Search

Automated metadata annotation is only as good as training dataset, or rules that are available for the domain. It's important to learn what type of data content a pre-trained machine learning algorithm has been trained on to understand its limitations and potential biases. Consider what type of content is readily available to train an algorithm—what's popular and what's available. However, scholarly and historical content is often not available in consumable, homogenized, and interoperable formats at the large volume that is required for machine learning. There are exceptions such as science and medicine, where large, well documented collections are available. This paper presents the current state of automated metadata annotation in cultural heritage and research data, discusses challenges identified from use cases, and proposes solutions.

The degradation of critical components inside large industrial assets, such as ball-bearings, has a negative impact on production facilities, reducing the availability of assets due to an unexpectedly high failure rate. Machine learning- based monitoring systems can estimate the remaining useful life (RUL) of ball-bearings, reducing the downtime by early failure detection. However, traditional approaches for predictive systems require run-to-failure (RTF) data as training data, which in real scenarios can be scarce and expensive to obtain as the expected useful life could be measured in years. Therefore, to overcome the need of RTF, we propose a new methodology based on online novelty detection and asymmetrical hidden Markov models (As-HMM) to work out the health assessment. This new methodology does not require previous RTF data and can adapt to natural degradation of mechanical components over time in data-stream and online environments. As the system is designed to work online within the electrical cabinet of machines it has to be deployed using embedded electronics. Therefore, a performance analysis of As-HMM is presented to detect the strengths and critical points of the algorithm. To validate our approach, we use real life ball-bearing data-sets and compare our methodology with other methodologies where no RTF data is needed and check the advantages in RUL prediction and health monitoring. As a result, we showcase a complete end-to-end solution from the sensor to actionable insights regarding RUL estimation towards maintenance application in real industrial environments. This study was supported partially by the Spanish Ministry of Economy and Competitiveness through the PID2019-109247GB-I00 project and by the Spanish Ministry of Science and Innovation through the RTC2019-006871-7 (DSTREAMS project). Also, by the H2020 IoTwins project (Distributed Digital Twins for industrial SMEs: a big-data platform) funded by the EU under the call ICT-11-2018- 2019, Grant Agreement No. 857191.

Glioma is the most common type of tumor in humans originating in the brain. According to the World Health Organization, gliomas can be graded on a four-stage scale, ranging from the most benign to the most malignant. The grading of these tumors from image information is a far from trivial task for radiologists and one in which they could be assisted by machine-learning-based decision support. However, the machine learning analytical pipeline is also fraught with perils stemming from different sources, such as inadvertent data leakage, adequacy of 2D image sampling, or classifier assessment biases. In this paper, we analyze a glioma database sourced from multiple datasets using a simple classifier, aiming to obtain a reliable tumor grading and, on the way, we provide a few guidelines to ensure such reliability. Our results reveal that by focusing on the tumor region of interest and using data augmentation techniques we significantly enhanced the accuracy and confidence in tumor classifications. Evaluation on an independent test set resulted in an AUC-ROC of 0.932 in the discrimination of low-grade gliomas from high-grade gliomas, and an AUC-ROC of 0.893 in the classification of grades 2, 3, and 4. The study also highlights the importance of providing, beyond generic classification performance, measures of how reliable and trustworthy the model’s output is, thus assessing the model’s certainty and robustness. Carla Pitarch is a fellow of Eurecat’s “Vicente López” PhD grant program.

Ever since the concept of the Internet was born, it has experienced a massive increase in the number of users. Similarly, there has been an exponential increase in the amount of information that they transmit, partly because of new technologies and services that have emerged, such as cloud and real-time communication. Customers do now demand a fast and reliable interaction, without any data loss or undesired delays. In order to meet these needs and keep up with the constantly-evolving technologies different non-profit organizations needed to devise a new technology, and this is when Software-Defined Networking (SDN) was born. Unlike traditional networks to date, it decouples the control layer from the data layer, leaving the act of forwarding traffic to the network device and delegating all network decisions to a controller, thus centralizing all the decisions, which improves network operability and agility. Despite all the improvements made with this technology, we still encounter the same problems of traditional networks regarding traffic routing. Conventional algorithms such as Dijkstra and Least Loaded (LL) based on the occupancy of the links, allocate traffic without considering the impact the chosen path could have if future traffic were to be introduced in the network. The objective of this project is to create an algorithm capable of routing traffic taking into account the future impact. To evaluate the proposed mechanism of Dijkstra, we will first obtain some experimental results using the default link weights and then base these weights on the occupancy of the links, using the Open Network Operating System (ONOS) as the SDN controller and the Multi-Generator (MGEN) tool to generate traffic. After these results, this research will use Reinforcement Learning (RL), a subcategory of Artificial Intelligence (AI), to train a RL model in Python using a network of eight interconnected switches. After the agent is trained, we have made a comparison between RL, LL, and SP, in which we have run a series of files containing flows of different rates. In conclusion, this research will exhibit that, on average, the RL algorithm consistently beats the other two algorithms by 16%, when it comes to reducing the data loss, which will improve the efficiency of the network.

