Research on modeling trees and plants has attracted a great deal of attention in recent years. Early procedural tree modeling can be divided into four main categories: rule-based algorithms, repetitive patterns, cellular automata, and particle systems. These methods offer a very high level of realism; however, creating millions of varied tree datasets manually is not logistically possible, even for professional 3D modeling artists. Trees created using these previous methods are typically static and the controllability of these procedural tree models is low. Deep generative models are capable of generating any type of shape automatically, making it possible to create 3D models at large scale. In this paper, we introduce a novel deep generative model that generates 3D (botanical) tree models, which are not only edible, but also have diverse shapes. Our proposed network, denoted BranchNet, trains the tree branch structures on a hierarchical Variational Autoencoder (VAE) that learns new generative model structures. By directly encoding shapes into a hierarchy graph, BranchNet can generate diverse, novel, and realistic tree structures. To assist the creation of tree models, we create a domain-specific language with a GUI for modeling 3D shape structures, in which the continuous parameters can be manually edited in order to produce new tree shapes. The trees are interpretable and the GUI can be edited to capture the subset of shape variability.

Nowadays, machine learning methods are used in many applications and deployed in manyelectronic devices to solve problems and predict future states. One of the challenges mostbig cities confront is traffic jams since the roads are crammed with more and more vehicles, which will easily cause traffic congestion. Traffic jams are not environment-friendly,but scientific planning can minimize their effect. Traffic prediction is one of the most interesting subjects for Intelligent transportation systems due to its ability to prevent trafficjams with the knowledge of the predictions. Traffic prediction is a very challenging taskfor researchers to find or implement a model to perform accurately in different scenarios.Accurate traffic forecasting has become an essential mission for intelligent transportationsystems, which improve transportation efficiency, safety, and sustainability using moderntechnology and data analysis. Capturing both temporal and spatial dependencies is one ofthe most essential key in traffic prediction. Combining two or several models is one way tocapture both dependencies. A temporal graph convolutional network (T-GCN) is a graphneural network model, a combination of a graph convolutional network and a gated recurrent unit (GRU). In T-GCN, a graph convolutional network (GCN) is used to capture spatialwhile recurrent gated units to capture temporal dependencies. One of the main issues ofT-GCN is long-term prediction failure, where the model’s accuracy decreases when the prediction length increases. In this paper, we propose a Decomposed Temporal Self-AttentionMulti-layer Graph Convolutional network (DTSA-3GCN) to enhance overall traffic prediction in different horizons based on Singular Value Decomposition (SVD), Self-Attention(SA), and a Temporal Multi-layer Graph Convolutional Network. The experiment resultdemonstrates that DTSA-3GCN outperforms the state-of-the-art models such as T-GCN,A3T-GCN, and STGODE.

This diploma thesis describes and implements the design of a graph neural network usedfor 2D segmentation of neural structure. The first chapter of the thesis briefly introduces the problem of segmentation. In this chapter, segmentation techniques are divided according to the principles of the methods they use. Each type of technique contains the essence of this category as well as a description of one representative. The second chapter of the diploma thesis explains graph neural networks (GNN for short). Here, the thesis divides graph neural networks in general and describes recurrent graph neural networks(RGNN for short) and graph autoencoders, that can be used for image segmentation, in more detail. The specific image segmentation solution is based on the message passing method in RGNN, which can replace convolution masks in convolutional neural networks.RGNN also provides a simpler multilayer perceptron topology. The second type of graph neural networks characterised in the thesis are graph autoencoders, which use various methods for better encoding of graph vertices into Euclidean space. The last part ofthe diploma thesis deals with the analysis of the problem, the proposal of its specific solution and the evaluation of results. The purpose of the practical part of the work was the implementation of GNN for image data segmentation. The advantage of using neural networks is the ability to solve different types of segmentation by changing training data. RGNN with messaging passing and node2vec were used as implementation GNNf or segmentation problem. RGNN training was performed on graphics cards provided bythe school and Google Colaboratory. Learning RGNN using node2vec was very memory intensive and therefore it was necessary to train on a processor with an operating memory larger than 12GB. As part of the RGNN optimization, learning was tested using various loss functions, changing topology and learning parameters. A tree structure method was developed to use node2vec to improve segmentation, but the results did not confirman improvement for a small number of iterations. The best outcomes of the practical implementation were evaluated by comparing the tested data with the convolutional neural network U-Net. It is possible to state comparable results to the U-Net network, but further testing is needed to compare these neural networks. The result of the thesisis the use of RGNN as a modern solution to the problem of image segmentation and providing a foundation for further research.

