The thesis addresses the topic of Deep Neural Networks, in particular the methods regar- ding the field of Deep Learning, which is used to initialize the weight and learning process s itself within Deep Neural Networks. The focus is also put to the basic theory of the classical Neural Networks, which is important to comprehensive understanding of the issue. The aim of this work is to determine the optimal set of optional parameters of the algori- thms on various complexity levels of image recognition tasks through experimenting with created application applying Deep Neural Networks. Furthermore, evaluation and analysis of the results and lessons learned from the experimentation with classical and Deep Neural Networks are integrated in the thesis.

Being able to accurately approximate outputs of computationally expensive simulations for arbitrary input parameters, also called missing points estimation, is central in many different areas of research and development with applications ranging from uncertainty propagation to control system design to name a few. This project investigates the potential of kernel transformations and nonlinear autoencoders as methods of improving the accuracy of the proper orthogonal decomposition method combined with regression. The techniques are applied on aerodynamic pressure CFD data around airplane wings in both two- and three-dimensional settings. The novel methods show potential in select situations, but cannot at this stage be generally considered superior. Their performances are similar although the procedure of design and training of a nonlinear autoencoder is less straight forward and more time demanding than using kernel transformations. The results demonstrate the regression bottleneck of the proper orthogonal decomposition method, which partially is improved with the new methods. Future studies should focus on adapting the autoencoder training strategy to the architecture and data as well as improving the regression stage of all methods.

We consider a mechanistic non-linear machine learning approach to learning signals in financial time series data. A modularised and decoupled algorithm framework is established and is proven on daily sampled closing time-series data for JSE equity markets. The input patterns are based on input data vectors of data windows preprocessed into a sequence of daily, weekly and monthly or quarterly sampled feature measurement changes (log feature fluctuations). The data processing is split into a batch processed step where features are learnt using a Stacked AutoEncoder (SAE) via unsupervised learning, and then both batch and online supervised learning are carried out on Feedforward Neural Networks (FNNs) using these features. The FNN output is a point prediction of measured time-series feature fluctuations (log differenced data) in the future (ex-post). Weight initializations for these networks are implemented with restricted Boltzmann machine pretraining, and variance based initializations. The validity of the FNN backtest results are shown under a rigorous assessment of backtest overfitting using both Combinatorially Symmetrical Cross Validation and Probabilistic and Deflated Sharpe Ratios. Results are further used to develop a view on the phenomenology of financial markets and the value of complex historical data under unstable dynamics.

