Search

In the era of data-driven decision-making and innovation, synthetic data serves as a promising tool that bridges the need for vast datasets in machine learning (ML) and the imperative necessity of data privacy. By simulating real-world data while preserving privacy, synthetic data generators have become more prevalent instruments in AI and ML development. A key challenge with synthetic data lies in accurately estimating its utility. For such purpose, Population Fidelity (PF) measures have shown to be good candidates, a category of metrics that evaluates how well the synthetic data mimics the general distribution of the original data. With this setting, we aim to answer: "How well are different population fidelity measures able to indicate the utility of synthetic data for machine learning based classification models?" We designed a reusable six-step experiment framework to examine the correlation between nine PF measures and the performance of four ML for training classification models over five datasets. The six-step approach includes data preparation, training, testing on original and synthetic datasets, and PF measures computation. The study reveals non-linear relationships between the PF measures and synthetic data utility. The general analysis, meaning the monotonic relationship between the PF measure and performance over all models, yielded at most moderate correlations, where the Cluster measure showed the strongest correlation. In the more granular model-specific analysis, Random Forest showed strong correlations with three PF measures. The findings show that no PF measure shows a consistently high correlation over all models to be considered a universal estimator for model performance.This highlights the importance of context-aware application of PF measures and sets the stage for future research to expand the scope, including support for a wider range of types of data and integrating privacy evaluations in synthetic data assessment. Ultimately, this study, I eran av datadriven beslutsfattning och innovation, fungerar syntetiska data som ett lovande verktyg som bryggar behovet av omfattande dataset inom maskininlärning (ML) och nödvändigheten för dataintegritet. Genom att simulera verklig data samtidigt som man bevarar integriteten, har generatorer av syntetiska data blivit allt vanligare verktyg inom AI och ML-utveckling. En viktig utmaning med syntetiska data är att noggrant uppskatta dess användbarhet. För detta ändamål har mått under kategorin Populations Fidelity (PF) visat sig vara goda kandidater, det är mätvärden som utvärderar hur väl syntetiska datan efterliknar den generella distributionen av den ursprungliga datan. Med detta i åtanke strävar vi att svara på följande: Hur väl kan olika population fidelity mätvärden indikera användbarheten av syntetisk data för maskininlärnings baserade klassifikationsmodeller? För att besvara frågan har vi designat ett återanvändbart sex-stegs experiment ramverk, för att undersöka korrelationen mellan nio PF-mått och prestandan hos fyra ML klassificeringsmodeller, på fem dataset. Sex-stegs strategin inkluderar datatillredning, träning, testning på både ursprungliga och syntetiska dataset samt beräkning av PF-mått. Studien avslöjar förekommandet av icke-linjära relationer mellan PF-måtten och användbarheten av syntetiska data. Den generella analysen, det vill säga den monotona relationen mellan PF-måttet och prestanda över alla modeller, visade som mest medelmåttiga korrelationer, där Cluster-måttet visade den starkaste korrelationen. I den mer detaljerade, modell-specifika analysen visade Random Forest starka korrelationer med tre PF-mått. Resultaten visar att inget PF-mått visar konsekvent hög korrelation över alla modeller för att betraktas som en universell indikator för modellprestanda. Detta understryker vikten av kontextmedveten tillämpning av PF-mått och banar väg för framtida forskning för att utöka omfånget, inklusive stöd för ett bredare utbud för data av olika typ

The following research provides a detailed analysis of the factors that influence the purchase intention of vegetable beverages. The study focuses on the Ecuadorian Generation Z and analyzes the variables that influence their satisfaction and, therefore, their purchase decision. The study uses a model based on synthetic data and structural equations to analyze the relationships between variables. Data were collected from 630 synthetic observations generated from a base of 50 Generation Z respondents and the SEM PLS technique was used to interpret the responses. The findings indicated that intrinsic, introjected and social pressure motivations have no significant effect while natural content, sensory appeal and identified motivation variables influence purchase intention and thus positively affect Satisfaction. These results can guide that companies producing vegetable beverages should focus on highlighting the natural content and sensory appeal of their products in their marketing strategies, La siguiente investigación proporciona un análisis detallado de los factores que influyen en la intención de compra de bebidas vegetales. El estudio se centra en la generación Z ecuatoriana y analiza las variables que influyen en su satisfacción y, por ende, en su decisión de compra. El estudio utiliza un modelo basado en datos sintéticos y ecuaciones estructurales para analizar las relaciones entre las variables. Se recolectaron datos de 630 observaciones sintéticas generadas en una base de 50 respondientes de la generación Z y la técnica SEM PLS se utilizó para interpretar las respuestas. Los hallazgos indicaron que las motivaciones intrínsecas, introyectadas y presión social no tienen un efecto significativo mientras que las variables de contenido natural, atractivo sensorial y motivación identificada influyen en la intención de compra y por ende inciden de manera positiva en la Satisfacción. Estos resultados pueden guiar a que las empresas que producen bebidas vegetales deben enfocarse en resaltar el contenido natural y el atractivo sensorial de sus productos en sus estrategias de marketing.

This thesis investigates the application of Denoising Diffusion Probabilistic Models (DDPM) for synthesizing 12-lead Electrocardiogram (ECG) signals. Utilizing classifier-free guidance along with the bidirectional Mamba State Space Model (SSM) within a DiffWave framework, we developed a model capable of both unconditional and conditional ECG signal generation.Despite the promising potential of Mamba for enhancing temporal signal encoding, its performance compared to the time-invariant SSM model, S4, was either worse or inconclusive, likely due to the limitations of current metrics. Visual assessments often contradicted the automated metrics, indicating a significant gap in current evaluation methods. We also explored the feasibility of training models on all 12 leads, contrasting previous studies that used fewer leads. Our findings indicate that diffusion models can adequately learn the linear relationships between leads without significantly increasing model size. Overall, our work reduces the need for extensive data pre- or post-processing, streamlines ECG data generation, and highlights the limitations of existing evaluation methodologies, suggesting the need for further evaluation.

Generative diffusion models have recently become extremely popular in a variety of domains,but especially in image generation. These models are capable of generating a wide variety of high-quality images and can be guided by text prompts, depth maps, and more. Despite these impressive capabilities, these models typically generate images with poor projective geometry. As a result, generated images differ significantly from real images, decreasing the photo-realism of generated images. In addition, since perspective is crucial for representing 3D information in 2D images, discrepancies in projective geometry limit the use of generative models as synthetic data generators. In this work, we introduce a geometric constraint to improve the projective geometry of diffusion models and show that outputs of models trained with this constraint both appear more photo-realistic and serve as useful synthetic data by improving the performance of downstream models fine-tuned on generated images.

Differential privacy has seen a lot of growth in the last decade and has been accepted as a rigorous definition of privacy. A key use case of differential privacy is to generate synthetic data that can be released to the public without concerns about revealing sensitive information. Much of the research in differential privacy has focused on offline applications with the assumption that all data is available at once. When these algorithms are applied in practice to streams where data is collected over time, this either violates the privacy guarantees or results in poor utility.In this dissertation, we propose three methods that target the task of generating differentially private synthetic data for three different use cases. In Chapters 3 and 4 we look at streams in spatial and tabular spaces but with the assumption that the number of contributions by any particular user is limited. In Chapter 5 we consider the task of generating trajectories that preserve the correlation between points submitted by a user without restricting the number of contributions they make. We show that these proposed methods perform better than existing baselines and have the potential to be adopted for real-world datasets.

