Search

La irrupció de la intel·ligència artificial (IA) generativa ha obert un nou escenari en el qual s’ha de replantejar el model educatiu actual. Aquests models d’IA utilitzen algoritmes d'aprenentatge automàtic per generar dades noves a partir d'un conjunt d'exemples d'entrenament. Una de les aplicacions que poden fer ús d’aquestes IA generatives són els bots conversacionals o xatbots. Aquests programes informàtics simulen una conversa a través de missatges de text. Per tal de respondre les peticions introduïdes fan servir intel·ligència artificial, que seguint arquitectures de dades complexes són capaces de simular entendre les preguntes i generar respostes coherents. Aquest tipus d’eines s’han popularitzat recentment a partir d’aplicacions com ChatGPT i tenen el potencial d’ajudar als alumnes com a assistents d’aprenentatge. L’assignatura de primer curs, Expressió Gràfica, s’imparteix pel professorat del departament d’Enginyeria Gràfica i de Disseny i segons la guia docent, la competència transversal que es treballa a l’assignatura és l’autoaprenentatge. L'objectiu principal de la comunicació és presentar el projecte d’innovació educativa desenvolupat dins de la convocatòria de Galàxia Aprenentatge 2023. Aquest projecte pretén reforçar l’autoaprenentatge de la teoria de l’assignatura amb un bot conversacional, que ajudarà als alumnes durant el seu estudi. Com a objectiu secundari es mostraran els detalls de l’experiment dissenyat per tal d’avaluar la satisfacció de l’alumnat en l’assignatura

The extensive use of HPC infrastructures and frameworks for running data-intensive applications has led to a growing interest in data partitioning techniques and strategies. In fact, application performance can be heavily affected by how data are partitioned, which in turn depends on the selected size for data blocks, i.e. the block size. Therefore, finding an effective partitioning, i.e. a suitable block size, is a key strategy to speed-up parallel data-intensive applications and increase scalability. This paper describes a methodology, namely BLEST-ML (BLock size ESTimation through Machine Learning), for block size estimation that relies on supervised machine learning techniques. The proposed methodology was evaluated by designing an implementation tailored to dislib, a distributed computing library highly focused on machine learning algorithms built on top of the PyCOMPSs framework. We assessed the effectiveness of the provided implementation through an extensive experimental evaluation considering different algorithms from dislib, datasets, and infrastructures, including the MareNostrum 4 supercomputer. The results we obtained show the ability of BLEST-ML to efficiently determine a suitable way to split a given dataset, thus providing a proof of its applicability to enable the efficient execution of data-parallel applications in high performance environments., This work has been partially supported by the European Commission through the Horizon 2020 Research and Innovation program and the EuroHPC JU under contract 955558 (eFlows4HPC project) and by MCIN/AEI/10.13039/501100011033 and the European Union NextGenerationEU/PRTR (PCI2021‑121957 and CEX2021‑ 001148‑S) and by the Spanish Government (PID2019‑107255GB), and Generalitat de Catalunya (contract 2021‑SGR‑00412). We also acknowledge financial support from “National Centre for HPC, Big Data and Quantum Computing”, CN00000013 ‑ CUP H23C22000360005, and from “FAIR ‑ Future Artificial Intelligence Research” project ‑ CUP H23C22000860006., Peer Reviewed, Postprint (published version)

Machine Learning is entering a phase of maturity, but its medical applications still lag behind in terms of practical use. The field of oncological radiology (and neuro-oncology in particular) is at the forefront of these developments, now boosted by the success of Deep-Learning methods for the analysis of medical images. This paper reviews in detail some of the most recent advances in the use of Deep Learning in this field, from the broader topic of the development of Machine-Learning-based analytical pipelines to specific instantiations of the use of Deep Learning in neuro-oncology; the latter including its use in the groundbreaking field of ultra-low field magnetic resonance imaging., This research was funded by H2020-EU.1.3.—EXCELLENT SCIENCE—Marie Skłodowska-Curie Actions, grant number H2020-MSCA-ITN-2018-813120; Proyectos de investigación en salud 2020, grant number PI20/00064. PID2019-104551RB-I00; Centro de Investigación Biomédica en Red en Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN (http://www.ciber-bbn.es/en, accessed on 3 November 2023), CB06/01/0010), an initiative of the Instituto de Salud Carlos III (Spain) co-funded by EU Fondo Europeo de Desarrollo Regional (FEDER); Spanish Agencia Española de Investigación (AEI) PID2022-143299OB-I00 grant; XartecSalut 2021-XARDI-00021. Carla Pitarch is a fellow of Eurecat’s “Vicente López” Ph.D. grant program., Peer Reviewed, Postprint (published version)

The risk factors for post-COVID-19 cognitive impairment have been poorly described. This study aimed to identify the sociodemographic, clinical, and lifestyle characteristics that characterize a group of post-COVID-19 condition (PCC) participants with neuropsychological impairment. The study sample included 426 participants with PCC who underwent a neurobehavioral evaluation. We selected seven mental speed processing and executive function variables to obtain a data-driven partition. Clustering algorithms were applied, including K-means, bisecting K-means, and Gaussian mixture models. Different machine learning algorithms were then used to obtain a classifier able to separate the two clusters according to the demographic, clinical, emotional, and lifestyle variables, including logistic regression with least absolute shrinkage and selection operator (LASSO) (L1) and Ridge (L2) regularization, support vector machines (linear/quadratic/radial basis function kernels), and decision tree ensembles (random forest/gradient boosting trees). All clustering quality measures were in agreement in detecting only two clusters in the data based solely on cognitive performance. A model with four variables (cognitive reserve, depressive symptoms, obesity, and change in work situation) obtained with logistic regression with LASSO regularization was able to classify between good and poor cognitive performers with an accuracy and a weighted averaged precision of 72%, a recall of 73%, and an area under the curve of 0.72. PCC individuals with a lower cognitive reserve, more depressive symptoms, obesity, and a change in employment status were at greater risk for poor performance on tasks requiring mental processing speed and executive function., Open Access funding provided thanks to the CRUE-CSIC agreement with Springer Nature. This research was supported by the Agency for Management of University and Research Grants (AGAUR) from the Generalitat de Catalunya (Pandemies, 2020PANDE00053), the La Marató de TV3 Foundation (202111–30-31–32), the Ministerio de Ciencia e Innovación (TED2021-130409B-C55)., Peer Reviewed, Membres del NAUTILUS-Project Collaborative Group: Jose A. Bernia, Servei d’Anestesia Reanimació i Clinica del Dolor, Consorci Sanitari de Terrassa (CST) (Terrassa, Barcelona, Spain). Vanesa Arauzo, Servei de Medicina Intensiva. Consorci Sanitari de Terrassa (CST) (Terrassa, Barcelona, Spain). Marta Balague Marmaña, Hospital Sant Joan Despí Moisès Broggi, Consorci Sanitari Integral (Sant Joan Despí, Spain). Pérez-Pellejero, Cristian, Hospital Sant Joan Despí Moisès Broggi, Consorci Sanitari Integral (Sant Joan Despí, Barcelona, Spain). Silvia Cañizares. Hospital Clinic de Barcelona (Barcelona, Spain). Jose Antonio Lopez Muñoz. Occupational Health Care Service, Hospital Clínic (Barcelona, Spain). Jesús Caballero, Hospital Universitari Arnau de Vilanova (Lleida, Spain). Anna Carnes-Vendrell, Hospital Universitari de Santa Maria (Lleida, Spain). Gerard Piñol-Ripoll, Hospital Universitari de Santa Maria (Lleida, Spain). Ester Gonzalez-Aguado, Consorci Sanitari Alt Penedès-Garraf (Vilafranca de Penedés, Barcelona, Spain). Mar Riera-Pagespetit, Consorci Sanitari Alt Penedès-Garraf (Vilafranca de Penedés, Barcelona, Spain). Eva Forcadell-Ferreres, Hospital Verge de la Cinta, (Tortosa, Tarragona, Spain). Silvia Reverte-Vilarroya, Hospital Verge de la Cinta, (Tortosa, Tarragona, Spain). Susanna Forné, Fundació Sant Hospital de la Seu d’Urgell (La Seu d’Urgell, Lleida, Spain). Jordina Muñoz-Padros, Consorci Hospitalari de Vic (Vic, Barcelona, Spain). Anna Bartes-Plan, Consorci Hospitalari de Vic (Vic, Barcelona, Spain). Jose A. Muñoz-Moreno, Servei de Malalties Infeccioses, Fundació Lluita contra les Infeccions—Hospital Universitari Germans Trias i Pujol (Badalona, Barcelona, Spain). Anna Prats-Paris, Servei de Malalties Infeccioses, Fundació Lluita contra les Infeccions—Hospital Universitari Germans Trias i Pujol (Badalona, Barcelona, Spain). Inmaculada Rico Pons, Hospital Universitari de Bellvitge (L’Hospitalet de Llobregat, Barcelona, Spain). Judit Martínez Molina, Hospital, Postprint (published version)

