In this thesis, a comparison of three different pre-processing methods for imbalanced classification data, is conducted. Variational Autoencoder, Random Under-Sampling Boosting and a hybrid approach of the two, are applied to three imbalanced classification data sets with different class imbalances. A logistic regression (LR) model is fitted to each pre-processed data set and based on its classification performance, the pre-processing methods are evaluated. All three methods shows indications of different advantages when handling class imbalances. For each pre-processed data, the LR-model has is better at correctly classifying minority class observations, compared to a LR-model fitted to the original class imbalanced data sets. Evaluating the overall classification performance, both VAE and RUSBoost shows improving classification results while the hybrid method performs worse for the moderate class imbalanced data and best for the highly imbalanced data.

While many of the currently used network intrusion detection systems (NIDS) employ signature-based approaches, there is an increasing research interest in the examination of anomaly-based detection methods, which seem to be more suited for recognizing zero-day attacks. Nevertheless, requirements for their practical deployment, as well as objective and reproducible evaluation methods, are hereby often neglected. The following thesis defines aspects that are crucial for a practical evaluation of anomaly-based NIDS, such as the focus on modern attack types, the restriction to one-class classification methods, the exclusion of known attacks from the training phase, a low false detection rate, and consideration of the runtime efficiency. Based on those principles, a framework dedicated to developing, testing and evaluating models for the detection of network anomalies is proposed. It is applied to two datasets featuring modern traffic, namely the UNSW-NB15 and the CIC-IDS-2017 datasets, in order to compare and evaluate commonly-used network intrusion detection methods. The implemented approaches include, among others, a highly configurable network flow generator, a payload analyser, a one-hot encoder, a one-class support vector machine, and an autoencoder. The results show a significant difference between the two chosen datasets: While for the UNSW-NB15 dataset several reasonably well performing model combinations for both the autoencoder and the one-class SVM can be found, most of them yield unsatisfying results when the CIC-IDS-2017 dataset is used. / Obwohl viele der derzeit genutzten Systeme zur Erkennung von Netzwerkangriffen (engl. NIDS) signaturbasierte Ansätze verwenden, gibt es ein wachsendes Forschungsinteresse an der Untersuchung von anomaliebasierten Erkennungsmethoden, welche zur Identifikation von Zero-Day-Angriffen geeigneter erscheinen. Gleichwohl werden hierbei Bedingungen für deren praktischen Einsatz oft vernachlässigt, ebenso wie objektive und reproduzierbare Evaluationsmethoden. Die folgende Arbeit definiert Aspekte, die für eine praxisorientierte Evaluation unabdingbar sind. Dazu zählen ein Schwerpunkt auf modernen Angriffstypen, die Beschränkung auf One-Class Classification Methoden, der Ausschluss von bereits bekannten Angriffen aus dem Trainingsdatensatz, niedrige Falscherkennungsraten sowie die Berücksichtigung der Laufzeiteffizienz. Basierend auf diesen Prinzipien wird ein Rahmenkonzept vorgeschlagen, das für das Entwickeln, Testen und Evaluieren von Modellen zur Erkennung von Netzwerkanomalien bestimmt ist. Dieses wird auf zwei Datensätze mit modernem Netzwerkverkehr, namentlich auf den UNSW-NB15 und den CIC-IDS- 2017 Datensatz, angewendet, um häufig genutzte NIDS-Methoden zu vergleichen und zu evaluieren. Die für diese Arbeit implementierten Ansätze beinhalten, neben anderen, einen weit konfigurierbaren Netzwerkflussgenerator, einen Nutzdatenanalysierer, einen One-Hot-Encoder, eine One-Class Support Vector Machine sowie einen Autoencoder. Die Resultate zeigen einen großen Unterschied zwischen den beiden ausgewählten Datensätzen: Während für den UNSW-NB15 Datensatz verschiedene angemessen gut funktionierende Modellkombinationen, sowohl für den Autoencoder als auch für die One-Class SVM, gefunden werden können, bringen diese für den CIC-IDS-2017 Datensatz meist unbefriedigende Ergebnisse.

