Parkinson’s disease is a brain disorder associated with reduced dopamine levels in the brain, affecting cognition and motor control in the human brain. One of the motor controls that can be affected is eye movements and can therefore be critically affected in patients with Parkinson’s disease. Eye movement can be measured using eye trackers, and this data can be used for analyzing the eye movement characteristics in Parkinson’s disease. The eye movement analysis provides the possibility of diagnostics and can therefore lead to further insights into Parkinson’s disease. In this thesis, feature extraction of clinical relevance in diagnosing Parkinson’s patients from eye movement data is studied. We have used an autoencoder (AE) constructed to learn micro and macro-scaled representation for eye movements and constructed three different models. Learning of the AEs was evaluated using the F1 score, and differences were statistically assessed using the Wilcoxon sign rank test. Extracted features from data based on patients and healthy subjects were visualized using t-SNE. Using the extracted features, we have measured differences in features using cosine and Mahalanobis distances. We have furthermore clustered the features using fuzzy c-means. Qualities of the generated clusters were assessed by F1-score, fuzzy partition coefficient, Dunn’s index and silhouette index. Based on successful tests using a test data set of a previous publication, we believe that the network used in this thesis has learned to represent natural eye movement from subjects allowed to move their eye freely. However, distances, visualizations, clustering all suggest that latent representations from the autoencoder do not provide a good separation of data from patients and healthy subjects. We, therefore, conclude that a micro-macro autoencoder does not suit the purpose of generating a latent representation of saccade movements of the type used in this thesis. / Parkinsons sjukdom är en hjärnsjukdom orsakad av minskade dopaminnivåer i hjärnan, vilket påverkar kognition och motorisk kontroll i människans hjärna. En av de motoriska kontrollerna som kan påverkas är ögonrörelser och kan därför vara kritiskt påverkat hos patienter diagnostiserade med Parkinsons sjukdom. Ögonrörelser kan mätas med hjälp av ögonspårare, som i sin tur kan användas för att analysera ögonrörelsens egenskaper vid Parkinsons sjukdom. Ögonrörelseanalysen ger möjlighet till diagnostik och kan därför leda till ytterligare förståelse för Parkinsons sjukdom. I denna avhandling studeras särdragsextraktion av ögonrörelsedata med en klinisk relevans vid diagnos av Parkinsonpatienter. Vi har använt en autoencoder (AE) konstruerad för att lära sig mikro- och makrosackadrepresentation för ögonrörelser och konstruerat tre olika modeller. Inlärning av AE utvärderades med hjälp av F1-poängen och skillnader bedömdes statistiskt med hjälp av Wilcoxon rank test. Särdragsextraktionen visualiserades med t-SNE och med hjälp av resultatet ifrån särdragsextraktion har vi mätt skillnader med cosinus- och Mahalanobis- avstånd. Vi har dessutom grupperat resultatet ifrån särdragsextraktionen med fuzzy c-means. Kvaliteten hos de genererade klusterna bedömdes med F1- poäng, suddig fördelningskoefficient, Dunns index och silhuettindex.Sammanfattningsvis finner vi att en mikro-makro-autokodare inte passar syftet med att analysera konstgjorda ögonrörelsesdata. Vi tror att nätverket som används i denna avhandling har lärt sig att representera naturlig ögonrörelse ifrån en person som fritt får röra sina ögon.