En els darrers anys, les tècniques generatives d'aprenentatge profund han vist un ampli ús, especialment en la generació d'imatges i text. Aquests mètodes també s'han mostrat prometedors en la generació de dades de sèries temporals, fins i tot en dominis continus. No obstant això, tenen dificultats per capturar certes distribucions de probabilitat, com les distribucions de cua pesada comúment observades en els rendiments d'actius financers. Aquesta limitació sorgeix de la continuïtat Lipschitz de les xarxes neuronals generadores, que impedeix la generació de distribucions amb cues més pesades que el soroll previ utilitzat en aquests mètodes. Utilitzar senzillament un soroll previ més pesat porta a problemes d'estabilitat i extrapolació. Per abordar aquest repte, el nostre treball proposa un enfocament innovador. Reduïm les dades d'entrenament de sèries temporals a l'espai de paràmetres d'una funció de distribució acumulativa de cua pesada, on els paràmetres evolucionen amb el temps. Després, s'utilitzen tècniques d'aprenentatge profund per generar aquesta evolució de paràmetres, aprofitant la distribució de probabilitat més favorable de les dades transformades. Finalment, es generen sèries temporals en l'espai original de les dades mitjançant la transformació inversa de l'espai de paràmetres utilitzant la funció quantil, el que resulta en un modelatge més precís de la distribució de la cua., En los últimos años, las técnicas generativas de aprendizaje profundo han visto un amplio uso, especialmente en la generación de imágenes y texto. Estos métodos también se han mostrado prometedores en la generación de datos de series temporales, incluso en dominios continuos. Sin embargo, tienen dificultades para capturar ciertas distribuciones de probabilidad, como las distribuciones de cola pesada comúnmente observadas en los rendimientos de activos financieros. Esta limitación surge de la continuidad Lipschitz de las redes neuronales generadoras, que impide la generación de distribuciones con colas más pesadas que el ruido previo utilizado en estos métodos. Simplemente usar un ruido previo más pesado conduce a problemas de estabilidad y extrapolación. Para abordar este desafío, nuestro trabajo propone un enfoque novedoso. Reducimos los datos de entrenamiento de series temporales al espacio de parámetros de una función de distribución acumulativa de cola pesada, donde los parámetros evolucionan con el tiempo. Luego, se emplean técnicas de aprendizaje profundo para generar esta evolución de parámetros, aprovechando la distribución de probabilidad más favorable de los datos transformados. Finalmente, se generan series temporales en el espacio original de los datos mediante la transformación inversa del espacio de parámetros utilizando la función cuantil, lo que resulta en un modelado más preciso de la distribución de la cola., In recent years, generative deep learning techniques have seen widespread use, particularly in image and text generation. These methods have also shown promise in generating time series data, even in continuous domains. However, they struggle to capture certain probability distributions, such as the heavy-tailed distributions commonly observed in financial asset returns. This limitation arises from the Lipschitz continuity of the generator neural networks, which prevents the generation of distributions with heavier tails than the prior noise used in these methods. Simply using a heavier prior noise leads to stability and extrapolation issues. To address this challenge, our work proposes a novel approach. We reduce time series training data to the parameter space of a heavy-tailed cumulative distribution function, where the parameters evolve over time. Deep learning techniques are then employed to generate this parameter evolution, leveraging the more favourable probability distribution of the transformed data. Finally, time series on the sample space are generated by inversely transforming the parameter space using the quantile function, resulting in a more accurate modelling of the tail distribution., Outgoing

Les Xarxes Neuronals sobre Feixos (SNNs) estenen de forma natural les Xarxes Neuronals sobre Grafs (GNNs) otorgant un feix cel·lular sobre el graf, equipant els seus nodes i arestes amb espais vectorials i definint aplicacions lineals entre aquests. Tot i que s'ha demostrat que l'estructura geomètrica afegida al graf és útil per analitzar l'heterofília i el sobre allisament, fins ara els mètodes pels quals s'ha inferit el feix no sempre garanteixen un bon rendiment en aquestes circumstàncies. Aquest projecte agafa inspiració de conceptes de dinàmica d'opinions per proposar nous mètodes per aprendre el feix per tal que sigui més intuitiu entendre'l i tractin el sobre allisament i l'heterofília de forma més explícita. Aquests es centren en explorar modificacions del Laplacià per grafs amb conjunts de punts d'equilibri més diversos. Amb aquest propòsit, en primer lloc tenim una explicació exhaustiva de tots els conceptes matemàtics que hi ha darrere els feixos i el sobre allisament. Després repassem algunes de les GNNs més populars, incloent-hi les SNNs, i cobrim els conceptes de sobre allisament i heterofília en el context de les GNNs. Llavors, ens endinsem a la teoria de dinàmica d'opinions per extreure'n processos útils per dades heterofíliques. Un cop tota la base teòrica ha estat coberta, proposem dos nous models, JdSNN i RiSNN, per ser utilitzats en dades heterofíliques on les GNNs tinguin problemes de sobre allisament. Per comparar aquests models amb les SNNs originals en un problema controlat proposem una nova forma de generar dades sintètiques per evaluar xarxes neuronals basades en feixos. Fent servir aquestes dades, evaluem els nostres models i les SNNs per treure conclusions significatives sobre el seu comportament. Després, evaluem els nostres models en dades referents que s'havien fet servir prèviament, noves dades referents, un problema de detecció de notícies falses i algunes dades referents d'hipergrafs., Las Redes Neuronales sobre Fajos (SNNs) extienden de forma natural las Redes Neuronales sobre Grafos (GNNs) otorgando un fajo celular sobre el grafo, equipando sus nodos y aristas con espacios vectoriales y definiendo aplicaciones lineales entre estos. Aunque se ha demostrado que la estructura geométrica añadida al grafo es útil para analizar la heterofilia y el sobre alisamiento, hasta ahora los métodos por los cuales el fajo se ha calculado no siempre garantizan un buen rendimiento en estas circumstancias. Este proyecto coge inspiración de conceptos de dinámica de opiniones para proponer nuevos métodos para aprender el fajo para que sea más intuitivo entenderlo y traten el sobre alisamiento y la heterofilia de forma más explícita. Estos se focalizan en explorar modificaciones del Laplaciano para grafos con conjuntos de puntos de equilibrio más ricos. Con este propósito, en primer lugar tenemos una explicación exhaustiva de todos los conceptos matemáticos que hay destrás de los fajos y el sobre alisamiento. Después repasamos algunas de las GNNs más populares, incluyendo las SNNs, y cubrimos los conceptos de sobre alisamiento y heterofília en el contexto de las GNNs. Entonces entramos en la teoria de dinámica de opiniones para extraer procesos útiles para datos heterofílicos. Con toda la base teórica cubierta, proponemos dos nuevos modelos, JdSNN y RiSNN, para ser usados en datos heterofílicos donde las GNNs tengan problemas de sobre alisamiento. Para comparar estos modelos con las SNNs originales en un problema controlado proponemos una nueva forma de generar datos sintéticos para evaluar redes neuronales basades en fajos. Usando estos datos, evaluamos nuestros modelos y las SNNs para extraer conclusiones significativas sobre su comportamiento. Después, evaluamos nuestros modelos en datos referentes que se habían usado previamente, nuevos datos referentes, un problema de detección de notícias falsas y algunos datos referentes de hipergrafos., Sheaf Neural Networks (SNNs) naturally extend Graph Neural Networks (GNNs) by endowing a cellular sheaf over the graph, equipping nodes and edges with vector spaces and defining linear mappings between them. While the attached geometric structure has proven to be useful in analyzing heterophily and oversmoothing, so far the methods by which the sheaf is computed do not always guarantee a good performance in such settings. This work draws inspiration from opinion dynamics concepts to propose novel sheaf learning methods to provide a more intuitive understanding of the sheaf and address oversmoothing and heterophily more explicitly than previous works. These focus on exploring modifications of the graph Laplacian with richer sets of equilibrium points. For this purpose, an exhaustive explanation of the mathematical concepts behind sheaves and oversmoothing is done. Then we go over some of the most popular GNN models, including SNNs, and cover the concepts of oversmoothing and heterophily in the context of GNNs. After this, we dive into the concepts of opinion dynamics theory that are useful for heterophilic data. With all the background done, two new models, JdSNN and RiSNN, are proposed to be used on heterophilic data where oversmoothing might be a problem. To test these models against regular SNNs in a controlled environment we propose a new framework to generate synthetic data to test sheaf-based graph neural networks. Using said framework we test our models against regular SNNs and draw meaningful conclusions from said results. Then these new models are tested on previously used benchmarks, new benchmarks, a fake news detection dataset and some hypergraph datasets. Finally, there's a methodology and planning section which explains how this thesis was developed as well as a sustainability report., Outgoing

Automotive production systems are designed to produce large quantities in high quality and short throughput times and are therefore organized as line production. This places high quality requirements on the joining processes in automotive body shops, in which automated, robot-guided welding is a key process. The quality of these thermal joining processes depends on various physical and chemical influencing factors, whose interactions cannot be explicitly modelled. This leads to enormous quality assurance efforts in several quality control loops, which may include visual inspections, non-destructive testing of samples to assess the internal structure and destructive testing of samples for the assessment of mechanical properties such as tensile strength. Due to the increasing availability of data in automated processes and the complexity of welding processes, the application of Deep Learning has a great potential to reduce quality control efforts in automotive body shops. Using Deep Learning to leverage process data and accurately predict quality parameters in welding processes is investigated in research, yet model training requires a large, balanced and annotated dataset, whose generation is time and cost intensive, particularly for production data. However, there are generative AI methods such as Generative Adversarial Networks (GANs) that are able to generate synthetic data and thus offer the potential to generate a large amount of annotated production data with relatively little effort. This paper presents a systematic approach to evaluate the potential of incorporating synthetic data in a real-world production dataset to improve quality control using Deep Learning. The approach is validated for the analysis of real-world ultrasound images of resistance spot welding (RSW) processes from the automotive industry. Different Deep Learning architectures to generate synthetic data are compared. Results show that adding synthetic data to the training dataset can improve