To combat fake news, researchers mostly focused on detecting fake news and journalists built and maintained fact-checking sites (e.g., Snopes.com and Politifact.com). However, fake news dissemination has been greatly promoted by social media sites, and these fact-checking sites have not been fully utilized. To overcome these problems and complement existing methods against fake news, in this thesis, we propose a deep-learning based fact-checking URL recommender system to mitigate impact of fake news in social media sites such as Twitter and Facebook. In particular, our proposed framework consists of a multi-relational attentive module and a heterogeneous graph attention network to learn complex/semantic relationship between user-URL pairs, user-user pairs, and URL-URL pairs. Extensive experiments on a real-world dataset show that our proposed framework outperforms seven state-of-the-art recommendation models, achieving at least 3~5.3% improvement.

To combat fake news, researchers mostly focused on detecting fake news and journalists built and maintained fact-checking sites (e.g., Snopes.com and Politifact.com). However, fake news dissemination has been greatly promoted by social media sites, and these fact-checking sites have not been fully utilized. To overcome these problems and complement existing methods against fake news, in this thesis, we propose a deep-learning based fact-checking URL recommender system to mitigate impact of fake news in social media sites such as Twitter and Facebook. In particular, our proposed framework consists of a multi-relational attentive module and a heterogeneous graph attention network to learn complex/semantic relationship between user-URL pairs, user-user pairs, and URL-URL pairs. Extensive experiments on a real-world dataset show that our proposed framework outperforms seven state-of-the-art recommendation models, achieving at least 3~5.3% improvement.

<p>Roller bearings are the common components used in the mechanical systems for mechanical processing and production. The running state of roller bearings often determines the machining accuracy and productivity on a manufacturing line. Roller bearing failure may lead to the shutdown of production lines, resulting in serious economic losses. Therefore, the research on roller bearing fault diagnosis has a great value. This thesis research first proposes a method of signal frequency spectral resampling to tackle the problem of bearing fault detection at different rotating speeds using a single speed dataset for training the network such as the one dimensional convolutional neural network (1D CNN). Second, this research work proposes a technique to connect the graph structures constructed from spectral components of the different bearing fault frequency bands into a sparse graph structure, so that the fault identification can be carried out effectively through a graph neural network in terms of the computation load and classification rate. Finally, the frequency spectral resampling method for feature extraction is validated using our self-collected datasets. The performance of the graph neural network with our proposed sparse graph structure is validated using the Case Western Reserve University (CWRU) dataset as well as our self-collected datasets. The results show that our proposed method achieves higher bearing fault classification accuracy than those recently proposed by other researchers using machine learning approaches and neural networks.</p>

Road accidents pose a significant threat, leading to fatalities and injuries with far-reaching consequences. This study addresses two crucial challenges in road safety: analyzing traffic intersections to enhance safety by predicting potentially risky situations, and monitoring driver activity to prevent distracted driving accidents. Focusing on Virginia's intersections, we thoroughly examine traffic participant interactions to identify and mitigate conflicts, employing graph-based modeling of traffic scenarios to evaluate contributing parameters. Additionally, we leverage graph neural networks to detect and track potential crash situations from intersection videos, offering practical recommendations to enhance intersection safety. To understand the causes of risky behavior, we specifically investigate accidents resulting from distracted driving, which has become more prevalent due to advanced driver assistance systems in semi-autonomous vehicles. For monitoring driver activity inside vehicles, we propose the use of Video Transformers on challenging secondary driver activity datasets, incorporating grayscale and low-quality data to overcome limitations in capturing overall image context. Finally, we validate our predictions by studying attention modules and introducing explainability into the computer vision model. This research contributes to improving road safety by providing comprehensive analysis and recommendations for intersection safety enhancement and prevention of distracted driving accidents. / Master of Science / Road accidents are a serious problem causing numerous deaths and injuries each year. By studying driver behavior, we can uncover common causes of accidents like distracted driving, impaired driving, speeding, and not following traffic rules. New vehicle technologies aim to assist drivers, raising concerns about driver attentiveness. It is crucial for car manufacturers to develop systems that can detect and prevent accidents, especially in semi-autonomous vehicles. This study focuses on intersections in Virginia and examines driver behavior within vehicles to identify and prevent dangerous situations. We create models of different traffic scenarios using graphs/networks and utilize machine learning to identify potential accidents. Our objective is to provide practical recommendations for improving intersection safety. Existing datasets and algorithms for recognizing driver activities often fail to capture common distractions like eating, drinking, and phone use. To address this, we introduce two challenging datasets specifically designed to capture distracted driving activities. Finally, we try to understand the predictions bade by the chosen deep learning model by visualizing the inner workings.