Intelligent fault detection systems using machine learning can be applied to learn to spot anomalies in signals sampled directly from machinery. As a result, expensive repair costs due to mechanical breakdowns and potential harm to humans due to malfunctioning equipment can be prevented. In recent years, Autoencoders have been applied for fault detection in areas such as industrial manufacturing. It has been shown that they are well suited for the purpose as such models can learn to recognize healthy signals that facilitate the detection of anomalies. The content of this thesis is an investigation into the applicability of Autoencoders for fault detection in mobile robotics by assigning anomaly scores to sampled torque signals based on the Autoencoder reconstruction errors and the Mahalanobis distance to a known distribution of healthy errors. An experiment was carried out by training a model with signals recorded from a four-wheeled mobile robot executing a pre-defined diagnostics routine to stress the motors, and datasets of healthy samples along with three different injected faults were created. The model produced overall greater anomaly scores for one of the fault cases in comparison to the healthy data. However, the two other cases did not yield any difference in anomaly scores due to the faults not impacting the pattern of the signals. Additionally, the Autoencoders ability to isolate a fault to a location was studied by examining the reconstruction errors faulty samples determine whether the errors of signals originating from the faulty component could be used for this purpose. Although we could not confirm this based on the results, fault isolation with Autoencoders could still be possible given more representative signals.

Smart wearables, known as emerging paradigms for vital big data capturing, have been attracting intensive attentions. However, one crucial problem is their power-hungriness, i.e., the continuous data streaming consumes energy dramatically and requires devices to be frequently charged. Targeting this obstacle, we propose to investigate the biodynamic patterns in the data and design a data-driven approach for intelligent data compression. We leverage Deep Learning (DL), more specifically, Convolutional Autoencoder (CAE), to learn a sparse representation of the vital big data. The minimized energy need, even taking into consideration the CAE-induced overhead, is tremendously lower than the original energy need. Further, compared with state-of-the-art wavelet compression-based method, our method can compress the data with a dramatically lower error for a similar energy budget. Our experiments and the validated approach are expected to boost the energy efficiency of wearables, and thus greatly advance ubiquitous big data applications in era of smart health.<br><div>In recent years, there has also been a growing interest in edge intelligence for emerging instantaneous big data inference. However, the inference algorithms, especially deep learning, usually require heavy computation requirements, thereby greatly limiting their deployment on the edge. We take special interest in the smart health wearable big data mining and inference. <br></div><div><br></div><div>Targeting the deep learning’s high computational complexity and large memory and energy requirements, new approaches are urged to make the deep learning algorithms ultra-efficient for wearable big data analysis. We propose to leverage knowledge distillation to achieve an ultra-efficient edge-deployable deep learning model. More specifically, through transferring the knowledge from a teacher model to the on-edge student model, the soft target distribution of the teacher model can be effectively learned by the student model. Besides, we propose to further introduce adversarial robustness to the student model, by stimulating the student model to correctly identify inputs that have adversarial perturbation. Experiments demonstrate that the knowledge distillation student model has comparable performance to the heavy teacher model but owns a substantially smaller model size. With adversarial learning, the student model has effectively preserved its robustness. In such a way, we have demonstrated the framework with knowledge distillation and adversarial learning can, not only advance ultra-efficient edge inference, but also preserve the robustness facing the perturbed input.</div>

Python is the most popular language when it comes to prototyping and developing machine learning algorithms. Python is an interpreted language that causes it to have a significant performance loss compared to compiled languages. Julia is a newly developed language that tries to bridge the gap between high performance but cumbersome languages such as C++ and highly abstracted but typically slow languages such as Python. However, over the years, the Python community have developed a lot of tools that addresses its performance problems. This raises the question if choosing one language over the other has any significant performance difference. This thesis compares the performance, in terms of execution time, of the two languages in the machine learning domain. More specifically, image processing with GPU-accelerated deep neural networks and classification with k-nearest neighbor on the MNIST and EMNIST dataset. Python with Keras and Tensorflow is compared against Julia with Flux for GPU-accelerated neural networks. For classification Python with Scikit-learn is compared against Julia with Nearestneighbors.jl. The results point in the direction that Julia has a performance edge in regards to GPU-accelerated deep neural networks. With Julia outperforming Python by roughly 1.25x − 1.5x. For classification with k-nearest neighbor the results were a bit more varied with Julia outperforming Python in 5 out of 8 different measurements. However, there exists some validity threats and additional research is needed that includes all different frameworks available for the languages in order to provide a more conclusive and generalized answer.