Este trabajo de final de grado aborda la problemática de la falta de acceso a información de primera mano sobre parques de atracciones, ya que las grandes cantidades de reseñas disponibles no siempre reflejan de manera clara la opinión de los clientes ni los motivos detrás de las puntuaciones otorgadas. La nota general de una empresa, no siempre es un reflejo directo de la opinión de los clientes sobre los distintos servicios que ofrece la misma. El objetivo principal de la investigación es desarrollar métodos para extraer los puntos clave de una empresa a partir de estas reseñas y poder ofrecerlos de una forma clara y sencilla a la empresa y los futuros clientes de la misma. El proceso implica una primera fase de clasificación de textos según su sentimiento, seguida de un análisis más detallado que identifica las quejas en las reseñas negativas, los temas discutidos en las reseñas neutrales y los aspectos elogiados por los clientes en las reseñas positivas. Este enfoque busca proporcionar una herramienta efectiva para la obtención de feedback por parte de los clientes, facilitando una comprensión más profunda de las experiencias y percepciones en los parques de atracciones., This final degree project addresses the problem of lack of access to first-hand information about amusement parks, since the large amounts of reviews available do not always clearly reflect the opinions of customers or the reasons behind the scores granted. The general rating of a company is not always a direct reflection of the customer's opinion on the different services it offers. The main objective of the research is to develop methods to extract the key points of a company from these reviews and to be able to offer them in a clear and simple way to the company and its future clients. The process involves a first phase of classifying texts according to sentiment, followed by a more detailed analysis that identifies complaints in negative reviews, topics discussed in neutral reviews, and aspects praised by customers in positive reviews. This approach seeks to provide an effective tool for obtaining feedback from customers, facilitating a deeper understanding of experiences and perceptions in amusement parks.

Aquest estudi realitza una exploració experimental en la combinació de tècniques d'enginyeria civil i aprenentatge automàtic amb l'objectiu de desenvolupar dissenys de paviment de formigó permeable sostenibles per mitigar els efectes del canvi climàtic. L'estudi utilitza xarxes neuronals artificials (XNA) per a la modelització predictiva per avaluar l'efecte de diferents característiques d'entrada sobre la resistència a la compressió i la permeabilitat a l'aigua del formigó permeable, que són ambdós paràmetres clau per a la durabilitat i el rendiment ambiental del formigó porós. Utilitzant un conjunt de dades limitat, aquest estudi adopta un enfocament de validació creuada k-fold i entrena nombroses configuracions de XNA per optimitzar la precisió i prevenir el sobreajustament. Es troba que la porositat i el tipus d'agregat gruixut tenen un efecte significatiu en el rendiment del paviment, la qual cosa guia els enginyers a optimitzar les formulacions de formigó segons els requeriments del projecte. Tot i enfrontar-se a limitacions del conjunt de dades i la "caixa negra" interna de les xarxes neuronals profundes, els models XNA obtinguts mostren una alta precisió predictiva, amb efectes potencials en el disseny d'infraestructures urbanes i l'adequació climàtica. Reflexionant sobre les limitacions d'aquest estudi i la possibilitat d'integrar dades més diverses i algorismes avançats en el futur, aquest estudi destaca el potencial de l'aprenentatge automàtic per jugar un paper clau en la millora de la sostenibilitat i l'eficiència de les solucions d'enginyeria civil, així com la seva contribució al canvi climàtic global i al desenvolupament urbà sostenible., Este estudio realiza una exploración experimental en la combinación de técnicas de ingeniería civil y aprendizaje automático con el objetivo de desarrollar diseños sostenibles de pavimentos de concreto permeable para mitigar los efectos del cambio climático. El estudio utiliza redes neuronales artificiales (RNA) para modelado predictivo con el fin de evaluar el efecto de diferentes características de entrada en la resistencia a la compresión y la permeabilidad al agua del concreto permeable, que son ambos parámetros clave para la durabilidad y el rendimiento ambiental del concreto poroso. Utilizando un conjunto de datos limitado, este estudio adopta un enfoque de validación cruzada k-fold y entrena numerosas configuraciones de RNA para optimizar la precisión y prevenir el sobreajuste. Se encuentra que la porosidad y el tipo de agregado grueso tienen un efecto significativo en el rendimiento del pavimento, lo que guía a los ingenieros a optimizar las formulaciones de concreto de acuerdo con los requisitos del proyecto. A pesar de enfrentarse a limitaciones del conjunto de datos y la "caja negra" interna de las redes neuronales profundas, los modelos de RNA obtenidos muestran una alta precisión predictiva, con efectos potenciales en el diseño de infraestructuras urbanas y la adecuación climática. Reflexionando sobre las limitaciones de este estudio y la posibilidad de integrar datos más diversos y algoritmos avanzados en el futuro, este estudio destaca el potencial del aprendizaje automático para jugar un papel clave en la mejora de la sostenibilidad y eficiencia de las soluciones de ingeniería civil, así como su contribución al cambio climático global y el desarrollo urbano sostenible., This study makes an experimental exploration in combining civil engineering and machine learning techniques with the aim of developing sustainable pervious concrete pavement designs for mitigating the effects of climate change. The study utilizes artificial neural networks (ANN) for predictive modeling to evaluate the effect of different input characteristics on the compressive strength and water permeability of pervious concrete, which are both key parameters for the durability and environmental performance of porous concrete. By using a limited dataset, this study adopts a k-fold cross-validation approach and trains numerous ANN configurations to optimize accuracy and prevent overfitting. It is found that porosity and coarse aggregate type have a significant effect on pavement performance, which guides engineers to optimize concrete formulations according to project requirements. Although faced with dataset limitations and the internal "black box" of deep neural networks, the obtained ANN models show high predictive accuracy, with potential effects on urban infrastructure design and climate suitability. By reflecting on the limitations of this study and the possibility of integrating more diverse data and advanced algorithms in the future, this study highlights the potential of machine learning to play a key role in improving the sustainability and efficiency of civil engineering solutions, as well as contributing to global climate change and sustainable urban development.

The full deployment of sixth-generation (6G) networks is inextricably connected with a holistic network redesign able to deal with various emerging challenges, such as integration of heterogeneous technologies and devices, as well as support of latency and bandwidth demanding applications. In such a complex environment, resource optimization, and security and privacy enhancement can be quite demanding, due to the vast and diverse data generation endpoints and associated hardware elements. Therefore, efficient data collection mechanisms are needed that can be deployed at any network infrastructure. In this context, the network data analytics function (NWDAF) has already been defined in the fifth-generation (5G) architecture from Release 15 of 3GPP, that can perform data collection from various network functions (NFs). When combined with advanced machine learning (ML) techniques, a full-scale network optimization can be supported, according to traffic demands and service requirements. In addition, the collected data from NWDAF can be used for anomaly detection and thus, security and privacy enhancement. Therefore, the main goal of this paper is to present the current state-of-the-art on the role of the NWDAF towards data collection, resource optimization and security enhancement in next generation broadband networks. Furthermore, various key enabling technologies for data collection and threat mitigation in the 6G framework are identified and categorized, along with advanced ML approaches. Finally, a high level architectural approach is presented and discussed, based on the NWDAF, for efficient data collection and ML model training in large scale heterogeneous environments., This work was supported in part by the HORSE Project funded by the Smart Networks and Services Joint Undertaking (SNS JU) through the European Union’s Horizon Europe Research and Innovation Program under Grant 101096342 (www.horse-6g.eu)., Peer Reviewed, Postprint (published version)