This thesis introduces methods for accurate 3D reconstruction of FIB tomography data using machine learning. It addresses challenges in obtaining large datasets and ground truth values, proposing a method for virtual FIB tomography data generation and an isotropy-based validation method. ML-based segmentation methods tackle issues in BSE images, like the shine-through effect and image intensity ambiguities. Integrating the slice repositioning method with image inpainting resolves inconsistencies in slice thicknesses in FIB tomography data. Additionally, a multimodal machine learning approach using multivoltage images further enhances 3D reconstruction accuracy., In dieser Arbeit werden Methoden für die 3D-Rekonstruktion von FIB-Tomographiedaten mithilfe von maschinellem Lernen vorgestellt. Es befasst sich mit den Herausforderungen bei der Datenerfassung und dem Fehlen von Wahrheitsdaten; und schlägt eine virtuelle Datengenerierung und eine auf Isotropie basierende Validierung vor. Die ML-basierte Segmentierung löst Probleme wie den shine-through effekt und die Mehrdeutigkeit der Bildintensität in BSE-Bildern. Durch die Integration der Neupositionierung von Schichten und der Methode der Bildverbesserung werden Inkonsistenzen in der Schichtdicke beseitigt. Ein multimodaler Lernansatz mit Multivoltage-Bildern verbessert die Genauigkeit der 3D-Rekonstruktion weiter.

Automotive production systems are designed to produce large quantities in high quality and short throughput times and are therefore organized as line production. This places high quality requirements on the joining processes in automotive body shops, in which automated, robot-guided welding is a key process. The quality of these thermal joining processes depends on various physical and chemical influencing factors, whose interactions cannot be explicitly modelled. This leads to enormous quality assurance efforts in several quality control loops, which may include visual inspections, non-destructive testing of samples to assess the internal structure and destructive testing of samples for the assessment of mechanical properties such as tensile strength. Due to the increasing availability of data in automated processes and the complexity of welding processes, the application of Deep Learning has a great potential to reduce quality control efforts in automotive body shops. Using Deep Learning to leverage process data and accurately predict quality parameters in welding processes is investigated in research, yet model training requires a large, balanced and annotated dataset, whose generation is time and cost intensive, particularly for production data. However, there are generative AI methods such as Generative Adversarial Networks (GANs) that are able to generate synthetic data and thus offer the potential to generate a large amount of annotated production data with relatively little effort. This paper presents a systematic approach to evaluate the potential of incorporating synthetic data in a real-world production dataset to improve quality control using Deep Learning. The approach is validated for the analysis of real-world ultrasound images of resistance spot welding (RSW) processes from the automotive industry. Different Deep Learning architectures to generate synthetic data are compared. Results show that adding synthetic data to the training dataset can improve

The purpose of the study is to demonstrate the feasibility of combining traffic simulator technology with machine learning (ML) methods to create realistic and comprehensive synthetic traffic data. Synthetic data alleviates many ethical and privacy concerns, significantly reduces the costs associated with data collection, and enables researchers to study scenarios and conditions that are difficult or impossible to replicate in real-world environments. Access to large amounts of diverse and controlled data is essential for developing and testing artificial intelligence (AI) models and leads to more reliable and robust results. Traffic simulators like SUMO have been successfully used for that purpose in the past, creating realistic vehicular traces. One drawback is that, without coupling them with complex physics emulators, they are not capable of generating internal vehicle parameters. Such parameters, on the other hand, are crucial for many purposes, from understanding energy efficiency and optimizing driver behavior to predictive maintenance and monitoring the degradation of key components, such as driveline batteries. In this paper, we propose Synthetic Traffic Data Generator (STDG) and demonstrate that an ML model that is trained on the internal parameters of a vehicle in one set of conditions (Sweden) can be used to generate synthetic data corresponding to another setting (Monaco). The proposed method promises to eliminate the need for an expensive collection of the original vehicle parameters across many different settings. Moreover, sharing the synthetic data with additional stakeholders is easier due to the reduced security and integrity risk of exposing the vehicle’s privacy-sensitive original parameters. This study compares several ML techniques, including deep learning (DL) based, for generating internal parameters of vehicles, such as fuel rate, engine speed, and wet tank air pressure. Using the actual bus data from a small city to train our ML models, we

Recent studies have shown how synthetic data generation methods can be applied to electronic health records (EHRs) to obtain synthetic versions that do not violate privacy rules. This growing body of research has resulted in the emergence of numerous methods for evaluating the quality of generated data, with new publications often introducing novel evaluation methods. This work presents a detailed review of synthetic EHRs, focusing on the various evaluation methods used to assess the quality of the generated EHRs. We discuss the existing evaluation methods, offering insights into their use as well as providing an interpretation of the evaluation metrics from the perspectives of achieving fidelity, utility and privacy. Furthermore, we highlight the key factors influencing the selection of evaluation methods, such as the type of data (e.g., categorical, continuous, or discrete) and the mode of application (e.g., patient level, cohort level, and feature level). To assess the effectiveness of current evaluation measures, we conduct a series of experiments to shed light on the potential limitations of these measures. The findings from these experiments reveal notable shortcomings, including the need for meticulous application of methods to the data to reduce inconsistent evaluations, the qualitative nature of some assessments subject to individual judgment, the need for clinical validations, and the absence of techniques to evaluate temporal dependencies within the data. This highlights the need to place greater emphasis on evaluation measures, their application, and the development of comprehensive evaluation frameworks as it is crucial for advancing progress in this field. © 2024 The Author(s)

Synthetic data generation has been proposed as a potential solution to accessing Electronic Health Records (EHRs) while minimizing the privacy risks associated with real EHRs. Nevertheless, the practical use of synthetic EHRs rests on their ability to resemble the quality of real EHRs. Existing evaluations of synthetic EHRs often focus on assessing them as static snapshots frozen in time, neglecting temporal dependencies and varying temporal patterns. Moreover, some of these methods rely on subjective judgments, are limited to segmentable time-series, and employ methods that adopt a one-to-one approach. This study employs a comprehensive approach to evaluating fidelity in synthetic time-series EHRs to address these challenges. We extend the functionality of time-series analysis methods such as temporal clustering, time-series similarity measures, Sample Entropy, and trend analysis, to evaluate varying temporal patterns in synthetic time-series EHRs. Our findings provide valuable insights into how synthetic EHRs align with real EHRs in the temporal context, considering aspects such as patient groupings, temporal dynamics, predictability, and directional change. We empirically demonstrate the feasibility of assessing temporal fidelity with these methods, offering an understanding of the quality of synthetic EHRs in capturing the varying temporal patterns inherent in EHRs. © 2024 IEEE.

In the manufacturing industry, automatic quality inspections can lead to improved product quality and productivity. Deep learning-based computer vision technologies, with their superior performance in many applications, can be a possible solution for automatic quality inspections. However, collecting a large amount of annotated training data for deep learning is expensive and time-consuming, especially for processes involving various products and human activities such as assembly. To address this challenge, we propose a method for automated assembly quality inspection using synthetic data generated from computer-aided design (CAD) models. The method involves two steps: automatic data generation and model implementation. In the first step, we generate synthetic data in two formats: two-dimensional (2D) images and three-dimensional (3D) point clouds. In the second step, we apply different state-of-the-art deep learning approaches to the data for quality inspection, including unsupervised domain adaptation, i.e., a method of adapting models across different data distributions, and transfer learning, which transfers knowledge between related tasks. We evaluate the methods in a case study of pedal car front-wheel assembly quality inspection to identify the possible optimal approach for assembly quality inspection. Our results show that the method using Transfer Learning on 2D synthetic images achieves superior performance compared with others. Specifically, it attained 95% accuracy through fine-tuning with only five annotated real images per class. With promising results, our method may be suggested for other similar quality inspection use cases. By utilizing synthetic CAD data, our method reduces the need for manual data collection and annotation. Furthermore, our method performs well on test data with different backgrounds, making it suitable for different manufacturing environments., CC BY 4.0 DEED© The Author(s) 2024Correspondence Address: X. Zhu; Scania CV AB (publ), Södertälje, 151 87, Sweden; email: xiazhu@kth.se; CODEN: JIMNEOpen access funding provided by Royal Institute of Technology. This work is partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.

This study addresses the issue of black-grass, a herbicide-resistant weed that threatens wheat yields in Western Europe, through the use of high- resolution Unmanned Aerial Vehicles (UAVs) and synthetic data augmentation in precision agriculture. We mitigate challenges such as the need for large labeled datasets and environmental variability by employing synthetic data augmentations in training a Mask R-CNN model. Using a minimal dataset of 43 black-grass and 12 wheat field images, we achieved a 37% increase in Area Under the Curve (AUC) over the non-augmented baseline, with scaling as the most effective augmentation. The best model attained a recall of 53% at a precision of 64%, offering a promising approach for future precision agriculture applications. © NLDL 2024. All rights reserved.

