical delay line, which feeds the values to the underlying chemical perceptron. That allows the system to learn, e.g., the linear moving-average and to some degree predict a highly nonlinear NARMA benchmark series. Another important contribution to the area of chemical learning, which I have helped to shape, is the composability of perceptrons into larger multi-compartment networks. Each compartment hosts a single chemical perceptron and compartments communicate with each other through a channel-mediated exchange of molecular species. Besides the feedforward pass, I implemented the chemical error backpropagation analogous to that of feedforward neural networks. Also, after applying mass-action kinetics for the catalytic reactions, I succeeded to systematically analyze the ODEs of my models and derive the closed exact and approximative formulas for both the input-weight integration and the weight update with a learning rate annealing. I proved mathematically that the formulas of certain chemical perceptrons equal the formal linear and sigmoid neurons, essentially bridging neural networks and adaptive CRNs. For all my models the basic methodology was to first design species and reactions, and then set the rate constants either "empirically" by hand, automatically by a standard genetic algorithm (GA), or analytically if possible. I performed all simulations in my COEL framework, which is the first cloud-based chemistry modeling tool, accessible at http://coel-sim.org. I minimized the amount of required molecular species and reactions to make wet chemical implementation possible. I applied an automatized mapping technique, Soloveichik's CRN-to-DNA-strand-displacement transformation, to the chemical linear perceptron and the manual signalling delay line and obtained their full DNA-strand specified implementations. As an alternative DNA-based substrate, I mapped these two models also to deoxyribozyme-mediated cleavage reactions reducing the size of the displacement variant to a third. Both DNA-based incarnations could directly serve as blue-prints for wet biochemicals. Besides an actual synthesis of my models and conducting an experiment in a biochemical laboratory, the most promising future work is to employ so-called reservoir computing (RC), which is a novel machine learning method based on recurrent neural networks. The RC approach is relevant because for time-series prediction it is clearly superior to classical recurrent networks. It can also be implemented in various ways, such as electrical circuits, physical systems, such as a colony of Escherichia Coli, and water. RC's loose structural assumptions therefore suggest that it could be expressed in a chemical form as well. This could further enhance the expressivity and capabilities of chemically-embedded learning. My chemical learning systems may have applications in the area of medical diagnosis and smart medication, e.g., concentration signal processing and monitoring, and the detection of harmful species, such as chemicals produced by cancer cells in a host (cancer miRNAs) or the detection of a severe event, defined as a linear or nonlinear temporal concentration pattern. My approach could replace hard-coded solutions and would allow to specify, train, and reuse chemical systems without redesigning them. With time-series integration, biochemical computers could keep a record of changing biological systems and act as diagnostic aids and tools in preventative and highly personalized medicine.

Submitted by Maike Costa (maiksebas@gmail.com) on 2017-05-30T13:07:14Z No. of bitstreams: 1 - Geraldo Moura -.pdf: 5891648 bytes, checksum: 9ca720ba25a16879f0da2a9a837d1f4b (MD5)<br>Made available in DSpace on 2017-05-30T13:07:14Z (GMT). No. of bitstreams: 1 - Geraldo Moura -.pdf: 5891648 bytes, checksum: 9ca720ba25a16879f0da2a9a837d1f4b (MD5) Previous issue date: 2016-11-21<br>This work deals with pressure control in water distribution networks to promote the optimization of hydraulic loads in order to minimize water losses in the pipes and energy in the corresponding pumping system. Therefore, a neural fuzzy control system (NFCS) beyond the adaptive neural control system (ANCS) were developed. These control systems have been tested and evaluated on experimental bench. The neural fuzzy control system (NFCS) involves techniques of artificial neural network (ANN) and fuzzy logic. The adaptive neural control system (ANCS) used a ANN Perceptron type multilayer by backpropagation technique and gradient descent with Levenberg-Marquardt optimizer. The pressure control will be through the frequency inverter with frequency adjustments in real time, which will act on pump motor assembly installed in the trial bench hydraulic network. Control systems NFCS and ANCS, in this work, were confronted in order to promote a comparative analysis between controllers. The results showed that the ANCS reached a performance index greater than NFCS almost entirely. Finally it was added a logic filter to supervisory control and data acquisition system (SCADA) to make the ANCS able to alternately control the minimum pressure points from the distribution network of experimental bench. Both control systems, ANCS and NFCS were developed in programming environment LabVIEW®<br>Este trabalho tem como objetivo o controle de pressão em redes de distribuição de água, a fim promover a otimização das cargas hidráulicas, buscando minimizar as perdas de água nas tubulações e de energia no correspondente sistema de bombeamento. Para tanto foram elaborados um sistema de controle fuzzy neural (SCFN) e um sistema de controle neural adaptativo (SCNA). Esses sistemas de controle foram testados e avaliados em uma bancada experimental. O sistema de controle fuzzy neural (SCFN) envolve técnicas de rede neural artificial (RNA) e lógica fuzzy. O sistema de controle neural adaptativo (SCNA) utilizou uma RNA do tipo Perceptron de múltiplas camadas, através da técnica de retropropagação (backpropagation) e gradiente descendente com otimizador de Levenberg-Marquardt. O controle de pressão é realizado através do conversor de frequência, com ajustes da frequência, em tempo real (on-line), que atuará sobre conjunto motor bomba (CMB) instalado na rede hidráulica da bancada experimental. Os sistemas de controle SCFN e o SCNA, apresentados neste trabalho, foram confrontados a fim de promover uma análise comparativa entre os controladores. Os resultados demonstraram que o SCNA apresentou especificações superiores ao SCFN em quase sua totalidade. Finalmente foi acrescentado um filtro lógico ao SCADA (supervisory control system and data acquisition) para tornar o SCNA capaz de controlar alternadamente a pressão mínima dentre pontos da rede de distribuição da bancada experimental. Ambos os sistemas de controle, SCFN e SCNA foram desenvolvidos em ambiente de programação LabVIEW®.

