This document describes image compression using different types of neural networks. Features of neural networks like convolutional and recurrent networks are also discussed here. The document contains detailed description of various neural network architectures and their inner workings. In addition, experiments are carried out on various neural network structures and parameters in order to find the most appropriate properties for image compression. Also, there are proposed new concepts for image compression using neural networks that are also immediately tested. Finally, a network of the best concepts and parts discovered during experimentation is designed.

Lorsqu'on effectue une tâche, les circuits neuronaux doivent représenter et manipuler des stimuli continus à l'aide de potentiels d'action discrets. On suppose communément que les neurones représentent les quantités continues à l'aide de leur fréquence de décharge et ceci indépendamment les un des autres. Cependant, un tel codage indépendant est inefficace puisqu'il exige la génération d'un très grand nombre de potentiels d'action pour atteindre un certain niveau de précision. Dans ces travaux, on montre que les neurones d'un réseau récurrent peuvent apprendre - à l'aide d'une règle de plasticité locale - à coordonner leurs potentiels d'actions afin de représenter l'information avec une très haute précision tout en déchargeant de façon minimale. La règle d'apprentissage qui agit sur les connexions récurrentes, conduit à un codage efficace en imposant au niveau de chaque neurone un équilibre précis entre excitation et inhibition. Cet équilibre est un phénomène fréquemment observer dans le cerveau et c'est un principe central de notre théorie. On dérive également deux autres règles d'apprentissages biologiquement plausibles qui permettent respectivement au réseau de s'adapter aux statistiques de ses entrées et d'effectuer des transformations complexes et dynamiques sur elles. Finalement, dans ces réseaux, le stochasticité du temps de décharge d'un neurone n'est pas la signature d'un bruit mais au contraire de précision et d'efficacité. Le caractère aléatoire du temps de décharge résulte de la dégénérescence de la représentation. Ceci constitue donc une interprétation radicalement différente et nouvelle de l'irrégularité trouvée dans des trains de potentiels d'actions. / When performing a task, neural circuits must represent and manipulate continuous stimuli using discrete action potentials. It is commonly assumed that neurons represent continuous quantities with their firing rate and this independently from one another. However, such independent coding is very inefficient because it requires the generation of a large number of action potentials in order to achieve a certain level of accuracy. We show that neurons in a spiking recurrent network can learn - using a local plasticity rule - to coordinate their action potentials in order to represent information with high accuracy while discharging minimally. The learning rule that acts on recurrent connections leads to such an efficient coding by imposing a precise balance between excitation and inhibition at the level of each neuron. This balance is a frequently observed phenomenon in the brain and is central in our work. We also derive two biologically plausible learning rules that respectively allows the network to adapt to the statistics of its inputs and to perform complex and dynamic transformations on them. Finally, in these networks, the stochasticity of the spike timing is not a signature of noise but rather of precision and efficiency. In fact, the random nature of the spike times results from the degeneracy of the representation. This constitutes a new and a radically different interpretation of the irregularity found in spike trains.

We demonstrate the success of recurrent neural networks in gene network inference and expression prediction using a hybrid of particle swarm optimization and differential evolution to overcome the classic obstacle of local minima in training recurrent neural networks. We also provide an improved validation framework for the evaluation of genetic network modeling systems that will result in better generalization and long-term prediction capability. Success in the modeling of gene regulation and prediction of gene expression will lead to more rapid discovery and development of therapeutic medicine, earlier diagnosis and treatment of adverse conditions, and vast advancements in life science research.