<div>Leaf Area Index (LAI) is an important variable for both for characterizing plant canopy and as an input to many crop models. It is a dimensionless quantity broadly defined as the total one-sided leaf area per unit ground area, and is estimated over agriculture row crops by both direct and indirect methods. Direct methods, which involve destructive sampling, are laborious and time-consuming, while indirect methods such as remote sensing-based approaches have multiple sources of uncertainty. LiDAR (Light Detection and Ranging) remotely sensed data acquired from manned aircraft and UAVs’ have been investigated to estimate LAI based on physical/geometric features such as canopy gap fraction. High-resolution point cloud data acquired with a laser scanner from any platform, including terrestrial laser scanning and mobile mapping systems, contain random noise and outliers. Therefore, outlier detection in LiDAR data is often useful prior to analysis. Applications in agriculture are particularly challenging, as there is typically no prior knowledge of the statistical distribution of points, description of plant complexity, and local point densities, which are crop dependent. This dissertation first explores the effectiveness of using LiDAR data to estimate LAI for row crop plants at multiple times during the growing season from both a wheeled vehicle and an Unmanned Aerial Vehicle (UAV). Linear and nonlinear regression models are investigated for prediction utilizing statistical and plant structure-based features extracted from the LiDAR point cloud data and ground reference obtained from an in-field plant canopy analyzer and leaf area derived from destructive sampling. LAI estimates obtained from support vector regression (SVR) models with a radial basis function (RBF) kernel developed using the wheel-based LiDAR system and UAVs are promising, based on the value of the coefficient of determination (R2) and root mean squared error (RMSE) of the residuals. </div><div>This dissertation also investigates approaches to minimize the impact of outliers on discrete return LiDAR acquired over crops, and specifically for sorghum and maize breeding experiments, by an unmanned aerial vehicle (UAV) and a wheel-based ground platform. Two methods are explored to detect and remove the outliers from the plant datasets. The first is based on surface fitting to noisy point cloud data based on normal and curvature estimation in a local neighborhood. The second utilizes the deep learning framework PointCleanNet. Both methods are applied to individual plants and field-based datasets. To evaluate the method, an F-score and LAI are calculated both before and after outlier removal for both scenarios. Results indicate that the deep learning method for outlier detection is more robust to changes in point densities, level of noise, and shapes. Also, the predicted LAI was improved for the wheel-based vehicle data based on the R2 value and RMSE of residuals. </div><div>The quality of the extracted features depends on the point density and laser penetration of the canopy. Extracting appropriate features is a critical step to have accurate prediction models. Deep learning frameworks are increasingly being used in remote sensing applications. In the last objective of this study, a feature extraction approach is investigated for encoding LiDAR data acquired by UAV platforms multiple times during the growing season over sorghum and maize plant breeding experiments. LAI estimates obtained with these inputs are used to develop support vector regression (SVR) models using plant canopy analyzer data as the ground reference. Results are compared to models based on estimates from physically-based features and evaluated in terms of the coefficient determination (R2). The effects of experimental conditions, including flying height, sensor characteristics, and crop type, are also investigated relative to the estimates of LAI.</div><div><br></div>

Nuclear power plants are inherently complex systems. While the technology has been used to generate electrical power for many decades, process monitoring continuously evolves. There is always room for improvement in terms of maximizing the availability by reducing the risks of problems and errors. In this context, automated monitoring systems have become important tools – not least with the rapid progress being made in the field of data analytics thanks to ever increasing amounts of processing power. There are many different types of models that can be utilized for identifying anomalies. Some rely on physical properties and theoretical relations, while others rely more on the patterns of historical data. In this thesis, a data-driven approach using a hierarchical autoencoder framework has been developed for the purposes of anomaly detection at the Swedish nuclear power plant Forsmark. The model is first trained to recognize normal operating conditions. The trained model then creates reference values and calculates the deviations in relation to real data in order to identify any issues. This proof-of-concept has been evaluated and benchmarked against a currently used hybrid model with more physical modeling properties in order to identify benefits and drawbacks. Generally speaking, the created model has performed in line with expectations. The currently used tool is more flexible in its understanding of different plant states and is likely better at determining root causes thanks to its physical modeling properties. However, the created autoencoder framework does bring other advantages. For instance, it allows for a higher time resolution thanks to its relatively low calculation intensity. Additionally, thanks to its purely data-driven characteristics, it offers great opportunities for future reconfiguration and adaptation with different signal selections.

Histogram data are widely used for compressing high-frequency time-series signals due to their ability to capture distributional informa-tion. However, this compression comes at the cost of increased di-mensionality and loss of contextual details from the original features.This study addresses the challenge of effectively capturing changesin distributions over time and their contribution to failure prediction.Specifically, we focus on the task of predicting Time to Event (TTE) forturbocharger failures.In this thesis, we propose a novel approach to improve failure pre-diction by fine-tuning lower-dimensional representations of bi-variatehistograms. The goal is to optimize these representations in a waythat enhances their ability to predict component failure. Moreover, wecompare the performance of our learned representations with hand-crafted histogram features to assess the efficacy of both approaches.We evaluate the different representations using the Weibull Time ToEvent - Recurrent Neural Network (WTTE-RNN) framework, which isa popular choice for TTE prediction tasks. By conducting extensive ex-periments, we demonstrate that the fine-tuning approach yields supe-rior results compared to general lower-dimensional learned features.Notably, our approach achieves performance levels close to state-of-the-art results.This research contributes to the understanding of effective failureprediction from time series histogram data. The findings highlightthe significance of fine-tuning lower-dimensional representations forimproving predictive capabilities in real-world applications. The in-sights gained from this study can potentially impact various indus-tries, where failure prediction is crucial for proactive maintenanceand reliability enhancement.