Orientador: Douglas Eduardo Zampieri<br>Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Mecanica<br>Made available in DSpace on 2018-08-04T03:12:52Z (GMT). No. of bitstreams: 1 Goncalves_JoaoBosco_D.pdf: 9689870 bytes, checksum: db2fb279a1080765ab5bddf9068a356d (MD5) Previous issue date: 2004<br>Resumo: Este trabalho concebeu um robô bípede composto por uma sucessão de elos rígidos interconectados por 12 articulações rotativas, permitindo movimentos tridimensionais. O robô bípede é constituído por dois subsistemas: tronco e membros inferiores. A modelagem matemática foi realizada em separado para cada um dos subsistemas, que são integrados pelas forças reativas de vínculo. Nossa proposta permite ao robô bípede executar a andadura dinâmica utilizando o tronco para fornecer o balanço dinâmico (estabilidade postural). De forma inédita, foi desenvolvido um gerador automático de trajetória para o tronco que processa as informações de posições e acelerações impostas aos membros inferiores, dotando o robô bípede de reflexos. Foi desenvolvido um gerador de marcha que utiliza a capacidade do robô bípede de executar movimentos tridimensionais, implicando andadura dinamicamente estável sem a efetiva utilização do tronco. O gerador automático de trajetória para o tronco entra em ação se a marcha gerada não mantiver o balanço dinâmico, restabelecendo uma marcha estável. Foi projetado um sistema de controle adaptativo por modelo de referência que utiliza redes neurais artificiais. A avaliação de estabilidade é feita segundo o critério de Lyapunov. O sistema de controle e o gerador automático de trajetórias para o tronco são integrados, compondo os mecanismos adaptativos desenvolvidos para solucionar o modo de andar dinâmico<br>Abstract: The main objective of this work is to project a biped robot machine with a trunk. The mathematical model was realized by considering two sub-systems: the legs and the trunk. The trajectories of the trunk are planned to compensate torques inherent to the dynamic gait, permitting to preserve the dynamic balance of the biped robot. An automatic generator of trajectory for the trunk was developed that processes the infonnation of positions and accelerations imposed to the legs. A gait generator was developed that uses the capacity of the biped robot to execute three-dimensional movements, causing a steady dynamic gait without the effective use of the trunk. The automatic generator of trajectory for the trunk actuates, if the generated do not keep the dynamic balance, reestablishing he steady dynamic gait. A neural network reference model for the adaptive control was projected, which utilizes an RBF neural network and a stability evaluation is based on the criterion of Lyapunov. The system of control and the automatic generator of trajectories for the trunk are integrated, composing the adaptive mechanisms developed to solve the way of dynamic walking<br>Doutorado<br>Mecanica dos Sólidos e Projeto Mecanico<br>Doutor em Engenharia Mecânica

Dans cette thèse, nous étudions le vaste domaine des neurosciences à travers des outils théoriques, numériques et expérimentaux. Nous étudions comment les modèles à taux de décharge peuvent être utilisés pour capturer différents phénomènes observés dans le cerveau. Nous étudions les régimes dynamiques des réseaux couplés de neurones excitateurs (E) et inhibiteurs (I): Nous utilisons une description fournie par un modèle à taux de décharge et la comparons avec les simulations numériques des réseaux de neurones à potentiel d'action décrits par le modèle EIF. Nous nous concentrons sur le régime où le réseau EI présente des oscillations, puis nous couplons deux de ces réseaux oscillants pour étudier la dynamique résultante. La description des différents régimes pour le cas de deux populations est utile pour comprendre la synchronisation d'une chaine de modules E-I et la propagation d'ondes observées dans le cerveau. Nous examinons également les modèles à taux de décharge pour décrire l'adaptation sensorielle: Nous proposons un modèle de ce type pour décrire l'illusion du mouvement consécutif («motion after effect», (MAE)) dans la larve du poisson zèbre. Nous comparons le modèle à taux de décharge avec des données neuronales et comportementales nouvelles<br>In this thesis, we investigate the vast field of neuroscience through theoretical, numerical and experimental tools. We study how rate models can be used to capture various phenomena observed in the brain. We study the dynamical regimes of coupled networks of excitatory (E) and inhibitory neurons (I) using a rate model description and compare with numerical simulations of networks of neurons described by the EIF model. We focus on the regime where the EI network exhibits oscillations and then couple two of these oscillating networks to study the resulting dynamics. The description of the different regimes for the case of two populations is helpful to understand the synchronization of a chain of E-I modules and propagation of waves observed in the brain. We also look at rate models of sensory adaptation. We propose one such model to describe the illusion of motion after effect in the zebrafish larva. We compare this rate model with newly obtained behavioural and neuronal data in the zebrafish larva