Deep learning-based dose prediction is a promising approach to automated radiotherapy planning but carries with it the risk of failing silently when the inputs are highly abnormal compared to the training data. One way to address this issue is to develop a dedicated outlier detector capable of detecting anomalous patient geometries. I examine the potential of so-called generative models to handle this task. These models are promising due to being able to model the distribution of the input data regardless of the downstream task, but they have also been shown to suffer from serious biases when applied to outlier detection. No consensus has been reached regarding the root cause of these biases, or how to address them. I investigate this by attempting to design a variational autoencoder-based outlier detector trained to detect anomalous samples of shapes represented in a binary voxel format. I find the standard procedure application to suffer from severe bias when encountering cropped shapes, leading to systematic misclassification of some outlier patient cases. I overcome this by adopting a segmentation metric as an out-of-distribution metric and show that this outperforms recently proposed general-purpose solutions to the likelihood bias issue. I then benchmark my proposed method on clinical samples and conclude that this approach achieves performance comparable to a one-class support vector machine model that uses handcrafted domain-specific features. / Djupinlärningsbaserad dosprediktion är en mycket lovande metod för att automatiskt generera behandlingsplaner för strålterapi. Djupinlärningsmodeller kan dock endast förväntas fungera på data som är tillräckligt lik träningsdatan, vilket skapar en säkerhetsrisk i kliniska miljöer. Ett möjlig lösning på detta problem är att använda en särskild detektor som klarar av att identifiera avvikande data. I denna uppsats undersöker jag om en generativa djupinlärningsmodell kan användas som en sådan detektor. Generativa modeller är särskilt intressanta för detta ändamål då de är både kraftfulla och flexibla. Dessvärre har generativa modeller visats kunna vilseledas av vissa typer av data. Orsakerna och de underliggande faktorerna till detta har ännu inte identifierats. Jag undersöker denna problematik genom att designa en detektor baserad på en variationell autokodare. Jag upptäcker att den en naiv applikation av denna modell inte är tillräcklig för den kliniska datan, då modellen systematiskt felvärderar beskärda former. Jag löser detta problem genom att nyttja ett modifierat segmenteringsmått som detektionsmått, och visar att denna metod fungerar bättre än mer allmänna lösningar på vilseledningsproblemet. Jag evaluerar metoderna på klinisk data och finner att min metod fungerar lika bra som en en-klass stödvektormaskin som använder sig av handgjorda domänspecifika features.

The strive for cost reduction of services and repairs combined with a desire for increased vehicle reliability has led to the development of predictive maintenance programs. In maintenance plans, accurate forecasts and predictions regarding which components in a vehicle is in risk of a breakdown is bene_cial to obtain since this enables components to be predictively exchanged or serviced before they break down and cause unnecessary downtime. Previous works in data driven predictive maintenance models typically utilize customer and operational data to predict component wear trough regressive or classi_er models. In this thesis the possibilities and bene_ts associated with utilizing vehicle repair and service history data for trucks in a predictive model is investigated. The repair and service data is a time series of irregularly sampled visits to a service centre and is used in conjunction with operational data and chassis con_guration data collected by a truck manufacturer. To tackle the problem a Random Forest, a Neural Network as well as a Recurrent Neural Network model was tested on the various datasets. The Recurrent Neural Network model made it possible to utilize the entire vehicle repair time series data whereas the Random Forest model used a condensed form of the repair data. The Recurrent model proved to perform signi_cantly better than the Neural Network model trained on operational data however it was not proven signi_cantly better than a Random Forest model trained on the condensed form of repair data. A conclusion that can be drawn is that repair history data can increase the performance of a predictive model, however it is unclear if the time sequence plays a part or if a list of previously exchanged parts works equally well. / Strävan efter att reducera kostnader av reparationer och service samt att öka fordons pålitlighet har lett till utvecklingen av prediktiva underhållsprogram. Träffsäkra förutsägeleser och prediktioner kring vilka delar som riskerar att fallera möjliggör prediktiva utbytelser eller service av delar innan de går sönder. Tidigare arbeten i prediktivt underhåll använder sig vanligen av kunddata och operationell data för att generera en prediktion genom regressions eller klassificeringsmetoder. I det här examensarbetet utforskas möjligheterna och fördelarna med att använda verkstadsdata från lastbilar i en prediktiv modell. Verkstadsdatan består av en oregelbundet genererad tidsserie av besök till en serviceanläggning och används i kombination med operationell data samt chassiutförandedata. För att angripa problemet användes en Random Forest, en Neuronnäts samt en Recurrent (Återkommande) Neuronnätsmodell på de olika datakällorna. Recurrent Neuronnätsmodellen möjliggjorde användandet av kompletta tidserieverkstadsdatan och denna modell visade sig ge bäst resultat men kunde inte påvisas  vara signifikant bättre än en Random Forest modell som tränades på en komprimerad variant av verkstadsdatan.  En slutsats som kan dras av arbetet är att verkstadsdatan kan öka prestandan i en prediktiv model men att det är oklart om det är tidssekvensen av datat som ger ökningen eller om det fungerar lika bra med en lista över tidigare utbytta delar.