El estudio de fin de grado que el lector se dispone a leer se centra en la utilización de la ciencia de datos para analizar la percepción ciudadana sobre la ciudad de Barcelona y la obtención de relaciones entre los datos de los habitantes encuestados. Con tal de desarrollar el trabajo se han utilizado la mineria de datos, análisis estadísticos exploratorios y analisis de Clustering. En esta memoria se presenta la documentación teórica relacionada con el estudio, los archivos de datos utilizados, los métodos estadísticos realizados, los resultados obtenidos y las conclusiones extraídas. Todos los procesos analíticos de este trabajo han sido realizados con el lenguaje de programación Python con las librerías de Pandas y Kmodes L'estudi de fi de grau que el lector es disposa a llegir es centra en la utilització de la ciència de dades per a analitzar la percepció ciutadana sobre la ciutat de Barcelona i l'obtenció de relacions entre les dades dels habitants enquestats. Per tal de desenvolupar el treball s'han utilitzat la mineria de dades, anàlisis estadístiques exploratòries i anàlisis de Clustering. En aquesta memòria es presenta la documentació teòrica relacionada amb l'estudi, els arxius de dades utilitzades, els mètodes estadístics realitzats, els resultats obtinguts i les conclusions extretes. Tots els processos analítics d'aquest treball han estat realitzats amb el llenguatge de programació Python amb les llibreries de Pandes i Kmodes The final-degree project that the reader is about to read focuses on the use of data science to analyse citizen perception of the city of Barcelona and to obtain relationships in the data of the inhabitants surveyed. In order to develop the project, data mining, exploratory statistical analysis and clustering analysis have been used. This report presents the theoretical documentation related to the study, the data files used, the statistical methods used, the results obtained and the conclusions drawn. All the analytical processes in this work have been done using the Python programming language with the Pandas and Kmodes libraries

Traditional psychiatric hospitalization models have been criticized for being restrictive and for violating fundamental rights, which can be counterproductive for long-term patient recovery. In response to this situation, new hospitalization models have been developed that aim to create a safe therapeutic environment that promotes patient recovery, while minimizing the need for coercive measures. The psychiatry unit from Germans Trias i Pujol Hospital (HGTiP) has designed a dual model that combines a care method, called Safewards, with a risk prediction tool in emergencies and a risk categorization and workload tool on the ward through a global score. The objective of the model is to safely open the unit doors, taking into account the risks and the level of work of the care team. The tools currently used are not scientifically validated, thus hindering the scalability of the model. Therefore, the purpose of this project is to develop a proposal for intervention to develop two Artificial Intelligence (AI) algorithms, using validated variables, that predict the unit environment and patient risk at the time of admission. First, a qualitative study was conducted with expert personnel from the psychiatric unit at HGTiP to objectify the criteria by which risk level and workload on the ward are defined. A total of 35 items were extracted, and a digital platform was developed to store them. These will be used, along with other dynamic variables, as input into a regression algorithm. In addition, classification algorithms will be developed to predict the level of risk in the first 24 hours of admission. The input data will have variables related to symptoms and conflictive events in the last month. For both cases, the performance of different algorithms will be evaluated, and the one with the best predictive character will be chosen. This dual model, which uses Artificial Intelligence as a tool to support clinical decision-making, sets a precedent in risk control in open door models. Once the model is fully validated, it will be scalable to other units and will allow for an improvement in the quality of care, while maintaining the safety of all individuals involved in the recovery process Los modelos de hospitalización psiquiátrica tradicionales han sido criticados por ser restrictivos y vulnerar derechos fundamentales, lo que puede ser contraproducente para la recuperación de los pacientes a largo plazo. En respuesta a esta situación, se han desarrollado nuevos modelos de hospitalización que pretenden crear un ambiente terapéutico seguro y que promueva la recuperación del paciente, minimizando la necesidad de medidas coercitivas. El Hospital Germans Trias i Pujol, ha diseñado un modelo dual, que combina un método asistencial, llamado Safewards, con una herramienta de predicción de riesgos en urgencias y una categorización de riesgos y de carga asistencial en planta a través de una puntuación global. El objetivo del modelo es abrir las puertas de la unidad de manera segura, teniendo en cuenta los riesgos y el nivel de trabajo del equipo asistencial. Las herramientas que se usan actualmente no están validadas científicamente, dificultando así la escalabilidad del modelo. Por lo que la finalidad de este proyecto es elaborar una propuesta de intervención para desarrollar dos algoritmos de Inteligencia Artificial, con variables validadas, que predigan el ambiente de la unidad y el riesgo de un paciente en el momento de ingreso. En primer lugar, se ha realizado un estudio cualitativo con personal experto de la unidad de psiquiatría del HGTiP para objetivar los criterios por los que se define el nivel de riesgo y la carga asistencial en planta. Se han extraído un total de 35 ítems y se ha desarrollado una plataforma digital para almacenarlos. Estos serán utilizados, junto a otras variables dinámicas, como entrada en un algoritmo de regresión. Por otra parte, se desarrollarán algoritmos de clasificación para predecir el nivel de riesgo en las primeras 24 horas de ingreso. Los datos de entrada serán variables relacionadas con sintomatología y eventos conflictivos en el último mes. Para ambos casos, se evaluará el rendimiento de diferentes algoritmos y se elegirá el que mejor carácter predictivo tenga. Este modelo dual, que utiliza la Inteligencia Artificial como herramienta de apoyo a la decisión clínica, supone un precedente en el control de riesgos en modelos de puertas abiertas. Una vez validado el modelo en su totalidad, sería escalable a otras unidades y permitiría mejorar la calidad asistencial, manteniendo la seguridad de todas las personas implicadas en el proceso de recuperación Els models d'hospitalització psiquiàtrica tradicionals han estat criticats per ser restrictius i vulnerar drets fonamentals, cosa que pot ser contraproduent per a la recuperació dels pacients a llarg termini. En resposta a aquesta situació, s'han desenvolupat nous models d'hospitalització que pretenen crear un ambient terapèutic segur i que promogui la recuperació del pacient, minimitzant la necessitat de mesures coercitives. L'Hospital Germans Trias i Pujol, ha dissenyat un model dual, que combina un mètode assistencial, anomenat Safewards, amb una eina de predicció de riscos a urgències i una categorització de riscos i de càrrega assistencial a planta a través d'una puntuació global. L'objectiu del model és obrir les portes de la unitat de manera segura, tenint en compte els riscos i el nivell de treball de l'equip assistencial. Les eines que es fan servir actualment no estan validades científicament, dificultant així l'escalabilitat del model. La finalitat d'aquest projecte és elaborar una proposta d'intervenció per desenvolupar dos algorismes d'Intel·ligència Artificial, amb variables validades, que prediguin l'ambient de la unitat i el risc d'un pacient a l'ingrés. En primer lloc, s'ha fet un estudi qualitatiu amb personal expert de la unitat de psiquiatria de l'HGTiP per objectivar els criteris pels quals es defineix el nivell de risc i la càrrega assistencial a planta. Se n'han extret un total de 35 ítems i s'ha desenvolupat una plataforma digital per emmagatzemar-los. Aquests seran utilitzats, juntament amb altres variables dinàmiques, com a entrada en un algorisme de regressió. D'altra banda, es desenvoluparan algorismes de classificació per predir el nivell de risc en les primeres 24 hores d’ingrés. Les dades d'entrada seran variables relacionades amb simptomatologia i esdeveniments conflictius a l'últim mes. Per a tots dos casos, s'avaluarà el rendiment de diferents algorismes i es triarà el que millor caràcter predictiu tingui. Aquest model dual, que utilitza la Intel·ligència Artificial com a eina de suport a la decisió clínica, suposa un precedent en el control de riscos en models de portes obertes. Una vegada validat el model en la seva totalitat, seria escalable a altres unitats i permetria millorar la qualitat assistencial, mantenint la seguretat de totes les persones implicades en el procés de recuperació