Les méthodes d’apprentissage automatique qui reposent sur les Réseaux de Neurones (RNs) ont démontré des performances de prédiction qui s'approchent de plus en plus de la performance humaine dans plusieurs applications du Traitement Automatique de la Langue (TAL) qui bénéficient de la capacité des différentes architectures des RNs à généraliser à partir des régularités apprises à partir d'exemples d'apprentissage. Toutefois, ces modèles sont limités par leur dépendance aux données annotées. En effet, pour être performants, ces modèles neuronaux ont besoin de corpus annotés de taille importante. Par conséquent, uniquement les langues bien dotées peuvent bénéficier directement de l'avancée apportée par les RNs, comme par exemple les formes formelles des langues. Dans le cadre de cette thèse, nous proposons des méthodes d'apprentissage par transfert neuronal pour la construction d'outils de TAL pour les langues peu dotées en exploitant leurs similarités avec des langues bien dotées. Précisément, nous expérimentons nos approches pour le transfert à partir du domaine source des textes formels vers le domaine cible des textes informels (langue utilisée dans les réseaux sociaux). Tout au long de cette thèse nous proposons différentes contributions. Tout d'abord, nous proposons deux approches pour le transfert des connaissances encodées dans les représentations neuronales d'un modèle source, pré-entraîné sur les données annotées du domaine source, vers un modèle cible, adapté par la suite sur quelques exemples annotés du domaine cible. La première méthode transfère des représentations contextuelles pré-entraînées sur le domaine source. Tandis que la deuxième méthode utilise des poids pré-entraînés pour initialiser les paramètres du modèle cible. Ensuite, nous effectuons une série d'analyses pour repérer les limites des méthodes proposées ci-dessus. Nous constatons que, même si l'approche d'apprentissage par transfert proposée améliore les résultats du domaine cible, un transfert négatif « dissimulé » peut atténuer le gain final apporté par l'apprentissage par transfert. De plus, une analyse interprétative du modèle pré-entraîné, montre que les neurones pré-entraînés peuvent être biaisés par ce qu'ils ont appris du domaine source, et donc peuvent avoir des difficultés à apprendre des « patterns » spécifiques au domaine cible. Issu de notre analyse, nous proposons un nouveau schéma d'adaptation qui augmente le modèle cible avec des neurones normalisés, pondérés et initialisés aléatoirement qui permettent une meilleure adaptation au domaine cible tout en conservant les connaissances apprises du domaine source. Enfin, nous proposons une approche d’apprentissage par transfert qui permet de profiter des similarités entre différentes tâches, en plus des connaissances pré-apprises du domaine source<br>Recent approaches based on end-to-end deep neural networks have revolutionised Natural Language Processing (NLP), achieving remarkable results in several tasks and languages. Nevertheless, these approaches are limited with their "gluttony" in terms of annotated data, since they rely on a supervised training paradigm, i.e. training from scratch on large amounts of annotated data. Therefore, there is a wide gap between NLP technologies capabilities for high-resource languages compared to the long tail of low-resourced languages. Moreover, NLP researchers have focused much of their effort on training NLP models on the news domain, due to the availability of training data. However, many research works have highlighted that models trained on news fail to work efficiently on out-of-domain data, due to their lack of robustness against domain shifts. This thesis presents a study of transfer learning approaches, through which we propose different methods to take benefit from the pre-learned knowledge on the high-resourced domain to enhance the performance of neural NLP models in low-resourced settings. Precisely, we apply our approaches to transfer from the news domain to the social media domain. Indeed, despite the importance of its valuable content for a variety of applications (e.g. public security, health monitoring, or trends highlight), this domain is still poor in terms of annotated data. We present different contributions. First, we propose two methods to transfer the knowledge encoded in the neural representations of a source model pretrained on large labelled datasets from the source domain to the target model, further adapted by a fine-tuning on few annotated examples from the target domain. The first transfers contextualised supervisedly pretrained representations, while the second method transfers pretrained weights, used to initialise the target model's parameters. Second, we perform a series of analysis to spot the limits of the above-mentioned proposed methods. We find that even if the proposed transfer learning approach enhances the performance on social media domain, a hidden negative transfer may mitigate the final gain brought by transfer learning. In addition, an interpretive analysis of the pretrained model, show that pretrained neurons may be biased by what they have learned from the source domain, thus struggle with learning uncommon target-specific patterns. Third, stemming from our analysis, we propose a new adaptation scheme which augments the target model with normalised, weighted and randomly initialised neurons that beget a better adaptation while maintaining the valuable source knowledge. Finally, we propose a model, that in addition to the pre-learned knowledge from the high-resource source-domain, takes advantage of various supervised NLP tasks

Cette thèse présente différentes méthodes d'identification d'une loi de commande pour le contrôle de systèmes dynamiques. Ces méthodes sont basées sur l'utilisation de réseaux de neurones artificiels pour l'approximation de fonctions à partir d'exemples. Une synthèse bibliographique des différentes applications des réseaux de neurones pour le contrôle de processus est présentée. Trois types d'utilisation des réseaux de neurones sont décrits : l'identification directe d'un système ou d'un contrôleur à partir d'exemples, l'identification d'un contrôleur grâce à l'algorithme de «rétropropagation à travers le temps» et, enfin, les méthodes d'apprentissage par renforcement. Cette dernière famille d'algorithmes est analysée en détail. Un nouvel algorithme d'apprentissage par renforcement baptisé «B-Learning» est proposé. L'originalité de cet algorithme réside dans l'estimation de «bénéfices» associés aux commandes. Ces bénéfices sont définis comme la variation au cours du temps de la qualité à long terme de l'état du système. Le B-Learning ainsi que d'autres algorithmes d'apprentissage par renforcement sont expérimentés sur un cas d'école, le pendule inverse, ainsi que sur une application industrielle : le contrôle d'une usine de production d'eau potable