Drinking water is one of our most important re- sources, so the ability to reliably monitor harmful contaminations in our water distribution network is vital. In order to minimize false alarms for water monitoring, while keeping a high sensitivity, a machine learning approach was evaluated in this project. Measurement data captured with a new kind of sensor, an electronic tongue, was provided by Linköping university. The solution was an artificial neural network, in the structure of an Autoencoder, which could learn the dynamic behaviour of natural deviations and with a false alarm rate of approximately one false alarm per week. This was done by evaluating the data and assembling an input structure to account for daily cyclic phenomena, which then was used to train the neural network. The solution could detect anomalies as small as 1.5% by comparing the input with the reconstructed vector, and raise an alarm. In conclusion, an Autoencoder is a viable method for detecting anomalies in water quality. / Drickvatten är en av våra mest värdefulla tillgångar, det är därför mycket viktigt att det finns sätt att pålitligt övervaka om dricksvattennätet blivit förorenat. För att kunna minimera antalet falsklarm och samtidigt ha hög känslighet mot dessa föroreningar undersöktes och implementerades en lösning med maskininlärningsalgoritmer. Mätdata tillhandahölls av Linköpings universitet och kom från en ny sensor kallad elektronisk tunga. Lösningen var ett artificiellt neuralt nätverk i form av en Autoencoder, som kunde lära sig det dynamiska beteende som ofarliga avvikelser utgjorde. Detta gav en lösning som i medel gav ett falsklarm per sju dagar. Detta gjordes genom att utvärdera rådata och konstruera en struktur på indata som tar hänsyn till dygnsbunda naturliga fenomen. Denna struktur användes sedan för att träna det neurala nätverket. Lösningen kunde upptäcka fel ner till 1.5% genom att jämföra indata med den rekonstruerade vektorn, och på så sätt ge ett alarm. / Kandidatexjobb i elektroteknik 2020, KTH, Stockholm

This dissertation addresses multi-layer optimization aspects of multiple input multiple output (MIMO) and deep learning-based communication systems. The initial focus is on the rate optimization for multi-user MIMO (MU-MIMO) configurations; specifically, multiple access channel (MAC) and interference channel (IC). First, the ergodic sum rates of MIMO MAC and IC configurations are determined by jointly integrating the error and overhead effects due to channel estimation (training) and feedback into the rate optimization. Then, we investigated methods that will increase the achievable rate for parallel Gaussian IC (PGIC) which is a special case of MIMO IC where there is no interference between multiple antenna elements. We derive a generalized iterative waterfilling algorithm for power allocation that maximizes the ergodic achievable rate. We verified the sum rate improvement with our proposed scheme through extensive simulation tests. Next, we introduce a novel physical layer scheme for single user MIMO spatial multiplexing systems based on unsupervised deep learning using an autoencoder. Both transmitter and receiver are designed as feedforward neural networks (FNN) and constellation diagrams are optimized to minimize the symbol error rate (SER) based on the channel characteristics. We first evaluate the SER in the presence of a constant Rayleigh-fading channel as a performance upper bound. Then, we quantize the Gaussian distribution and train the autoencoder with multiple quantized channel matrices. The channel is provided as an input to both the transmitter and the receiver. The performance exceeds that of conventional communication systems both when the autoencoder is trained and tested with single and multiple channels and the performance gain is sustained after accounting for the channel estimation error. Moreover, we evaluate the performance with increasing number of quantization points and when there is a difference between training and test channels. We show that the performance loss is minimal when training is performed with sufficiently large number of quantization points and number of channels. Finally, we develop a distributed and decentralized MU-MIMO link selection and activation protocol that enables MU-MIMO operation in wireless networks. We verified the performance gains with the proposed protocol in terms of average network throughput. / Doctor of Philosophy / Multiple Input Multiple Output (MIMO) wireless systems include multiple antennas both at the transmitter and receiver and they are widely used today in