Spatiotemporal Predictive Learning (SPL) is an essential research topic involving many practical and real-world applications, e.g., motion detection, video generation, precipitation forecasting, and traffic flow prediction. The problems and challenges of this field come from numerous data characteristics in both time and space domains, and they vary depending on the specific task. For instance, spatial analysis refers to the study of spatial features, such as spatial location, latitude, elevation, longitude, the shape of objects, and other patterns. From the time domain perspective, the temporal analysis generally illustrates the time steps and time intervals of data points in the sequence, also known as interval recording or time sampling. Typically, there are two types of time sampling in temporal analysis: regular time sampling (i.e., the time interval is assumed to be fixed) and the irregular time sampling (i.e., the time interval is considered arbitrary) related closely to the continuous-time prediction task when data are in continuous space. Therefore, an efficient spatiotemporal predictive method has to model spatial features properly at the given time sampling types. In this thesis, by taking advantage of Machine Learning (ML) and Deep Learning (DL) methods, which have achieved promising performance in many complicated computational tasks, we propose three DL-based models used for Spatiotemporal Sequence Prediction (SSP) with several types of time sampling. First, we design the Trajectory Gated Recurrent Unit Attention (TrajGRU-Attention) with novel attention mechanisms, namely Motion-based Attention (MA), to improve the performance of the standard Convolutional Recurrent Neural Networks (ConvRNNs) in the SSP tasks. In particular, the TrajGRU-Attention model can alleviate the impact of the vanishing gradient, which leads to the blurry effect in the long-term predictions and handle both regularly sampled and irregularly sampled time series. Consequently, this model can work effectively with different scenarios of spatiotemporal sequential data, especially in the case of time series with missing time steps. Second, by taking the idea of Neural Ordinary Differential Equations (NODEs), we propose Trajectory Gated Recurrent Unit integrating Ordinary Differential Equation techniques (TrajGRU-ODE) as a continuous time-series model. With Ordinary Differential Equation (ODE) techniques and the TrajGRU neural network, this model can perform continuous-time spatiotemporal prediction tasks and generate resulting output with high accuracy. Compared to TrajGRU-Attention, TrajGRU-ODE benefits from the development of efficient and accurate ODE solvers. Ultimately, we attempt to combine those two models to create TrajGRU-Attention-ODE. NODEs are still in their early stage of research, and recent ODE-based models were designed for many relatively simple tasks. In this thesis, we will train the models with several video datasets to verify the ability of the proposed models in practical applications. To evaluate the performance of the proposed models, we select four available spatiotemporal datasets based on the complexity level, including the MovingMNIST, MovingMNIST++, and two real-life datasets: the weather radar HKO-7 and KTH Action. With each dataset, we train, validate, and test with distinct types of time sampling to justify the prediction ability of our models. In summary, the experimental results on the four datasets indicate the proposed models can generate predictions properly with high accuracy and sharpness. Significantly, the proposed models outperform state-of-the-art ODE-based approaches under SSP tasks with different circumstances of interval recording.

No description available.

This thesis examines so called folding neural networks as a mechanism for machine learning. Folding networks form a generalization of partial recurrent neural networks such that they are able to deal with tree structured inputs instead of simple linear lists. In particular, they can handle classical formulas - they were proposed originally for this purpose. After a short explanation of the neural architecture we show that folding networks are well suited as a learning mechanism in principle. This includes three parts: the proof of their universal approximation ability, the aspect of information theoretical learnability, and the examination of the complexity of training. Approximation ability: It is shown that any measurable function can be approximated in probability. Explicit bounds on the number of neurons result if only a finite number of points is dealt with. These bounds are new results in the case of simple recurrent networks, too. Several restrictions occur if a function is to be approximated in the maximum norm. Afterwards, we consider briefly the topic of computability. It is shown that a sigmoidal recurrent neural network can compute any mapping in exponential time. However, if the computation is subject to noise almost the capability of tree automata arises. Information theoretical learnability: This part contains several contributions to distribution dependent learnability: The notation of PAC and PUAC learnability, consistent PAC/ PUAC learnability, and scale sensitive versions are considered. We find equivalent characterizations of these terms and examine their respective relation answering in particular an open question posed by Vidyasagar. It is shown at which level learnability only because of an encoding trick is possible. Two approaches from the literature which can guarantee distribution dependent learnability if the VC dimension of the concept class is infinite are generalized to function classes: The function class is stratified according to the input space or according to a so-called luckiness function which depends on the output of the learning algorithm and the concrete training data. Afterwards, the VC, pseudo-, and fat shattering dimension of folding networks are estimated: We improve some lower bounds for recurrent networks and derive new lower bounds for the pseudodimension and lower and upper bounds for folding networks in general. As a consequence, folding architectures are not distribution independent learnable. Distribution dependent learnability can be guaranteed. Explicit bounds on the number of examples which guarantee valid generalization can be derived using the two approaches mentioned above. We examine in which cases these bounds are polynomial. Furthermore, we construct an explicit example for a learning scenario where an exponential number of examples is necessary. Complexity: It is shown that training a fixed folding architecture with perceptron activation function is polynomial. Afterwards, a decision problem, the so-called loading problem, which is correlated to neural network training is examined. For standard multilayer feed-forward networks the following situations turn out to be NP-hard: Concerning the perceptron activation function, a classical result from the literature, the NP-hardness for varying input dimension, is generalized to arbitrary multilayer architectures. Additionally, NP-hardness can be found if the input dimension is fixed but the number of neurons may vary in at least two hidden layers. Furthermore, the NP-hardness is examined if the number of patterns and number of hidden neurons are correlated. We finish with a generalization of the classical NP result as mentioned above to the sigmoidal activation function which is used in practical applications.