Les xarxes neuronals es poden utilitzar per classificar dades, que poden tenir diverses classes. Així, donada una entrada, la sortida d'una xarxa neuronal és la probabilitat assignada que l'entrada pertanyi a cada classe. L'estratègia habitual per seleccionar la classe en la classificació multiclasse és obtenir la que tingui més probabilitat, cosa que pot provocar una pèrdua de precisió. En aquesta tesi, pretenem oferir una selecció millorada de la classe d'entrada utilitzant distribucions de llei de potències i teoria de valors extrems. Per aconseguir-ho, estudiarem l'encaix dels diferents tipus de powerlaw i la validació d'aquest encaix. Finalment, s'utilitzaran algunes tècniques d'aprenentatge automàtic per estudiar els resultats obtinguts i la seva millora aportada respecte a aquesta opció de predicció d'ús habitual., Las redes neuronales pueden utilizarse para clasificar datos, que pueden tener múltiples clases. Así, dada una entrada, la salida de una red neuronal es la probabilidad asignada de que la entrada pertenezca a cada clase. La estrategia habitual para seleccionar la clase en la clasificación multiclase es obtener aquella con la probabilidad más alta, lo que puede llevar a una pérdida de precisión. En este TFM, pretendemos proporcionar una selección mejorada de la clase de entrada utilizando distribuciones de ley de potencias y la teoría de los valores extremos. Para ello, estudiaremos el ajuste de los distintos tipos de powerlaw y la validación de dicho ajuste. Finalmente, se utilizarán algunas técnicas de aprendizaje automático para estudiar los resultados obtenidos y la mejora aportada respecto a esta elección de predicción comúnmente utilizada., Neural networks can be used to classify data, which can have multiple classes. Thus, given an input, the output of a neural network is the assigned probability that the input belongs to each class. The usual strategy for selecting the class in multiclass classification is to get the one with the highest probability, which can lead to a loss of accuracy. In this thesis, we aim to provide an improved selection of the input class using powerlaw distributions and extreme value theory. To achieve this, we will study the fitting of the different types of powerlaw and the validation of this fitting. Finally, some machine learning techniques will be used to study the results obtained and their improvement provided with respect to this commonly used choice of prediction.

This work is a companion reproducibility paper of the experiments and results reported in Giovanelli et al. (2022), where data pre-processing pipelines are evaluated in order to find pipeline prototypes that reduce the classification error of supervised learning algorithms. With the recent shift towards data-centric approaches, where instead of the model, the dataset is systematically changed for better model performance, data pre-processing is receiving a lot of attention. Yet, its impact over the final analysis is not widely recognized, primarily due to the lack of publicly available experiments that quantify it. To bridge this gap, this work introduces a set of reproducible experiments on the impact of data pre-processing by providing a detailed reproducibility protocol together with a software tool and a set of extensible datasets, which allow for all the experiments and results of our aforementioned work to be reproduced. We introduce a set of strongly reproducible experiments based on a collection of intermediate results, and a set of weakly reproducible experiments (Lastra-Diaz, 0000) that allows reproducing our end-to-end optimization process and evaluation of all the methods reported in our primary paper. The reproducibility protocol is created in Docker and tested in Windows and Linux. In brief, our primary work (i) develops a method for generating effective prototypes, as templates or logical sequences of pre-processing transformations, and (ii) instantiates the prototypes into pipelines, in the form of executable or physical sequences of actual operators that implement the respective transformations. For the first, a set of heuristic rules learned from extensive experiments are used, and for the second techniques from Automated Machine Learning (AutoML) are applied., This work is partially supported by the EU’s Horizon Programme call, under Grant Agreements No. 101093164 (ExtremeXP) and the DOGO4ML project, funded by the Spanish Ministerio de Ciencia e Innovación under the project/funding scheme PID2020-117191RB-I00 / AEI / 10.13039/501100011033., Peer Reviewed, Postprint (author's final draft)

El setembre de 2021, hi va haver una exposició experimental i nova a la galeria Crone de Viena, Àustria. Totes les obres d’art que s’hi varen mostrar van ser generades per un precursor de la intel·ligència artificial actual, el generador de text Poetry Machine, el qual ha estat desenvolupant David Link des de 2001. L’article detalla els textos produïts i mostra alguns exemples de les obres d’art de l’exposició. Per tant, l’article demostra de manera pràctica un nou enfocament en la curació d’exposicions i en l’organització del flux de treball entre curadors, artistes i màquines., In September 2021, a novel, experimental exhibition took place at Crone Gallery in Vienna, Austria. The art-works shown here were all generated by a precursor of today’s Artificial Intelligence: the text generator Poetry Machine, which David Link has been actively developing since 2001. The article details the texts produced and shows some examples of the artworks in the exhibition. The article thereby practically demonstrates a new approach to curating exhibitions and organizing the workflow between curators, artists and machines., En septiembre de 2021, tuvo lugar una exposición experimental y novedosa en la galería Crone de Viena, Austria. Todas las obras de arte que se muestran aquí fueron generadas por un precursor de la inteli-gencia artificial actual, el generador de texto Poetry Machine, el cual ha estado desarrollando David Link desde 2001. El artículo detalla los textos producidos y muestra algunos ejemplos de las obras de arte de la exposición. Por lo tanto, el artículo demuestra de manera práctica un nuevo enfoque en la curación de exposiciones y en la organización del flujo de trabajo entre curadores, artistas y máquinas.

The Barcelona Expert Center (BEC) has become an international reference in the generation of high resolution soil moisture (SM) maps using disaggregation algorithms. More than a decade ago, a semi-empirical approach was developed to produce Soil Moisture and Ocean Salinity (SMOS) SM at 1 km. This method has been refined over the years to obtain cloud free maps, and modified to further improve the spatial resolution up to 300 m. More recently, a machine-learning approach has been developed to derive European Space Agency (ESA)’s Climate Change Initiative (CCI) SM at 60 m.Thanks to the multi-spectral information added during the disaggregation process, the downscaled SM maps have an overall accuracy similar to the coarse ones, but provide additional information about the SM spatial variability. In this regard, three different applications over the north of Africa are presented here to exemplify the added-value of SM data at high resolution., This research was supported by the Spanish Ministry of Science, Innovation and Universities (MCIN/AEI/10.13039/501100011033) through the project INTERACT (PID2020-114623RB-C32) and by the Institute of Space Studies of Catalonia (IEEC)., Peer Reviewed, Postprint (author's final draft)

El proyecto es una investigación comparativa en el campo del Procesamiento del Lenguaje Natural y Aprendizaje Automático, enfocándose en la evaluación y comparación de diferentes algoritmos para la clasificación de textos. Busca determinar la eficacia, eficiencia y aplicabilidad de estos algoritmos en contextos reales, con un enfoque en la optimización de recursos y el impacto económico y social. La meta es establecer directrices para la integración efectiva de estas tecnologías en entornos empresariales e institucionales.

In recent years, the financial markets have increased popularity, driven in large part by internet access and online trading platforms. This surge has not only increased the number of market participants but has also amplified the impact of information dissemination on stock prices. This thesis explores to relate fundamental analysis and technical analysis thanks to artificial intelligence (AI). The fundamental analysis traditionally associated with financial health, and market prediction with technical analysis which is the price chart movement. Our models will try to predict financial market with help of news articles. Currently, AI models have emerged as indispensable tools for processing massive information sources. Machine learning algorithms (ML), natural language processing (NLP) and sentiment analysis enable the automate extraction in massive text data information. The importance of news articles in financial market prediction cannot be underestimated. News items, both traditional and social media, have the power to rapidly influence investor sentiment and, consequently, asset prices. This thesis delves into the role of sentiment expressed in news articles translated to the price movement. Online articles platforms can shape investor behavior converting either bullish (up trend) or bearish (down trend) trends.

The evolution of wireless communication technologies, moving from the established fifth-generation (5G) to beyond 5G and ultimately the sixth-generation (6G), highlights the need for a significant shift in architectural approach. This transformation is essential to effectively accommodate the significant increase in multi-connectivity and on-demand services that is expected in the near future. Machine Learning (ML) and more specifically Deep Reinforcement Learning (DRL) are a promising approach to solving the aforementioned challenges. In this project, an approach for dynamic resource allocation and management of both the Radio Acess Network (RAN) and Multi-Access Edge Computing (MEC) leveraging Deep Q-Network (DQN) is proposed. Two DQN models are implemented for admission control and maintenance of RAN-level slicing, in order to be deployed as Extended Applications (xApps) within the O-RAN architectural framework. This methodology ensures effective resource allocation while maintaining the Quality of Services (QoS). The proposed solution is validated through simulation results, demonstrating its effectiveness in improving network efficiency and performance in future 5G and 6G networks. Further stages include the implementation of Federated Learning to deploy the proposed models in multiple mobile scenarios and the correspondent emulation in real-scale frameworks.

