In der Arbeit wird ein neuartiges Lernverfahren für Gauß-Mischverteilungen entwickelt. Es basiert auf der Technik der Markov-Chain Monte-Carlo Verfahren und ist in der Lage, in einem Zuge die Größe der Mischverteilung sowie deren Parameter zu bestimmen. Das Verfahren zeichnet sich sowohl durch eine gute Anpassung an die Trainingsdaten als auch durch eine gute Generalisierungsleistung aus. Ausgehend von einer Beschreibung der stochastischen Grundlagen und einer Analyse der Probleme, die beim Lernen von Gauß-Mischverteilungen auftreten, wird in der Abeit das neue Lernverfahren schrittweise entwickelt und seine Eigenschaften untersucht. Ein experimenteller Vergleich mit bekannten Lernverfahren für Gauß-Mischverteilungen weist die Eignung des neuen Verfahrens auch empirisch nach.

No description available.

No description available.

Traditional image processing solutions have been applied in the past to detect and count strawberries. These methods typically involve feature extraction followed by object detection using one or more features. Some object detection problems can be ambiguous as to what features are relevant and the solutions to many problems are only fully realized when the modern approach has been applied and tested, such as deep learning. In this work, we investigate the use of R-CNN for green strawberry detection. The object detection involves finding regions of interest (ROIs) in field images using the selective segmentation algorithm and inputting these regions into a pre-trained deep neural network (DNN) model. The convolutional neural networks VGG, MobileNet and ResNet were implemented to detect subtle differences between green strawberries and various background elements. Downscaling factors, intersection over union (IOU) thresholds and non-maxima suppression (NMS) values can be tweaked to increase recall and reduce false positives while data augmentation and negative hardminging can be used to increase the amount of input data. The state of the art model is sufficient in locating the green strawberries with an overall model accuracy of 74%. The R-CNN model can then be used for crop yield prediction to forecast the actual red strawberry count one week in advance with a 90% accuracy.

During the broadcast of a golf tournament, a way to show the audience what a player's swing looks like would be to draw a trace following the movement of the club head. A computer vision model can be trained to identify the position of the club head in an image, but due to the high speed at which professional players swing their clubs coupled with the low frame rate of a typical broadcast camera, the club head is not discernible whatsoever in most frames. This means that the computer vision model is only able to deliver a few sparse detections of the club head. This thesis project aims to develop a machine learning model that can predict the complete motion of the club head, in the form of a swing trace, based on the sparse club head detections. Slow motion videos of golf swings are collected, and the club head's position is annotated manually in each frame. From these annotations, relevant data to describe the club head's motion, such as position and time parameters, is extracted and used to train the machine learning models. The dataset contains 256 annotated swings of professional and competent amateur golfers. The two models that are implemented in this project are XGBoost and a feed forward neural network. The input given to the models only contains information in specific parts of the swing to mimic the pattern of the sparse detections. Both models learned the underlying physics of the golf swing, and the quality of the predicted traces depends heavily on the amount of information provided in the input. In order to produce good predictions with only the amount of input information that can be expected from the computer vision model, a lot more training data is required. The traces predicted by the neural network are significantly smoother and thus look more realistic than the predictions made by the XGBoost model.