Tracab’s optical tracking system allows to track the 2-dimensional trajectories of players and ball during a football game. Using this data it is possible to train machine learning models to identify events that happen during the match. In this thesis, we explore the detection of corners, free kicks, and throw-in events by means of neural networks. Training a model to solve this task is not easy; the neural network needs to model the spatio-temporal interactions between different agents moving in a 2-dimensional space. We decided to address this problem using graph neural networks in combination with recurrent neural networks, which allow us to model respectively the spatial and temporal components of the data. Tracking the position of the ball is difficult, which makes the dataset noisy. In this thesis, we mainly work with a version of the dataset where the position of the ball has been manually corrected. However, to study how the noisy position of the ball affects the results we also train the models on the original data. The results show that detecting the corner and the throw-in is much easier than detecting the free kick. Moreover, the noisy position of the ball affects significantly the performance of the model. We conclude that to train the model on the original data it is necessary to use a much larger training set. Since the amount of training data for these events is limited, we also train the model on the more generic ball-dead-to-alive event, for which much more data is available, and we observe that by increasing the amount of training data the results can improve significantly. In this report, we also provide an in-depth discussion about all the challenges faced during the project and how different hyperparameters and design choices can affect the results. / Tracabs optiska spårningssystem gör det möjligt att spåra de 2-dimensionella banorna för spelare och boll under en fotbollsmatch. Med hjälp av dessa data är det möjligt att träna maskininlärningsmodeller för att identifiera händelser som inträffar under matchen. I denna avhandling utforskar vi upptäckten av hörnor, frisparkar och inkastningshändelser med hjälp av neurala nätverk. Att träna en modell för att lösa denna uppgift är inte lätt; det neurala nätverket behöver modellera de rums-temporala interaktionerna mellan olika agenter som rör sig i ett 2-dimensionellt rum. Vi bestämde oss för att ta itu med detta problem med hjälp av grafiska neurala nätverk i kombination med återkommande neurala nätverk, vilket gör att vi kan modellera de rumsliga respektive temporala komponenterna i datan. Det är svårt att spåra bollens position, vilket gör datauppsättningen bullrig. I detta examensarbete arbetar vi främst med en version av datamängden där bollens position har korrigerats manuellt. Men för att studera hur bollens bullriga position påverkar resultaten tränar vi också modellerna på originaldata. Resultaten visar att det är mycket lättare att upptäcka hörna och inkastet än att upptäcka frisparken. Dessutom påverkar bollens bullriga position avsevärt modellens prestanda. Vi drar slutsatsen att för att träna modellen på originaldata är det nödvändigt att använda en mycket större träningsuppsättning. Eftersom mängden träningsdata för dessa evenemang är begränsad, tränar vi också modellen på den mer generiska bollen död-till-levande-händelsen, för vilken mycket mer data finns tillgänglig, och vi observerar att genom att öka mängden träningsdata resultaten kan förbättras avsevärt. I denna rapport ger vi också en fördjupad diskussion om alla utmaningar som ställs inför under projektet och hur olika hyperparametrar och designval kan påverka resultaten.