Brain-age can be inferred from structural neuroimaging and compared to chronological age (brain-age delta) as a marker of biological brain aging. Accelerated aging has been found in neurodegenerative disorders like Alzheimer's disease (AD), but its validation against markers of neurodegeneration and AD is lacking. Here, imaging-derived measures from the UK Biobank dataset (N=22,661) were used to predict brain-age in 2,314 cognitively unimpaired (CU) individuals at higher risk of AD and mild cognitive impaired (MCI) patients from four independent cohorts with available biomarker data: ALFA+, ADNI, EPAD and OASIS. Brain-age delta was associated with abnormal amyloid-b, more advanced stages (AT) of AD pathology and APOE-e4 status. Brain-age delta was positively associated with plasma neurofilament light, a marker of neurodegeneration, and sex differences in the brain effects of this marker were found. These results validate brain-age delta as a non-invasive marker of biological brain aging in non-demented individuals with abnormal levels of biomarkers of AD and axonal injury. Marc Suárez-Calvet: Horizon 2020 - Research and Innovation Framework Programme 948677, Instituto de Salud Carlos III PI19/00155, La Caixa Foundation 100010434, Horizon 2020 - Research and Innovation Framework Programme 847648 / Juan Domingo Gispert: EU/EFPIA Innovative Medicines Initiative Joint Undertaking AMYPAD 115952 / Marc Suárez-Calvet: La Caixa Foundation 100010434; LCF/PR/ GN17/50300004 / Irene Cumplido-Mayoral: TriBEKa Imaging Platform project TriBEKa-17-519007, Universities and Research Secretariat, Ministry of Business and Knowledge of the Catalan Government 2017-SGR-892 / Juan Domingo Gispert: EIT Digital Grant 2021, Spanish Ministry of Science and Innovation (MCIN)/Spanish Research Agency (AEI) MCIN/AEI /10.13039/501100011033 RTI2018-102261, cofunded by the European Regional Development Fund (FEDER) / Marc Suárez-Calvet: NIHR biomedical research center at UCLH Frederik Barkhof Instituto de Salud Carlos III PI22/00456 / Raffaele Cacciaglia: Spanish Ministry of Science and Innovation (MCIN)/Spanish Research Agency (AEI) MCIN/ AEI/10.13039/501100011033 PID2021-125433OA-100 / Verónica Vilaplana: Spanish Research Agency (AEI) PID2020-116907RB-I00 of the call MCIN/ AEI /10.13039/501100011033 Peer Reviewed Article signat per 27 autors/es: Irene Cumplido-Mayoral, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Marina García-Prat, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Grégory Operto, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Carles Falcon, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Mahnaz Shekari, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Raffaele Cacciaglia, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Marta Milà-Alomà, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Luigi Lorenzini, Department of Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Amsterdam, Netherlands; Silvia Ingala, Department of Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Amsterdam, Netherlands; Alle Meije Wink, Department of Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Amsterdam, Netherlands; Henk JMM Mutsaerts, Department of Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Amsterdam, Netherlands; Carolina Minguillón, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Karine Fauria, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; José Luis Molinuevo, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Sven Haller, CIRD Centre d'Imagerie Rive Droite, Geneva, Switzerland; Gael Chetelat, Normandie Univ, UNICAEN, INSERM, U1237, Caen, France; Adam Waldman, Centre for Dementia Prevention, University of Edinburgh, Edinburgh, United Kingdom; Adam J Schwarz, Takeda Pharmaceutical Company Ltd, Cambridge, United States; Frederik Barkhof, Department of Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Amsterdam, Netherlands; Ivonne Suridjan, Roche Diagnostics International Ltd, Rotkreuz, Switzerland; Gwendlyn Kollmorgen, Roche Diagnostics GmbH, Penzberg, Germany; Anna Bayfield, Roche Diagnostics GmbH, Penzberg, Germany; Henrik Zetterberg, Institute of Neuroscience and Physiology, University of Gothenburg, Gothenburg, Sweden; Kaj Blennow, Institute of Neuroscience and Physiology, University of Gothenburg, Mölndal, Sweden; Marc Suárez-Calvet, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain; Verónica Vilaplana, Department of Signal Theory and Communications, Universitat Politècnica de Catalunya, Barcelona, Spain; Juan Domingo Gispert López, Barcelonaβeta Brain Research Center, Pasqual Maragall Foundation, Barcelona, Spain