One of the main goals of human genetics is to understand the connections between genomic variation and the predisposition to develop a complex disorder. These disease–variant associations are usually studied in a single independent manner, disregarding the possible effect derived from the interaction between genomic variants. In particular, in a background of complex diseases, these interactions can be directly linked to the disorder and may play an important role in disease development. Although their study has been suggested to help complete the understanding of the genetic bases of complex diseases, this still represents a big challenge due to large computing demands. Here, we take advantage of high-performance computing technologies to tackle this problem by using a combination of machine learning methods and statistical approaches. As a result, we created a containerized framework that uses multifactor dimensionality reduction (MDR) to detect pairs of variants associated with type 2 diabetes (T2D). This methodology was tested on the Northwestern University NUgene project cohort using a dataset of 1,883,192 variant pairs with a certain degree of association with T2D. Out of the pairs studied, we identified 104 significant pairs: two of which exhibit a potential functional relationship with T2D. These results place the proposed MDR method as a valid, efficient, and portable solution to study variant interaction in real reduced genomic datasets., This work was partially financed by the European Commission (EU-HORIZON NEARDATA GA.101092644) and by the Universitat Politècnica de Catalunya (45-FPIUPC2018); it was also partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478; it was also partially financed by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI), and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, MCIN/AEI/10.13039/501100011033/FEDER, UE., Peer Reviewed, Postprint (published version)

En los últimos años, se han hecho descubrimientos en nanotecnología y materiales, que están dando paso a nanodispositivos con capacidades de detección, almacenamiento y procesamiento de datos. Se prevé que estos nanodispositivos naveguen por el torrente sanguíneo, con el fin de detectar eventos médicos de interés. La denominada "flow-guided localization" pretende asignar estos eventos a sus posiciones físicas, lo que permitiría un rápido diagnóstico y una reducción en costes e invasividad. Dado que este campo de investigación se encuentra aún en sus primeras fases, actualmente no hay un estándar común en las evaluaciones de este tipo de localización. Este hecho dificulta comparaciones objetivas entre este tipo de soluciones, lo que obstaculiza el avance en el ámbito. Como consecuencia, los investigadores se han centrado en el desarrollo de marcos estandarizados para evaluar los sistemas de "flow-guided localization". Sin embargo, aún no se ha investigado el impacto que la selección de los eventos tiene en su localización. Este estudio aborda esta laguna de conocimiento, analizando cómo la cantidad y la ubicación de los eventos a detectar afectan al rendimiento de los sistemas de localización. Para ello, se presenta una metodología de muestreo de puntos de evaluación, que garantiza estimaciones representativas de la precisión de estos algoritmos en todo el sistema cardiovascular. El conjunto de eventos obtenido reduce los recursos computacionales necesarios para las valoraciones, a la vez que mantiene la objetividad y fiabilidad de los resultados., Recent advances in the field of nanotechnology and novel materials are paving the way towards nanodevices that integrate sensing, storage and processing capabilities. These nanodevices are envisioned to navigate through the bloodstream, while sensing for medical events of interest. Flow-guided localization seeks to map these events to physical sites, resulting in early diagnosis while lowering costs and invasiveness. As the field is still in its early stages, existing evaluations of flow-guided localization are non-standardized, making objective comparisons challenging. To promote advancement in the field, researchers have focused on developing standardized frameworks for evaluating flow-guided localization systems. However, research into the impact of target event selection on localization is still lacking in the literature. This study addresses this knowledge gap by looking at how the number and placement of medical target events affect the performance of flow-guided localization systems. To address this, a methodology for sampling evaluation points is presented, ensuring representative assessments of performance indicators across the cardiovascular system. The derived subset of target events reduces the computational resources required for testing while maintaining objectivity and reliability of the results.

Due to the abundance of data, noisy, irrelevant, or redundant features often need to be identified and discarded. Feature selection is a collection of methods used to ensure that only relevant data are used for a data analysis task. Extracting and using only useful data for analysis promotes model understanding and performance and reduces the model training time and variance, i.e., overfitting. There is an abundance of methods for feature selection, and they can be categorised by various perspectives and are applicable to differing use cases. In this tutorial, we introduce the feature selection problem and present it from three perspectives of categorisation: search strategy, model reliance, and relevance definition. Furthermore, we propose a guideline for the use of the various methods. Lastly, we discuss current challenges and opportunities for research on feature selection., The project leading to this publication has received funding from 2020 research and innovation programme (grant agreement No 955895)., Peer Reviewed, Postprint (published version)

Artificial Intelligence has become increasingly important for aiding human decision making in various fields such as healthcare, automotive industry, education, and economy. Hence, in the last years, the interest of researchers and practitioners on developing and deploying Machine Learning (ML) models that are trustworthy is increasing. In this paper, we present TrustML, a simple to use, extensible and yet comprehensive Python package that allows specifying and computing the trustworthiness of ML models meant to conform AI systems. This package may be used for evaluating ML model's trustworthiness either during their development process and in production environments to support the continuous monitoring of their trustworthiness, benefiting taking correction actions when needed., This work was supported by “Agència de Gestió d’Ajuts Universitaris i de Recerca (Generalitat de Catalunya)” under grant agreement 2020 FI SDUR 00279, and by the “Spanish Ministerio de Ciencia e Innovación” under project / funding scheme PID2020–117191RB-I00 / AEI/ 10.13039/501100011033., Peer Reviewed, Postprint (published version)

The rapid expansion of Machine Learning (ML) and Artificial Intelligence (AI) has profoundly influenced the technological landscape, reshaping various industries and applications. This surge in computational demands has led to the widespread adoption of Cloud data centers, crucial for supporting the storage and processing requirements of these advanced technologies. However, this expansion poses significant challenges, particularly in terms of energy consumption and associated carbon emissions. As the reliance on cloud data centers intensifies, there is a growing concern about the environmental impact, necessitating innovative solutions to enhance energy efficiency and reduce the ecological footprint of these computational infrastructures. This paper focuses on addressing the challenges linked to training ML and AI applications, emphasizing the importance of energy-efficient solutions. The proposed framework integrates components from the AI-SPRINT project toolchain, such as Krake, Space4AI-R, and PyCOMPSs. Our reference application involves training a Random Forest model for electrocardiogram classification, profiling available resources to obtain a performance model able to predict the training time, and dynamically migrating the workload to sites with cleaner energy sources providing guarantees on the training process due date. Results demonstrate the framework’s capacity to estimate execution time and resource requirements with low error, highlighting its potential for establishing an environmentally sustainable AI ecosystem., This work has been funded by the European Commission under the H2020 grant N. 101016577 AI-SPRINT: AI in Secure Privacy pReserving computINg conTinuum., Peer Reviewed, Postprint (author's final draft)

El ràpid creixement de l'aprenentatge automàtic ha transformat significativament diverses indústries, incloent-hi la salut, les finances i els sistemes autònoms. Comprendre les tendències en aquest camp dinàmic és crucial per guiar la investigació, assignar recursos i anticipar desenvolupaments futurs. Aquest estudi aborda la necessitat d'una anàlisi de tendències exhaustiva en la investigació d'aprenentatge automàtic des de 2014 fins a 2024 mitjançant l'examen dels títols i resums d'articles científics. En extreure qualificadors descriptius, hem classificat els articles en àrees específiques i n'hem analitzat l'evolució al llarg del temps. La nostra metodologia inclou un estudi detallat de qualificadors, l'estudi de la co-ocurrència dels qualificadors amb regles d'associació, la classificació dels articles en àrees específiques i la predicció de tendències per a cada àrea. Les troballes clau destaquen la prominència contínua de temes com "Xarxes Neuronals Artificials i Aprenentatge Profund" i l'aparició de noves àrees com els "Models Generatius". L'anàlisi ha identificat tendències consistents, proporcionant valuoses perspectives sobre el desenvolupament del camp. Aquest estudi demostra l'efectivitat de les tècniques de mineria de textos en el seguiment i la predicció de tendències de recerca., The rapid growth of machine learning has significantly transformed various industries, including healthcare, finance, and autonomous systems. Understanding trends in this dynamic field is crucial for guiding research, allocating resources, and anticipating future developments. This study addresses the need for a comprehensive trend analysis in machine learning research from 2014 to 2024 by examining the titles and abstracts of scientific articles. By extracting descriptive qualifiers, we classified articles into specific topics and analyzed their evolution over time. Our methodology includes a detailed study of qualifiers, the study of the co-occurrence of this qualifiers with association rules, topic classification of the articles, and trend prediction for each topic. Key findings highlight the continued prominence of topics such as "Artificial Neural Networks and Deep Learning" and the emergence of new areas like "Generative Models." The analysis revealed significant shifts in research focus and identified consistent trends, providing valuable insights into the development of the field. This study demonstrates the effectiveness of text mining techniques in tracking and predicting research trends.