In the manufacturing industry, automatic quality inspections can lead to improved product quality and productivity. Deep learning-based computer vision technologies, with their superior performance in many applications, can be a possible solution for automatic quality inspections. However, collecting a large amount of annotated training data for deep learning is expensive and time-consuming, especially for processes involving various products and human activities such as assembly. To address this challenge, we propose a method for automated assembly quality inspection using synthetic data generated from computer-aided design (CAD) models. The method involves two steps: automatic data generation and model implementation. In the first step, we generate synthetic data in two formats: two-dimensional (2D) images and three-dimensional (3D) point clouds. In the second step, we apply different state-of-the-art deep learning approaches to the data for quality inspection, including unsupervised domain adaptation, i.e., a method of adapting models across different data distributions, and transfer learning, which transfers knowledge between related tasks. We evaluate the methods in a case study of pedal car front-wheel assembly quality inspection to identify the possible optimal approach for assembly quality inspection. Our results show that the method using Transfer Learning on 2D synthetic images achieves superior performance compared with others. Specifically, it attained 95% accuracy through fine-tuning with only five annotated real images per class. With promising results, our method may be suggested for other similar quality inspection use cases. By utilizing synthetic CAD data, our method reduces the need for manual data collection and annotation. Furthermore, our method performs well on test data with different backgrounds, making it suitable for different manufacturing environments., CC BY 4.0 DEED© The Author(s) 2024Correspondence Address: X. Zhu; Scania CV AB (publ), Södertälje, 151 87, Sweden; email: xiazhu@kth.se; CODEN: JIMNEOpen access funding provided by Royal Institute of Technology. This work is partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.

Generative Adversarial Neural nets (GANs) are a new branch of machine learning techniques. A GAN learns to generate new data from the training data set. We examine the characteristics of the fake financial data using GANs trained on samples of daily S&P 500 and FTSE 100 index values. GANs feature two competing neural networks in a game theoretic context. The Generator net generates pseudo data that is presented to the discriminator net which then attempts to distinguish between the real and the fake data. This facilitates unsupervised learning on the dataset. The generative network generates data sets, while the discriminative network evaluates them. Equilibrium is reached when the generator can fool the discriminator half the time. Potential convergence difficulties led to the development of Wasserstein GANs, which we use in the analysis. We examine the characteristics of the generated fake S&P500 and FTSE 100 data sets. A key issue is how closely does the fake series mimic the real series? We explore this using a variety of metrics including regression analysis, applications of moments and characteristic functions, plus Random Forest analysis. We provide a practical application using the fake data to calculate Value at Risk (VaR).

In the last five years, transformer-based large language models have not only made tremendous advancements in fluent and accurate text generation, but also show unparalleled performance as zero- or few-shot shot learners of numerous downstream tasks. As their arsenal continues to expand, so must our evaluation methods. This thesis project contributes to an existing linguistically motivated template-based question-answering behavior testing framework, the Multilingual Mor-phological Checklist, by developing Japanese templates for a number of logical reasoning tasks. Furthermore, we add politeness as a new linguistic dimension within the framework. The Japanese question-answering prompts generated by our templates are evaluated on three pre-trained transformer-based multilingual generative language models alongside English, German and Finnish prompts. The model performance is evaluated in three experimental settings: one zero-shot experiment and two one-shot experiments. Using a number of automatic partial string-matching strategies, we find that all models achieve reasonable accuracy on the Japanese prompts for a majority of the reasoning tasks. The accuracy on Japanese is, unsurprisingly, worse than for English on average. Moreover, the best performance is observed in the zero-shot setting, suggesting that the introduction of additional elements in prompts may confuse the model more than the example helps clarify the task format. Additionally, we evaluate the reliability of the automated string-matching strategies, comparing them to a sample of human annotation. This reveals that the English scores align closer to the human scores than Japanese scores do, suggesting that different languages may benefit from different automated scoring approaches. Lastly, we find that although the performance across the politeness dimensionis quite scattered, formal Japanese prompts sometimes outperform the polite and familiar language variants. On average, however, the formal langua

Synthetic generation of biophysical signals caters to diverse applications of test vectors generation, exploration, data augmentation and machine learning training. Arrhythmiator, our contribution is a Python framework to generate synthetically, action potential for explain-ability and explore-ability of pathology (Arrhythmia) influenced cardiac episodes. In this work, we address the problem of large cardiac data-sets need, for deep learning models which are hindered by privacy laws by synthetic generation. We have addressed the problem of explainable computations, multi-scale, by virtue of novel stack sliced polynomial signatures, piece-wise linear and application context driven fitting/filtering in algorithmic framework. To account the problem of lack of standardised pathology specifications we developed PDL (Pathology descriptive language) portal which can be customised. We also addressed the challenge of high turnaround development cycle, inter-operate-ability computation issue by developing a single unified Python framework. To address computational costs, we leverage domain-specific heuristics through sliced state space control algorithms and a constant temporal resolution. We demonstrate the utility of our framework by generating two cases of Tachycardia, one influenced by sodium pump inhibition of spontaneous diastolic depolarisation (SDD) slope and other by ischaemic episode causing maximum diastolic depolarisation (MDP) shift. The concept of the pacemaker hierarchy is encompassed by developing pathology explainable computation as a structure (PECS). We showcase the flexibility and explore-ability of our framework through ease of filter connect profile plug-ins for action potential states upstroke and plateau. We validated the generated waveforms by developing trace and debug algorithms and a reference deviation score which marks the pathology influence. The validator module is a vectored attribute of action potential based on state space control algorithms

Large-scale self-supervised pretraining of deep learning models is known to be critical in several fields, such as language processing, where its has led to significant breakthroughs. Indeed, it is often more impactful than architectural designs. However, the use of self-supervised pretraining lags behind in several domains, such as hyperspectral images, due to data scarcity. This paper addresses the challenge of data scarcity in the development of methods for spatial super-resolution of hyperspectral images (HSI-SR). We show that state-of-the-art HSI-SR methods are severely bottlenecked by the small paired datasets that are publicly available, also leading to unreliable assessment of the architectural merits of the models. We propose to capitalize on the abundance of high resolution (HR) RGB images to develop a self-supervised pretraining approach that significantly improves the quality of HSI-SR models. In particular, we leverage advances in spectral reconstruction methods to create a vast dataset with high spatial resolution and plausible spectra from RGB images, to be used for pretraining HSI-SR methods. Experimental results, conducted across multiple datasets, report large gains for state-of-the-art HSI-SR methods when pretrained according to the proposed procedure, and also highlight the unreliability of ranking methods when training on small datasets.

The fashion industry is a significant contributor to greenhouse gas emissions and textile waste, prompting the need for sustainable practices. This thesis explores the use of diffusion models for generating synthetic data to enhance datasets used in machine learning, specifically focusing on second-hand fashion. Diffusion models, known for their ability to create high-quality images, offer potential solutions to the imbalance and quality issues in existing datasets. The study investigates how image generation and editing through diffusion models can improve datasets, the effectiveness of different prompting strategies, and the performance of synthetic data in machine learning models compared to real data. The methodology involves using the Kandinsky 2.2 inpainting model to generate and edit images, followed by manual and automated classification to evaluate image quality. Experiments demonstrate that diffusion models can plausibly improve dataset quality by adding and removing damage in images, although fully automating this process remains challenging. The results indicate that augmenting the datasets with synthetic images can potentially enhance the performance of the model, although the variability of the results suggests the need for further research. This thesis contributes to the field of sustainable fashion by proposing innovative methods for dataset augmentation using state-of-the-art generative models, aiming to support the development of efficient and automated sorting processes in the textile industry.

