Deep learning has dominated the computer vision field since 2012, but a common criticism of deep learning methods is their dependence on large amounts of data. To combat this criticism research into data-efficient deep learning is growing. The foremost success in data-efficient deep learning is transfer learning with networks pre-trained on the ImageNet dataset. Pre-trained networks have achieved state-of-the-art performance on many tasks. We consider the pre-trained network method for a new task where we have to collect the data. We hypothesize that the data efficiency of pre-trained networks can be improved through informed data collection. After exhaustive experiments on CaffeNet and VGG16, we conclude that the data efficiency indeed can be improved. Furthermore, we investigate an alternative approach to data-efficient learning, namely adding domain knowledge in the form of a spatial transformer to the pre-trained networks. We find that spatial transformers are difficult to train and seem to not improve data efficiency.

Atrial fibrillation is an arrhythmia commonly detected from ECG using its specific characteristics. An early detection of this arrhythmia is a key to prevention of more serious conditions. Nowadays, atrial fibrillation detection is being implemented more often using deep learning. This work presents detection of atrial fibrillation from 12lead ECG using deep convolutional network. In the first section, there is a theoretical context of this work, then there is a description of proposed algorithm. Detection is implemented by a program in Python in two variations and their accuracy is rated by Accuracy and F1 measure. Results of the work are being discussed, mutually compared and compared to other similar publications.

This thesis attempts to find an alternative method for biometric identification using retinal images. First part is focused on the introduction to biometrics, human eye anatomy and methods used for retinal biometry. The essence of neural networks and deep learning methods is described as it will be used practically. In the last part of the thesis a chosen identification algorithm and its implementation is described and the results are presented.

The clue of learning to recognize objects using neural network lies in imitation of animal neural network's behavior. In spite the details of how brain works is not known yet, the teams consisting of scientists from various medical or technical professions are trying to search for them. Thanks to giants like Geoffrey Hinton science made a big progress in this domain. The convolutional networks which are based on animal model of optical system can be advantageously used for image segmentation and therefore they ware chosen for segmentation of tumor and edema from images of magnetic resonance. The models of artificial neural networks used in this work had achieved the 41\% of success in edema segmentation and 79\% in segmentation of tumor from brain issue.

Knowledge Graphs are fascinating concepts in machine learning as they can hold usefully structured information in the form of entities and their relations. Despite the valuable applications of such graphs, most knowledge bases remain incomplete. This missing information harms downstream applications such as information retrieval and opens a window for research in statistical relational learning tasks such as node classification and link prediction. This work proposes a deep learning framework based on existing relational convolutional (R-GCN) layers to learn on highly multi-relational data characteristic of realistic knowledge graphs for node property classification tasks. We propose a deep and improved variant, Deep-RGCNs, with dense and residual skip connections between layers. These skip connections are known to be very successful with popular deep CNN-architectures such as ResNet and DenseNet. In our experiments, we investigate and compare the performance of Deep-RGCN with different baselines on multi-relational graph benchmark datasets, AIFB and MUTAG, and show how the deep architecture boosts the performance in the task of node property classification. We also study the training performance of Deep-RGCNs (with N layers) and discuss the gradient vanishing and over-smoothing problems common to deeper GCN architectures.

This report analyzes the difference between discriminative and generative image classifiers when tested on noise. The generative classifier was a maximum-likelihood based classifier using a normalizing flow as the generative model. In this work, a coupling flow such as RealNVP was used. For the discriminative classifier a convolutional network was implemented. A detailed description of how these classifiers were implemented is given in the report. The report shows how this generative classifier outperforms the discriminative classifier when tested on adversarial noise. However, tests are also conducted on salt and pepper noise and Gaussian noise, here the results show that the generative classifier gets outperformed by the discriminative classifier. Tests were also conducted on Gaussian noise once both classifiers had been trained on Gaussian noise, the results from these tests show that the discriminative classifier performs significantly better once trained on Gaussian noise. However, the generative classifier does only show marginal increases in performance and performs worse on clean data once trained on Gaussian noise. / Den här rapporten analyserar skillnaden mellan diskriminativa och generativa modellklasser för bildigenkänning när de testas på brus. Den generativa modellklassen var en maximum-likelihood baserad generativ klassifikationsmodell. Inom detta arbete användes kopplingsflödet RealNVP. För den diskriminativa bildigenkänningsmodellen så implementerades ett faltningsnätverk. En detaljerad beskrivning för hur dessa bildigenkänningsmodeller genomfördes är given i rapporten. Rapporten visar hur den generativa modellklassen överträffar den diskriminativa modellklassen när de testas på adversarialt brus. Testerna utförs emellertid med salt och peppar brus och Gaussiskt brus, för dessa visar resultaten att den generativa modellklassen överträffas av den diskriminativa modellklassen. Den generativa modellklassen visar emellertid endast marginella ökningar i prestanda, och har en sämre prestanda på ren data efter att den tränats på Gaussiskt brus. / Kandidatexjobb i elektroteknik 2022, KTH, Stockholm