cellular and wireless local area network systems to increase robustness, reliability and data rate. Multi-user MIMO (MU-MIMO) configurations include multiple access channel (MAC) where multiple transmitters communicate simultaneously with a single receiver; interference channel (IC) where multiple transmitters communicate simultaneously with their intended receivers; and broadcast channel (BC) where a single transmitter communicates simultaneously with multiple receivers. Channel state information (CSI) is required at the transmitter to precode the signal and mitigate interference effects. This requires CSI to be estimated at the receiver and transmitted back to the transmitter in a feedback loop. Errors occur during both channel estimation and feedback processes. We initially analyze the achievable rate of MAC and IC configurations when both channel estimation and feedback errors are taken into account in the capacity formulations. We treat the errors associated with channel estimation and feedback as additional noise. Next, we develop methods to maximize the achievable rate for IC by using interference cancellation techniques at the receivers when the interference is very strong. We consider parallel Gaussian IC (PGIC) which is a special case of MIMO IC where there is no interference between multiple antenna elements. We develop a power allocation scheme which maximizes the ergodic achievable rate of the communication systems. We verify the performance improvement with our proposed scheme through simulation tests. Standard optimization techniques are used to determine the fundamental limits of MIMO communications systems. However, there is still a gap between current operational systems and these limits due to complexity of these solutions and limitations in their assumptions. Next, we introduce a novel physical layer scheme for MIMO systems based on machine learning; specifically, unsupervised deep learning using an autoencoder. An autoencoder consists of an encoder and a decoder that compresses and decompresses data, respectively. We designed both the encoder and the decoder as feedforward neural networks (FNNs). In our case, encoder performs transmitter functionalities such as modulation and error correction coding and decoder performs receiver functionalities such as demodulation and decoding as part of the communication system. Channel is included as an additional layer between the encoder and decoder. By incorporating the channel effects in the design process of the autoencoder and jointly optimizing the transmitter and receiver, we demonstrate the performance gains over conventional MIMO communication schemes. Finally, we develop a distributed and decentralized MU-MIMO link selection and activation protocol that enables MU-MIMO operation in wireless networks. We verified the performance gains with the proposed protocol in terms of average network throughput.

Master of Science / There has been an astronomical increase in data at the network edge due to the rapid development of 5G infrastructure and the proliferation of the Internet of Things (IoT). In order to improve the network controller's decision-making capabilities and improve the user experience, it is of paramount importance to properly analyze this data. However, transporting such a large amount of data from edge devices to the network controller requires large bandwidth and increased latency, presenting a significant challenge to resource-constrained wireless networks. By using information processing techniques, one could effectively address this problem by sending only pertinent and critical information to the network controller. Nevertheless, finding critical information from high-dimensional observation is not an easy task, especially when large amounts of background information are present. Our thesis proposes to extract critical but low-dimensional information from high-dimensional observations using an information-theoretic deep learning framework. We focus on two distinct problems where critical information extraction is imperative. In the first problem, we study the problem of feature extraction from video frames collected in a dynamic environment and showcase its effectiveness using a video game simulation experiment. In the second problem, we investigate the detection of anomaly signals in the spectrum by extracting and analyzing useful features from spectrograms. Using extensive simulation experiments based on a practical data set, we conclude that our proposed approach is highly effective in detecting anomaly signals in a wide range of signal-to-noise ratios.