Orientador: Rubens Maciel Filho<br>Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Química<br>Made available in DSpace on 2018-08-17T00:14:12Z (GMT). No. of bitstreams: 1 Ribeiro_PleycienneTrajano_D.pdf: 5889372 bytes, checksum: fb40a08b4781dc281483bc320321f3ff (MD5) Previous issue date: 2010<br>Resumo: Este trabalho teve como principal objetivo o desenvolvimento e implementação de um controlador adaptativo do tipo regulador auto-ajustável (STR - Self Tuning Regulator), com a subsequente comparação de seu desempenho com um controlador PID (proporcionalintegrativo-derivativo) e dois controladores preditivos: um preditivo baseado em redes neurais artificiais e um controlador DMC (Dynamic Matrix Control). Esses esquemas de controle foram todos implementados na ferramenta de simulação desenvolvida, o FCCGUI (Fluid Catalytic Cracking Graphical User Interface). Como modelo para estimativa dos parâmetros do controlador adaptativo foi treinada e validada uma rede neural. Esse modelo caixa-preta forneceu uma abordagem eficiente para identificação e controle não-linear do processo de craqueamento catalítico. Para implementação do controlador adaptativo foram estruturadas três novas malhas de controle PID a partir de estudos estatísticos desenvolvidos para a análise dos efeitos das variáveis de processo e suas interações. Dentre essas novas malhas de controle, optou-se pela implementação do controle adaptativo no par manipulada-controlada CTCV-SEVER (abertura de catalisador regenerado - severidade da reação). Após aperfeiçoamentos e reestruturações no simulador FCCGUI, foram realizadas várias simulações para avaliação gráfica e numérica do desempenho do controlador através do critério de desempenho dinâmico ITAE (Integral of Time and Absolute Error). O controlador adaptativo apresentou bons resultados, tanto para testes servo quanto para regulatórios em comparação com a estratégia PID sem adaptação, bem como para as demais estratégias disponíveis no simulador, MPC-RNA (Model Predictive Control baseado em uma Rede Neural Artificial) e DMC. A capacidade de ajuste dos parâmetros do controlador torna-o uma estratégia promissora para sistemas que sofrem com alterações contínuas em suas variáveis de processo ou mudanças de setpoint<br>Abstract: This work had as main objective the development and implementation of an selftuning regulator (STR) adaptive controller, with subsequent comparison of its performance with a PID (proportional-integral-derivative) controller and two predictive controllers, namely a predictive based on artificial neural networks (MPC-ANN) and a dynamic matrix controller (DMC). These control schemes were all implemented in the developed simulation tool, the FCCGUI - Fluid Catalytic Cracking Graphical User Interface. An artificial neural network, used as a model to estimate controller parameters, was trained and validated. This black box model provided an efficient approach for identification and nonlinear control of the catalytic cracking process. To implement the adaptive controller, three new PID control loops were structured based on statistical studies designed to analyze the effects of process variables and their interactions. The implementation of adaptive control was chosen to be in the manipulated-controlled pair CTCV-SEVER (regenerated catalyst valve opening - reaction severity). After restructuring and improvements in the simulator FCCGUI, several simulations were performed for graphical and numerical evaluation of controller performance through ITAE (Integral of Time and Absolute Error) dynamic performance criterion. The adaptive controller presented good results for both tests: servo and regulatory, in comparison with PID strategy without adaptation and other strategies available to the simulator, MPC-ANN and DMC. The ability to adjust the parameters of the controller makes it a promising strategy for systems that suffer from continuous changes in their process variables or setpoints<br>Doutorado<br>Desenvolvimento de Processos Químicos<br>Doutor em Engenharia Química

Made available in DSpace on 2014-12-17T14:56:18Z (GMT). No. of bitstreams: 1 TiagoFBS_DISSERT.pdf: 1617004 bytes, checksum: 989c53485329a28af611291f87ca09f0 (MD5) Previous issue date: 2014-02-28<br>Coordena??o de Aperfei?oamento de Pessoal de N?vel Superior<br>Nowadays, optic fiber is one of the most used communication methods, mainly due to the fact that the data transmission rates of those systems exceed all of the other means of digital communication. Despite the great advantage, there are problems that prevent full utilization of the optical channel: by increasing the transmission speed and the distances involved, the data is subjected to non-linear inter symbolic interference caused by the dispersion phenomena in the fiber. Adaptive equalizers can be used to solve this problem, they compensate non-ideal responses of the channel in order to restore the signal that was transmitted. This work proposes an equalizer based on artificial neural networks and evaluates its performance in optical communication systems. The proposal is validated through a simulated optic channel and the comparison with other adaptive equalization techniques<br>A fibra ?ptica ? um dos meios de comunica??o mais utilizados atualmente, principalmente devido ao fato da taxa de transmiss?o de dados desses sistemas excederem as de todos os outros meios de comunica??o digital. Apesar desta grande vantagem, existem problemas que impedem o total aproveitamento do canal ?ptico: com o aumento da velocidade de transmiss?o e das dist?ncias envolvidas, os dados ficam sujeitos a interfer?ncia intersimb?lica n?o linear, causada pelos fen?menos de dispers?o na fibra ?ptica. Para solucionar esse problema podem ser utilizados equalizadores adaptativos, que compensam respostas n?o ideais do canal, com o intuito de restaurar o sinal que foi transmitido. Neste trabalho apresentamos uma proposta de equalizador baseado em redes neurais artificiais e avaliamos seu desempenho em sistemas de comunica??o ?ptica. A proposta ? validada em um canal ?ptico simulado e comparada a outras t?cnicas de equaliza??o adaptativa

