<p dir="ltr">In the realm of autonomous vehicle technology, the Multimodal Vehicle Detection Network (MVDNet) represents a significant leap forward, particularly in the challenging context of weather conditions. This paper focuses on the enhancement of MVDNet through the integration of a multi-head attention layer, aimed at refining its performance. The integrated multi-head attention layer in the MVDNet model is a pivotal modification, advancing the network's ability to process and fuse multimodal sensor information more efficiently. The paper validates the improved performance of MVDNet with multi-head attention through comprehensive testing, which includes a training dataset derived from the Oxford Radar Robotcar. The results clearly demonstrate that the Multi-Head MVDNet outperforms the other related conventional models, particularly in the Average Precision (AP) estimation, under challenging environmental conditions. The proposed Multi-Head MVDNet not only contributes significantly to the field of autonomous vehicle detection but also underscores the potential of sophisticated sensor fusion techniques in overcoming environmental limitations.</p>

Work aimed at analyzing methods and algorithms for protection against spam and email attacks using machine learning and neural networks. The main focus is on the development of a hybrid model combining convolutional neural networks (CNN) and long-term short-term memory (LSTM) for classifying spam messages. The study includes the stages of data preprocessing, model development and training, as well as evaluating its performance on test data. The results show that the proposed model demonstrates high accuracy and efficiency in detecting spam messages. / Проведен анализ методов и алгоритмов защиты от спама и атак по электронной почте с использованием машинного обучения и нейронных сетей. Основное внимание уделено разработке гибридной модели, сочетающей сверточные нейронные сети (CNN) и долгосрочную краткосрочную память (LSTM) для классификации спам-сообщений. Исследование включает этапы предварительной обработки данных, разработки и обучения модели, а также оценку ее производительности на тестовых данных. Результаты показывают, что предложенная модель демонстрирует высокую точность и эффективность в обнаружении спам-сообщений.

In dieser Doktorarbeit wird die Frage untersucht, wie erfolgreich deep learning bei der Diagnostik von neurodegenerativen Erkrankungen unterstützen kann. In 5 experimentellen Studien wird die Anwendung von Convolutional Neural Networks (CNNs) auf Daten der Magnetresonanztomographie (MRT) untersucht. Ein Schwerpunkt wird dabei auf die Erklärbarkeit der eigentlich intransparenten Modelle gelegt. Mit Hilfe von Methoden der erklärbaren künstlichen Intelligenz (KI) werden Heatmaps erstellt, die die Relevanz einzelner Bildbereiche für das Modell darstellen. Die 5 Studien dieser Dissertation zeigen das Potenzial von CNNs zur Krankheitserkennung auf neurologischen MRT, insbesondere bei der Kombination mit Methoden der erklärbaren KI. Mehrere Herausforderungen wurden in den Studien aufgezeigt und Lösungsansätze in den Experimenten evaluiert. Über alle Studien hinweg haben CNNs gute Klassifikationsgenauigkeiten erzielt und konnten durch den Vergleich von Heatmaps zur klinischen Literatur validiert werden. Weiterhin wurde eine neue CNN Architektur entwickelt, spezialisiert auf die räumlichen Eigenschaften von Gehirn MRT Bildern. / Deep learning and especially convolutional neural networks (CNNs) have a high potential of being implemented into clinical decision support software for tasks such as diagnosis and prediction of disease courses. This thesis has studied the application of CNNs on structural MRI data for diagnosing neurological diseases. Specifically, multiple sclerosis and Alzheimer’s disease were used as classification targets due to their high prevalence, data availability and apparent biomarkers in structural MRI data. The classification task is challenging since pathology can be highly individual and difficult for human experts to detect and due to small sample sizes, which are caused by the high acquisition cost and sensitivity of medical imaging data. A roadblock in adopting CNNs to clinical practice is their lack of interpretability. Therefore, after optimizing the machine learning models for predictive performance (e.g. balanced accuracy), we have employed explainability methods to study the reliability and validity of the trained models. The deep learning models achieved good predictive performance of over 87% balanced accuracy on all tasks and the explainability heatmaps showed coherence with known clinical biomarkers for both disorders. Explainability methods were compared quantitatively using brain atlases and shortcomings regarding their robustness were revealed. Further investigations showed clear benefits of transfer-learning and image registration on the model performance. Lastly, a new CNN layer type was introduced, which incorporates a prior on the spatial homogeneity of neuro-MRI data. CNNs excel when used on natural images which possess spatial heterogeneity, and even though MRI data and natural images share computational similarities, the composition and orientation of neuro-MRI is very distinct. The introduced patch-individual filter (PIF) layer breaks the assumption of spatial invariance of CNNs and reduces convergence time on different data sets without reducing predictive performance. The presented work highlights many challenges that CNNs for disease diagnosis face on MRI data and defines as well as tests strategies to overcome those.

