I en värld som digitaliseras allt mer blir databaser mer komplexa än någonsin. För att kunna lita att information är korrekt uppdaterad i system behöver man kunna lita på underliggande integrationer. Det sker fel i många databaser och detta kan resultera i problem i framtiden då fel kan orsaka skador både för kunden och ett företags rykte. Det är viktigt att en databas är säker och att det uppstår så lite fel som möjligt. I detta examensarbete undersöks det hur machine learning kan användas för att förutspå fel i en databas för att enkelt kunna förebygga dessa. Olika typer av inlärningsalgoritmer analyseras för att finna den som passar bäst in för arbetet. Fyra olika algoritmer som ML.NETbidrar med analyseras sedan för att presentera vilken algoritm som är mest passande till problemet och som har bäst prestanda med hänsyn till noggrannhet och körtid.

No description available.

Community detection is an important graph mining task and one of the most researched problems in its field of study. One reason for this is its applicability in a variety of disciplines ranging from biology to computer science. Community detection methods differ, however, and a major reason for this is due to the fact that there does not exist a unique definition for what a community is. This project involved creating a new method for detecting and building communities for a specific type of network represented by multi-layered graphs, where each layer represents an edge type/ type of relation. More specifically, an algorithm for detecting and building communities in multi-layered graphs was built by first implementing a method to detect semi-cliques spanning multiple graphs, and then integrating the method with a second method which builds communities using multi-layered cliques. The new method wasthen tested on synthetic as well as real-world data to demonstrate functionality and test validity.

Visual and lidar-based odometry for mobile robots has been thoroughlyinvestigated and performs very well in good weather conditions. However,both are sensitive to bad weather conditions with atmospheric disturbancessuch as rain and snow. Recently Radar sensors specialized for mobilerobot use have become available. Radar sensors are much more robustagainst atmospheric disturbances, which makes them an exciting alternative.This thesis presents a radar odometry pipeline that can handle both lidar andradar data with minor modifications. The results show that it outperformsthe current state of the art radar odometry solutions. While also being able tohandle 3d lidar odometry with good performance.

As more and more organizations digitize their records, the need for automatic document processing software increases. In particular, the rise of ‘digital humanities’ precede a new set of problems on how to digitize historical archival material in an efficient and accurate manner. The transcription of archival material to formats fit for research purposes, such as handwritten spreadsheets, is still expensive and plagued by tedious manual labor. Over the decades, research in handwritten text recognition has focused on text line extraction and recognition. In this thesis, we examine document images that contain complex details, contain more categories of text than handwriting, and handwritten text that is not separated easily to lines. The thesis examines the sub-problem of handwritten text segmentation in detail. We propose a broad definition of text segmentation that requires both text detection and text classification, since this enables us to detect multiple kinds of text within the same image. The aim is to design a system which can detect and identify both handwriting and machine-text within the same image. Working with photographs of spreadsheet documents from the years 1871-1951, a topdown layout-agnostic image processing pipeline is developed. Different kinds of preprocessing are examined, to correct illumination and enhance contrast before binarization, and to detect and clear line contours. To achieve text region detection, we evaluate connected components labeling and MSER as region detectors, extracting textual and non-textual sub-images. On detected sub-images, we perform a Bag-of-Visual-Words quantization of k-means clustered feature descriptor vectors and perform categorical classification by training a Naïve Bayesclassifier on the feature distances to the cluster centroids. Results include a novel two-stage illumination correction and contrast enhancement algorithm that improves document quality as a precursor to binarization, increasing the mean grayscale values of an image while retaining low grayscale variance. Region detectors are evaluated on images with different types of preprocessing and the results show that clearing document outlines influences text region detection. Training on a small sample of sub-images, the categorical classification model proves viable for discrimination between machine-text and handwriting, enabling the use of this model for further recognition purposes.

Background: There has been a consistent increase in the used cars industry from the past decade as there is an increase in the usage of cars. Usedcars are attracting more attention as they are affordable than new ones.This situation demands high-performance algorithms that can be used topredict prices for the used cars. Many machine learning algorithms are usedto predict the price of cars. Objectives: This thesis aims in detecting features that impact predicting the price of used cars, and experiments are performed to investigatean optimal algorithm for price prediction of used cars. Algorithms selectedfor experimenting are Linear Regression (LR), Light Gradient Boosted Machine (LGBM), Random Forest Regression (RFR), Decision Tree Regression(DTR). These algorithms are further compared using performance metricsof regression models. Methods: The initial step in this study is to gather a suitable dataset andapplying preprocessing techniques to that data. Feature selection is performed using a correlation matrix with the heat map. Label Encoding isperformed on the data to change the categorical values into numerical values. A new dataset is created based on the feature "region" from the originaldataset. train-test-split technique is used to divide the original dataset intotrain and test data in the ratio of 80:20. The new dataset is manually divided into unique regions of train and test data. Selected Machine Learningalgorithms were trained using both datasets. The accuracy score of selectedalgorithms is derived using performance metrics. An optimal algorithm isachieved by comparing the accuracy scores derived. Results: Light Gradient Boosted Machine is considered as optimal algorithm based on R2score, for the original dataset, it obtained 91.12% on testdata. Light Gradient Boosted Machine achieved 85.30% on test data for thenew dataset. The feature "region" has the highest feature importance overthe remaining features. It has a feature importance of 55220 with respectto number of instances i.e, 568654. Conclusions: Among selected algorithms, Light Gradient Boosted Machine obtained a high R2score over other algorithms on both original andnew datasets. Feature "region" has a significant impact on predicting theprice of the used car, and this is justified by performing feature importanceon Light Gradient Boosted Machine.