Dissertação (mestrado)—Universidade de Brasília, Faculdade de Tecnologia, Departamento de Engenharia Elétrica, 2018.<br>Submitted by Fabiana Santos (fabianacamargo@bce.unb.br) on 2018-09-27T20:19:43Z No. of bitstreams: 1 2018_LucasGuilhemdeMatos_RESUMO.pdf: 59314 bytes, checksum: fcca0edce88c4bbe8975f96cfb38ac6f (MD5)<br>Rejected by Fabiana Santos (fabianacamargo@bce.unb.br), reason: O arquivo PDF está errado. on 2018-10-08T21:05:29Z (GMT)<br>Submitted by Fabiana Santos (fabianacamargo@bce.unb.br) on 2018-10-08T21:06:22Z No. of bitstreams: 1 2018_LucasGuilhemdeMatos.pdf: 22624129 bytes, checksum: 8246b13bfdcea5ae4862864196f406be (MD5)<br>Approved for entry into archive by Fabiana Santos (fabianacamargo@bce.unb.br) on 2018-10-09T20:24:27Z (GMT) No. of bitstreams: 1 2018_LucasGuilhemdeMatos.pdf: 22624129 bytes, checksum: 8246b13bfdcea5ae4862864196f406be (MD5)<br>Made available in DSpace on 2018-10-09T20:24:27Z (GMT). No. of bitstreams: 1 2018_LucasGuilhemdeMatos.pdf: 22624129 bytes, checksum: 8246b13bfdcea5ae4862864196f406be (MD5) Previous issue date: 2018-08-24<br>Fundação de Apoio a Pesquisa do Distrito Federal (FAP-DF).<br>Este trabalho propõe um contolador adaptativo utilizando redes neuras e aprendizado por reforço para lidar com não-linearidades e variância no tempo. Para a realização de testes, um sistema de nível de líquidos de quarta ordem foi escolhido por apresentar uma gama de constantes de tempo e por possibilitar a mudança de parâmetros. O sistema foi identificado com redes neurais para prever estados futuros com o objetivo de compensar o atraso e melhorar a performance do controlador. Diversos testes foram realizados com diversas redes neurais para decidir qual rede neural seria utilizada para cada tarefa pertinente ao controlador. Os parâmetros do controlador foram ajustados e testados para que o controlador pudesse alcançar parâmetros arbitrários de performance. O controlador foi testado e comparado com o PI tradicional para validação e mostrou caracteristicas adaptativas e melhoria de performance ao longo do tempo, além disso, o controlador desenvolvido não necessita de informação prévia do sistema.<br>This work presents a proposal of an adaptive controller using reinforcement learning and neural networks in order to deal with non-linearities and time-variance. To test the controller a fourth-order fluid level system was chosen because of its great range of time constants and the possibility of varying the system parameters. System identification was performed to predict future states of the system, bypass delay and enhance the controller’s performance. Several tests with different neural networks were made in order to decide which network would be assigned to which task. Various parameters of the controller were tested and tuned to achieve a controller that satisfied arbitrary specifications. The controller was tested against a conventional PI controller used as reference and has shown adaptive features and improvement during execution. Also, the proposed controller needs no previous information on the system in order to be designed.

Made available in DSpace on 2014-06-11T19:22:34Z (GMT). No. of bitstreams: 0 Previous issue date: 2013-02-28Bitstream added on 2014-06-13T19:28:06Z : No. of bitstreams: 1 santosjunior_cr_me_ilha.pdf: 674616 bytes, checksum: 0eb6d5afdb466f77cd53efea2c4e4db7 (MD5)<br>O problema de mensagens não solicitadas pelos usuários em meios de comunicação eletrônica, apesar de ter surgido antes mesmo da popularização da Internet, ainda é um assunto preocupante. Desperdício de largura de banda, perda de tempo, de produtividade e de dados, ou atraso na leitura de e-mails legítimos, são alguns dos problemas que as mensagens não solicitadas, ou Spams, podem causar. Diversas técnicas de filtragem automática de e-mails são apresentadas na literatura, porém muitas destas não oferecem a possibilidade de adaptação, já que o problema em sistemas reais tem como um de seus principais aspectos ser dinâmico, ou seja, mudar constantemente de características com intuito de evadir as técnicas de filtragem. Neste trabalho é desenvolvido um filtro anti-spam utilizando uma técnica de préprocessamento disponível na literatura, no qual os e-mails são submetidos à extração e seleção de características; e uma Rede Neural Artificial baseada na Teoria da Ressonância Adaptativa, para detecção e classificação de Spams. Tais redes neurais possuem grande capacidade de generalização e adaptabilidade, características importantes para um bom desempenho de filtros anti-spam. O modelo proposto neste trabalho é testado a fim de se validar a eficiência do filtro.<br>The problem in receiving non desired messages in electronic communication systems is a very hard task; even it has begun before the popularization of Internet. The problems that these kinds of messages can cause are among others: waste of time, waste of band width, productivity and data or delay in reading the real e-mails. Several e-mail automatic filtering techniques are presented in the literature, however many of them without capacity of adaptation, while the problem in real systems must be dynamical, i.e. avoid filtering techniques. This work develops a SPAM filtering using a pre processing technique available in the literature, where the e-mails are submitted to extract and select the characteristics; and a neural network based on the resonance adaptive theory to detect and classify the SPAMS. These neural networks have capacity in generalization and adaptation, important characteristics of good performance of SPAM filters. The proposed model is submitted to several tests to validate the efficiency of the filter.