The forestry and the sawmill industry have a lot of incoming and outgoing piles of wood. It's important to maintain quality and efficiency. This motivates an examination of whether machine learning- or more specifically, anomaly detection techniques can be implemented and used to detect faulty shipments. This thesis presents and evaluates some computer vision techniques and some deep learning techniques. Deep learning can be divided into groups; supervised, semi-supervised and unsupervised. In this thesis, all three groups were examined and it covers supervised methods such as Convolutional Neural Networks, semi-supervised methods such as a modified Convolutional Autoencoder (CAE) and lastly, an unsupervised technique such as Generative Adversarial Network (GAN) was being tested and evaluated.  A version of a GAN model proved to perform best for this thesis in terms of the accuracy of faulty detecting shipments with an accuracy rate of 68.2% and 79.8\% overall, which was satisfactory given the problems that were discovered during the progress of the thesis.

The thesis proposes an unsupervised representation learning method to predict 3D human pose from a 2D skeleton via a VAEGAN (Variational Autoencoder Generative Adversarial Network) hybrid network. The method learns to lift poses from 2D to 3D using selfsupervision and adversarial learning techniques. The method does not use images, heatmaps, 3D pose annotations, paired/unpaired 2Dto3D skeletons, 3D priors, synthetic 2D skeletons, multiview or temporal information in any shape or form. The 2D skeleton input is taken by a VAE that encodes it in a latent space and then decodes that latent representation to a 3D pose. The 3D pose is then reprojected to 2D for a constrained, selfsupervised optimization using the input 2D pose. Parallelly, the 3D pose is also randomly rotated and reprojected to 2D to generate a ’novel’ 2D view for unconstrained adversarial optimization using a discriminator network. The combination of the optimizations of the original and the novel 2D views of the predicted 3D pose results in a ’realistic’ 3D pose generation. The thesis shows that the encoding and decoding process of the VAE addresses the major challenge of erroneous and incomplete skeletons from 2D detection networks as inputs and that the variance of the VAE can be altered to get various plausible 3D poses for a given 2D input. Additionally, the latent representation could be used for crossmodal training and many downstream applications. The results on Human3.6M datasets outperform previous unsupervised approaches with less model complexity while addressing more hurdles in scaling the task to the real world. / Uppsatsen föreslår en oövervakad metod för representationslärande för att förutsäga en 3Dpose från ett 2D skelett med hjälp av ett VAE GAN (Variationellt Autoenkodande Generativt Adversariellt Nätverk) hybrid neuralt nätverk. Metoden lär sig att utvidga poser från 2D till 3D genom att använda självövervakning och adversariella inlärningstekniker. Metoden använder sig vare sig av bilder, värmekartor, 3D poseannotationer, parade/oparade 2D till 3D skelett, a priori information i 3D, syntetiska 2Dskelett, flera vyer, eller tidsinformation. 2Dskelettindata tas från ett VAE som kodar det i en latent rymd och sedan avkodar den latenta representationen till en 3Dpose. 3D posen är sedan återprojicerad till 2D för att genomgå begränsad, självövervakad optimering med hjälp av den tvådimensionella posen. Parallellt roteras dessutom 3Dposen slumpmässigt och återprojiceras till 2D för att generera en ny 2D vy för obegränsad adversariell optimering med hjälp av ett diskriminatornätverk. Kombinationen av optimeringarna av den ursprungliga och den nya 2Dvyn av den förutsagda 3Dposen resulterar i en realistisk 3Dposegenerering. Resultaten i uppsatsen visar att kodningsoch avkodningsprocessen av VAE adresserar utmaningen med felaktiga och ofullständiga skelett från 2D detekteringsnätverk som indata och att variansen av VAE kan modifieras för att få flera troliga 3D poser för givna 2D indata. Dessutom kan den latenta representationen användas för crossmodal träning och flera nedströmsapplikationer. Resultaten på datamängder från Human3.6M är bättre än tidigare oövervakade metoder med mindre modellkomplexitet samtidigt som de adresserar flera hinder för att skala upp uppgiften till verkliga tillämpningar.