The last few years have seen an increasing wave of attacks with serious economic and privacy damages, which evinces the need for accurate Network Intrusion Detection Systems (NIDS). Recent works propose the use of Machine Learning (ML) techniques for building such systems (e.g., decision trees, neural networks). However, existing ML-based NIDS are barely robust to common adversarial attacks, which limits their applicability to real networks. A fundamental problem of these solutions is that they treat and classify flows independently. In contrast, in this paper we argue the importance of focusing on the structural patterns of attacks, by capturing not only the individual flow features, but also the relations between different flows (e.g., the source/destination hosts they share). To this end, we use a graph representation that keeps flow records and their relationships, and propose a novel Graph Neural Network (GNN) model tailored to process and learn from such graph-structured information. In our evaluation, we first show that the proposed GNN model achieves state-of-the-art results in the well-known CIC-IDS2017 dataset. Moreover, we assess the robustness of our solution under two common adversarial attacks, that intentionally modify the packet size and inter-arrival times to avoid detection. The results show that our model is able to maintain the same level of accuracy as in previous experiments, while state-of-the-art ML techniques degrade up to 50% their accuracy (F1-score) under these attacks. This unprecedented level of robustness is mainly induced by the capability of our GNN model to learn flow patterns of attacks structured as graphs., Comment: 7 pages, 4 figures, 1 table

The goal of the presented thesis is to find a classifier able to recognize cycles in different vehicle models from a company. The problem faced is an unsupervised type, so there is no prior information on how many cycles exist. To deal with it, two contrastive methodologies were used. Once the model is trained and tuned, the generated encoder is used for downstream tasks, in this case, the space generated is used to apply a KMeans model to find clusters and therefore, the cycles within. With this cluster, a classifier is trained to have a final model able to be used in real-time. Previous work has been done by the company without applying any contrastive methodology and without any automation. Therefore, automation of the processes is required by the company. Silhouette score is used to check the performance and compare the results. Contrastive learning methodologies outperform the current solution of the company when it is applied to a specific vehicle model

Machine learning has received increasing attention in computer science in recent years and many types of methods have been proposed. In computer networks, little attention has been paid to the use of ML for fault detection, the main reason being the lack of datasets. This is motivated by the reluctance of network operators to share data about their infrastructure and network failures. In this paper, we attempt to fill this gap using anomaly detection techniques to discern hardware failure events in wireless community networks. For this purpose we use 4 unsupervised machine learning, ML, approaches based on different principles. We have built a dataset from a production wireless community network, gathering traffic and non-traffic features, e.g. CPU and memory. For the numerical analysis we investigated the ability of the different ML approaches to detect an unprovoked gateway failure that occurred during data collection. Our numerical results show that all the tested approaches improve to detect the gateway failure when non-traffic features are also considered. We see that, when properly tuned, all ML methods are effective to detect the failure. Nonetheless, using decision boundaries and other analysis techniques we observe significant different behavior among the ML methods. This work has received funding through the DiPET CHIST-ERA under grant agreement PCI2019-111850-2; Spanish grant PID2019-106774RB-C21; Romanian DIPET (62652/15.11.2019) project funded via PN 124/2020; and has been partially supported by the EU research project SERRANO (101017168) and hardware resources courtesy of the Romanian Ministry of Research and Innovation UEFISCDI COCO research project PN III-P4-ID-PCE-2020-0407.