Malware detection plays and important role in modern digital systems. Protecting against the fast-paced evolving cyber attacks is critical to safeguard sensitive information and preserve the integrity of digital infrastructure. Traditional signature-based detection methods are not effective when detecting new or altered versions of malware, such as polymorphic or metamorphic malware. Machine learning approaches have been proven to be much more effective at detecting such malware. Runtime behavior can be captured using the most fundamental part of a program, its instructions, also referred as the opcodes. This study presents both static and dynamic analysis using opcodes as the main feature for machine learning models., Funded by the European Union. Project number: 101093062., Preprint

We present an empirical approach to identify the key factors affecting the execution performance of task-based workflows on a High Performance Computing (HPC) infrastructure composed of heterogeneous CPU-GPU clusters. Our results reveal that the execution performance in distributed GPU-accelerated task-based workflows highly depends on several interrelated factors regarding the task algorithm, dataset, resources, and system employed. In addition, our analysis identifies key correlations among these factors, presents novel observations, and offers guidelines toward designing an automated method to handle task-based workflows in modern, high-compute capacity, CPU-GPU engines., This work has been partially supported by DEDS (H2020-MSCAITN2020) with grant agreement No. 955895, the EU-HORIZON programme CREXDATA under GA.101092749, the EU-HORIZON programme FAIR-CORE4EOSC under GA.101057264, the EUHORIZON programme EXTREMEXP under GA.101093164, the Spanish Government projects PID2019-107255GB and PID2020117191RB-I00/AEI/10.13039/501100011033andMCIN/AEI/10.13039 /501100011033 (CEX2021-001148-S), and by the Departament de Recerca i Universitats de la Generalitat de Catalunya (2021 SGR 00412, MPiEDist)., Peer Reviewed, Postprint (published version)

El treball consta de dues parts principals. La primera, la realització d'un algorisme d'inteligència artificial que permet a un agent virtual (2 braços robòtics fent servir el simulador Gazebo i el software de control ROS) aprendre a completar una tasca per si sol. La segona part consisteix en construir físicament els robots i implementar-hi el mateix algorisme alhora que s'intenta minimitzar l'error que posa la realitat., El trabajo se compone de dos partes principales. La primera, la realización de un algoritmo de inteligencia artificial que permite a un agente virtual (en este caso 2 brazos robóticos simulados mediante Gazebo i el software de control ROS) aprender a completar una actividad por sí solo. La segunda parte consiste en construir físicamente los robots e implementar el mismo algoritmo a la vez que se intenta reducir el error que impone el medio real., This research addresses some of the challenges concerning Multi-Agent Reinforcement Learning. Specifically, it focuses on reducing the reality gap in a MARL environment through the training of a Deep Reinforcement Learning policy. The study also aims to enhance the skills and knowledge learned during the bachelors’ degree in electronics engineering while contributing to the academic goal of bridging the sim-to-real-transfer issue. Simulation outcomes demonstrate successful learning by the agent while implementation falls short due to critical oversights in the early phases of the project concerning time management design. Personal goals, including the familiarization of several frameworks such as Gazebo, ROS 2 and Linux, as well as enhancing Python and C++ programming skills, are fully achieved. Despite challenges, resources developed while pursuing this project, such as the policy (DDPG), test files and other classes, are robust and reusable. Therefore, it is encouraged that future works in similar domains make use of them. In conclusion, the research provides valuable insights into the challenges of MARL implementation, highlighting the need for careful project management and offering reusable resources for future endeavours., Outgoing

This thesis presents an exploratory data analysis (EDA) of a medical treatment for psychosis, specifically focusing on patients who underwent Metacognitive Training (MCT), as part of the PERMEPSY European research project. Psychosis, characterized by a detachment from reality manifesting in delusions and hallucinations, significantly impairs the quality of life, necessitating effective treatment modalities. Despite the benefits of psychological interventions, access to specialized mental health services and the personalization of treatments are still remaining challenges. The primary objective of the research presented in this thesis is to leverage unsupervised machine learning techniques to analyze clinical data from MCT-treated patients, aiming to identify distinct patient profiles that could improve treatment personalization. The study utilizes the PERMEPSY dataset, comprising sociodemographic information and Positive and Negative Syndrome Scale (PANSS) scores of 698 patients before and after MCT intervention. The methodology involves statistical analysis for data understanding, univariate and multivariate outlier detection for data refinement, and clustering algorithms for patient cohort identification. Techniques such as K-Means, agglomerative hierarchical Clustering, and Gaussian Mixture Models (GMM) are employed to discover meaningful patient groupings. The analyses reveal significant insights into the sociodemographic and clinical characteristics of patients, the impact of treatment on symptomatology, and the efficacy of clustering methods in identifying patient subgroups. These findings should contribute to the broader goal of personalized medicine by providing actionable insights for tailored psychosis treatment strategies.

Currently, many internal processes in companies are carried out mechanically and repetitively by employees during their working day. However, with the advent of artificial intelligence, these processes can be significantly optimized. This thesis combines these two factors and uses artificial intelligence to intelligently streamline the business classification process, especially following the reference to NAICS codes. The resulting application includes a classification model trained with artificial intelligence and various features aimed at improving the user experience. Among these features, there is generative artificial intelligence that expands the information provided by the user and, if necessary, translates responses into the language of the provided text. This combination of technologies seeks to optimize business processes and provide a comprehensive solution to users., A día de hoy, muchos procesos internos en las empresas se llevan a cabo de manera mecánica y repetitiva por parte de los trabajadores durante su jornada laboral. Sin embargo, con la aparición de la inteligencia artificial, estos procesos pueden ser optimizados de manera significativa. Este trabajo combina estos dos factores, utiliza la inteligencia artificial para agilizar de manera inteligente el proceso de clasificación de empresas, especialmente siguiendo la referencia de los códigos NAICS. La aplicación resultante incluye un modelo de clasificación entrenado con inteligencia artificial y diversas funcionalidades destinadas a mejorar la experiencia del usuario. Entre estas funcionalidades destaca una Inteligencia Artificial generativa que amplía la información proporcionada por el usuario y, si es necesario, traduce las respuestas al idioma del texto proporcionado. Esta combinación de tecnologías busca optimizar los procesos empresariales y proporcionar una solución integral a los usuarios., En l'actualitat, molts processos interns en empreses es duen a terme de manera mecànica i repetitiva per part dels treballadors durant la seva jornada laboral. No obstant això, amb l'aparició de la intel·ligència artificial, aquests processos poden ser optimitzats de manera significativa. Aquest treball combina aquests dos factors, utilitza la intel·ligència artificial per agilitzar de manera intel·ligent el procés de classificació d'empreses, especialment seguint la referència dels codis NAICS. L'aplicació resultant inclou un model de classificació entrenat amb intel·ligència artificial i diverses funcionalitats destinades a millorar l'experiència de l'usuari. Entre aquestes funcionalitats destaca una Intel·ligència Artificial generativa que amplia la informació proporcionada per l'usuari i, si és necessari, tradueix les respostes a l'idioma del text proporcionat. Aquesta combinació de tecnologies busca optimitzar els processos empresarials i proporcionar una solució integral als usuaris.

In the realm of cybersecurity, the constant evolution of malware poses a significant challenge for detection methods. Traditional signature-based approaches struggle to keep pace with emerging threats, necessitating innovative solutions. This master thesis explores the application of advanced machine learning techniques, specifically Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) architectures, enhanced by word embedding methodologies, to address the intricate task of detecting concealed malware within executable files. The research commences with a thorough exploration of fundamental machine learning principles and robust data processing methodologies, establishing a solid foundation. Building upon this knowledge, the study develops a specialized deep learning model tailored for accurate malware detection. Every aspect of model construction undergoes meticulous attention, encompassing data collection, preprocessing, rigorous experimentation, and hyperparameter optimization (HPO). The HPO process systematically refines model configurations, unveiling the most effective setups through comprehensive analysis of performance metrics. The evaluation of the final model provides a comprehensive assessment of its capabilities in malware detection. In summary, this research presents an adaptive and robust deep learning model for malware detection, leveraging LSTM and BiLSTM architectures enriched by word embedding techniques. It offers a detailed account of the research process, covering data collection, preprocessing, hyperparameter optimization, and model evaluation. This work contributes valuable insights to the dynamic field of cybersecurity, highlighting the potential of machine learning in fortifying the security of digital systems.