[pt] Este trabalho propõe e avalia arquiteturas profundas para o reconhecimento de culturas agrícolas a partir de seqüências de imagens multitemporais de sensoriamento remoto. Essas arquiteturas combinam a capacidade de modelar contexto espacial prórpia de redes totalmente convolucionais com a capacidade de modelr o contexto temporal de redes recorrentes para a previsão prever culturas agrícolas em cada data de uma seqüência de imagens multitemporais. O desempenho destes métodos é avaliado em dois conjuntos de dados públicos. Ambas as áreas apresentam alta dinâmica espaçotemporal devido ao clima tropical/subtropical e a práticas agrícolas locais, como a rotação de culturas. Nos experimentos verificou-se que as arquiteturas propostas superaram os métodos recentes baseados em redes recorrentes em termos de Overall Accuracy (OA) e F1-score médio por classe. / [en] This work proposes and evaluates deep learning architectures for multi-date agricultural crop recognition from remote sensing image sequences. These architectures combine the spatial modelling capabilities of fully convolutional networks and the sequential modelling capabilities of recurrent networks into end-to-end architectures so-called fully convolutional recurrent networks, configured to predict crop type at multiple dates from a multitemporal image sequence. Their performance is assessed over two publicly available datasets. Both datasets present highly spatio-temporal dynamics due to their tropical/sub-tropical climate and local agricultural practices such as crop rotation. The experiments indicated that the proposed architectures outperformed state of the art methods based on recurrent networks in terms of Overall Accuracy (OA) and per-class average F1 score.

Diese Arbeit ist ein erster Versuch, mit Modellbildung und mathematischer Analyse die Experimente zu verstehen, die zeigten, dass die Stimulation eines einzelnen Neurons im Cortex eine Verhaltensreaktion auslösen kann. Dieser Befund stellt die verbreitete Ansicht infrage, dass viele Neurone nötig sind, um Information zuverlässig kodieren zu können. Der Ausgangspunkt der vorliegenden Untersuchung ist die Stimulation einer zufällig ausgewählten Zelle in einem Zufallsnetzwerk exzitatorischer und inhibitorischer Neuronmodelle. Es wird dann nach einem plausiblen Ausleseverfahren gesucht, das die Einzelzellstimulation mit einer mit den Experimenten vergleichbaren Zuverlässigkeit detektieren kann. Das erste Ausleseschema reagiert auf Abweichungen vom spontanen Zustand in der Aktivität einer Auslesepopulation. Die Stimulation wird detektiert, wenn bei der Auswahl der Auslesepopulation denjenigen Neuronen ein Vorzug gegeben wird, die eine direkte Verbindung von der stimulierten Zelle bekommen. Im zweiten Teil der Arbeit wird das Ausleseschema erweitert, indem ein zweites Netzwerk als Ausleseschaltkreis dient. Interessanterweise erweist sich dieses Ausleseschema nicht nur als plausibler, sondern auch als effektiver. Diese Resultate basieren sowohl auf Simulationen als auch auf analytischen Rechnungen. Weitere Experimente zeigten, dass eine konstante Strominjektion einen Effekt auslöst, der kaum von Dauer und Intensität der Stimulation abhängt, der aber bei unregelmäßiger Stimulation zunimmt. Der letzte Teil der Arbeit befasst sich mit einer theoretischen Erklärung für diese Ergebnisse. Hierzu werden die biologischen Eigenschaften des Systems im Modell detaillierter beschrieben. Weiterhin wird die Funktionsweise des Ausleseschemas so modifiziert, dass es auf Veränderungen reagiert, anstatt den Input zu integrieren. Dieser Differenzierdetektor liefert Ergebnisse, die mit den Experimenten übereinstimmen, und könnte bei nichtstationärem Input vorteilhaft sein. / This thesis is a first attempt at developing a theoretical model of the experiments which show that the stimulation of a single cell in the cortex can trigger a behavioral reaction and that challenge the common belief that many neurons are needed to reliably encode information. As a starting point of the present work, one neuron selected at random within a random network of excitatory and inhibitory integrate-and-fire neurons is stimulated. One important goal of this thesis is to seek a readout scheme that can detect the single-cell stimulation in a plausible way with a reliability compatible with the experiments. The first readout scheme reacts to deviations from the spontaneous state in the activity of a readout population. When the choice of readout neurons is sufficiently biased towards those receiving direct links from the stimulated cell, the stimulation can be detected. In the second part of the thesis, the readout scheme is extended by employing a second network as a readout circuit. Interestingly, this new readout scheme is not only more plausible, but also more effective. These results are based both on numerical simulations of the network and on analytical approximations. Further experiments showed that the probability of the behavioral reaction is substantially independent of the length and intensity of the stimulation, but it increases when an irregular current is used. The last part of this thesis seeks a theoretical explanation for these findings. To this end, a recurrent network including more biological details of the system is considered. Furthermore, the functioning principle of the readout is modified to react to changes in the activity of the local network (a differentiator readout), instead of integrating the input. This differentiator readout yields results in accordance with the experiments and could be advantageous in the presence of nonstationarities.