This thesis explores the efficacy of synthetic image data for training convolutional neural networks (CNNs) in the domain of 2D pose estimation of horses. The study addresses the general lack of annotated datasets for horse pose estimation by creating the synthetic dataset HorSy15k using a 3D model in a computer graphics engine. This dataset comprises 15,000 images of horses in various scenes with automatically generated annotations for 20 keypoints. CNN models were trained using 10 different datasets: one with entirely synthetic images, two with real images, and seven with varying ratios of real and synthetic images. The performance of these models was evaluated on unseen real horse images. The results demonstrate that incorporating synthetic data can enhance model accuracy and generalization capabilities compared to models trained exclusively on real images. Notably, models trained with a balanced mix of synthetic and real images exhibited superior performance, suggesting that synthetic data can effectively complement real data in training robust CNNs for horse pose estimation. The findings underscore the potential of synthetic data in overcoming the limitations of real-world data scarcity and annotation challenges., I denna rapport undersöks effektiviteten av syntetisk bilddata för att träna faltande neurala nätverk (convolutional neural networks, CNNs) inom området 2D pose-estimering av hästar. Studien angriper den allmänna bristen på annoterade dataset för pose-estimering av hästar genom att med hjälp av en 3D-modell och en motor för 3D-grafik skapa HorSy15k, ett syntetiskt dataset. Detta dataset består av 15 000 bilder av hästar i olika miljöer med automatiskt genererade annotationer för 20 nyckelpunkter. CNN-modeller tränades på tio olika dataset: ett med helt syntetiska bilder, två med riktiga bilder och sju med varierande proportioner av riktiga och syntetiska bilder. Modellerna utvärderades sedan på osedda riktiga bilder av hästar. Resultatet visar att inkorporering av syntetiska data kan förbättra modellers noggrannhet och generaliseringsförmåga jämfört med modeller som enbart tränats på riktiga bilder. Särskilt visar studien att modeller som tränats med en balanserad blanding av syntetiska och riktiga bilder ger en överlägsen prestanda, vilket tyder på att syntetisk data effektivt kan komplettera reell data vid träning av robusta CNN-modeller för pose-estimering av hästar. Resultaten understryker potentialen hos syntetisk data att kunna övervinna begränsningar som brist på verkliga data och annoteringsutmaningar.

The development of new solutions incorporating artificial intelligence (AI) within the medical field is an area of great interest. However, access to comprehensive and diverse datasets is restricted due to the sensitive nature of the data. A potential solution to this is to generatesynthetic datasets based on real medical data. Synthetic data could protect the integrity of the subjects while preserving the inherent information necessary for training AI models and be generated in greater quantity than otherwise available. This thesis project aims to generate reliable clinical trial subject data using a generative adversarial network (GAN). The main data set used is a mock clinical trial dataset consisting of multiple subject visits, however an additional data set containing authentic medical data is also used for better insights into the model’s ability to learn underlying relationships. The thesis also investigates training strategies for simulating the temporal dimension and the missing values in the data. The GAN model used is an altered version of the Conditional Tabular GAN (CTGAN)made to be compatible with the preprocessed clinical trial mock data, and multiple model architectures and number of training epochs are examined. The results show great potential for GAN models on clinical trial datasets, especially for real-life data. One model, trained on the authentic dataset, generates near-perfect synthetic data with respect to column distributions and correlation between columns. The results also show that classification models trained on synthetic data and tested on real data have the potential to match the performance of classification models trained on real data. While the synthetic data replicates the missing values, no definitive conclusion can be drawn regarding the temporal characteristics due to the sparsity of the mock dataset and lack of real correlations in it. Although the results are promising, further experiments on authentic datasets with less spa

The development of high-performing object detection models requires extensive and varied datasets with accurately annotated images, a process that is traditionally labor-intensive and prone to errors. To address these challenges, this report explores the generation of synthetic data using domain randomization techniques to train object detection models. We present a pipeline that integrates synthetic data creation in Unity, and the training of YOLOv8 object detection models. Our approach uses the Unity Perception package to produce diverse and precisely annotated datasets, overcoming the domain gap typically associated with synthetic data. The pipeline was evaluated through a series of experiments, analyzing the impact of various parameters such as background textures, and training arguments on model performance. The results demonstrate that models trained with our synthetic data can achieve high accuracy and generalize well to real-world scenarios, offering a scalable and efficient alternative to manual data annotation.

Modern deep learning models require large amounts of data to train, and the acquisition of data can be challenging. Synthetic data provides an alternative to manually collecting real data, alleviating problems associated with real data acquisition. For recycling processes, classifying metal scrap piles containing hazardous objects is important, where hazardous objects can be damaging and costly if handled incorrectly. Automatically detecting hazardous objects in metal scrap piles using image classification models requires large amounts of data, and metal scrap piles contain large variations in objects, textures, and lighting. Furthermore, data acquisition can be challenging in the recycling domain, where positive objects can be scarce and manual acquisition setup can be challenging. In this thesis, synthetic images of metal scrap piles in a recycling process are created, intended for training image classification models to detect metal scrap piles containing fire extinguishers or hydraulic cylinders. Synthetic images are created with physically based rendering and domain randomization, rendered with either rasterization or ray tracing engines. Ablation studies are conducted to investigate the effect of using domain randomization. The performance of models trained on purely synthetic datasets is compared to models trained on datasets containing only real images. Furthermore, photorealistic rendering with ray tracing rendering is evaluated by comparing F1 scores between models trained on data sets created with rasterization or ray tracing. The F1 scores show that models trained on purely synthetic data outperform those trained solely on real data when classifying images containing fire extinguishers or hydraulic cylinders. Ablation studies show that domain randomization of textures is beneficial both for the classification of fire extinguishers and for the classification of hydraulic cylinders in metal scrap piles. High dynamic range image lighting randomization does

This thesis investigates the effectiveness of Tabular Variational Auto Encoders (TVAEs) in generating high-quality synthetic tabular data and assesses their compliance with differential privacy principles. The study shows that while TVAEs are better than VAEs at generating synthetic data that faithfully reproduces the distribution of real data as measured by the Synthetic Data Vault (SDV) metrics, the latter does not guarantee that the synthetic data is up to the task in practical industrial applications. In particular, models trained on TVAE-generated data from the Creditcards dataset are ineffective. The author also explores various optimisation methods on TVAE, such as Gumbel Max Trick, Drop Out (DO) and Batch Normalization, while pointing out that techniques frequently used to improve two-dimensional TVAE, such as Kullback–Leibler Warm-Up and B Disentanglement, are not directly transferable to the one-dimensional context. However, differential privacy to TVAE was not implemented due to time constraints and inconclusive results. The study nevertheless highlights the benefits of stabilising training with the Differential Privacy - Stochastic Gradient Descent (DP-SGD), as with a dropout, and the existence of an optimal equilibrium point between the constraints of differential privacy and the number of training epochs in the model., Denna avhandling undersöker hur effektiva Tabular Variational AutoEncoders (TVAE) är när det gäller att generera högkvalitativa syntetiska tabelldata och utvärderar deras överensstämmelse med differentierade integritetsprinciper. Studien visar att även om TVAE är bättre än VAE på att generera syntetiska data som troget återger fördelningen av verkliga data mätt med Synthetic Data Vault (SDV), garanterar det senare inte att de syntetiska data är upp till uppgiften i praktiska industriella tillämpningar. I synnerhet är modeller som tränats på TVAE-genererade data från Creditcards-datasetet ineffektiva. Författaren undersöker också olika optimeringsmetoder för TVAE, såsom Gumbel Max Trick, DO och Batch Normalization, samtidigt som han påpekar att tekniker som ofta används för att förbättra tvådimensionell TVAE, såsom Kullback-Leibler Warm-Up och B Disentanglement, inte är direkt överförbara till det endimensionella sammanhanget. På grund av tidsbegränsningar och redan ofullständiga resultat implementerades dock inte differentierad integritet för TVAE. Studien belyser ändå fördelarna med att stabilisera träningen med Differential Privacy - Stochastic Gradient Descent (DP-SGD), som med en drop-out, och förekomsten av en optimal jämviktspunkt mellan begränsningarna för differential privacy och antalet träningsepoker i modellen.

In the last five years, transformer-based large language models have not only made tremendous advancements in fluent and accurate text generation, but also show unparalleled performance as zero- or few-shot shot learners of numerous downstream tasks. As their arsenal continues to expand, so must our evaluation methods. This thesis project contributes to an existing linguistically motivated template-based question-answering behavior testing framework, the Multilingual Mor-phological Checklist, by developing Japanese templates for a number of logical reasoning tasks. Furthermore, we add politeness as a new linguistic dimension within the framework. The Japanese question-answering prompts generated by our templates are evaluated on three pre-trained transformer-based multilingual generative language models alongside English, German and Finnish prompts. The model performance is evaluated in three experimental settings: one zero-shot experiment and two one-shot experiments. Using a number of automatic partial string-matching strategies, we find that all models achieve reasonable accuracy on the Japanese prompts for a majority of the reasoning tasks. The accuracy on Japanese is, unsurprisingly, worse than for English on average. Moreover, the best performance is observed in the zero-shot setting, suggesting that the introduction of additional elements in prompts may confuse the model more than the example helps clarify the task format. Additionally, we evaluate the reliability of the automated string-matching strategies, comparing them to a sample of human annotation. This reveals that the English scores align closer to the human scores than Japanese scores do, suggesting that different languages may benefit from different automated scoring approaches. Lastly, we find that although the performance across the politeness dimensionis quite scattered, formal Japanese prompts sometimes outperform the polite and familiar language variants. On average, however, the formal langua

