APRÈN. Portal de la producció docent del professorat de la UPC

This thesis investigates the feasibility and potential of visual neural decoding across different neuroimaging modalities, with a particular focus on adapting magnetoencephalography (MEG) based approaches to electroencephalography (EEG) data. Building upon recent advancements in the field, the study validates and extends a crosssubject visual neural decoding approach, exploring its reproducibility, performance, and adaptability from MEG to EEG. The research demonstrates, for the first time, that neural visual decoding is feasible on EEG data using cross-subject training models. This finding represents a significant step towards more accessible and practical brain-computer interfaces, as EEG is a more widely available and cost-effective neuroimaging technique compared to MEG or fMRI. Through comprehensive experiments across multiple datasets, the study highlights the critical role of stimulus duration in EEG-based visual decoding. The results underscore the importance of EEG dataset design for visual neural decoding, revealing that longer stimuli presentations significantly enhance decoding performance. This insight provides valuable guidance for future EEG studies in this field. While challenges were encountered in fully replicating state-of-the-art MEG results, the models consistently performed above chance level on both MEG and EEG data. The thesis provides a detailed analysis of the performance differences between single-subject and cross-subject models, as well as a comparison of various image models across modalities.

This thesis investigates the intersection of generative artificial intelligence and music creation, with a focus on drum sample synthesis. It begins with a comprehensive literature review that examines various music representation techniques and traces the evolution of music generation methods, from early algorithmic composition to contemporary deep learning approaches. The core of the research involves the design, implementation, and evaluation of two generative models: a Convolutional Autoencoder (CAE) and a Convolutional Variational Autoencoder (CVAE). These models are developed with the primary aim of synthesizing high-quality drum samples from existing audio data. The study compares the performance of these architectures across different datasets, analyzing their capabilities in sample reconstruction, latent space representation, and novel sound generation. By exploring the strengths and limitations of each approach, this research contributes to the growing field of AI-driven music production tools and provides insights into the potential of deep learning for creative audio synthesis.

In the presented master's thesis, a detailed examination of the relationship between the dominant color on the cover of a musical album and the music genre was conducted, using machine learning methods. The aim of the study was to identify potential correlations and understand how visual aspects of covers can affect the perception and classification of music. In this paper it's being evaluated if this information ¿ and more specifically the color information ¿ is sufficient to differentiate music genres. The theoretical part of the thesis presents an extensive review of literature and databases on color analysis in the context of graphics on music album covers. Various methods of color feature extraction were discussed, taking into account their advantages and limitations. Particular attention was paid to color theory, analyzing its basic concepts and the relationships between colors, emotions, and perception. The phenomenon of synesthesia, which involves perceiving music through colors, was also examined, as it may be relevant in analyzing album covers. Preparations for the practical part began with the design and implementation of a dataset, essential for conducting the analysis. Challenges related to data collection and processing were discussed, including issues of quality, representativeness, and methods of preparing them to achieve the set goals. The practical part of the thesis focused on the process of extracting the dominant color from album covers and analyzing these data. Various image processing and machine learning techniques were used to create models capable of classifying music genres based on color analysis. A series of experiments were conducted to determine whether and to what extent the dominant color on a cover can be an indicator of the music genre, and the results were carefully analyzed. The results obtained from the model were confronted with the responses obtained in a survey conducted among respondents.

Wildfires are increasingly frequent and severe globally, threatening ecosystems, human lives, and property. Accurate prediction of wildfire spread is essential for effective mitigation and management. In this thesis, we develop a novel dataset using data from VIIRS and GOES satellite instruments to analyze the physical dynamics of wildfires in California. By sampling along active fire lines, we obtain detailed information that has not been extensively explored in existing literature. We apply both time series methods and spatio-temporal modeling techniques to understand and predict wildfire behavior. Our findings reveal that traditional statistical models, particularly the ARIMA model, outperform neural network architectures in predicting burned fire areas using historical data. When incorporating additional environmental covariates, ensemble methods like Extreme Gradient Boosting (XGBoost) achieve superior performance, highlighting the significance of factors such as Vapor Pressure Deficit over operational elements like suppression personnel. Although spatio-temporal models like Cellular Automata and ConvLSTM show potential in simulations, they face challenges in real-world applications due to complexity and data limitations. This research underscores the importance of enhancing data quality and developing tailored models to improve wildfire prediction, ultimately contributing to better wildfire management and risk mitigation.