Novelty detection is a statistical method that verifies new or unknown data, determines whether these data are inliers (within the norm) or outliers (outside the norm), and can be used, for example, in developing classification strategies in machine learning systems for industrial applications. To this end, two types of energy that have evolved over time are solar photovoltaic and wind power generation. Some organizations around the world have developed energy quality standards to avoid known electric disturbances; however, their detection is still a challenge. In this work, several techniques for novelty detection are implemented to detect different electric anomalies (disturbances), which are k-nearest neighbors, Gaussian mixture models, one-class support vector machines, self-organizing maps, stacked autoencoders, and isolation forests. These techniques are applied to signals from real power quality environments of renewable energy systems such as solar photovoltaic and wind power generation. The power disturbances that will be analyzed are considered in the standard IEEE-1159, such as sag, oscillatory transient, flicker, and a condition outside the standard attributed to meteorological conditions. The contribution of the work consists of the development of a methodology based on six techniques for novelty detection of power disturbances, under known and unknown conditions, over real signals in the power quality assessment. The merit of the methodology is a set of techniques that allow to obtain the best performance of each one under different conditions, which constitutes an important contribution to the renewable energy systems.

The diagnosis and treatment of epilepsy depend on accurate seizure detection. In clinical practice, the evaluation of seizures is done by visual inspection of an electroencephalogram (EEG). it is very time­consuming and requires trained experts. Automatic seizure detection is important. Machine learning approaches are intensely being applied to this problem due to their ability to classify seizure conditions from a large amount of data, and provide pre­screened results for neurologists. This work proposes a variety of experiments with different machine­learning architectures (support vector machine SVM, K nearest neighbour KNN, random forest RF, feef forward neural network FFNN and convolutional neural network CNN) for the detection of epileptic seizures using multichannel EEG signals from the CHBT­MIT Scalp EEG Database. The best model built in this work contains a combination of a feed­forward neural network (FFNN) and a convolutional neural network (CNN). CNN input images are constructed by applying short­time Fourier transform (STFT) to electroencephalography (EEG) signals and then merged with statistical metrics into a FFNN. The best model of this project showed an outstanding performance of 98.615% accuracy, 98.737% sensitivity and 98.425% specificity. This work also includes a discussion of other exciting ideas that could lead to future research investigations.

En aquest treball s’ha estudiat la viabilitat d’una xarxa neuronal siamesa encarregada de calcular l'índex de similitud entre les cares de dues persones. L’objectiu principal de l’estudi és obtenir una xarxa neuronal amb una precisió del 75% per a demostrar la possibilitat de crear una aplicació comercial que sigui capaç d’aparellar imatges d’usuaris amb imatges de famosos segons la seva semblança. Aquest estudi s’ha realitzat a través d’iteracions o experiments on s’han desenvolupat diversos mòduls per a obtenir diferents configuracions de xarxes neuronals. Finalment, s'ha obtingut un codi per a poder calcular un índex de similitud entre una imatge introduïda manualment i totes les imatges de famosos del conjunt de dades CelebA

The project consist of the study of the popularity of online news, with given data, the objective is to analyze the popularity index of the news by training Machine Learning models and evaluating their performance. El proyecto consiste en el estudio de la popularidad de las noticias online, con datos dados, el objetivo es analizar el índice de popularidad de las noticias entrenando modelos de Machine Learning y evaluando su desempeño. El projecte consisteix en l'estudi de la popularitat de les notícies en línia, amb dades donades, l'objectiu és analitzar l'índex de popularitat de les notícies formant models d'aprenentatge automàtic i avaluant-ne el rendiment.