Open Domain Question Answering (OpenQA) systems find an answer to a question from a large collection of unstructured documents. In this information era, we have an immense amount of data at our disposal. However, filtering all the content and trying to find the answers to our questions can be too time-consuming and ffdiicult. In addition, in such a globalised world, the information we look for to answer a question may be in a different language. Current research is focused on improving monolingual (English) OpenQA performance. This creates a disparity between the tools accessible between English and non-English speakers. The techniques explored in this study involve the combination of different methods, such as data augmentation and hard negative filtering for performance increase, and binary embeddings for improving the efficiency, with multilingual Transformers. The downstream performance is evaluated using sentiment multilingual datasets covering Cross-Lingual Transfer (XLT), question and answer in the same language, and Generalised Cross-Lingual Transfer (G-XLT), different languages for question and answer. The results show that data augmentation increased Recall by 37.0% and Mean Average Precision (MAP) by 67.0% using languages absent from the test set for XLT. Combining binary embeddings and hard negatives can reduce inference time and index size to 12.5% and 3.1% of the original, retaining 97.1% of the original Recall and 94.8% of MAP (averages of XLT and MAP). / Open Domain Question Answering (OpenQA)-system hittar svar på frågor till stora samlingar av ostrukturerade dokument. I denna informationsepok har vi en enorm mängd kunskap till vårt förfogande. Att filtrera allt innehåll för att försöka att hitta svar på våra frågor kan dock vara mycket tidskrävande och svårt. I en globaliserad värld kan informationen vi söker för att besvara en fråga dessutom vara på ett annat språk. Nuvarande forskning är primärt inriktad på att förbättra OpenQA:s enspråkiga (engelska) prestanda. Detta skapar ett gap mellan de verktyg som är tillgängliga för engelsktalande och icke-engelsktalande personer. De tekniker som undersöks i den här studien innebär en kombination av olika metoder, t.ex. dataförstärkning och hård negativ filtrering för att öka prestandan, och binära embeddings för att förbättra effektiviteten med flerspråkiga Transformatorer. Prestandan nedströms utvärderas med hjälp av flerspråkiga dataset som omfattar Cross-Lingual Transfer (XLT), fråga och svar på samma språk, och Generalised Cross-Lingual Transfer (G-XLT), olika språk för fråga och svar. Resultaten visar att dataförstärkning ökade recall med 37.0% och 67.0% för Mean Average Precision (MAP) med hjälp av språk som inte fanns med i testuppsättningen för XLT. Genom att kombinera binära embeddings och hårda negationer kan man minska tiden för inferens och indexstorleken till 12.5% och 3.1% av originalet, samtidigt som man behåller 97.1% av ursprunglig recall samt 94.8% av MAP (medelvärden av XLT och MAP).

No description available.

In this research the focus is on data-centric AI with a specific concentration on data quality evaluation and improvement for machine learning. We first present a practical framework for data quality evaluation and improvement, using a legal domain as a case study and build a corpus for legal argument mining. We first created an initial corpus with 4,937 instances that were manually labeled. We define five data quality evaluation dimensions: comprehensiveness, correctness, variety, class imbalance, and duplication, and conducted a quantitative evaluation on these dimensions for the legal dataset and two existing datasets in the medical domain for medical concept normalization. The first group of experiments showed that class imbalance and insufficient training data are the two major data quality issues that negatively impacted the quality of the system that was built on the legal corpus. The second group of experiments showed that the overlap between the test datasets and the training datasets, which we defined as "duplication," is the major data quality issue for the two medical corpora. We explore several widely used machine learning methods for data quality improvement. Compared to pseudo-labeling, co-training, and expectation-maximization (EM), generative adversarial network (GAN) is more effective for automated data augmentation, especially when a small portion of labeled data and a large amount of unlabeled data is available. The data validation process, the performance improvement strategy, and the machine learning framework for data evaluation and improvement discussed in this dissertation can be used by machine learning researchers and practitioners to build high-performance machine learning systems. All the materials including the data, code, and results will be released at: https://github.com/haihua0913/dissertation-dqei.

CNN-based (Convolutional Neural Network) visual object detectors often reach human level of accuracy but need to be trained with large amounts of manually annotated data. Collecting and annotating this data can frequently be time-consuming and financially expensive. Using generative models to augment the data can help minimize the amount of data required and increase detection per-formance. Many state-of-the-art generative models are Generative Adversarial Networks (GANs). This thesis investigates if and how one can utilize image data to generate new data through GANs to train a YOLO-based (You Only Look Once) object detector, and how CAD (Computer-Aided Design) models can aid in this process. In the experiments, different models of GANs are trained and evaluated by visual inspection or with the Fréchet Inception Distance (FID) metric. The data provided by Ericsson Research consists of images of antenna and baseband equipment along with annotations and segmentations. Ericsson Research supplied the YOLO detector, and no modifications are made to this detector. Finally, the YOLO detector is trained on data generated by the chosen model and evaluated by the Average Precision (AP). The results show that the generative models designed in this work can produce RGB images of high quality. However, the quality reduces if binary segmentation masks are to be generated as well. The experiments with CAD input data did not result in images that could be used for the training of the detector. The GAN designed in this work is able to successfully replace objects in images with the style of other objects. The results show that training the YOLO detector with GAN-modified data compared to training with real data leads to the same detection performance. The results also show that the shapes and backgrounds of the antennas contributed more to detection performance than their style and colour.