Network Calculus views a network system as a queuing framework and provides a series of mathematical functions for finding an upper bound of an end-to-end delay. It is crucial for the design of networks and applications with a hard delay guarantee, such as the emerging Time Sensitive Network. Even though several approaches in Network Calculus can be used directly to find bounds on the worst-case delay, these bounds are usually not tight, and making them tight is a hard problem due to the extremely intensive computing requirements. This problem has also been proven as NP-Hard. One newly introduced solution to tighten the delay bound is the so-called Flow Prolongation. It extends the paths of cross flows to new sink servers, which naturally increases the worst-case delay, but might at the same time decrease the delay bound. The most straightforward and the most rigorous solution to find the optimal Flow Prolongation combinations is by doing exhaustive searches. However, this approach is not scalable with the network size. Thus, a machine learning model, Graph Neural Network (GNN), has been introduced for the prediction of the optimal Flow Prolongation combinations, mitigating the scalability issue. However, early research also found out that machine learning models consistently misclassify adversarial examples. In this thesis, Fast Gradient Sign Method (FGSM) is used to benchmark how adversarial attacks will influence the delay bound achieved by the Flow Prolongation method. It is performed by slightly modifying the input network features based on their gradients. To achieve this, we first learned the usage of NetCal DNC, an Free and Open Source Software, to calculate the Pay Multiplexing Only Once (PMOO), one of the Network Calculus methods for the delay bound calculation. Then we reproduced the GNN model based on PMOO, and achieved an accuracy of 65%. Finally, the FGSM is implemented on a newly created dataset with a large number of servers and flows inside. Our results demonstrate that with at most 14% changes on the network features input, the accuracy of GNN drastically decreases to an average 9.45%, and some prominent examples are found whose delay bounds are largely loosened by the GNN Flow Prolongation prediction after the FGSM attack. / Nätverkskalkylen behandlar ett nätverkssystem som ett system av köer och tillhandahåller ett antal matematiska funktioner som används för att hitta en övre gräns för end-to-end förseningar. Det är mycket viktigt för designen av nätverk och applikationer med strikta begränsningar för förseningar, så som det framväxande Time Sensitive Network. Även om ett flertal tillvägagångssätt i nätverkskalkylen kan användas direkt för att finna gränsen för förseningar i det värsta fallet så är dessa vanligtvis inte snäva. Att göra gränserna snäva är svårt då det är ett NP-svårt problem som kräver extremt mycket beräkningar. En lösning för att strama åt förseningsgränserna som nyligen introducerats kallas Flow Prolongation. Den utökar vägarna av korsflöden till nya sink servrar, vilket naturligt ökar förseningen i värsta fallet, men kan eventuellt också sänka förseningsgränsen. Den enklaste och mest rigorösa lösningen för att hitta de optimala Flow Prolongation kombinationerna är att göra uttömmande sökningar. Detta tillvägagångssätt är dock inte skalbart för stora nätverk. Därför har en maskininlärningsmodell, ett Graph Neural Network (GNN), introducerats för att förutspå de optimala Flow Prolongation kombinationerna och samtidigt mildra problemen med skalbarhet. Dock så visar de tidiga fynden att maskininlärningsmodeller ofta felaktigt klassificerar motstridiga exempel. I detta projekt används Fast Gradient Sign Method (FGSM) för att undersöka hur motståndarattacker kan påverka förseningsgränsen som hittas med hjälp av Flow Prolongation metoden. Detta görs genom att modifiera indata-nätverksfunktionerna en aning baserat på dess gradienter. För att uppnå detta lärde vi oss först att använda NetCal DNC, en mjukvara som är gratis och Open Source, för att kunna beräkna Pay Multiplexinng Only Once (PMOO), en metod inom nätverkskalkylen för att beräkna förseningsgränser. Sedan reproducerade GNN modellen baserat på PMOO, och uppnådde en träffsäkerhet på 65%. Slutligen implementerades FGSM på ett nytt dataset med ett stort antal servrar och flöden. Våra resultat visar att förändringar på upp till 14% på indata-nätverksfunktionerna resulterar i att träffsäkerheten hos GNN minskar drastiskt till ett genomsnitt på 9.45%. Vissa exempel identifierades där förseningsgränsen utvidgas kraftfullt i GNN Flow Prolongation förutsägelsen efter FGSM attacken.

京都大学 / 0048 / 新制・課程博士 / 博士(薬科学) / 甲第22754号 / 薬科博第128号 / 新制||薬科||14(附属図書館) / 京都大学大学院薬学研究科医薬創成情報科学専攻 / (主査)教授 馬見塚 拓, 教授 緒方 博之, 教授 石濱 泰 / 学位規則第4条第1項該当 / Doctor of Pharmaceutical Sciences / Kyoto University / DFAM