The increase in the demand for eXtended Reality (XR)/Virtual Reality (VR) services in the recent years, poses a great challenge for Wi-Fi networks to maintain the strict latency requirements. In VR over Wi-Fi, latency is a significant issue. In fact, VR users expect instantaneous responses to their interactions, and any noticeable delay can disrupt user experience. Such disruptions can cause motion sickness, and users might end up quitting the service. Differentiating interactive VR traffic from Non-VR traffic within a Wi-Fi network can aim to decrease latency for VR users and improve Wi-Fi Quality of Service (QoS) with giving priority to VR users in the access point (AP) and efficiently handle VR traffic. In this paper, we propose a machine learning-based approach for identifying interactive VR traffic in a Cloud-Edge VR scenario. The correlation between downlink and uplink is crucial in our study. First, we extract features from single-user traffic characteristics and then, we compare six common classification techniques (i.e., Logistic Regression, Support Vector Machines, k-Nearest Neighbors, Decision Trees, Random Forest, and Naive Bayes). For each classifier, a process of hyperparameter tuning and feature selection, namely permutation importance is applied. The model created is evaluated using datasets generated by different VR applications, including both single and multi-user cases. Then, a Wi-Fi network simulator is used to analyze the VR traffic identification and prioritization QoS improvements. Our simulation results show that we successfully reduce VR traffic delays by a factor of 4.2x compared to scenarios without prioritization, while incurring only a 2.3x increase in delay for background (BG) traffic related to Non-VR services., This work is partially funded by Wi-XR PID2021-123995NB-I00 (MCIU/AEI/FEDER,UE), MAX-R ( 101070072 ) EU, SGR 00955-2021 AGAUR, and by MCIN/AEI under the Maria de Maeztu Units of Excellence Programme ( CEX2021-001195-M ). This paper has also been partially funded by the Spanish Ministry of Science and Innovation MCIN/AEI/10.13039/501100011033 under ARTIST project (ref. PID2020- 115104RB-I00 )., Peer Reviewed, Postprint (author's final draft)

The AI-SPRINT project, launched in 2021 and funded by the European Commission, focuses on the development and implementation of AI applications across the computing continuum. This continuum ensures the coherent integration of computational resources and services from centralized data centers to edge devices, facilitating efficient and adaptive computation and application delivery. AI-SPRINT has achieved significant scientific advances, including streamlined processes, improved efficiency, and the ability to operate in real time, as evidenced by three practical use cases. This paper provides an in-depth examination of these applications – Personalized Healthcare, Maintenance and Inspection, and Farming 4.0 – highlighting their practical implementation and the objectives achieved with the integration of AI-SPRINT technologies. We analyze how the proposed toolchain effectively addresses a range of challenges and refines processes, discussing its relevance and impact in multiple domains . After a comprehensive overview of the main AI-SPRINT tools used in these scenarios, the paper summarizes of the findings and key lessons learned., Peer Reviewed, Postprint (published version)

This paper presents a study on connection-blocking prediction in Elastic Optical Networks (EONs) using Convolutional Neural Networks (CNNs). In EONs, connections are established and torn down dynamically to fulfill the instantaneous requirements of the users. The dynamic allocation of the connections may cause spectrum fragmentation and lead to network performance degradation as connection blocking increases. Predicting potential blocking situations can be helpful during EON operations. For example, this prediction could be used in real networks to trigger proper spectrum defragmentation mechanisms at suitable moments, thereby enhancing network performance. Extensive simulations over the well-known NSFNET (National Science Foundation Network) backbone network topology were run by generating realistic traffic patterns. The obtained results are later used to train the developed machine learning models, which allow the prediction of connection-blocking events. Resource use was continuously monitored and recorded during the process. Two different Convolutional Neural Network models, a 1D CNN (One-Dimensional Convolutional Neural Network) and 2D CNN (Two-Dimensional Convolutional Neural Network), are proposed as the predicting methods, and their behavior is compared to other conventional models based on an SVM (Support Vector Machine) and KNN (K Nearest Neighbors). The results obtained show that the proposed 2D CNN predicts blocking with the best accuracy (92.17%), followed by the SVM, the proposed 1D CNN, and KNN. Results suggest that 2D CNN can be helpful in blocking prediction and might contribute to increasing the efficiency of future EON networks., This research was partially funded by the Catalonia Government through the Agency for Management of University and Research Grants, grant SGR-2021-00598., Peer Reviewed, Postprint (published version)

El objetivo de este proyecto es mejorar la capacidad de evitación de colisiones del robot terrestre no tripulado Jackal integrando algoritmos avanzados de percepción y toma de decisiones. El objetivo principal de es utilizar técnicas de aprendizaje automático para lograr una mayor resolución y mejorar el rendimiento de un modelo de evitación de colisiones basado en D3QN desarrollado inicialmente desarrollado para robots aéreos. Nuestro enfoque para adaptar la tecnología de aprendizaje profundo por refuerzo a un robot terrestre se centra en reducir el ruido causado por la prominencia irrelevante del suelo en los datos de los sensores. Una política de acción adicional de aborda retos de navegación específicos, garantizando trayectorias más suaves y seguras. También nos centramos en optimizar el tiempo de procesamiento de la estimación de profundidad, la fusión eficiente de los modelos de percepción y decisión, y el desarrollo de una función de evaluación para evaluar el rendimiento. Los experimentos clave validan nuestro planteamiento, y los resultados muestran mejoras significativas en la capacidad del robot para seguir trayectorias suaves y seguras., L'objectiu d'aquest projecte és millorar la capacitat d'evitació de col·lisions del robot terrestre no tripulat Jackal integrant algorismes avançats de percepció i presa de decisions. L'objectiu principal d'és utilitzar tècniques d'aprenentatge automàtic per a aconseguir una major resolució i millorar el rendiment d'un model d'evitació de col·lisions basat en D3QN desenvolupat inicialment desenvolupat per a robots aeris. El nostre enfocament per a adaptar la tecnologia d'aprenentatge profund per reforç a un robot terrestre se centra en reduir el soroll causat per la prominència irrellevant del sòl en les dades dels sensors. Una política d'acció addicional d'aborda reptes de navegació específics, garantint trajectòries més suaus i segures. També ens centrem en optimitzar el temps de processament de l'estimació de profunditat, la fusió eficient dels models de percepció i decisió, i el desenvolupament d'una funció d'avaluació per a avaluar el rendiment. Els experiments clau validen el nostre plantejament, i els resultats mostren millores significatives en la capacitat del robot per a seguir trajectòries suaus i segures., This project aims to enhance the collision avoidance capabilities of the Jackal unmanned ground robot by integrating advance perception and decision-making algorithms. The primary objective is use machine learning techniques to achieve a higher-resolution and quality results, seeking to improve the performance of a D3QN-based collision avoidance model initially developed for aerial robots. Our approach to tailoring the deep reinforcement learning technology to a ground robot focuses on reducing the noise caused by irrelevant ground prominence in sensor data. An additional action policy addresses specific navigation challenges, ensuring smoother and safer trajectories. We also focus on optimizing the processing time of depth estimation, efficiently merging perception and decision models, and developing an evaluation function to assess performance. Key experiments validate our approach, results show significant enhancements in the robot's ability to navigate safely and efficiently

The Dagstuhl Seminar 23352, titled “Integrating HPC, AI, and Workflows for Scientific Data Analysis,” held from August 27 to September 1, 2023, was a significant event focusing on the synergy between HighPerformance Computing (HPC), Artificial Intelligence (AI), and scientific workflow technologies. The seminar recognized that modern Big Data analysis in science rests on three pillars: workflow technologies for reproducibility and steering, AI and Machine Learning (ML) for versatile analysis, and HPC for handling large data sets. These elements, while crucial, have traditionally been researched separately, leading to gaps in their integration. The seminar aimed to bridge these gaps, acknowledging the challenges and opportunities at the intersection of these technologies. The event highlighted the complex interplay between HPC, workflows, and ML, noting how ML has increasingly been integrated into scientific workflows, thereby enhancing resource demands and bringing new requirements to HPC architectures, like support for GPUs and iterative computations. The seminar also addressed the challenges in adapting HPC for large-scale ML tasks, including in areas like deep learning, and the need for workflow systems to evolve to leverage ML in data analysis fully. Moreover, the seminar explored how ML could optimize scientific workflow systems and HPC operations, such as through improved scheduling and fault tolerance. A key focus was on identifying prestigious use cases of ML in HPC and understanding their unique, unmet requirements. The stochastic nature of ML and its impact on the reproducibility of data analysis on HPC systems was also a topic of discussion., Postprint (published version)

Despite the large impact chronic obstructive pulmonary disease (COPD) that has on the population, the implementation of new technologies for diagnosis and treatment remains limited. Current practices in ambulatory oxygen therapy used in COPD rely on fixed doses overlooking the diverse activities which patients engage in. To address this challenge, we propose a software architecture aimed at delivering patient-personalized edge-based artificial intelligence (AI)-assisted models that are built upon data collected from patients’ previous experiences along with an evaluation function. The main objectives reside in proactively administering precise oxygen dosages in real time to the patient (the edge), leveraging individual patient data, previous experiences, and actual activity levels, thereby representing a substantial advancement over conventional oxygen dosing. Through a pilot test using vital sign data from a cohort of five patients, the limitations of a one-size-fits-all approach are demonstrated, thus highlighting the need for personalized treatment strategies. This study underscores the importance of adopting advanced technological approaches for ambulatory oxygen therapy., This research was funded by the Spanish Ministry of Science, Innovation and Universities and FEDER, grant number PID2021-124463OB-100, and by the AGAUR Catalan Agency, grant number 2021_SGR_00326., Peer Reviewed, Postprint (published version)