<p dir="ltr">This thesis revisits self-organizing maps (SOM) and explores their applications as both a clustering and feature learning method, establishing SOM as a powerful tool for analyzing complex biological and temporal datasets. The first study integrates evolutionary information with co-fractionation mass spectrometry data to create enhanced benchmarks for protein complex predictions. The second study applies SOM to investigate the intricate relationships between mRNA and protein expression levels during cotton fiber development. By clustering mRNA-protein pairs based on their time-course profiles, SOM captures distinct, non-linear patterns that extend beyond traditional linear correlation methods, offering new insights into gene expression and protein production dynamics. In the third study, SOM is integrated with a generative model, the Variational Autoencoder (VAE), and a Long Short-Term Memory (LSTM) network to tackle challenges in time-series clustering. This framework combines VAE for learning latent representations, LSTM for capturing temporal dependencies, and SOM for clustering, achieving superior performance in modeling complex temporal patterns. Through these studies, this thesis redefines SOM as a versatile and effective tool for uncovering complex patterns in biological and temporal datasets, demonstrating its relevance and potential in modern data analysis.</p>

Construction of radiotherapy treatments for cancer is a laborious and time consuming task. At the same time, when presented with a treatment plan, an oncologist can quickly judge whether or not it is suitable. This means that the problem of constructing these treatment plans is well suited for automation. This thesis investigates a novel way of automatic treatment planning. The treatment planning system this pipeline is constructed for provides dose mimicking functionality with probability density functions of dose–volume histograms (DVHs) and spatial dose as inputs. Therefore this will be the output of the pipeline. The input is historically treated patient scans, segmentations and spatial doses. The approach involves three modules which are individually replaceable with little to no impact on the remaining two modules. The modules are: an autoencoder as a feature extractor to concretise important features of a patient segmentation, a distance optimisation step to learn a distance in the previously constructed feature space and, finally, a probabilistic spatial dose estimation module using sparse pseudo-input Gaussian processes trained on voxel features. Although performance evaluation in terms of clinical plan quality was beyond the scope of this thesis, numerical results show that the proposed pipeline is successful in capturing salient features of patient geometry as well as predicting reasonable probability distributions for DVH and spatial dose. Its loosely connected nature also gives hope that some parts of the pipeline can be utilised in future work. / Skapandet av strålbehandlingsplaner för cancer är en tidskrävande uppgift. Samtidigt kan en onkolog snabbt fatta beslut om en given plan är acceptabel eller ej. Detta innebär att uppgiften att skapa strålplaner är väl lämpad för automatisering. Denna uppsats undersöker en ny metod för att automatiskt generera strålbehandlingsplaner. Planeringssystemet denna metod utvecklats för innehåller funktionalitet för dosrekonstruktion som accepterar sannolikhetsfördelningar för dos–volymhistogram (DVH) och dos som input. Därför kommer detta att vara utdatan för den konstruerade metoden. Metoden är uppbyggd av tre beståndsdelar som är individuellt utbytbara med liten eller ingen påverkan på de övriga delarna. Delarna är: ett sätt att konstruera en vektor av kännetecken av en patients segmentering, en distansoptimering för att skapa en distans i den tidigare konstruerade känneteckensrymden, och slutligen en skattning av sannolikhetsfördelningar med Gaussiska processer tränade på voxelkännetecken. Trots att utvärdering av prestandan i termer av klinisk plankvalitet var bortom räckvidden för detta projekt uppnåddes positiva resultat. De estimerade sannolikhetsfördelningarna uppvisar goda karaktärer för både DVHer och doser. Den löst sammankopplade strukturen av metoden gör det dessutom möjligt att delar av projektet kan användas i framtida arbeten.