Made available in DSpace on 2014-06-11T19:30:52Z (GMT). No. of bitstreams: 0 Previous issue date: 2005-10-27Bitstream added on 2014-06-13T20:00:56Z : No. of bitstreams: 1 lopes_mlm_dr_ilha.pdf: 1509538 bytes, checksum: 3842df54e0429972a030219c885bd09a (MD5)<br>Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)<br>Nos dias atuais, principalmente pelo fato de alguns sistemas serem desregulamentados, o estudo dos problemas de análise, planejamento e operação de sistemas de energia elétrica é de extrema importância para o funcionamento do sistema. Para isso é necessário que se obtenha, com antecedência, o comportamento da carga elétrica com o propósito de garantir o fornecimento de energia aos consumidores de forma econômica, segura e contínua. Este trabalho propõe o desenvolvimento de redes neurais artificiais utilizadas para resolver o problema de previsão de cargas elétricas. Para tanto, inicialmente, propôs-se a introdução de melhorias na rede neural feedforward com treinamento realizado utilizando o algoritmo retropropagação. Neste caso, foi desenvolvida/implementada a adaptação dos parâmetros de inclinação e translação da função sigmóide (função de ativação da rede neural). A inclusão desta nova estrutura de redes neurais produziu melhores resultados, se comparado à rede neural retropropagação convencional. Essas arquiteturas proporcionam bons resultados, porém, são estruturas de redes neurais que possuem o problema de convergência. O problema de previsão de cargas elétricas a curto-prazo necessita de uma rede neural que forneça uma saída de forma rápida e eficaz. No intuito de solucionar os problemas encontrados com o algoritmo retropropagação foi desenvolvida/implementada uma rede neural baseada na arquitetura ART (Adaptive Rossonance Theory), denominada rede neural ART&ARTMAP nebulosa, aplicada ao problema de previsão de carga elétrica. Trata-se, por conseguinte, da principal contribuição desta tese. As redes neurais, baseadas na arquitetura ART, possuem duas características fundamentais que são de extrema importância para o desempenho da rede (estabilidade e plasticidade), que permite a implementação do treinamento de modo contínuo...<br>Nowadays due to the deregulamentation it is very important to study the problems of analyzing, planning and operation of electric power systems. For a reliable operation it is necessary to know previously the behavior of the load to guarantee the energy providing to the users with security and continuity and in an economic way. This work proposes to develop artificial neural networks to solve the problem of electric load forecasting. First, it is introduced some improvements on the feedforward neural network, with the training effectuated with the backpropagation algorithm. The improvement was the adaptation of the inclination and translation parameters of the sigmoid function (activation function of the neural network). The inclusion of this new structure provides better results if compared to the conventional backpropagation algorithm. These architectures provide good results, although they are structures that have some convergence problems. The short term electric load forecasting problem needs a neural network that provide a fast and efficient output. To solve this problem a neural network based on the ART (Adaptive Ressonance Theory), called_ fuzzy ART&ARTMAP applied to the load-forecasting problem, was developed and implemented._This is one of the contributions of this work. Neural networks based on the ART architecture have two important characteristics for the network performance, which are stability and plasticity, allowing the continuous training. The fuzzy ART&ARTMAP neural network reduces the imprecision of the results by a mechanism that separates the binary and analogical data and processing them separately. This represents a quality and an improvement on the results (reduction of the processing time and better precision), if compared to the neural network with backpropagation training (often considered as a benchmark in precision by the specialized...(Complete abastract click electronic access below)

Submitted by Rosivalda Pereira (mrs.pereira@ufma.br) on 2017-07-18T19:31:22Z No. of bitstreams: 1 MadsonMachado.pdf: 3046442 bytes, checksum: 71cc6800f83fdbf38b97607067653f63 (MD5)<br>Made available in DSpace on 2017-07-18T19:31:22Z (GMT). No. of bitstreams: 1 MadsonMachado.pdf: 3046442 bytes, checksum: 71cc6800f83fdbf38b97607067653f63 (MD5) Previous issue date: 2017-05-26<br>The need to increase industrial productivity coupled with quality and low cost requirements has generated a demand for the development of high performance controllers. Motivated by this demand, we presented in this work models, algorithms and a methodology for the online project of high-performance control systems. The models have characteristics of adaptability through adaptive control system architectures. The models developed were based on artificial neural networks of radial basis function type, for the online project of model reference adaptive control systems associated with the of sliding modes control. The algorithms and the embedded system developed for the online project were evaluated for tracking mobile targets, in this case, the solar radiation. The control system has the objective of keeping the surface of the photovoltaic module perpendicular to the solar radiation, in this way the energy generated by the module will be as high as possible. The process consists of a photovoltaic panel coupled in a structure that rotates around an axis parallel to the earth’s surface, positioning the panel in order to capture the highest solar radiation as function of its displacement throughout the day.<br>A necessidade de aumentar a produtividade industrial, associada com os requisitos de qualidade e baixo custo, gerou uma demanda para o desenvolvimento de controladores de alto desempenho. Motivado por esta demanda, apresentou-se neste trabalho modelos, algoritmos e uma metodologia para o projeto online de sistemas de controle de alto desempenho. Os modelos apresentam características de adaptabilidade por meio de arquiteturas de sistemas de controle adaptativo. O desenvolvimento de modelos, baseia-se em redes neurais artificiais (RNA), do tipo função de base radial (RBF, radial basis function), para o projeto online de sistemas de controle adaptativo do tipo modelo de referência associado com o controle de modos deslizantes (SMC, sliding mode control). Os algoritmos e o sistema embarcado desenvolvidos para o projeto online são avaliados para o rastreamento de alvos móveis, neste caso, o rastreamento da radiação solar. O sistema de controle tem o objetivo de manter a superfície do módulo fotovoltaico perpendicular à radiação solar, pois dessa forma a energia gerada pelo módulo será a maior possível. O processo consiste de um painel fotovoltaico acoplado em uma estrutura que gira em torno de um eixo paralelo à superfície da terra, posicionando o painel de forma a capturar a maior radiação solar em função de seu deslocamento ao longo do dia.