<p>Piping is an essential component in buildings, and its as-built information is critical to facility management tasks. Manually extracting piping information from legacy drawings that are in paper, PDF, or image format is mentally exerting, time-consuming, and error-prone. Symbol recognition and detection are core problems in the computer-based interpretation of piping drawings, and the main technical challenge is to determine robust features that are invariant to scaling, rotation, and translation. This thesis aims to use convolutional neural networks (CNNs) to automatically extract features from raw images, and consequently, to locate and recognize symbols in piping drawings.</p> <p>In this thesis, the Spatial Transformer Network (STN) is applied to improve the performance of a standard CNN model for recognizing piping symbols, and the Faster Region-based Convolutional Neural Network (Faster RCNN) is adopted to exploit its capacity in symbol detection. For experimentation, the synthetic data are generated as follows. Two datasets are generated for symbol recognition and detection, respectively. For recognition, eight types of symbols are synthesized based on the geometric constraints between the primitives. The drawing samples for detection are manually sketched using AutoCAD MEP software and its piping component library, and seven types of symbols are selected from the piping component library. Both sets of samples are augmented with various scales, rotations, and random noises.</p> <p>The experiment for symbol recognition is conducted and the accuracies of the recognition accuracy of the CNN + STN model and the standard CNN model are compared. It is observed that the spatial transformer layer improves the accuracy in classifying piping symbols from 95.39% to 98.26%. For the symbol detection task, the experiment is conducted using a public implementation of Faster RCNN. The mean Average Precision (mAP) is 82.8% when Intersection over Union (IoU) threshold equals to 0.5. Imbalanced data (i.e., imbalanced samples in each class) led to a decrease in the Average Precision in the minority class. Also, the symbol library, the small dataset, and the complex backbone network limit the generality of the model. Future work will focus on the collection of larger set of drawings and the improvement of the network’s geometric invariance.</p>

In large scale productions of metal sheets, it is important to maintain an effective way to continuously inspect the products passing through the production line. The inspection mainly consists of detection of defects and tracking of ID numbers. This thesis investigates the possibilities to create an automatic inspection system by evaluating different machine learning algorithms for defect detection and optical character recognition (OCR) on metal sheet data. Digit recognition and defect detection are solved separately, where the former compares the object detection algorithm Faster R-CNN and the classical machine learning algorithm NCGF, and the latter is based on unsupervised learning using a convolutional autoencoder (CAE). The advantage of the feature extraction method is that it only needs a couple of samples to be able to classify new digits, which is desirable in this case due to the lack of training data. Faster R-CNN, on the other hand, needs much more training data to solve the same problem. NCGF does however fail to classify noisy images and images of metal sheets containing an alloy, while Faster R-CNN seems to be a more promising solution with a final mean average precision of 98.59%. The CAE approach for defect detection showed promising result. The algorithm learned how to only reconstruct images without defects, resulting in reconstruction errors whenever a defect appears. The errors are initially classified using a basic thresholding approach, resulting in a 98.9% accuracy. However, this classifier requires supervised learning, which is why the clustering algorithm Gaussian mixture model (GMM) is investigated as well. The result shows that it should be possible to use GMM, but that it requires a lot of GPU resources to use it in an end-to-end solution with a CAE.