This project is focused on the development of an audio-denoising system to improve the robustness of audio fingerprinting algorithms. The state of the art on deep learning has advanced in the last years, specifically in multimedia processing, this opens a door to using this kind of tool for the described task. The goal is to develop software that learns to clean noisy audio samples so the cleaned samples are better matched. First, a review of the previous work on different topics is made. Then, different ways to create a dataset are presented as well as a deep neural network to clean the audio, also an evaluation step has been developed to compare fairly the different models trained. Results suggest that the neural network is learning the task to perform, despite this, there is a large margin to improve the evaluation results. Furthermore, the limitations in terms of memory usage and inference speed are described. Finally, a discussion introduces the necessity of a larger and more robust dataset and the idea of tailoring the dataset to the collection of songs to be denoised. Este proyecto se centra en el desarrollo de un sistema de eliminación de ruido de audio para mejorar la robustez de los algoritmos de detección basados en la extracción de huellas digitales acústicas. El estado del arte en el aprendizaje profundo ha avanzado en los últimos años, específicamente en el procesamiento multimedia, lo que abre una puerta para usar este tipo de herramienta para la tarea descrita. El objetivo es desarrollar un software que aprenda a limpiar muestras de audio ruidosas para que las muestras limpias se identifiquen mejor. Primeramente, realizamos una revisión del trabajo previo sobre diferentes temas. Luego, presentamos diferentes formas de crear un conjunto de datos, así como una red neuronal profunda para limpiar el audio, también hemos desarrollado un módulo de evaluación para comparar de manera justa los diferentes modelos entrenados. Los resultados sugieren que la red neuronal está aprendiendo la tarea a realizar, a pesar de esto, hay un gran margen para mejorar los resultados de evaluación. Además, describimos las limitaciones en términos de uso de memoria y velocidad de inferencia. Finalmente, se presenta la discusión de la necesidad de un conjunto de datos más grande y robusto y la idea de adaptar el conjunto de datos a la colección de canciones que queremos detectar. Aquest projecte se centra en el desenvolupament d'un sistema d'eliminació de soroll d'àudio per millorar la robustesa dels algoritmes d'empremtes digitals d'àudio. L'estat de l'art en l'aprenentatge profund ha avançat en els darrers anys, específicament en el processament multimèdia, això obre una porta per utilitzar aquest tipus d'eina per a la tasca descrita. L'objectiu és desenvolupar un programari que aprèn a netejar mostres d'àudio amb soroll perquè les mostres netes es detectin millor. En primer lloc, realitzem una revisió del treball previ sobre diferents temes. Després, presentem diferents formes de crear un conjunt de dades, així com una xarxa neuronal profunda per netejar l'àudio, també hem desenvolupat un pas d'avaluació per comparar de manera justa els diferents models entrenats. Els resultats suggereixen que la xarxa neuronal està aprenent la tasca a fer, malgrat això, hi ha un gran marge per millorar els resultats d'avaluació. A més, presentem les limitacions en termes d'ús de memòria i velocitat d'inferència. Finalment, presentem la discussió de la necessitat d'un conjunt de dades més gran i robust i la idea de personalitzar el conjunt de dades per a la col·lecció de cançons que es vulguin detectar.

Single-atom catalytic sites may have existed in all supported transition metal catalysts since their first application. Yet, interest in the design of single-atom heterogeneous catalysts (SACs) only really grew when advances in transmission electron microscopy (TEM) permitted direct confirmation of metal site isolation. While atomic-resolution imaging remains a central characterization tool, poor statistical significance, reproducibility, and interoperability limit its scope for deriving robust characteristics about these frontier catalytic materials. Here, we introduce a customized deep-learning method for automated atom detection in image analysis, a rate-limiting step toward high-throughput TEM. Platinum atoms stabilized on a functionalized carbon support with a challenging irregular three-dimensional morphology serve as a practically relevant test system with promising scope in thermo- and electrochemical applications. The model detects over 20,000 atomic positions for the statistical analysis of important properties for establishing structure–performance relations over nanostructured catalysts, like the surface density, proximity, clustering extent, and dispersion uniformity of supported metal species. Good performance obtained on direct application of the model to an iron SAC based on carbon nitride demonstrates its generalizability for single-atom detection on carbon-related materials. The approach establishes a route to integrate artificial intelligence into routine TEM workflows. It accelerates image processing times by orders of magnitude and reduces human bias by providing an uncertainty analysis that is not readily quantifiable in manual atom identification, improving standardization and scalability., Journal of the American Chemical Society, 144 (18), ISSN:0002-7863, ISSN:1520-5126

Recent advancements in vehicle-to-everything (V2X) communication have notably improved existing transport systems by enabling increased connectivity and driving autonomy levels. The remarkable benefits of V2X connectivity come inadvertently with challenges which involve security vulnerabilities and breaches. Addressing security concerns is essential for seamless and safe operation of mission-critical V2X use cases. This paper surveys current literature on V2X security and provides a systematic and comprehensive review of the most relevant security enhancements to date. An in-depth classification of V2X attacks is first performed according to key security and privacy requirements. Our methodology resumes with a taxonomy of security mechanisms based on their proactive/reactive defensive approach, which helps identify strengths and limitations of state-of-the-art countermeasures for V2X attacks. In addition, this paper delves into the potential of emerging security approaches leveraging artificial intelligence tools to meet security objectives. Promising data-driven solutions tailored to tackle security, privacy and trust issues are thoroughly discussed along with new threat vectors introduced inevitably by these enablers. The lessons learned from the detailed review of existing works are also compiled and highlighted. We conclude this survey with a structured synthesis of open challenges and future research directions to foster contributions in this prominent field. This work is supported by the H2020-INSPIRE-5Gplus project (under Grant agreement No. 871808), the ”Ministerio de Asuntos Económicos y Transformacion Digital” and the European Union-NextGenerationEU in the frameworks of the ”Plan de Recuperación, Transformación y Resiliencia” and of the ”Mecanismo de Recuperación y Resiliencia” under references TSI-063000-2021-39/40/41, and the CHIST-ERA-17-BDSI-003 FIREMAN project funded by the Spanish National Foundation (Grant PCI2019-103780).