Instruction Scheduling is a back-end compiler optimisation technique that can provide significant performance gains. It refers to ordering instructions in a particular order to reduce latency for processors with instruction-level parallelism. At the present typical compilers use heuristics to perform instruction scheduling and solve other related non-polynomial complete problems. This thesis aims to present a machine learning-based approach to challenge heuristic methods concerning performance. In this thesis, a novel reinforcement learning (RL) based model for the instruction scheduling problem is developed including modelling features of processors such as forwarding, resource utilisation and treatment of the action space. An efficient optimal scheduler is presented to be used for an optimal schedule length based reward function, however, this is not used in the final results as a heuristic based reward function was deemed to be sufficient and faster to compute. Furthermore, an RL agent that interacts with the model of the problem is presented using three different types of graph neural networks for the state processing: graph conventional networks, graph attention networks, and graph attention based on the work of Lee et al. A simple two-layer neural network is also used for generating embeddings for the resource utilisation stages. The proposed solution is validated against the modelled environment and favourable but not significant improvements were found compared to the most common heuristic method. Furthermore, it was found that having embeddings relating to resource utilisation was very important for the explained variance of the RL models. Additionally, a trained model was tested in an actual compiler, however, no informative results were found likely due to register allocation or other compiler stages that occur after instruction scheduling. Future work should include improving the scalability of the proposed solution. / Instruktionsschemaläggning är en optimeringsteknik för kompilatorer som kan ge betydande prestandavinster. Det handlar om att ordna instruktioner i en viss ordning för att minska latenstiden för processorer med parallellitet på instruktionsnivå. För närvarande använder vanliga kompilatorer heuristiker för att utföra schemaläggning av instruktioner och lösa andra relaterade ickepolynomiala kompletta problem. Denna avhandling syftar till att presentera en maskininlärningsbaserad metod för att utmana heuristiska metoder när det gäller prestanda. I denna avhandling utvecklas en ny förstärkningsinlärningsbaserad (RL) modell för schemaläggning av instruktioner, inklusive modellering av processorns egenskaper såsom vidarebefordran, resursutnyttjande och behandling av handlingsutrymmet. En effektiv optimal schemaläggare presenteras för att eventuellt användas för belöningsfunktionen, men denna används inte i de slutliga resultaten. Dessutom presenteras en RL-agent som interagerar med problemmodellen och använder tre olika typer av grafneurala nätverk för tillståndsprocessering: grafkonventionella nätverk, grafuppmärksamhetsnätverk och grafuppmärksamhet baserat på arbetet av Lee et al. Ett enkelt neuralt nätverk med två lager används också för att generera inbäddningar för resursanvändningsstegen. Den föreslagna lösningen valideras mot den modellerade miljön och gynnsamma men inte signifikanta förbättringar hittades jämfört med den vanligaste heuristiska metoden. Dessutom visade det sig att det var mycket viktigt för den förklarade variansen i RL-modellerna att ha inbäddningar relaterade till resursutnyttjande. Dessutom testades en tränad modell i en verklig kompilator, men inga informativa resultat hittades, sannolikt på grund av registerallokering eller andra kompilatorsteg som inträffar efter schemaläggning av instruktioner. Framtida arbete bör inkludera att förbättra skalbarheten hos den föreslagna lösningen.

[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.