The motivation for this work stems from the inefficiencies in resource allocation in serverless systems, which are particularly critical in data-intensive fields such as genomics. Efficient resource management can significantly reduce execution times and improve overall throughput, thus accelerating research progress. The goal is to develop an automatic recommender system that dynamically adjusts resource allocation based on predicted workload demands, minimising both over-provisioning and under-provisioning. Two primary approaches to telemetry forecasting are investigated: (1) a streaming method using Linear Regression or Linear SVM models trained on real-time telemetry data, and (2) a batch method using deep learning models like LSTMs trained on historical data from multiple workloads. The methodology involves collecting telemetry data, preprocessing it for modelling, and evaluating the performance of different models. The best-performing models from each category are compared against each other and a naive baseline model. The findings suggest that while deep learning models generally outperform linear models in predicting future resource usage, the choice of model and its configuration should be workload-specific. Iterative deep LSTM models showed the most promise in terms of predictive capabilities for multi-step ahead forecasting, although this approach incurs higher computational costs. The study also explores the trade-off between minimising under-provisioning and the resultant increase in over-provisioning. This thesis demonstrates that ML-based telemetry forecasting can optimise resource allocation in serverless environments, specifically by making a comparison to naive approaches. The recommended approach for large genomics tasks involves deploying an iterative deep LSTM model, though further validation is necessary for different workload types. This work contributes to the broader goal of enhancing serverless systems' efficiency, with potential applicatio

Aquest treball tracta sobre els canvis que la implementació de la Intel·ligència Artificial (IA) s’estan duent i es portaran a terme en l'àmbit de la salut. S'analitza com la IA, amb tècniques com l'aprenentatge automàtic i l’aprenentatge profund, canvien la manera com s'aborden els desafiaments mèdics actualment i millora la qualitat en l'atenció als pacients. El futur de la medicina dependrà en gran mesura de la interacció humana i la tecnologia, i farà que els professionals en l’ambit de la salut estiguin obligats a utilitzar aquestes noves eines i coneguin els seus avantatges i els seus inconvenients, perquè seran imprescindibles en la seva feina. Al treball es mostra com la IA oferirà als pacients més seguretat, autonomia i la possibilitat d'atenció mèdica a zones de difícil accés, i als metges els ajudarà a disminuir càrrega administrativa, temps observant pantalles i esgotament professional. La IA permetrà també reduir errors mèdics i millorar-ne la precisió diagnòstica a través de l'anàlisi i la interpretació d'informació per algorismes. Els tractaments podran ser personalitazts. L'automatització d'activitats repetitives alliberarà temps al personal de salut i potencialment millorarà la relació metge-pacient mitjançant la comunicació, l'empatia i la confiança durant la malaltia, activitats que mai seran reemplaçades per la IA. Però caldrà establir normes clares que assegurin la privacitat, l’ètica i les responsabilitats.

Aquesta tesi explora la integració de tècniques d'aprenentatge automàtic en la presa de decisions financeres per millorar les taxes de conversió de pagaments. L'estudi se centra en el desenvolupament d'una pipeline d'aprenentatge automàtic que utilitza tant dades històriques com de nova recopilació per predir el camí de routing òptim per a les transaccions, maximitzant així les taxes d'aprovació i minimitzant els costos. La investigació es duu a terme en col·laboració amb PayXpert, una empresa fintech especialitzada en serveis de pagament per a comerciants en línia i al detall a Europa. Les conclusions principals indiquen que, mentre que les xarxes neuronals proporcionen la màxima precisió, els boscos aleatoris ofereixen un rendiment equilibrat amb millor interpretabilitat i eficiència, cosa que els fa adequats per a un desplegament inicial. L'estudi també aborda els reptes de conjunts de dades desequilibrats i la integració de nous adquirents al sistema, proposant tècniques com l'aprenentatge incremental i la generació de dades sintètiques per mantenir la robustesa del model., This thesis explores the integration of machine learning techniques in financial decision-making to enhance payment conversion rates. The study focuses on the development of a machine learning pipeline that leverages both historical and newly collected transactional data to predict the optimal routing path for transactions, thereby maximizing approval rates and minimizing costs. The research is conducted in collaboration with PayXpert, a fintech company specializing in payment services for online and retail merchants across Europe. Key findings indicate that while neural networks provide the highest accuracy, random forests offer a balanced performance with better interpretability and efficiency, making them suitable for initial deployment. The study also addresses the challenges of imbalanced datasets and the integration of new acquirers into the system, proposing techniques such as incremental learning and synthetic data generation to maintain model robustness.

This project describes all the steps carried out for the understanding, design and validation of an O-RAN/MEC environment implemented in a private network located at an airport, where, by means of slices, an allocation will be carried out of resources where users will be classified in one slice or another depending on the requirements they demand. All this, to maintain the quality of service of the system. At first, all the elements and architectures that will be used further in the work are foing to be explained. Some of there are private networks, the core network, the O-RAN network, physical channels, the MEC and artificial intelligence, more in particular DQN. Secondly, once the theoretical principles are comprehended, an xApp, a component of O-RAN, will be developed to perform this slicing process. Moreover, all the tools and frameworks applied to the project will also be described. Finally, an amulation and validation of all the aspects incorporated into this work will be performed with the intention of determining if the application of slicing on the users truly preserves the quality of services while providing a better dedicated version of the resources., Objectius de Desenvolupament Sostenible::9 - Indústria, Innovació i Infraestructura, Objectius de Desenvolupament Sostenible::7 - Energia Assequible i No Contaminant

El nombre de ciberatacs és cada vegada més freqüent, afectant no només a individus sinó també a organitzacions. Això pot portar a nombrosos problemes, impulsant l'estudi i desenvolupament de tècniques per mitigar-los tant com sigui possible. En aquest treball, explorem l'ús de tècniques d'aprenentatge automàtic per crear sistemes intel·ligents capaços d'analitzar amenaces, centrant-nos específicament en amenaces relacionades amb dominis. Es recopilaran i processaran dades per entrenar un model d'aprenentatge automàtic supervisat capaç de classificar si un domini és maliciós o no. Es fa una comparació entre aquestes tècniques i les anteriors que utilitzaven regles predefinides. L'enfoc amb regles predefinides provoca un conflicte entre simplicitat i precisió perquè la sortida és binària. Amb el nou enfoc, volem obtenir una probabilitat de risc entre 0 i 1 en comtes del binàri. La conclusió destaca els resultats superiors obtinguts mitjançant l'ús de tècniques d'aprenentatge automàtic. Aquesta nova tècnica utilitzant el model d'aprenentatge automàtic supervisat mostra un rendiment molt millor que el punt de partida., The number of cyber attacks is increasingly common, affecting not only individuals but also organizations. This can lead to numerous problems, prompting the study and development of techniques to mitigate them as much as possible. In this thesis, we explore the use of machine learning techniques to create intelligent systems capable of analyzing threats, focusing specifically on domain-related threats. Data will be collected and processed to train a supervised machine learning model capable of classifying whether a domain is malicious or not. A comparison is drawn between these techniques and the previous one that used predefined rules. The approach with the predefined rules causes a conflict between simplicity and accuracy because the output is binary. With the new approach we want to output a risk probability between 0 and 1 instead. The conclusion highlights the superior results obtained through the use of machine learning techniques. The approach that used the supervised machine learning model shows way better performance than the baseline.