This thesis explores the efficacy of synthetic image data for training convolutional neural networks (CNNs) in the domain of 2D pose estimation of horses. The study addresses the general lack of annotated datasets for horse pose estimation by creating the synthetic dataset HorSy15k using a 3D model in a computer graphics engine. This dataset comprises 15,000 images of horses in various scenes with automatically generated annotations for 20 keypoints. CNN models were trained using 10 different datasets: one with entirely synthetic images, two with real images, and seven with varying ratios of real and synthetic images. The performance of these models was evaluated on unseen real horse images. The results demonstrate that incorporating synthetic data can enhance model accuracy and generalization capabilities compared to models trained exclusively on real images. Notably, models trained with a balanced mix of synthetic and real images exhibited superior performance, suggesting that synthetic data can effectively complement real data in training robust CNNs for horse pose estimation. The findings underscore the potential of synthetic data in overcoming the limitations of real-world data scarcity and annotation challenges., I denna rapport undersöks effektiviteten av syntetisk bilddata för att träna faltande neurala nätverk (convolutional neural networks, CNNs) inom området 2D pose-estimering av hästar. Studien angriper den allmänna bristen på annoterade dataset för pose-estimering av hästar genom att med hjälp av en 3D-modell och en motor för 3D-grafik skapa HorSy15k, ett syntetiskt dataset. Detta dataset består av 15 000 bilder av hästar i olika miljöer med automatiskt genererade annotationer för 20 nyckelpunkter. CNN-modeller tränades på tio olika dataset: ett med helt syntetiska bilder, två med riktiga bilder och sju med varierande proportioner av riktiga och syntetiska bilder. Modellerna utvärderades sedan på osedda riktiga bilder av hästar. Resultatet visar att inkorporering av syntetiska data kan förbättra modellers noggrannhet och generaliseringsförmåga jämfört med modeller som enbart tränats på riktiga bilder. Särskilt visar studien att modeller som tränats med en balanserad blanding av syntetiska och riktiga bilder ger en överlägsen prestanda, vilket tyder på att syntetisk data effektivt kan komplettera reell data vid träning av robusta CNN-modeller för pose-estimering av hästar. Resultaten understryker potentialen hos syntetisk data att kunna övervinna begränsningar som brist på verkliga data och annoteringsutmaningar.

In the era of data-driven decision-making and innovation, synthetic data serves as a promising tool that bridges the need for vast datasets in machine learning (ML) and the imperative necessity of data privacy. By simulating real-world data while preserving privacy, synthetic data generators have become more prevalent instruments in AI and ML development. A key challenge with synthetic data lies in accurately estimating its utility. For such purpose, Population Fidelity (PF) measures have shown to be good candidates, a category of metrics that evaluates how well the synthetic data mimics the general distribution of the original data. With this setting, we aim to answer: "How well are different population fidelity measures able to indicate the utility of synthetic data for machine learning based classification models?" We designed a reusable six-step experiment framework to examine the correlation between nine PF measures and the performance of four ML for training classification models over five datasets. The six-step approach includes data preparation, training, testing on original and synthetic datasets, and PF measures computation. The study reveals non-linear relationships between the PF measures and synthetic data utility. The general analysis, meaning the monotonic relationship between the PF measure and performance over all models, yielded at most moderate correlations, where the Cluster measure showed the strongest correlation. In the more granular model-specific analysis, Random Forest showed strong correlations with three PF measures. The findings show that no PF measure shows a consistently high correlation over all models to be considered a universal estimator for model performance.This highlights the importance of context-aware application of PF measures and sets the stage for future research to expand the scope, including support for a wider range of types of data and integrating privacy evaluations in synthetic data assessment. Ultimately, this study, I eran av datadriven beslutsfattning och innovation, fungerar syntetiska data som ett lovande verktyg som bryggar behovet av omfattande dataset inom maskininlärning (ML) och nödvändigheten för dataintegritet. Genom att simulera verklig data samtidigt som man bevarar integriteten, har generatorer av syntetiska data blivit allt vanligare verktyg inom AI och ML-utveckling. En viktig utmaning med syntetiska data är att noggrant uppskatta dess användbarhet. För detta ändamål har mått under kategorin Populations Fidelity (PF) visat sig vara goda kandidater, det är mätvärden som utvärderar hur väl syntetiska datan efterliknar den generella distributionen av den ursprungliga datan. Med detta i åtanke strävar vi att svara på följande: Hur väl kan olika population fidelity mätvärden indikera användbarheten av syntetisk data för maskininlärnings baserade klassifikationsmodeller? För att besvara frågan har vi designat ett återanvändbart sex-stegs experiment ramverk, för att undersöka korrelationen mellan nio PF-mått och prestandan hos fyra ML klassificeringsmodeller, på fem dataset. Sex-stegs strategin inkluderar datatillredning, träning, testning på både ursprungliga och syntetiska dataset samt beräkning av PF-mått. Studien avslöjar förekommandet av icke-linjära relationer mellan PF-måtten och användbarheten av syntetiska data. Den generella analysen, det vill säga den monotona relationen mellan PF-måttet och prestanda över alla modeller, visade som mest medelmåttiga korrelationer, där Cluster-måttet visade den starkaste korrelationen. I den mer detaljerade, modell-specifika analysen visade Random Forest starka korrelationer med tre PF-mått. Resultaten visar att inget PF-mått visar konsekvent hög korrelation över alla modeller för att betraktas som en universell indikator för modellprestanda. Detta understryker vikten av kontextmedveten tillämpning av PF-mått och banar väg för framtida forskning för att utöka omfånget, inklusive stöd för ett bredare utbud för data av olika typ

An ego-network is a graph representing the interactions of a node (ego) with its neighbors and the interactions among those neighbors. A sequence of ego-networks having the same ego can thus model the evolution of these interactions over time. We introduce the problem of segmenting a sequence of ego-networks into $k$ segments, for any given integer $k$. Each segment is represented by a summary network, and the goal is to minimize the total loss of representing $k$ segments by $k$ summaries. The problem allows partitioning the sequence into homogeneous segments with respect to the activities or properties of the ego (e.g., to identify time periods when a user acquired different circles of friends in a social network) and to compactly represent each segment with a summary. The main challenge is to construct a summary that represents a collection of ego-networks with minimum loss. To address this challenge, we employ Jaccard Median (JM), a well-known NP-hard problem for summarizing sets, for which, however, no effective and efficient algorithms are known. We develop a series of algorithms for JM offering different effectiveness/efficiency trade-offs: (I) an exact exponential-time algorithm, based on Mixed Integer Linear Programming; (II) exact and approximation polynomial-time algorithms for minimizing an upper bound of the objective function of JM; and (III) efficient heuristics for JM, which are based on an effective scoring scheme and one of them also on sketching. We also study a generalization of the segmentation problem, in which there may be multiple edges between a pair of nodes in an ego-network. To tackle this problem, we develop a series of algorithms, based on a

The increasing use of deep learning techniques has reduced interpretation time and, ideally, reduced interpreter bias by automatically deriving geological maps from digital outcrop models. However, accurate validation of these automated mapping approaches is a significant challenge due to the subjective nature of geological mapping and the difficulty in collecting quantitative validation data. Additionally, many state-of-the-art deep learning methods are limited to 2-D image data, which is insufficient for 3-D digital outcrops, such as hyperclouds. To address these challenges, we present Tinto, a multisensor benchmark digital outcrop dataset designed to facilitate the development and validation of deep learning approaches for geological mapping, especially for nonstructured 3-D data like point clouds. Tinto comprises two complementary sets: 1) a real digital outcrop model from Corta Atalaya (Spain), with spectral attributes and ground-truth data and 2) a synthetic twin that uses latent features in the original datasets to reconstruct realistic spectral data (including sensor noise and processing artifacts) from the ground truth. The point cloud is dense and contains 3242964 labeled points. We used these datasets to explore the abilities of different deep learning approaches for automated geological mapping. By making Tinto publicly available, we hope to foster the development and adaptation of new deep learning tools for 3-D applications in Earth sciences. The dataset can be accessed through this link: https://doi.org/10.14278/rodare.2256 .

Synthetic data is being used lately for training deep neural networks in computer vision applications such as object detection, object segmentation and 6D object pose estimation. Domain randomization hereby plays an important role in reducing the simulation to reality gap. However, this generalization might not be effective in specialized domains like a production environment involving complex assemblies. Either the individual parts, trained with synthetic images, are integrated in much larger assemblies making them indistinguishable from their counterparts and result in false positives or are partially occluded just enough to give rise to false negatives. Domain knowledge is vital in these cases and if conceived effectively while generating synthetic data, can show a considerable improvement in bridging the simulation to reality gap. This paper focuses on synthetic data generation procedures for parts and assemblies used in a production environment. The basic procedures for synthetic data generation and their various combinations are evaluated and compared on images captured in a production environment, where results show up to 15% improvement using combinations of basic procedures. Reducing the simulation to reality gap in this way can aid to utilize the true potential of robot assisted production using artificial intelligence.