Autoencoders, a type of artificial neural network, have gained recognition by researchers in various fields, especially machine learning due to their vast applications in data representations from inputs. Recently researchers have explored the possibility to extend the application of autoencoders to solve nonlinear differential equations. Algorithms and methods employed in an autoencoder framework include sparse identification of nonlinear dynamics (SINDy), dynamic mode decomposition (DMD), Koopman operator theory and singular value decomposition (SVD). These approaches use matrix multiplication to represent linear transformation. However, machine learning algorithms often use convolution to represent linear transformations. In our work, we modify these approaches to system identification and forecasting of solutions of nonlinear differential equations by replacing matrix multiplication with convolution transformation. In particular, we develop convolution-based approach to dynamic mode decomposition and discuss its application to problems not solvable otherwise.

Les auto-encodeurs sont des réseaux de neurones artificiels qui apprennent des représentations. Dans un auto-encodeur, l’encodeur transforme une entrée en une représentation, et le décodeur essaie de prédire l’entrée à partir de la représentation. Cette thèse compile trois applications de ces modèles au traitement automatique des langues : pour l’apprentissage de représentations de mots et de phrases, ainsi que pour mieux comprendre la compositionnalité. Dans le premier article, nous montrons que nous pouvons auto-encoder des définitions de dictionnaire et ainsi apprendre des vecteurs de définition. Nous proposons une nouvelle pénalité qui nous permet d’utiliser ces vecteurs comme entrées à l’encodeur lui-même, mais aussi de les mélanger des vecteurs distributionnels pré-entraînés. Ces vecteurs de définition capturent mieux la similarité sémantique que les méthodes distributionnelles telles que word2vec. De plus, l’encodeur généralise à un certain degré à des définitions qu’il n’a pas vues pendant l’entraînement. Dans le deuxième article, nous analysons les représentations apprises par les auto-encodeurs variationnels séquence-à-séquence. Nous constatons que les encodeurs ont tendance à mémo- riser les premiers mots et la longueur de la phrase d’entrée. Cela limite considérablement leur utilité en tant que modèles génératifs contrôlables. Nous analysons aussi des variantes architecturales plus simples qui ne tiennent pas compte de l’ordre des mots, ainsi que des mé- thodes basées sur le pré-entraînement. Les représentations qu’elles apprennent ont tendance à encoder plus nettement des caractéristiques globales telles que le sujet et le sentiment, et cela se voit dans les reconstructions qu’ils produisent. Dans le troisième article, nous utilisons des simulations d’émergence du langage pour étudier la compositionnalité. Un locuteur – l’encodeur – observe une entrée et produit un message. Un auditeur – le décodeur – tente de reconstituer ce dont le locuteur a parlé dans son message. Nous émettons l’hypothèse que faire des phrases impliquant plusieurs entités, telles que « Jean aime Marie », nécessite fondamentalement de percevoir chaque entité comme un tout. Nous dotons certains agents de cette capacité grâce à un mechanisme d’attention, alors que d’autres en sont privés. Nous proposons différentes métriques qui mesurent à quel point les langues des agents sont naturelles en termes de structure d’argument, et si elles sont davantage analytiques ou synthétiques. Les agents percevant les entités comme des touts échangent des messages plus naturels que les autres agents. / Autoencoders are artificial neural networks that learn representations. In an autoencoder, the encoder transforms an input into a representation, and the decoder tries to recover the input from the representation. This thesis compiles three different applications of these models to natural language processing: for learning word and sentence representations, as well as to better understand compositionality. In the first paper, we show that we can autoencode dictionary definitions to learn word vectors, called definition embeddings. We propose a new penalty that allows us to use these definition embeddings as inputs to the encoder itself, but also to blend them with pretrained distributional vectors. The definition embeddings capture semantic similarity better than distributional methods such as word2vec. Moreover, the encoder somewhat generalizes to definitions unseen during training. In the second paper, we analyze the representations learned by sequence-to-sequence variational autoencoders. We find that the encoders tend to memorize the first few words and the length of the input sentence. This limits drastically their usefulness as controllable generative models. We also analyze simpler architectural variants that are agnostic to word order, as well as pretraining-based methods. The representations that they learn tend to encode global features such as topic and sentiment more markedly, and this shows in the reconstructions they produce. In the third paper, we use language emergence simulations to study compositionality. A speaker – the encoder – observes an input and produces a message about it. A listener – the decoder – tries to reconstruct what the speaker talked about in its message. We hypothesize that producing sentences involving several entities, such as “John loves Mary”, fundamentally requires to perceive each entity, John and Mary, as distinct wholes. We endow some agents with this ability via an attention mechanism, and deprive others of it. We propose various metrics to measure whether the languages are natural in terms of their argument structure, and whether the languages are more analytic or synthetic. Agents perceiving entities as distinct wholes exchange more natural messages than other agents.