Cette thèse par article comprend quatre articles qui contribuent au domaine de l'apprentissage profond, en particulier à l'accélération de l’apprentissage par le biais de réseaux à faible précision et à l'application de réseaux de neurones profonds au traitement du langage naturel. Dans le premier article, nous étudions un schéma d’entraînement de réseau de neurones qui élimine la plupart des multiplications en virgule flottante. Cette approche consiste à binariser ou à ternariser les poids dans la propagation en avant et à quantifier les états cachés dans la propagation arrière, ce qui convertit les multiplications en changements de signe et en décalages binaires. Les résultats expérimentaux sur des jeux de données de petite à moyenne taille montrent que cette approche produit des performances encore meilleures que l’approche standard de descente de gradient stochastique, ouvrant la voie à un entraînement des réseaux de neurones rapide et efficace au niveau du matériel. Dans le deuxième article, nous avons proposé un mécanisme structuré d’auto-attention d’enchâssement de phrases qui extrait des représentations interprétables de phrases sous forme matricielle. Nous démontrons des améliorations dans 3 tâches différentes: le profilage de l'auteur, la classification des sentiments et l'implication textuelle. Les résultats expérimentaux montrent que notre modèle génère un gain en performance significatif par rapport aux autres méthodes d’enchâssement de phrases dans les 3 tâches. Dans le troisième article, nous proposons un modèle hiérarchique avec graphe de calcul dynamique, pour les données séquentielles, qui apprend à construire un arbre lors de la lecture de la séquence. Le modèle apprend à créer des connexions de saut adaptatives, ce qui facilitent l'apprentissage des dépendances à long terme en construisant des cellules récurrentes de manière récursive. L’entraînement du réseau peut être fait soit par entraînement supervisée en donnant des structures d’arbres dorés, soit par apprentissage par renforcement. Nous proposons des expériences préliminaires dans 3 tâches différentes: une nouvelle tâche d'évaluation de l'expression mathématique (MEE), une tâche bien connue de la logique propositionnelle et des tâches de modélisation du langage. Les résultats expérimentaux montrent le potentiel de l'approche proposée. Dans le quatrième article, nous proposons une nouvelle méthode d’analyse par circonscription utilisant les réseaux de neurones. Le modèle prédit la structure de l'arbre d'analyse en prédisant un scalaire à valeur réelle, soit la distance syntaxique, pour chaque position de division dans la phrase d'entrée. L'ordre des valeurs relatives de ces distances syntaxiques détermine ensuite la structure de l'arbre d'analyse en spécifiant l'ordre dans lequel les points de division seront sélectionnés, en partitionnant l'entrée de manière récursive et descendante. L’approche proposée obtient une performance compétitive sur le jeu de données Penn Treebank et réalise l’état de l’art sur le jeu de données Chinese Treebank. / This thesis by article consists of four articles which contribute to the field of deep learning, specifically in the acceleration of training through low-precision networks, and the application of deep neural networks on natural language processing. In the first article, we investigate a neural network training scheme that eliminates most of the floating-point multiplications. This approach consists of binarizing or ternarizing the weights in the forward propagation and quantizing the hidden states in the backward propagation, which converts multiplications to sign changes and binary shifts. Experimental results on datasets from small to medium size show that this approach result in even better performance than standard stochastic gradient descent training, paving the way to fast, hardware-friendly training of neural networks. In the second article, we proposed a structured self-attentive sentence embedding that extracts interpretable sentence representations in matrix form. We demonstrate improvements on 3 different tasks: author profiling, sentiment classification and textual entailment. Experimental results show that our model yields a significant performance gain compared to other sentence embedding methods in all of the 3 tasks. In the third article, we propose a hierarchical model with dynamical computation graph for sequential data that learns to construct a tree while reading the sequence. The model learns to create adaptive skip-connections that ease the learning of long-term dependencies through constructing recurrent cells in a recursive manner. The training of the network can either be supervised training by giving golden tree structures, or through reinforcement learning. We provide preliminary experiments in 3 different tasks: a novel Math Expression Evaluation (MEE) task, a well-known propositional logic task, and language modelling tasks. Experimental results show the potential of the proposed approach. In the fourth article, we propose a novel constituency parsing method with neural networks. The model predicts the parse tree structure by predicting a real valued scalar, named syntactic distance, for each split position in the input sentence. The order of the relative values of these syntactic distances then determine the parse tree structure by specifying the order in which the split points will be selected, recursively partitioning the input, in a top-down fashion. Our proposed approach was demonstrated with competitive performance on Penn Treebank dataset, and the state-of-the-art performance on Chinese Treebank dataset.