É investigado o uso de redes neurais aplicadas à equalização de canais de comunicação, sendo consideradas três tipos de redes: MLP (Multilayer Perceptron), RBF (Radial Basis Function) e RNN (Recurrent Neural Network). Os equalizadores não-lineares baseados nestas redes foram comparados com o equalizador linear transversal e com os equalizadores ótimos segundo os critérios de Bayes e da máxima verossimilhança. Nestas comparações foram utilizados um alfabeto binário e um quaternário transmitidos em modelos de canais cuja resposta ao pulso unitário é finita. Além das versões usuais de equalizadores, foram consideradas versões com realimentação de decisões sempre que isso se mostrou adequado. O treinamento desses equalizadores foi feito de forma supervisionada, ou seja, na fase de treinamento a seqüência de símbolos transmitida era conhecida no receptor. Além disso, foi realizado um estudo comparativo dos algoritmos de treinamento das redes. Neste âmbito, foi obtido um algoritmo do tipo acelerador para o treinamento de redes MLP. Com o intuito de se obter uma estrutura não-linear menos complexa e mais flexível, foi proposto ainda um equalizador híbrido constituído de uma combinação do equalizador linear e da rede RNN que faz uso de realimentação de decisões. Resultados de simulações indicam que o seu uso pode ser vantajoso tanto para canais não-lineares como lineares.<br>Equalization of communication channels using neural networks is investigated by considering three kinds of networks: MLP (Multilayer Perceptron), RBF (Radial Basis Function) and RNN (Recurrent Neural Network). The performance of the nonlinear equalizers based on these networks are compared with the linear transversal equalizer and the optimal equalizers given by the bayesian and maximum likelihood criteria. Binary and quaternary alphabets are used and transmitted over finite pulse response channel models. Decision feedback is considered whenever it is worthwhile. The training of these equalizers is considered in the supervised form and a comparison of some training algorithms has been performed. In this scope, a new algorithm based on parameter acceleration is introduced for the training of MLP networks. Moreover, a hybrid equalizer composed of a linear transversal equalizer and a RNN network is proposed. It is a simple and flexible nonlinear structure making use of decision feedback. imulation results show that it may be advantageously used to equalize linear and nonlinear channels.

Neste trabalho, o equacionamento dinâmico de um manipulador espacial de base livre flutuante é descrito a partir do conceito do manipulador dinamicamente equivalente para que as técnicas de controle desenvolvidas sejam experimentalmente validadas em um manipulador convencional de base fixa. Dois tipos de controle de movimento são considerados. O primeiro foi desenvolvido no espaço das juntas e realiza o comando direto de posicionamento das juntas do manipulador; o segundo foi desenvolvido no espaço inercial e o controle é direcionado para o posicionamento do efetuador no espaço Cartesiano. Nos dois casos, o problema de acompanhamento de trajetória de um manipulador espacial com base livre flutuante sujeito a incertezas na planta e perturbações externas é proposto e solucionado sob o ponto de vista do critério de desempenho H \'INFINITO\'. Considerando métodos de controle para sistemas subatuados, três técnicas adaptativas foram desenvolvidas a partir de um controlador H \'INFINITO\' não-linear baseado na teoria dos jogos. A primeira técnica foi proposta considerando a estrutura do modelo bem definida, porém calculada com base em parâmetros incertos. Uma lei adaptativa foi aplicada para estimar esses parâmetros utilizando parametrização linear. Redes neurais artificiais são aplicadas nas outras duas abordagens adaptativas. A primeira utiliza uma rede neural para aprender o comportamento dinâmico do sistema robótico, considerado totalmente desconhecido. Nenhum dado cinemático ou dinâmico da base é utilizado neste caso. A segunda abordagem considera a estrutura do modelo nominal do manipulador bem definida e a rede neural é aplicada para estimar o comportamento das incertezas paramétricas e da dinâmica não-modelada da base. O critério H \'INFINITO\' é aplicado nas três técnicas para atenuar o efeito dos erros de estimativa. Resultados experimentais foram obtidos com um robô manipulador de base fixa subatuado (UArmII) e apresentaram melhor desempenho no acompanhamento da trajetória e no consumo de energia para as abordagens baseadas em redes neurais.<br>In the present work, the dynamics of a free-floating space manipulator is described through the dynamically equivalent manipulator approach in order to obtain experimental results in a planar fixed base manipulator. Control in joint and Cartesian spaces are considered. The first acts directly on joints positioning; the second control scheme acts on positioning the end-effector in some inertially fixed position. In both cases, the problem of tracking control with a guaranteed H-infinity performance for free-floating manipulator systems with plant uncertainties and external disturbances is proposed and solved. Considering control methods for underactuated systems, three adaptive techniques were developed from a nonlinear H-infinity controller based on game theory. The first approach was proposed considering a well defined structure for the plant, however it was computed based on uncertain parameters. An adaptive law was applied to estimate these parameters using linear parametrization. Artificial neural networks were applied in the two other approaches. The first one uses a neural network to learn the dynamic behavior from the robotic system, which is considered totally unknown. No kinematics or dynamics data from the spacecraft are necessary in this case. The second approach considers the nominal model structure well defined and the neural network is applied to estimate the behavior of the parametric uncertainties and of the spacecraft non-modeled dynamics. The H-infinity criterion was applied to attenuate the effect of estimation errors in the three techniques. Experimental results were obtained with an underactuated fixed-base planar manipulator (UArmII) and presented better performance in tracking and energy consumption for the neural based approaches.