The performance of deep neural networks and estimations of their robustness has been rapidly developed. In contrast, despite the broad usage of deep convolutional neural networks (CNNs)[1] for medical image segmentation, research on their uncertainty estimations is being far less conducted. Deep learning tools in their nature do not capture the model uncertainty and in this sense, the output of deep neural networks needs to be critically analysed with quantitative measurements, especially for applications in the medical domain. In this work, epistemic uncertainty, which is one of the main types of uncertainties (epistemic and aleatoric) is analyzed and measured for volumetric medical image segmentation tasks (and possibly more diverse methods for 2D images) at pixel level and structure level. The deep neural network employed as a baseline is 3D U-Net architecture[2], which shares the essential structural concept with U-Net architecture[3], and various techniques are applied to quantify the uncertainty and obtain statistically meaningful results, including test-time data augmentation and deep ensembles. The distribution of the pixel-wise predictions is estimated by Monte Carlo simulations and the entropy is computed to quantify and visualize how uncertain (or certain) the predictions of each pixel are. During the estimation, given the increased network training time in volumetric image segmentation, training an ensemble of networks is extremely time-consuming and thus the focus is on data augmentation and test-time dropouts. The desired outcome is to reduce the computational costs of measuring the uncertainty of the model predictions while maintaining the same level of estimation performance and to increase the reliability of the uncertainty estimation map compared to the conventional methods. The proposed techniques are evaluated on publicly available volumetric image datasets, Combined Healthy Abdominal Organ Segmentation (CHAOS, a set of 3D in-vivo images) from Grand Challenge (https://chaos.grand-challenge.org/). Experiments with the liver segmentation task in 3D Computed Tomography (CT) show the relationship between the prediction accuracy and the uncertainty map obtained by the proposed techniques. / Prestandan hos djupa neurala nätverk och estimeringar av deras robusthet har utvecklats snabbt. Däremot, trots den breda användningen av djupa konvolutionella neurala nätverk (CNN) för medicinsk bildsegmentering, utförs mindre forskning om deras osäkerhetsuppskattningar. Verktyg för djupinlärning fångar inte modellosäkerheten och därför måste utdata från djupa neurala nätverk analyseras kritiskt med kvantitativa mätningar, särskilt för tillämpningar inom den medicinska domänen. I detta arbete analyseras och mäts epistemisk osäkerhet, som är en av huvudtyperna av osäkerheter (epistemisk och aleatorisk) för volymetriska medicinska bildsegmenteringsuppgifter (och möjligen fler olika metoder för 2D-bilder) på pixelnivå och strukturnivå. Det djupa neurala nätverket som används som referens är en 3D U-Net-arkitektur [2] och olika tekniker används för att kvantifiera osäkerheten och erhålla statistiskt meningsfulla resultat, inklusive testtidsdata-augmentering och djupa ensembler. Fördelningen av de pixelvisa förutsägelserna uppskattas av Monte Carlo-simuleringar och entropin beräknas för att kvantifiera och visualisera hur osäkra (eller säkra) förutsägelserna för varje pixel är. Under uppskattningen, med tanke på den ökade nätverksträningstiden i volymetrisk bildsegmentering, är träning av en ensemble av nätverk extremt tidskrävande och därför ligger fokus på dataaugmentering och test-time dropouts. Det önskade resultatet är att minska beräkningskostnaderna för att mäta osäkerheten i modellförutsägelserna samtidigt som man bibehåller samma nivå av estimeringsprestanda och ökar tillförlitligheten för kartan för osäkerhetsuppskattning jämfört med de konventionella metoderna. De föreslagna teknikerna kommer att utvärderas på allmänt tillgängliga volymetriska bilduppsättningar, Combined Healthy Abdominal Organ Segmentation (CHAOS, en uppsättning 3D in-vivo-bilder) från Grand Challenge (https://chaos.grand-challenge.org/). Experiment med segmenteringsuppgiften för lever i 3D Computed Tomography (CT) vissambandet mellan prediktionsnoggrannheten och osäkerhetskartan som erhålls med de föreslagna teknikerna.