<p dir="ltr">Since its inception, the field of heterogeneous catalysis has evolved to address the needs of the ever-growing human population. Necessity, after all, fosters innovation. Today, the world faces numerous challenges related to anthropogenic climate change, and that has necessitated, among other things, a search for new catalysts that can enable renewable energy conversion and storage, sustainable food and chemicals production, and a reduction in carbon emissions. This search has led to the emergence of many promising classes of materials, each having a unique set of catalytic properties. Among such candidate materials, high-entropy alloys (HEAs) have very recently shown the potential to be a new catalyst design paradigm. HEAs are multimetallic, disordered alloys containing more than four elements and, as a result, possess a higher configurational entropy, which gives them considerable stability. They have many conceivable benefits over conventional bimetallic alloy catalysts—greater site heterogeneity, larger design space, and higher stability, among others­. Consequently, there is a need to explore their application in a wide range of thermal and electrocatalytic reaction systems so that their potential can be realized.</p><p dir="ltr">In the past few decades, first principles-based approaches involving Density Functional Theory (DFT) calculations have proven to be effective in probing catalytic mechanisms at the atomic scale. Fundamental insights from first principles studies have also led to a detailed understanding of reactivity and stability trends for bimetallic alloy catalysts. However, the express application of first principles approaches to study HEA catalysts remains a challenge, due to the large computational cost incurred in performing DFT calculations for disordered alloys, which can be represented by millions of different configurations. A combination of first principles approaches and computationally efficient machine learning (ML) approaches can, however, potentially overcome this limitation.</p><p dir="ltr">In this thesis, combined workflows involving first principles and machine learning-based approaches are developed. To map catalyst structure to properties graph convolutional network (GCN) models are developed and trained on DFT-predicted target properties such as formation energies, surface energies, and adsorption energies. Further, the Monte Carlo dropout method is integrated into GCN models to provide uncertainty quantification, and these models are in turn used in active learning workflows that involve iterative model retraining to both improve model predictions and optimize the target property value. Dimensionality reduction methods, such as principal components analysis (PCA) and Diffusion Maps (DMaps), are used to glean physicochemical insights from the parameterization of the GCN.</p><p dir="ltr">These workflows are applied to the analysis of binary, ternary, and quaternary alloy catalysts, and a series of fundamental insights regarding their stability are elucidated. In particular, the origin and stability of “Pt skins” that form on Pt-based bimetallic alloys such as Pt₃Ni in the context of the oxygen reduction reaction (ORR) are investigated using a rigorous surface thermodynamic framework. The active learning workflow enables the study of Pt skin formation on stepped facets of Pt₃Ni (with a complex, low-symmetry geometry), and this analysis reveals a hitherto undiscovered relationship between surface coordination and surface segregation. In another study, an active learning workflow is used to identify the most stable bulk composition in the Pd-Pt-Sn ternary alloy system using a combination of exhaustively sampled binary alloy data and prudently sampled ternary alloy data. Lastly, a new GCN model architecture, called SlabGCN, is introduced to predict the sulfur poisoning characteristics of quaternary alloy catalysts, and to find an optimal sulfur tolerant composition.</p><p dir="ltr">On another front, the electrocatalytic activity of quinary HEAs towards the ORR is investigated by performing DFT calculations on HEA structures generated using the High-Entropy Alloy Toolbox (HEAT), an in-house code developed for the high-throughput generation and analysis of disordered alloy structures with stability constraints (such as Pt skin formation). DFT-predicted adsorption energies of key ORR intermediates are further deconvoluted into ligand, strain, and surface relaxation effects, and the influence of the number of Pt skins on these effects is expounded. A Sabatier volcano analysis is performed to calculate the ORR activities of selected HEA compositions, and correspondence between theoretical predictions and experimental results is established, to pave the way for rational design of HEA catalysts for oxygen reduction.</p><p dir="ltr">In summary, this thesis examines stability and reactivity trends of a multitude of alloy catalysts, from conventional bimetallic alloys to high-entropy alloys, using a combination of first principles approaches (involving Density Functional Theory calculations) and machine learning approaches comprising graph convolutional network models.</p>

A better understanding of the traffic flow in a city helps to smooth transport resulting in a better street environment, affecting not only road users and people in proximity. Good predictions of the flow of traffic helps to control and further develop the road network in order to avoid congestion and unneccessary time spent while traveling. This study investigates three different machine learning models with the purpose of predicting traffic flow on different road types inurban Stockholm using loop sensor data between 2013 and 2023. The models used was Long short term memory (LSTM), Temporal convolutional network (TCN) and a hybrid model of LSTM and TCN. The results from the hybrid model indicates a slightly better mean absolute error than TCN suggesting that a hybrid model might be advantagous when predicting traffic flow using loop sensor data. LSTM struggled to capture the complexity of the data and was unable to provide a proper prediction as a result. TCN produced a mean absolute error slightly bigger than the hybrid model and was to an extent able to capture the trends of the traffic flow, but struggled with capturing the scale of the traffic flow suggesting the need for further data preprocessing. Furthermore, this study suggests that the loop sensor data was able to act as a foundation for predicting the traffic flow using machine learning methods. However, it suggest that improvements to the data itself such as incorporating more related parameters might be advantageous to further improve traffic flow prediction.