El mercado de activos y criptomonedas ha crecido significativamente en los últimos años al igual que el número de jóvenes interesados en invertir y ganar dinero "fácil". Este aumento, sin embargo, no se ha acompañado de un incremento en formación, resultando en un gran número de errores evitables y la pérdida de grandes cantidades de dinero. Una metodología comúnmente utilizada para minimizar dichos riesgos consiste en consultar las redes sociales en busca de expertos que orienten o de identificar tendencias. El uso de la inteligencia artificial puede ser muy beneficioso para automatizar los pasos seguidos en esta metodología, en combinación con métricas como la cantidad de menciones o el sentimiento asociado a los comentarios de las redes sociales. En este proyecto se pretende desarrollar una aplicación capaz de obtener y procesar datos de redes sociales en tiempo real con la que fomentar la inversión consciente y no especulativa, valorando también el rendimiento de la base de datos utilizada y creando una página web donde los usuarios puedan consultar las distintas métricas generadas. The asset market has grown significantly in the last few years as well as the number of young people interested in investing and making "easy" money. This increase in demand, however, has not been met with an increase in formation or education, resulting in lots of avoidable mistakes and the loss of considerable amounts of money. A common approach on minimizing losses consists of using social media to look for the guidance of specialists or try to identify the latest trends. The use of artificial intelligence can be highly beneficial to automate the steps followed by this approach along with the use of metrics such as the number of mentions or the sentiment of the comments. The purpose of this project is to develop an app capable of obtaining and processing data from social media in real time, automating the process of identifying trends and promoting conscious investment against speculation. Besides the AI engine behind the app, this project also focuses on the efficiency and performance of the chosen database and the implementation of a webpage where users can see the processed data and generated metrics.

L'aprenentatge automàtic, un tipus particular d'intel·ligència artificial, s'utilitza cada vegada més en els sistemes de software modern. A més de requerir molta energia, el seu ús augmenta les emissions de CO2, té un elevat cost financer i consumeix molts recursos. Encara que existeixen nombrosos estudis que analitzen les possibilitats del ML, molt pocs d'ells se centren en les seves característiques ecològiques, com l'ús d'energia o la seva petjada de carboni. Aquest estudi tracta d'augmentar la conscienciació dels costos associats al desplegament d'un sistema de ML en algunes de les plataformes d'allotjament en el núvol més avançades. Per a això, duem a terme un ampli estudi empíric per a avaluar i contrastar les emissions i la latencia de quatre proveïdors de núvol diferents, Amazon Web Services, Azure, Heroku i Railway Machine learning, a particular type of artificial intelligence, is increasingly being used in modern software systems. In addition to requiring a lot of energy, its use raises CO2 emissions, has a high financial cost, and consumes many resources. Even while there are numerous studies that look at the possibilities of ML, very few of them concentrate on its green features, like energy consumption or its carbon footprint. This bachelor thesis tries to increase understanding of the costs associated with deploying an ML system into some of the state-of-the-art cloud hosting platforms. In order to do this, we conduct an extensive empirical study to assess and contrast the emissions and latency of four different cloud providers: Amazon Web Services, Azure, Heroku, and Railway.

Aplicar pruning en una xarxa neuronal disminueix la complexitat de la representació interna, permetent així un augment de la velocitat d'entrenament i inferència. A més, en ser un model més senzill, pot millorar la seva capacitat de generalització, per això l'aplicació de tècniques de pruning en xarxes preentrenades pot ajudar a facilitar l'aplicació de transfer learning i aconseguir millors models. En reduir la complexitat del model, també reduïm la quantitat de paràmetres necessaris per entrenar en la nova tasca i, per tant, reduïm la capacitat computacional necessària, facilitant l'aplicació de transfer learning. El nostre treball es basarà a aplicar múltiples mètodes de pruning sobre models preentrenats per tal de conèixer com afecta posteriorment a l'aplicació de transfer learning. Applying pruning on a neural network decreases the complexity of the internal representation, allowing an increase in the speed of training and inference. In addition, producing a simpler model, it can improve its generalization capacity, which is why applying pruning techniques on pre-trained neural networks can help improve the application of transfer learning and achieve better models. By reducing the complexity of the model, we also reduce the number of parameters needed to train on the new task, therefore reducing the computational capacity required, and improving the application of transfer learning. Our work will be based on applying multiple pruning methods on pre-trained models in order to know how it affects the application of transfer learning.