Cette thèse par article se compose de quatre articles qui contribuent au domaine de l’apprentissage profond, en particulier dans la compréhension et l’apprentissage des ap- proches neuronales des systèmes de dialogue. Le premier article fait un pas vers la compréhension si les architectures de dialogue neuronal couramment utilisées capturent efficacement les informations présentes dans l’historique des conversations. Grâce à une série d’expériences de perturbation sur des ensembles de données de dialogue populaires, nous constatons que les architectures de dialogue neuronal couramment utilisées comme les modèles seq2seq récurrents et basés sur des transformateurs sont rarement sensibles à la plupart des perturbations du contexte d’entrée telles que les énoncés manquants ou réorganisés, les mots mélangés, etc. Le deuxième article propose d’améliorer la qualité de génération de réponse dans les systèmes de dialogue de domaine ouvert en modélisant conjointement les énoncés avec les attributs de dialogue de chaque énoncé. Les attributs de dialogue d’un énoncé se réfèrent à des caractéristiques ou des aspects discrets associés à un énoncé comme les actes de dialogue, le sentiment, l’émotion, l’identité du locuteur, la personnalité du locuteur, etc. Le troisième article présente un moyen simple et économique de collecter des ensembles de données à grande échelle pour modéliser des systèmes de dialogue orientés tâche. Cette approche évite l’exigence d’un schéma d’annotation d’arguments complexes. La version initiale de l’ensemble de données comprend 13 215 dialogues basés sur des tâches comprenant six domaines et environ 8 000 entités nommées uniques, presque 8 fois plus que l’ensemble de données MultiWOZ populaire. / This thesis by article consists of four articles which contribute to the field of deep learning, specifically in understanding and learning neural approaches to dialog systems. The first article takes a step towards understanding if commonly used neural dialog architectures effectively capture the information present in the conversation history. Through a series of perturbation experiments on popular dialog datasets, wefindthatcommonly used neural dialog architectures like recurrent and transformer-based seq2seq models are rarely sensitive to most input context perturbations such as missing or reordering utterances, shuffling words, etc. The second article introduces a simple and cost-effective way to collect large scale datasets for modeling task-oriented dialog systems. This approach avoids the requirement of a com-plex argument annotation schema. The initial release of the dataset includes 13,215 task-based dialogs comprising six domains and around 8k unique named entities, almost 8 times more than the popular MultiWOZ dataset. The third article proposes to improve response generation quality in open domain dialog systems by jointly modeling the utterances with the dialog attributes of each utterance. Dialog attributes of an utterance refer to discrete features or aspects associated with an utterance like dialog-acts, sentiment, emotion, speaker identity, speaker personality, etc. The final article introduces an embedding-free method to compute word representations on-the-fly. This approach significantly reduces the memory footprint which facilitates de-ployment in on-device (memory constraints) devices. Apart from being independent of the vocabulary size, we find this approach to be inherently resilient to common misspellings.