Magnetic resonance imaging (MRI) is a commonly used non-invasive imaging technique that provides high resolution images of soft tissue. One problem with MRI is that it is sensitive to signal quality issues. The issues can arise for various reasons, for example by metal located either inside or outside of the body. Another common signal quality issue is caused by the patient being partly placed outside field of view of the MRI scanner.   This thesis aims to investigate the possibility to automatically detect these signal quality issues using deep neural networks. More specifically, two different 3D CNN network types were studied, a classification-based approach and a reconstruction-based approach. The datasets used consist of MRI volumes from UK Biobank which have been processed and manually annotated by operators at AMRA Medical. For the classification method four different network architectures were explored utilising supervised learning with multi-label classification. The classification method was evaluated using accuracy and label-based evaluation metrics, such as macro-precision, macro-recall and macro-F1. The reconstruction method was based on anomaly detection using an autoencoder which was trained to reconstruct volumes without any artefacts. A mean squared prediction error was calculated for the reconstructed volume and compared against a threshold in order to classify a volume with or without artefacts. The idea was that volumes containing artefacts should be more difficult to reconstruct and thus, result in a higher prediction error. The reconstruction method was evaluated using accuracy, precision, recall and F1-score.  The results show that the classification method has overall higher performance than the reconstruction method. The achieved accuracy for the classification method was 98.0% for metal artefacts and 97.5% for outside field of view artefacts. The best architecture for the classification method proved to be DenseNet201. The reconstruction method worked for metal artefacts with an achieved accuracy of 75.7%. Furthermore, it was concluded that reconstruction method did not work for detection of outside field of view artefacts.    The results from the classification method indicate that there is a possibility to automatically detect artefacts with deep neural networks. However, it is needed to further improve the method in order to completely replace a manual quality control step before using the volumes for calculation of biomarkers.

Anomaly detection in high-dimensional financial transaction data is challenging and resource-intensive, particularly when the dataset is unlabeled. Sometimes, one can alleviate the computational cost and improve the results by utilizing a pre-trained model, provided that the features learned from the pre-training are useful for learning the second task. Investigating this issue was the main purpose of this thesis. More specifically, it was to explore the potential gain of pre-training a detection model on one trader's transaction history and then retraining the model to detect anomalous trades in another trader's transaction history. In the context of transfer learning, the pre-trained and the retrained model are usually referred to as the source model and target model, respectively.  A deep LSTM autoencoder was proposed as the source model due to its advantages when dealing with sequential data, such as financial transaction data. Moreover, to test its anomaly detection ability despite the lack of labeled true anomalies, synthetic anomalies were generated and included in the test set. Various experiments confirmed that the source model learned to detect synthetic anomalies with highly distinctive features. Nevertheless, it is hard to draw any conclusions regarding its anomaly detection performance due to the lack of labeled true anomalies. While the same is true for the target model, it is still possible to achieve the thesis's primary goal by comparing a pre-trained model with an identical untrained model. All in all, the results suggest that transfer learning offers a significant advantage over traditional machine learning in this context.

Style transfer using Generative adversarial networks (GANs) has been successful in recent publications. One field in style transfer is music style transfer, in which a piece of music is transformed in some way, be it through genre-, harmonic-, rhythmic transfer, etc. In this thesis, we have performed genre style transfer using a CycleGAN architecture and symbolic representation of data. Previous work using the same architecture and representation has focused solely on transferring the arrangement of the notes (composition). We have improved this work by including the transfer of multiple instruments (timbre) to create more convincing results. Additional discriminators were added to the CycleGAN architecture to achieve this, and they are individually tasked with enforcing the timbre and composition of a song. Previous works have also used variable autoencoders (VAEs) with sequential data representation for style transfer. The use of VAEs for genre style transfer using symbolic data representation instead of sequential was explored, and recommendations for future work include omitting faults found during exploration. Two different classifiers were created to evaluate the results of the CycleGAN model. One uses symbolic representation, in which all instruments are merged into one, thus evaluating the composition of the generated songs. The other classifier uses a spectrogram representation which evaluates the transfer as a whole, both timbre and composition. The evaluation of the improved CycleGAN model using the classifiers showed that it could perform genre style transfer successfully even when adding timbre to the style transfer.