Advancements in dermatological artificial intelligence research require high-quality and comprehensive datasets that mirror real-world clinical scenarios. We introduce a collection of 18,946 dermoscopic images spanning from 2010 to 2016, collated at the Hospital Clínic in Barcelona, Spain. The BCN20000 dataset aims to address the problem of unconstrained classification of dermoscopic images of skin cancer, including lesions in hard-to-diagnose locations such as those found in nails and mucosa, large lesions which do not fit in the aperture of the dermoscopy device, and hypo-pigmented lesions. Our dataset covers eight key diagnostic categories in dermoscopy, providing a diverse range of lesions for artificial intelligence model training. Furthermore, a ninth out-of-distribution (OOD) class is also present on the test set, comprised of lesions which could not be distinctively classified as any of the others. By providing a comprehensive collection of varied images, BCN20000 helps bridge the gap between the training data for machine learning models and the day-to-day practice of medical practitioners. Additionally, we present a set of baseline classifiers based on state-of-the-art neural networks, which can be extended by other researchers for further experimentation., We acknowledge the support of the International Skin Imaging Collaboration (ISIC). This research was supported by the Spanish Research Agency (AEI) under project PID2020-116907RB-I00 of the call MCIN/ AEI /10.13039/501100011033 and the project 718/C/2019 with id 201923-30 and 201923-31, funded by Fundació la Marató de TV3, iTOBOs grant from the European Union’s Horizon 2020 research and innovation programme num 965221. Other funding sources include the Melanoma Research Alliance Young Investigator Award 614197. This research was funded in part through the NIH/NCI Cancer Center Support Grant P30 CA008748., Peer Reviewed, Postprint (published version)

As companies store, process, and analyse bigger and bigger volumes of highly heterogeneous data, novel research and technological challenges are emerging. Traditional and rigid data integration and processing techniques become inadequate for a new class of data-intensive applications. There is a need for new architectural, software, and hardware solutions that are capable of providing dynamic data integration, assuring high data quality, and offering safety and security mechanisms, while facilitating online data analysis. In this context, we propose moduli, a novel disaggregated data management reference architecture for data-intensive applications that organizes data processing in various zones. Working on moduli allowed us also to identify open research and technological challenges., Peer Reviewed, Postprint (author's final draft)

El nombre de sistemes que incorporen components d'intel·ligència artificial (IA) està creixent notablement els últims anys. Aquestes tecnologies poden tenir un fort impacte mediambiental, a causa del seu consum energètic, durant el seu desenvolupament i funcionament. Aquest projecte presenta l'eina GAISSALabel. Una eina que permet avaluar el consum energètic dels components d'IA i proporcionar una etiqueta d'eficiència energètica. Aquesta etiqueta permet entendre les implicacions mediambientals i, eventualment, facilitar la disminució d'aquest impacte. Seguint una metodologia àgil en el desenvolupament, s'analitzen els requisits de GAISSALabel, s'especifica, dissenya, implementa i desplega aquesta eina. S'analitza com l'eina es pot integrar en el procés de desenvolupament de sistemes basats en IA. D'aquesta manera, es presenten exemples d'ús que inclouen la generació d'etiquetes energètiques i recomanacions per reduir el consum d'energia. Per acabar, es conclou que GAISSALabel facilita la conscienciació i reducció de l'impacte mediambiental dels sistemes basats en IA., The number of systems incorporating artificial intelligence (AI) components has been growing significantly in recent years. These technologies can have a strong environmental impact due to their energy consumption during development and operation. This project introduces the tool GAISSALabel. A tool that allows evaluating the energy consumption of AI components and providing an energy efficiency label. This label helps understand the environmental implications and, ultimately, facilitates the reduction of this impact. Following an agile methodology in development, the requirements of GAISSALabel are analyzed, and the tool is specified, designed, implemented, and deployed. The tool's integration into the development process of AI-based systems is examined. In this way, usage examples are presented, including the generation of energy labels and recommendations to reduce energy consumption. In conclusion, it is noted that GAISSALabel facilitates awareness and reduction of the environmental impact of AI-based systems.

Machine learning (ML) and Deep learning (DL) are increasingly pivotal in the design of advanced metamaterials, seamlessly integrated with material or topology optimization. Their intrinsic capability to predict and interconnect material properties across vast design spaces, often computationally prohibitive for conventional methods, has led to groundbreaking possibilities. This paper introduces an innovative machine learning approach for the optimization of acoustic metamaterials, focusing on Multiresonant Layered Acoustic Metamaterial (MLAM), designed for targeted noise attenuation at low frequencies (below 1000 Hz). This method leverages ML to create a continuous model of the Representative Volume Element (RVE) effective properties essential for evaluating sound transmission loss (STL), and subsequently used to optimize the overall topology configuration for maximum sound attenuation using a Genetic Algorithm (GA). The significance of this methodology lies in its ability to deliver rapid results without compromising accuracy, significantly reducing the computational overhead of complete topology optimization by several orders of magnitude. To demonstrate the versatility and scalability of this approach, it is extended to a more intricate RVE model, characterized by a higher number of parameters, and is optimized using the same strategy. In addition, to underscore the potential of ML techniques in synergy with traditional topology optimization, a comparative analysis is conducted, comparing the outcomes of the proposed method with those obtained through direct numerical simulation (DNS) of the corresponding full 3D MLAM model. This comparative analysis highlights the transformative potential of this combination, particularly when addressing complex topological challenges with significant computational demands, ushering in a new era of metamaterial and component design., The authors acknowledge the Spanish Ministry of Science and Innovation (MCIN/AEI/10.13039/501100011033, Grant CEX2018-000797-S-19-1, TED2021-129413B-C21) for supporting G. Sal-Anglada with the PhD grant PRE2019-088777 under the FPI programme. This research has also been funded by the Ministry of Research and Universities of the Government of Catalonia, through the research grant 2021-PROD-00016 for the project METACOUSTECH. The authors D. Roca and D. Yago are Serra Húnter Fellows., Peer Reviewed, Postprint (published version)

In this thesis, an advanced order model reduction for the incompressible steady-state Navier-Stokes equations is investigated. Initially, some basic principles of the finite element an approach are reviewed. This initial theoretical section offers a chance to connect elegant and intuitive ideas, while also helping to frame the problem into a matrix formulation. Moreover, this primary phase will help acquire a solid understanding on the implementation of the Finite Element Method (FEM) in a vectorized Matlab code, which reduces significantly the computation cost of the MATLAB program. In the fifth chapter, which examines the foundations of reduced order modeling, the thesis makes its first contribution. Moreover, the code will be used to obtain essential training data sets which will serve as the foundational material for constructing velocity and pressure modes for the reduced order model. Finally, these modes will be used as comprehensive shape functions for the original finite element problem, integrating them into the framework to enhance its efficiency and accuracy.

Graph Neural Networks (GNNs) have achieved great success in solving many machine learning tasks, and many different neural architectures have been proposed over the past few years. However, the lack of robust public benchmarks hinders the assessment of GNN architectures, which has made most research papers rely on the same 5-10 datasets. Our first contribution is Graphlaxy++, a novel generative graph learning model designed to emulate GNN performance and that can be used to generate artificial benchmarks with user-defined metrics. Leveraging Graphlaxy++, we generate a diverse synthetic graph dataset that serves as input to a predictive statistical model that can predict the performance of GNNs. This model is able to successfully predict the relative performance of 4 different architectures: GCN, GAT, GIN, and MLP. The predictive capacity of this model offers potential applications in speeding up neural architecture searches and optimizing graph modeling strategies.

Durant els últims anys, la majoria de les indústries tecnològiques han optat per l'automatització dels problemes de classificació de dades mitjançant algorismes d'aprenentatge automàtic. A mesura que els problemes es tornen més intrincats, els models han de ser més complexos per superar el repte. Això ha portat als desenvolupadors a utilitzar models complexes per defecte, com les xarxes neuronals artificials, que funcionen molt bé quan tracten problemes que són massa difícils per als mètodes clàssics d'aprenentatge automàtic, però no tenen en compte la interpretabilitat del model i l'explicació de la solució. Explainable AI (XAI) aborda la importància dels models transparents en situacions crítiques, on el raonament darrere d'una decisió ha de ser accessible, i proposa mètodes d'explicabilitat per a models complexos opacs. En particular, aquesta tesi pretén apropar XAI a l'algorisme de reconeixement de vehicles utilitzat en el projecte Mobilytics. Investigarem com funcionen els algorismes populars d'aprenentatge automàtic, la seva complexitat i com s'adapten al nostre problema particular. Després, construirem un espai de treball basat en mètodes agnóstics d'explicació post-hoc populars per tal d'explicar les nostres prediccions i obtenir informació sobre com funciona el nostre model final., Over the last few years, the automation of data classification problems using machine learning algorithms has been widely adopted by most tech industries. As the problems get more intricate, the models need to get more complex in order to overcome the challenge. This has led developers to default to complex models, such as artificial neural networks, that perform great when dealing with problems that are too convoluted for classical machine learning methods, but disregard model interpretability and solution explanation. Explainable AI (XAI) addresses the importance of transparent models in critical situations, where the reasoning behind a decision must be accessible, and proposes explainability methods for complex opaque models. In particular, this thesis aims to bring XAI to the vehicle recognition algorithm used in the Mobilytics project. We will investigate how popular machine learning algorithms work, their complexity and how suited they are for our particular problem. Then, we will build a framework based on popular model-agnostic post-hoc explainability methods in order to explain our predictions and gain insight and how our final model works.