<p>Mental health significantly impacts issues like gun violence, school shootings, and suicide. There is a strong connection between mental health and emotional states. By monitoring emotional changes over time, we can identify triggering events, detect early signs of instability, and take preventive measures. This thesis focuses on the development of a generalized and modular system for human emotion recognition and explanation based on visual information. The aim is to address the challenges of effectively utilizing different cues (modalities) available in the data for a reliable and trustworthy emotion recognition system. Our face is one of the most important medium through which we can express our emotion. Therefore We first propose SAFER, A novel facial emotion recognition system with background and place features. We provide a detailed evaluation framework to prove the high accuracy and generalizability. However, relying solely on facial expressions for emotion recognition can be unreliable, as faces can be covered or deceptive.  To enhance the system's reliability, we introduce EMERSK, a multimodal emotion recognition system that integrates various modalities, including facial expressions, posture, gait, and scene background, in a flexible and modular manner. It employs convolutional neural networks (CNNs), Long Short-term Memory (LSTM), and denoising auto-encoders to extract features from facial images, posture, gait, and scene background. In addition to multimodal feature fusion, the system utilizes situational knowledge derived from place type and adjective-noun pairs (ANP) extracted from the scene, as well as the spatio-temporal average distribution of emotions, to generate comprehensive explanations for the recognition outcomes. Extensive experiments on different benchmark datasets demonstrate the superiority of our approach over existing state-of-the-art methods. The system achieves improved performance in accurately recognizing and explaining human emotions. Moreover, we investigate the impact of novelty, such as face masks during the Covid-19 pandemic, on the emotion recognition. The study critically examines the limitations of mainstream facial expression datasets and proposes a novel dataset specifically tailored for facial emotion recognition with masked subjects. Additionally, we propose a continuous learning-based approach that incorporates a novelty detector working in parallel with the classifier to detect and properly handle instances of novelty. This approach ensures robustness and adaptability in the automatic emotion recognition task, even in the presence of novel factors such as face masks. This thesis contributes to the field of automatic emotion recognition by providing a generalized and modular approach that effectively combines multiple modalities, ensuring reliable and highly accurate recognition. Moreover, it generates situational knowledge that is valuable for mission-critical applications and provides comprehensive explanations of the output. The findings and insights from this research have the potential to enhance the understanding and utilization of multimodal emotion recognition systems in various real-world applications.</p> <p><br></p>

No description available.