Modelling music with mathematical and statistical methods in general, and with neural networks in particular, has a long history and has been well explored in the last decades. Exactly when the first attempt at strictly systematic music took place is hard to say; some would say in the days of Mozart, others would say even earlier, but it is safe to say that the field of algorithmic composition has a long history. Even though composers have always had structure and rules as part of the writing process, implicitly or explicitly, following rules at a stricter level was well investigated in the middle of the 20th century at which point also the first music writing computer program based on mathematics was implemented. This work in computer science focuses on the history of musical composition with computers, also known as algorithmic composition, using machine learning and neural networks and consists of two parts: a literature survey covering in-depth the last decades in the field from which is drawn inspiration and experience to construct MahlerNet, a neural network based on the previous architectures MusicVAE, BALSTM, PerformanceRNN and BachProp, capable of modelling polyphonic symbolic music with up to 23 instruments. MahlerNet is a new architecture that uses a custom preprocessor with musical heuristics to normalize and filter the input and output files in MIDI format into a data representation that it uses for processing. MahlerNet, and its preprocessor, was written altogether for this project and produces music that clearly shows musical characteristics reminiscent of the data it was trained on, with some long-term structure, albeit not in the form of motives and themes. / Matematik och statistik i allmänhet, och maskininlärning och neurala nätverk i synnerhet, har sedan långt tillbaka använts för att modellera musik med en utveckling som kulminerat under de senaste decennierna. Exakt vid vilken historisk tidpunkt som musikalisk komposition för första gången tillämpades med strikt systematiska regler är svårt att säga; vissa skulle hävda att det skedde under Mozarts dagar, andra att det skedde redan långt tidigare. Oavsett vilket, innebär det att systematisk komposition är en företeelse med lång historia. Även om kompositörer i alla tider följt strukturer och regler, medvetet eller ej, som en del av kompositionsprocessen började man under 1900-talets mitt att göra detta i högre utsträckning och det var också då som de första programmen för musikalisk komposition, baserade på matematik, kom till. Den här uppsatsen i datateknik behandlar hur musik historiskt har komponerats med hjälp av datorer, ett område som också är känt som algoritmisk komposition. Uppsatsens fokus ligger på användning av maskininlärning och neurala nätverk och består av två delar: en litteraturstudie som i hög detalj behandlar utvecklingen under de senaste decennierna från vilken tas inspiration och erfarenheter för att konstruera MahlerNet, ett neuralt nätverk baserat på de tidigare modellerna MusicVAE, BALSTM, PerformanceRNN och BachProp. MahlerNet kan modellera polyfon musik med upp till 23 instrument och är en ny arkitektur som kommer tillsammans med en egen preprocessor som använder heuristiker från musikteori för att normalisera och filtrera data i MIDI-format till en intern representation. MahlerNet, och dess preprocessor, är helt och hållet implementerade för detta arbete och kan komponera musik som tydligt uppvisar egenskaper från den musik som nätverket tränats på. En viss kontinuitet finns i den skapade musiken även om det inte är i form av konkreta teman och motiv.

Data augmentation is a technique that can be performed in various ways to improve the training of discriminative models. The recent developments in deep generative models offer new ways of augmenting existing data sets. In this thesis, a framework for augmenting annotated data sets with deep generative models is proposed together with a method for quantitatively evaluating the quality of the generated data sets. Using this framework, two data sets for pupil localization was generated with different generative models, including both well-established models and a novel model proposed for this purpose. The unique model was shown both qualitatively and quantitatively to generate the best data sets. A set of smaller experiments on standard data sets also revealed cases where this generative model could improve the performance of an existing discriminative model. The results indicate that generative models can be used to augment or replace existing data sets when training discriminative models. / Dataaugmentering är en teknik som kan utföras på flera sätt för att förbättra träningen av diskriminativa modeller. De senaste framgångarna inom djupa generativa modeller har öppnat upp nya sätt att augmentera existerande dataset. I detta arbete har ett ramverk för augmentering av annoterade dataset med hjälp av djupa generativa modeller föreslagits. Utöver detta så har en metod för kvantitativ evaulering av kvaliteten hos genererade data set tagits fram. Med hjälp av detta ramverk har två dataset för pupillokalisering genererats med olika generativa modeller. Både väletablerade modeller och en ny modell utvecklad för detta syfte har testats. Den unika modellen visades både kvalitativt och kvantitativt att den genererade de bästa dataseten. Ett antal mindre experiment på standardiserade dataset visade exempel på fall där denna generativa modell kunde förbättra prestandan hos en existerande diskriminativ modell. Resultaten indikerar att generativa modeller kan användas för att augmentera eller ersätta existerande dataset vid träning av diskriminativa modeller.