As the feasibility of autonomous vehicles becoming commonplace in our world increases, effective methods to model and simulate their implementations become imperative. Furthermore, the reality of self-driving vehicles acting and coexisting gives rise to numerous ethical dilemmas. This work explores the potential of leveraging CARLA, an open-source driving simulator for autonomous vehicles in urban environments, to simulate ethical dilemmas that could occur in real-world self-driving car scenarios. More specifically, this work seeks to fill in an apparent gap in scientific research by investigating the potential to simulate more likely and pragmatic ethical trade-offs, as opposed to morally-focused ones that dominate the academic literature and public eye, such as the Trolley Problem. The Overview, Design concepts, Details (ODD) Protocol is used to model the simulation of an ethical trade-off in an easily understandable and repeatable way. The current state of Autonomous Vehicles, ethical concerns surrounding them, and leading simulation technology are detailed. Then, a highway exit scenario is created within the CARLA simulator, with the goal of facilitating an ambulance's safe and efficient exit. The vehicle-to-vehicle communication levels of agents are modified, and the resultant changes to the emergency vehicle's ability to achieve the goal of exiting the highway are analyzed with the trade-off of privacy versus efficiency and/or safety in mind. Finally, the results of this study are used to advocate for further research in this area, contending that ethical issues such as privacy in the industry of Autonomous Vehicles will have widespread implications on its practical implementation, legislation, and public adoption. While worst-case scenarios involving moral dilemmas are interesting and important to consider, this thesis argues that ethical dilemmas affecting users daily require further attention and research in academia sooner rather than later.

This thesis presents an investigation into the optimization of Graph Neural Network (GNN) architectures using the Graph Neural Architecture Search (GraphNAS) framework, an automated system that employs reinforcement learning techniques. The primary goal of this research is to enhance the performance of GNNs by systematically discovering and evaluating potential architectures tailored for specific types of graph data, such as those found in social networks, communication systems, and chemical structures. We initiated our study by evaluating various architecture search libraries and selected GraphNAS for its promising initial results and compatibility with our project objectives. Our methodology involved configuring GraphNAS to generate potential GNN architectures, beginning with basic Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT), and later expanding the search space to include diverse layer types and activation functions. We introduced a ”None” layer to investigate the potential for more sparse and efficient network structures and adjusted our strategy based on the initial performance insights, focusing on optimizing smaller, more efficient networks. Our results demonstrate that GraphNAS can effectively identify high-performing GNN architectures, with most configurations achieving remarkable accuracy levels above 0.9. We also explored the challenges of building a regressor to predict the performance of these architectures using a synthetically generated dataset, addressing issues related to data imbalance and model underfitting. This work contributes to the field of machine learning by demonstrating the efficacy of automated architecture search in optimizing GNNs, offering insights into the complexities of designing neural networks for graph-based data, and suggesting directions for future research to further improve the automation and efficiency of GNN design.

Machine Translation (MT), is one of the most important tasks related to natural language processing which facilitates seamless communication across different languages by automating the translation process, thereby significantly reducing communication costs. In this context, a multilingual neural machine translation (MNMT) system emerges as a cost-effective solution, capable of handling translations across multiple language pairs. An interesting property of MNMT systems is their ability to perform zero-shot translation, where the model translates between language pairs it has not explicitly seen before. This feature is particularly promising as it could potentially reduce the extensive resources and parameters typically required for MNMT models. Traditionally, MNMT models were thought to utilize a representation space independent of any specific language. This interlingua representation would allow the model to encode phrases conveying the same meaning similarly, regardless of the language, thereby aiding the translation process. However, in practice, it has been shown that this interlingua representation space is not as robust as previously assumed. In this project, we aim to study how MNMT representations still exhibit language dependent information and devise techniques to improve the model's representation. To achieve this goal, we propose some methods based on language projections.

The proliferation of Internet of Things (IoT) devices has significantly increased data traffic, necessitating robust security measures to protect against latent threats. Traditional anomaly detection methods often struggle to keep pace with the dynamic and diverse nature of IoT environments, particularly in use cases with limited accessible training data. This necessitates adaptive and efficient solutions. In this work, we propose the use of fine-tuned Transformer models for anomaly detection in an IoT traffic use-case. These techniques allow adjusting a pre-trained model to a new domain where scarce training data is accessible. Specifically, we propose a framework in which we explore the usage of fine-tuning Transformer architectures using the CIC IoMT 2024 dataset and evaluate it with the Aposemat IoT-23 dataset. The significance of the results lies in demonstrating that training on one dataset and applying the knowledge to another is feasible, though it involves aligning certain specifications between datasets and employing techniques to adapt to new knowledge. This research bridges the gap in developing new Transformer-based architectures capable of providing supervised fraud detection capabilities, even with highly limited datasets for training. This process, including both successful and unsuccessful methods, is described in this thesis.

This thesis presents the development and integration of a conversational AI within Non-Player Characters (NPCs) in video games. The main objective is to elevate player-NPC dialogues, making them more dynamic and reflective of human-like interactions. Utilizing OpenAI's GPT-4 API, the project pushes the boundaries of NPC interaction by enabling complex conversation and emotional engagement throughout the conversation. A comprehensive dataset was constructed, including environment descriptions, game context, and a broad range of human interactions such as humor, empathy, and problem-solving. These elements inform the AI's dialogue generation process, enhancing the NPCs' ability to provide tailored and context-aware responses. The technical implementation integrates these advanced conversational capabilities within the gaming environment, ensuring seamless interactions without disrupting the game's narrative flow. Initial testing phases focused on refining the AI's response mechanisms and reducing latency to improve real-time interaction quality. Subsequent evaluations involved user surveys to gauge the AI's impact on game experience and narrative engagement. Findings indicate a significant improvement in user satisfaction, confirming the potential of advanced conversational AIs to change NPC interactions in video games. The findings of this research not only showcase the potential enhancements to gameplay but also provide a framework for future advancements in the application of AI within interactive digital storytelling and virtual character development.