Nowadays, Machine Learning (ML) systems are widely used in various businesses and are increasingly being adopted to make decisions that can significantly impact people’s lives. However, these decision-making systems rely on data-driven learning, which poses a risk of propagating the bias embedded in the data. Despite various attempts by the algorithmic fairness community to outline different types of bias in data and algorithms, there is still a limited understanding of how these biases relate to the fairness of ML-based decision-making systems. In addition, efforts to mitigate bias and unfairness are often agnostic to the specific type(s) of bias present in the data. This paper explores the nature of fundamental types of bias, discussing their relationship to moral and technical frameworks. To prevent harmful consequences, it is essential to comprehend how and where bias is introduced throughout the entire modelling pipeline and possibly how to mitigate it. Our primary contribution is a framework for generating synthetic datasets with different forms of biases. We use our proposed synthetic data generator to perform experiments on different scenarios to showcase the interconnection between biases and their effect on performance and fairness evaluations. Furthermore, we provide initial insights into mitigating specific types of bias through post-processing techniques. The implementation of the synthetic data generator and experiments can be found at https://github.com/rcrupiISP/BiasOnDemand.

This study investigates the influence of synthetic data on the accuracy of 6D pose estimation in RGB images compared to RGB-Depth image-based methods. Additionally, it aims to examine how this performance varies across different types of small chess pieces during a picking task with a robotic arm. The methodology involves 3D scanning the chess pieces and generating a dataset of 61,910 synthetic images with diverse domain randomization. Using this synthetic dataset, six unique 6D pose estimation models were trained for each type of chess piece. The models were validated using a real-world 6D pose estimation dataset, and the obtained results were compared with those from the estimations on a synthetic dataset. It was observed that synthetic data can be used for bridging the visual simulation to reality gap. However there is superior performance on synthetic data compared to real-world data. This implies that results obtained in a synthetic environment cannot be directly projected to real-world scenarios. Also a noticeable decrease in accuracy in both object detection and 6D pose estimation was observed with respect to camera distance for real-world images, primarily linked to the reduced size of chess pieces in the images. Notably, each chess piece exhibited improvement after depth refinement, with accuracy closely linked to the performance of the depth camera. A picking experiment was also conducted, revealing that models relying solely on RGB data achieved a positive grasping rate of approximately 52%, while RGB-Depth-based methods reached around 70%. The findings underscore the potential for successfully picking chess pieces with and without depth refinement, emphasizing the feasibility of bridging the visual simulation to reality gap. Additionally, the study suggests several avenues for future research, including comparing synthetic data with real-world data, further exploring the training process, and introducing domain randomization in the picking experiment, su, Mechanical Engineering | Vehicle Engineering | Cognitive Robotics

This study addresses the issue of black-grass, a herbicide-resistant weed that threatens wheat yields in Western Europe, through the use of high- resolution Unmanned Aerial Vehicles (UAVs) and synthetic data augmentation in precision agriculture. We mitigate challenges such as the need for large labeled datasets and environmental variability by employing synthetic data augmentations in training a Mask R-CNN model. Using a minimal dataset of 43 black-grass and 12 wheat field images, we achieved a 37% increase in Area Under the Curve (AUC) over the non-augmented baseline, with scaling as the most effective augmentation. The best model attained a recall of 53% at a precision of 64%, offering a promising approach for future precision agriculture applications. © NLDL 2024. All rights reserved.

Was sind Diffusionsmodelle? Wie funktionieren sie? Können sie basierend auf Text Prompts Bilder generieren die zum Trainieren von Klassifizierungsalgorithmen genutzt werden? Dies sind die Hauptfragen, die ich in dieser Arbeit zu beantworten versuchen werde. Ich werde versuchen, verschiedene Bilder von statistischen Diagrammen, die aus der Open-Source-Bibliothek Wikimedia Commons stammen, zu klassfizieren. Insgesamt enthält diese Bibliothek fast 100 Millionen Medienelemente. Eine laissez - faire Politik bezüglich der Erstellung von Kategorien und dem Hochladen von Bildern, hat zu vielen unbeschrifteten oder falsch klassifizierten Bildern sowie zum Teil sehr kleinen Nischenklassen geführt. Ich werde diese Probleme angehen, indem ich Bildklassifizierungsalgorithmen auf bereinigten Datensätzen von Instanzen ausgewählter Kategorien trainiere. Ich werde daraufhin die Trainings- und Validierungssets mit variierend vielen generitren Bildern erweitern und untersuchen wie sich verschiedene Performanzmetriken in Abhängigkeit von der Anzahl bzw. dem Anteil der generierten Bilddaten verändern. Ich werde einige Ergebnisse präsentieren. Zwei Haupterkenntnisse sind, dass die generative Dateierweiterung sehr gut für konzeptionell unterschiedliche Kategorien funktioniert und bei ähnlicheren Klassen weniger gute Ergebnisse liefert., What are diffusion models? How do they work? Can they improve image classification algorithms? These are the main questions I will try to answer in this thesis. I will try to label different images of statistical charts that originate from the open-source library Wikimedia Commons. In total, this library contains almost 100 million media elements. A loose policy regarding the creation of categories and the uploading of images has led to a lot of unlabeled or falsely classified images. Another challenge when labeling the images on Wikimedia Commons is that some niche classes are very small. I will tackle these problems by training image classification algorithms on cleaned data sets of selected categories. I will enhance the training and validation sets with different amounts of artificial images. While there are many exciting results, the most important one is that generative data augmentation seems to work well for major categories but not that well for niche categories, which are similar conceptually.

In the manufacturing industry, automatic quality inspections can lead to improved product quality and productivity. Deep learning-based computer vision technologies, with their superior performance in many applications, can be a possible solution for automatic quality inspections. However, collecting a large amount of annotated training data for deep learning is expensive and time-consuming, especially for processes involving various products and human activities such as assembly. To address this challenge, we propose a method for automated assembly quality inspection using synthetic data generated from computer-aided design (CAD) models. The method involves two steps: automatic data generation and model implementation. In the first step, we generate synthetic data in two formats: two-dimensional (2D) images and three-dimensional (3D) point clouds. In the second step, we apply different state-of-the-art deep learning approaches to the data for quality inspection, including unsupervised domain adaptation, i.e., a method of adapting models across different data distributions, and transfer learning, which transfers knowledge between related tasks. We evaluate the methods in a case study of pedal car front-wheel assembly quality inspection to identify the possible optimal approach for assembly quality inspection. Our results show that the method using Transfer Learning on 2D synthetic images achieves superior performance compared with others. Specifically, it attained 95% accuracy through fine-tuning with only five annotated real images per class. With promising results, our method may be suggested for other similar quality inspection use cases. By utilizing synthetic CAD data, our method reduces the need for manual data collection and annotation. Furthermore, our method performs well on test data with different backgrounds, making it suitable for different manufacturing environments., CC BY 4.0 DEED© The Author(s) 2024Correspondence Address: X. Zhu; Scania CV AB (publ), Södertälje, 151 87, Sweden; email: xiazhu@kth.se; CODEN: JIMNEOpen access funding provided by Royal Institute of Technology. This work is partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.

The fashion industry is a significant contributor to greenhouse gas emissions and textile waste, prompting the need for sustainable practices. This thesis explores the use of diffusion models for generating synthetic data to enhance datasets used in machine learning, specifically focusing on second-hand fashion. Diffusion models, known for their ability to create high-quality images, offer potential solutions to the imbalance and quality issues in existing datasets. The study investigates how image generation and editing through diffusion models can improve datasets, the effectiveness of different prompting strategies, and the performance of synthetic data in machine learning models compared to real data. The methodology involves using the Kandinsky 2.2 inpainting model to generate and edit images, followed by manual and automated classification to evaluate image quality. Experiments demonstrate that diffusion models can plausibly improve dataset quality by adding and removing damage in images, although fully automating this process remains challenging. The results indicate that augmenting the datasets with synthetic images can potentially enhance the performance of the model, although the variability of the results suggests the need for further research. This thesis contributes to the field of sustainable fashion by proposing innovative methods for dataset augmentation using state-of-the-art generative models, aiming to support the development of efficient and automated sorting processes in the textile industry.