In this thesis, we work towards bridging the gap between two distinct areas: noisy text handling and text summarisation. The overall goal of the paper is to examine the effects of noise in the training of convolutional neural networks for text summarisation, with a view to understanding how to effectively create a noise-robust text-summarisation system. We look specifically at the problem of abstractive text summarisation of noisy data in the context of summarising error-containing documents from automatic speech recognition (ASR) output. We experiment with adding varying levels of noise (errors) to the 4 million-article Gigaword corpus and training an encoder-decoder CNN on it with the aim of producing a noise-robust text summarisation system. A total of six text summarisation models are trained, each with a different level of noise. We discover that the models with a high level of noise are indeed able to aptly summarise noisy data into clean summaries, despite a tendency for all models to overfit to the level of noise on which they were trained. Directions are given for future steps in order to create an even more noise-robust and flexible text summarisation system.

Weld joint penetration determination is the key factor in welding process control area. Not only has it directly affected the weld joint mechanical properties, like fatigue for example. It also requires much of human intelligence, which either complex modeling or rich of welding experience. Therefore, weld penetration status identification has become the obstacle for intelligent welding system. In this dissertation, an innovative method has been proposed to detect the weld joint penetration status using machine-learning algorithms. A GTAW welding system is firstly built. Project a dot-structured laser pattern onto the weld pool surface during welding process, the reflected laser pattern is captured which contains all the information about the penetration status. An experienced welder is able to determine weld penetration status just based on the reflected laser pattern. However, it is difficult to characterize the images to extract key information that used to determine penetration status. To overcome the challenges in finding right features and accurately processing images to extract key features using conventional machine vision algorithms, we propose using convolutional neural network (CNN) to automatically extract key features and determine penetration status. Data-label pairs are needed to train a CNN. Therefore, an image acquiring system is designed to collect reflected laser pattern and the image of work-piece backside. Data augmentation is performed to enlarge the training data size, which resulting in 270,000 training data, 45,000 validation data and 45,000 test data. A six-layer convolutional neural network (CNN) has been designed and trained using a revised mini-batch gradient descent optimizer. Final test accuracy is 90.7% and using a voting mechanism based on three consequent images further improve the prediction accuracy.

How do people discuss mental health on social media? Can we train a computer program to recognize differences between discussions of depression and other topics? Can an algorithm predict that someone is depressed from their tweets alone? In this project, we collect tweets referencing “depression” and “depressed” over a seven year period, and train word embeddings to characterize linguistic structures within the corpus. We find that neural word embeddings capture the contextual differences between “depressed” and “healthy” language. We also looked at how context around words may have changed over time to get deeper understanding of contextual shifts in the word usage. Finally, we trained a deep learning network on a much smaller collection of tweets authored by individuals formally diagnosed with depression. The best performing model for the prediction task is Convolutional LSTM (CNN-LSTM) model with a F-score of 69% on test data. The results suggest social media could serve as a valuable screening tool for mental health.

The thesis presents named-entity recognition in Czech historical newspapers from Modern Access to Historical Sources Project. Our goal was to create a specific corpus and annotation manual for the project and evaluate neural networks methods for named-entity recognition within the task. We created the corpus using scanned Czech historical newspapers. The scanned pages were converted to digitize text by optical character recognition (OCR) method. The data were preprocessed by deleting some OCR errors. We also defined specific named entities types for our task and created an annotation manual with examples for the project. Based on that, we annotated the final corpus. To find the most suitable neural networks model for our task, we experimented with different neural networks architectures, namely long short-term memory (LSTM), bidirectional LSTM and CNN-BiLSTM models. Moreover, we experimented with randomly initialized word embeddings that were trained during the training process and pretrained word embeddings for contemporary Czech published as open source by fastText. We achieved the best result F1 score 0.444 using CNN-BiLSTM model and the pretrained word embeddings by fastText. We found out that we do not need to normalize spelling of our historical texts to get closer to contemporary language if we use the neural network model. We provided a qualitative analysis of observed linguistics phenomena as well. We found out that some word forms and pair of words which were not frequent in our training data set were miss-tagged or not tagged at all. Based on that, we can say that larger data sets could improve the results.

Jon Stewart, the anchor and purveyor of “fake news,” has catapulted television's The Daily Show into prominence. The show functions as both a source of political humor and a vehicle for political commentary. This thesis explores how the program visually and rhetorically problematizes the hegemonic model of traditional television news, and how it tips the balance between what is considered serious news and what has become cliché about the broadcast industry.