This study aims to advance the application of Space Syntax by integrating it with Agent-Based System (ABS) features within the NetLogo environment, building upon the foundational work done using UCL DepthMap X. Initially, the research replicates Alasdair Turner's early 2000s experiments at the Tate Britain to validate the results within this new framework, establishing a baseline using the original DepthMap outputs from Turner ad Penn. The methodology progresses by constructing a parallel simulation framework in NetLogo, initially confirming functional equivalence through integration diagrams before introducing key ABS elements-namely agent communication and dynamic attractors in the form of spatial pins in a museum setting. These features are hypothesized to enrich the simulation of spatial dynamics by facilitating more nuanced interactions and emergent behaviors, potentially providing deeper insights into human spatial behavior. A comparative analysis will then be conducted, focusing on the variance in agent behaviors with and without the new ABS features. This will involve mimicking Turner's original analytic approaches to generate path diagrams and aiming to achieve correlations like the R2 correlation of 0.76 reported in Turner and Penn's 2001 study, which validated their DepthMap implementation against actual visitor movements. Findings will show the proposed NETLogo framework is validated by the same correlation, 0.89, Turner obtained in 2007 along his research on Through Vision mathematical definition. First implemented proper Agent-based-system features, message passing and goals/attractors, will show a clear influence on agents behavior as a priori guessed by Turner in 2011 along his evolved automaton proposal. The study, while not addressing Turner's subsequent proposals on fitness functions and goal validations due to their unfinished guidelines, sets the stage for further research. It proposes exploring these aspects as logical next steps in enhancing the NetLogo-based space syntax framework. This research not only seeks to validate the integration of ABS features into space syntax studies but also aims to broaden the methodological toolkit available for urban and architectural analysis, paving the way for more dynamic and comprehensive modeling techniques in the field.

This thesis explores the analysis of human disease comorbidity through the lens of network science. Recognizing the significance of comorbidities in human health, we emphasize the need to approach diseases as complex and interconnected entities rather than isolated conditions. This paradigm shift underscores the value of leveraging network science and artificial intelligence tools to facilitate research and analysis in this field. As medicine transitions towards a patient-centered perspective, patients are viewed as dynamic systems influenced by environmental factors, genetic information, and molecular interactions. Imbalances within these intricate relationships can contribute to disease development and subsequent comorbidities, highlighting the interconnected nature of health conditions. Our study focuses on analyzing comorbidity patterns using data extracted from the MIMIC-III dataset, encompassing patients treated at Beth Israel Deaconess Medical Center critical care unit between 2001 and 2012. We construct a bipartite network from this dataset and apply various network science techniques, including the examination of degree distribution, clustering coefficients, node centrality, and community structure analysis. To delve deeper into the community structure of the network, we employ inference methods to derive statistically and mathematically grounded insights into how the network is structured from a community perspective. This approach contributes to advancing the field by demonstrating the potential of AI techniques, particularly network science, in medical research. Overall, this thesis serves as a modest contribution to the field, showcasing the transformative impact of AI methodologies like network science on understanding disease comorbidities and fostering innovative approaches in medical research.

This thesis aims to evaluate the utility of quantitative biomarkers derived from optical coherence tomography angiography (OCTA) images for classifying diabetic retinopathy (DR) severity in patients with Type 1 diabetes mellitus (DM). An intensive image transformation process was implemented. We first applied two distinct vessel segmentation approaches, Fiji ImageJ for large vessels and Python libraries for comprehensive segmentation, to extract retinal vasculature from OCTA images of the superficial and deep capillary plexuses. A wide range of quantitative vascular biomarkers, including tortuosity metrics, foveal avascular zone measurements, vessel density, and novel geometric and morphological metrics, were then calculated to characterize retinal vascular changes associated with DR progression. Binary classification tasks were formulated to detect the presence of DM, discriminate DR stages, and identify early DR development. Feature selection techniques and robust machine learning models were employed to classify DR severity and evaluate the diagnostic performance of the OCTA-derived biomarkers. The Fiji ImageJ segmentation approach, focusing on large vessels, achieved higher accuracies across all tasks compared to the comprehensive segmentation approach. In feature transformation, the combination of multiple biomarkers consistently outperformed using only tortuosity metrics. The most informative biomarkers identified include vessel perimeter, parafoveal vessel density, and average tortuosity index. The strengths of the thesis lie in the large Type 1 DM cohort, the application of two segmentation approaches, and the extensive evaluation of a wide range of biomarkers. Limitations include potential limited generalizability to Type 2 DM, the influence of image quality on biomarker extraction and dataset imbalance. In conclusion, this study demonstrates the potential of OCTA-derived quantitative biomarkers in classifying DR severity and highlights the importance of a comprehensive assessment of retinal vascular health. The findings contribute to advancing early DR detection and management strategies, aiming to improve clinical outcomes for patients with Type 1 DM.