Context: Importance of facial emotion recognition is increasing significantly as it's applications play a key role in several sectors and fields. Deep learning techniques in machine learning provide good performance in facial recognition tasks, Where as deep neural networks like convolutional neural networks are most widely used for image recognition and classification tasks. However, these neural networks depend on configuration parameters called hyper-parameters. So, tuning these parameters play a vital role in facial emotion recognition. Moreover, it is challenging and time consuming to tune the hyper-parameters of neural networks since it involves many parameters. Tuning these hyper-parameters is considered as optimization task where evolutionary algorithms like genetic algorithms play a major role. Studying and experimenting different genetic algorithm concepts not only provide interesting insights for facial recognition tasks but also provide significant progresses in deep learning, gaming, and virtual reality. Objectives: The thesis aims to develop a model for facial emotion recognition by applying evolutionary mechanisms like genetic algorithms on convolutional neural networks. The developed model recognizes seven basic emotions in images of human beings such as fear,happy, surprise, sad, neutral, disgust and angry using FER-2013(facial emotion recognition) dataset. Methods: Emotion recognition of the facial images is done by hyper-tuning of convolutional neural network using evolutionary mechanisms. Literature review is performed for studying the working mechanism of genetic algorithm, techniques, best methods of genetic algorithms, genetic operators for hyper-parameter tuning of neural network. After studying the methods, experiment is conducted to evaluate and study the impact of applying genetic algorithm methods in hyper-parameter tuning which in turn helps in facial emotion recognition. Results: Genetic algorithm concepts which are identified from literature review improved the performance of convolutional neural network. Elitism and multiparent recombination concepts of genetic algorithm showed satisfying results by significantly boosting the performance of neural networks. Multipoint cross-over established a new theme in genetic algorithm by introducing sharp variations and gave scope for genetic diversity which results in increasing efficiency of neural network. Performed experimental model portrayed these concepts and has improved the performance of convolutional neural network. Conclusions: The genetic algorithm worked constructively for the improvement of performance of convolutional neural networks. Results from experimental model portrayed improvement of neural network and has helped in increasing accuracy of the images of facial emotion recognition. Variable length genetic algorithm helped the model in tracing out important variable parameters thus helping the neural networks to perform better. Different genetic mechanisms have different functions for effective functioning of neural network. Key observations, new insights gained from the experimental results of the current research are helpful and expand the scope of deep learning applications with evolutionary mechanisms.

The advances made in mobile and low-powered computing within the last decade has made mobility a key term of today, where one cannot imagine a daily life without a mobile phone. This is largely due to the availability of smaller and faster hardware, such as multi-core processors and high-speed mobile networks. However, heavy computations or applications performed on multiple cores can be very power consuming and as a result, leave the user with an unusable device. This thesis explores and measures the battery usage when performing parallel tasks on an Android device. This is done by developing an application and algorithm able to be executed on a chosen number of cores. The result is presented with both a system-wide battery usage as well as an app-specific battery usage of the developed application

Introduction: Machine Learning as a Service (MLaaS) is a capture term for a range of cloud-based platforms that use machine learning tools to produce solutions that help machine learning professionals. Many cloud-based service providers have led the road in recent years to provide I.T. specialists with comparatively cheap and instantly available machine learning solutions to simplify machine learning solutions. As a result, there is a greater need to compare cloud-based services that deliver machine learning solutions. As a result, we chose to evaluate AWS sagemaker and Azure ML cloud services in terms of user experience when loading, training, and serving ML models.                                            Background: The use of cloud computing has been increased these days, the companies that provide these services have been gradually increased. Although there are many cloud services available on the market, users should always select the more flexible and efficient ones to use. As a result, our research is focused on comparing cloud services in terms of user experience. Assessment approaches and concepts such as Cognitive Walkthrough and Usability heuristics apply to our study as we delve deeper into user interaction and experience. In this case, the user interfaces of Microsoft Azure Machine Learning Studio and Amazon Web Services sagemaker are compared while loading, training, and serving machine learning models. Objectives: The main objective of this thesis is to compare and evaluate the two cloud services such as AWS sagemaker and Microsoft Azure ML while loading, training, and serving machine learning models to decide which of these two cloud services has the best user interface from the users' perspective using Cognitive Walkthrough. Methods: Determining the best cloud service in terms of user experience between AWS Sage Maker and Microsoft Azure ML is done using Cognitive Walkthrough by executing selected tasks in both cloud services, and comparison is done using Usability heuristics to reach our research conclusions. Results: The results originated from the cognitive walkthrough, and comparison with Usability heuristics are presented in graphical formats such as pie charts. The results of cognitive walkthrough are obtained after completion of each task and best cloud service in users’ perspective is obtained. Conclusions: Finally, we conclude Microsoft Azure machine learning studio is better than AWS sagemaker in terms of user-experience while performing the specified tasks such as loading, training and serving ML models in both the cloud services.

<p>It was an online defence session.</p>