In this thesis, we propose a new framework for the generation of training data for machine learning techniques used for classification in communications applications. Machine learning-based signal classifiers do not generalize well when training data does not describe the underlying probability distribution of real signals. The simplest way to accomplish statistical similarity between training and testing data is to synthesize training data passed through a permutation of plausible forms of noise. To accomplish this, a framework is proposed that implements arbitrary channel conditions and baseband signals. A dataset generated using the framework is considered, and is shown to be appropriately sized by having $11\%$ lower entropy than state-of-the-art datasets. Furthermore, unsupervised domain adaptation can allow for powerful generalized training via deep feature transforms on unlabeled evaluation-time signals. A novel Deep Reconstruction-Classification Network (DRCN) application is introduced, which attempts to maintain near-peak signal classification accuracy despite dataset bias, or perturbations on testing data unforeseen in training. Together, feature transforms and diverse training data generated from the proposed framework, teaching a range of plausible noise, can train a deep neural net to classify signals well in many real-world scenarios despite unforeseen perturbations.

L'analyse de l'imagerie satellitaire et aérienne d'observation de la Terre nous permet d'obtenir des informations précises sur de vastes zones. Une analyse multitemporelle de telles images est nécessaire pour comprendre l'évolution de ces zones. Dans cette thèse, les réseaux de neurones convolutifs sont utilisés pour détecter et comprendre les changements en utilisant des images de télédétection provenant de diverses sources de manière supervisée et faiblement supervisée. Des architectures siamoises sont utilisées pour comparer des paires d'images recalées et identifier les pixels correspondant à des changements. La méthode proposée est ensuite étendue à une architecture de réseau multitâche qui est utilisée pour détecter les changements et effectuer une cartographie automatique simultanément, ce qui permet une compréhension sémantique des changements détectés. Ensuite, un filtrage de classification et un nouvel algorithme de diffusion anisotrope guidée sont utilisés pour réduire l'effet du bruit d'annotation, un défaut récurrent pour les ensembles de données à grande échelle générés automatiquement. Un apprentissage faiblement supervisé est également réalisé pour effectuer une détection de changement au niveau des pixels en utilisant uniquement une supervision au niveau de l'image grâce à l'utilisation de cartes d'activation de classe et d'une nouvelle couche d'attention spatiale. Enfin, une méthode d'adaptation de domaine fondée sur un entraînement adverse est proposée. Cette méthode permet de projeter des images de différents domaines dans un espace latent commun où une tâche donnée peut être effectuée. Cette méthode est testée non seulement pour l'adaptation de domaine pour la détection de changement, mais aussi pour la classification d'images et la segmentation sémantique, ce qui prouve sa polyvalence<br>The analysis of satellite and aerial Earth observation images allows us to obtain precise information over large areas. A multitemporal analysis of such images is necessary to understand the evolution of such areas. In this thesis, convolutional neural networks are used to detect and understand changes using remote sensing images from various sources in supervised and weakly supervised settings. Siamese architectures are used to compare coregistered image pairs and to identify changed pixels. The proposed method is then extended into a multitask network architecture that is used to detect changes and perform land cover mapping simultaneously, which permits a semantic understanding of the detected changes. Then, classification filtering and a novel guided anisotropic diffusion algorithm are used to reduce the effect of biased label noise, which is a concern for automatically generated large-scale datasets. Weakly supervised learning is also achieved to perform pixel-level change detection using only image-level supervision through the usage of class activation maps and a novel spatial attention layer. Finally, a domain adaptation method based on adversarial training is proposed, which succeeds in projecting images from different domains into a common latent space where a given task can be performed. This method is tested not only for domain adaptation for change detection, but also for image classification and semantic segmentation, which proves its versatility

Orientador: Anna Diva Plasencia Lotufo<br>Coorientador: Maria do Carmo Gomes da Silveira<br>Banca: Mara Lúcia Martins Lopes<br>Banca: Benedito Isaias de Lima Lopes<br>Resumo: O problema de mensagens não solicitadas pelos usuários em meios de comunicação eletrônica, apesar de ter surgido antes mesmo da popularização da Internet, ainda é um assunto preocupante. Desperdício de largura de banda, perda de tempo, de produtividade e de dados, ou atraso na leitura de e-mails legítimos, são alguns dos problemas que as mensagens não solicitadas, ou Spams, podem causar. Diversas técnicas de filtragem automática de e-mails são apresentadas na literatura, porém muitas destas não oferecem a possibilidade de adaptação, já que o problema em sistemas reais tem como um de seus principais aspectos ser dinâmico, ou seja, mudar constantemente de características com intuito de evadir as técnicas de filtragem. Neste trabalho é desenvolvido um filtro anti-spam utilizando uma técnica de préprocessamento disponível na literatura, no qual os e-mails são submetidos à extração e seleção de características; e uma Rede Neural Artificial baseada na Teoria da Ressonância Adaptativa, para detecção e classificação de Spams. Tais redes neurais possuem grande capacidade de generalização e adaptabilidade, características importantes para um bom desempenho de filtros anti-spam. O modelo proposto neste trabalho é testado a fim de se validar a eficiência do filtro.<br>Abstract: The problem in receiving non desired messages in electronic communication systems is a very hard task; even it has begun before the popularization of Internet. The problems that these kinds of messages can cause are among others: waste of time, waste of band width, productivity and data or delay in reading the real e-mails. Several e-mail automatic filtering techniques are presented in the literature, however many of them without capacity of adaptation, while the problem in real systems must be dynamical, i.e. avoid filtering techniques. This work develops a SPAM filtering using a pre processing technique available in the literature, where the e-mails are submitted to extract and select the characteristics; and a neural network based on the resonance adaptive theory to detect and classify the SPAMS. These neural networks have capacity in generalization and adaptation, important characteristics of good performance of SPAM filters. The proposed model is submitted to several tests to validate the efficiency of the filter.<br>Mestre