The development of high-performing object detection models requires extensive and varied datasets with accurately annotated images, a process that is traditionally labor-intensive and prone to errors. To address these challenges, this report explores the generation of synthetic data using domain randomization techniques to train object detection models. We present a pipeline that integrates synthetic data creation in Unity, and the training of YOLOv8 object detection models. Our approach uses the Unity Perception package to produce diverse and precisely annotated datasets, overcoming the domain gap typically associated with synthetic data. The pipeline was evaluated through a series of experiments, analyzing the impact of various parameters such as background textures, and training arguments on model performance. The results demonstrate that models trained with our synthetic data can achieve high accuracy and generalize well to real-world scenarios, offering a scalable and efficient alternative to manual data annotation.

The development of new solutions incorporating artificial intelligence (AI) within the medical field is an area of great interest. However, access to comprehensive and diverse datasets is restricted due to the sensitive nature of the data. A potential solution to this is to generatesynthetic datasets based on real medical data. Synthetic data could protect the integrity of the subjects while preserving the inherent information necessary for training AI models and be generated in greater quantity than otherwise available. This thesis project aims to generate reliable clinical trial subject data using a generative adversarial network (GAN). The main data set used is a mock clinical trial dataset consisting of multiple subject visits, however an additional data set containing authentic medical data is also used for better insights into the model’s ability to learn underlying relationships. The thesis also investigates training strategies for simulating the temporal dimension and the missing values in the data. The GAN model used is an altered version of the Conditional Tabular GAN (CTGAN)made to be compatible with the preprocessed clinical trial mock data, and multiple model architectures and number of training epochs are examined. The results show great potential for GAN models on clinical trial datasets, especially for real-life data. One model, trained on the authentic dataset, generates near-perfect synthetic data with respect to column distributions and correlation between columns. The results also show that classification models trained on synthetic data and tested on real data have the potential to match the performance of classification models trained on real data. While the synthetic data replicates the missing values, no definitive conclusion can be drawn regarding the temporal characteristics due to the sparsity of the mock dataset and lack of real correlations in it. Although the results are promising, further experiments on authentic datasets with less spa

Modern deep learning models require large amounts of data to train, and the acquisition of data can be challenging. Synthetic data provides an alternative to manually collecting real data, alleviating problems associated with real data acquisition. For recycling processes, classifying metal scrap piles containing hazardous objects is important, where hazardous objects can be damaging and costly if handled incorrectly. Automatically detecting hazardous objects in metal scrap piles using image classification models requires large amounts of data, and metal scrap piles contain large variations in objects, textures, and lighting. Furthermore, data acquisition can be challenging in the recycling domain, where positive objects can be scarce and manual acquisition setup can be challenging. In this thesis, synthetic images of metal scrap piles in a recycling process are created, intended for training image classification models to detect metal scrap piles containing fire extinguishers or hydraulic cylinders. Synthetic images are created with physically based rendering and domain randomization, rendered with either rasterization or ray tracing engines. Ablation studies are conducted to investigate the effect of using domain randomization. The performance of models trained on purely synthetic datasets is compared to models trained on datasets containing only real images. Furthermore, photorealistic rendering with ray tracing rendering is evaluated by comparing F1 scores between models trained on data sets created with rasterization or ray tracing. The F1 scores show that models trained on purely synthetic data outperform those trained solely on real data when classifying images containing fire extinguishers or hydraulic cylinders. Ablation studies show that domain randomization of textures is beneficial both for the classification of fire extinguishers and for the classification of hydraulic cylinders in metal scrap piles. High dynamic range image lighting randomization does

This thesis investigates the effectiveness of Tabular Variational Auto Encoders (TVAEs) in generating high-quality synthetic tabular data and assesses their compliance with differential privacy principles. The study shows that while TVAEs are better than VAEs at generating synthetic data that faithfully reproduces the distribution of real data as measured by the Synthetic Data Vault (SDV) metrics, the latter does not guarantee that the synthetic data is up to the task in practical industrial applications. In particular, models trained on TVAE-generated data from the Creditcards dataset are ineffective. The author also explores various optimisation methods on TVAE, such as Gumbel Max Trick, Drop Out (DO) and Batch Normalization, while pointing out that techniques frequently used to improve two-dimensional TVAE, such as Kullback–Leibler Warm-Up and B Disentanglement, are not directly transferable to the one-dimensional context. However, differential privacy to TVAE was not implemented due to time constraints and inconclusive results. The study nevertheless highlights the benefits of stabilising training with the Differential Privacy - Stochastic Gradient Descent (DP-SGD), as with a dropout, and the existence of an optimal equilibrium point between the constraints of differential privacy and the number of training epochs in the model., Denna avhandling undersöker hur effektiva Tabular Variational AutoEncoders (TVAE) är när det gäller att generera högkvalitativa syntetiska tabelldata och utvärderar deras överensstämmelse med differentierade integritetsprinciper. Studien visar att även om TVAE är bättre än VAE på att generera syntetiska data som troget återger fördelningen av verkliga data mätt med Synthetic Data Vault (SDV), garanterar det senare inte att de syntetiska data är upp till uppgiften i praktiska industriella tillämpningar. I synnerhet är modeller som tränats på TVAE-genererade data från Creditcards-datasetet ineffektiva. Författaren undersöker också olika optimeringsmetoder för TVAE, såsom Gumbel Max Trick, DO och Batch Normalization, samtidigt som han påpekar att tekniker som ofta används för att förbättra tvådimensionell TVAE, såsom Kullback-Leibler Warm-Up och B Disentanglement, inte är direkt överförbara till det endimensionella sammanhanget. På grund av tidsbegränsningar och redan ofullständiga resultat implementerades dock inte differentierad integritet för TVAE. Studien belyser ändå fördelarna med att stabilisera träningen med Differential Privacy - Stochastic Gradient Descent (DP-SGD), som med en drop-out, och förekomsten av en optimal jämviktspunkt mellan begränsningarna för differential privacy och antalet träningsepoker i modellen.

In the manufacturing industry, automatic quality inspections can lead to improved product quality and productivity. Deep learning-based computer vision technologies, with their superior performance in many applications, can be a possible solution for automatic quality inspections. However, collecting a large amount of annotated training data for deep learning is expensive and time-consuming, especially for processes involving various products and human activities such as assembly. To address this challenge, we propose a method for automated assembly quality inspection using synthetic data generated from computer-aided design (CAD) models. The method involves two steps: automatic data generation and model implementation. In the first step, we generate synthetic data in two formats: two-dimensional (2D) images and three-dimensional (3D) point clouds. In the second step, we apply different state-of-the-art deep learning approaches to the data for quality inspection, including unsupervised domain adaptation, i.e., a method of adapting models across different data distributions, and transfer learning, which transfers knowledge between related tasks. We evaluate the methods in a case study of pedal car front-wheel assembly quality inspection to identify the possible optimal approach for assembly quality inspection. Our results show that the method using Transfer Learning on 2D synthetic images achieves superior performance compared with others. Specifically, it attained 95% accuracy through fine-tuning with only five annotated real images per class. With promising results, our method may be suggested for other similar quality inspection use cases. By utilizing synthetic CAD data, our method reduces the need for manual data collection and annotation. Furthermore, our method performs well on test data with different backgrounds, making it suitable for different manufacturing environments., CC BY 4.0 DEED© The Author(s) 2024Correspondence Address: X. Zhu; Scania CV AB (publ), Södertälje, 151 87, Sweden; email: xiazhu@kth.se; CODEN: JIMNEOpen access funding provided by Royal Institute of Technology. This work is partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.

This study addresses the issue of black-grass, a herbicide-resistant weed that threatens wheat yields in Western Europe, through the use of high- resolution Unmanned Aerial Vehicles (UAVs) and synthetic data augmentation in precision agriculture. We mitigate challenges such as the need for large labeled datasets and environmental variability by employing synthetic data augmentations in training a Mask R-CNN model. Using a minimal dataset of 43 black-grass and 12 wheat field images, we achieved a 37% increase in Area Under the Curve (AUC) over the non-augmented baseline, with scaling as the most effective augmentation. The best model attained a recall of 53% at a precision of 64%, offering a promising approach for future precision agriculture applications. © NLDL 2024. All rights reserved.

This study addresses the issue of black-grass, a herbicide-resistant weed that threatens wheat yields in Western Europe, through the use of high- resolution Unmanned Aerial Vehicles (UAVs) and synthetic data augmentation in precision agriculture. We mitigate challenges such as the need for large labeled datasets and environmental variability by employing synthetic data augmentations in training a Mask R-CNN model. Using a minimal dataset of 43 black-grass and 12 wheat field images, we achieved a 37% increase in Area Under the Curve (AUC) over the non-augmented baseline, with scaling as the most effective augmentation. The best model attained a recall of 53% at a precision of 64%, offering a promising approach for future precision agriculture applications. © NLDL 2024. All rights reserved.