Debido a la globalización que hemos sufrido durante los últimos años, el estudio de la opinión de la población y como esta es influenciada ha resultado en un ámbito cada vez más estudiado. Hasta la fecha, multitud de estudios han observado comportamientos de polarización de la opinión en el ámbito político. Analizando pequeños grupos de preguntas para ver cómo ha cambiado la opinión a lo largo del tiempo. Esta tesis pretende estudiar la incorporación de herramientas de aprendizaje automático para el estudio de la opinión. Estas nos permites trabajar con mayores volúmenes de datos y encontrar relaciones que podrían no encontrarse de otra forma. Estudiaremos el efecto de varios atributos en la opinión de la gente, centrándonos sobre todo en el efecto que tiene la inclinación política en la ideología de la población. También estudiaremos como diferentes tópicos encapsulan una ideología y como estas han evolucionado con el tiempo. Due to the globalization we have undergone in recent years, the study of public opinion and how it is influenced has become an increasingly studied field. To date, many studies have observed behaviors that polarize opinion in the political sphere. Analyzing how small groups of questions have changed over time. This thesis aims to study the incorporation of machine learning tools for the study of opinion. These will allow us to work with larger volumes of data and find relationships that otherwise might not be found. We will study the effect of various attributes on people's opinions, focusing mainly on the effect of political leaning on people's ideology. We will also study how different topics encapsulate an ideology and how these ideologies have varied through time.

This project is my Degree Final Project for my Degree in Informatics Engineering in the Computing specialization, which I am studying at the Facultat d'Informàtica de Barcelona (FIB). I will be supervised by Felix Freitag, who is part of the Distributed Systems Group. In this project I will develop a machine learning application for an embedded device, specifically the Arduino Portenta H7, which is a board with a built in microcontroller and the necessary circuits to be able to add many accessories, like cameras, microphones or LEDs.

Existe una fuerte tendencia a ejecutar modelos de aprendizaje automático (ML) en dispositivos cada vez más pequeños, principalmente gracias al aumento de la capacidad de cómputo de los dispositivos IoT. El Federated Learning (FL) es un enfoque reciente para entrenar modelos de ML de forma distribuida en el que muchos nodos participan en el entrenamiento del modelo [2]. La combinación de estos dos factores, nos brinda la oportunidad de implementar algoritmos de ML en el borde de la red, gracias al esfuerzo conjunto de dichos nodos. Sin embargo, faltan estudios que prueben el aprendizaje federado en entornos reales. A continuación, se presenta un testbed con el que se puede experimentar de forma práctica una implementación de Federated Learning. Para ello, se utilizan dispositivos de baja capacidad como nodos del testbed en una red inalámbrica mallada de ciudad. Se utiliza el testbed para llevar a cabo experimentación sobre protocolos y algoritmos de Federated Learning así como para la orquestación de sus componentes. El proyecto culmina con la realización de una aplicación que cumple con todas las directrices y requisitos del Federated Learning, y se utilizan herramientas de orquestación que ayudan a gestionar los diversos elementos que la componen. La aplicación es ejecutada en un entorno real, mediante el uso de varios dispositivos de baja capacidad, por lo que la computación se realiza en el borde o Edge de la red, como estipula la técnica de FL. There is a strong trend to run machine learning (ML) models on increasingly smaller devices, mainly due to the increase in computing capacity of IoT devices. Federated Learning (FL) is a recent approach for training ML models in a distributed manner in which many nodes participate in the model training [2]. The combination of these two factors provides us with the opportunity to implement ML algorithms at the edge of the network, thanks to the joint effort of these nodes. However, there are few studies that have proven Federated Learning in real-world environments. A testbed is presented here that allows for practical experimentation of a Federated Learning implementation. Low-capacity devices are used as nodes in the testbed in a wireless mesh city network. The testbed is used to conduct experimentation on Federated Learning protocols and algorithms, as well as for orchestration of its components. The project concludes with the development of an application that meets all Federated Learning guidelines and requirements, using orchestration tools to manage the various elements that make it up. The application is run in a real environment using several low-capacity devices, so the computation is done at the edge of the network, as stipulated by the FL technique.

Machine learning algorithms and the increasing availability of data have radically changed the way how decisions are made in today’s Industry. A wide range of algorithms are being used to monitor industrial processes and predict process variables that are difficult to be measured. Maintenance operations are mandatory to tackle in all industrial equipment. It is well known that a huge amount of money is invested in operational and maintenance actions in industrial gas turbines (IGTs). In this paper, two variations of autoencoders were used to analyse the performance of an IGT after major maintenance. The data used to analyse IGT conditions were ambient factors, and measurements were performed using several sensors located along the compressor. The condition assessment of the industrial gas turbine compressor revealed significant changes in its operation point after major maintenance; thus, this indicates the need to update the internal operating models to suit the new operational mode as well as the effectiveness of autoencoder-based models in feature extraction. Even though the processing performance was not compromised, the results showed how this autoencoder approach can help to define an indicator of the compressor behaviour in long-term performance. This research was funded by Siemens Energy.

18th International Conference on Automation Science and Engineering (CASE), 20-24 August 2022.-- 8 pages, 7 figures, 1 table, Underwater target localization using range-only and single-beacon (ROSB) techniques with autonomous vehicles has been used recently to improve the limitations of more complex methods, such as long baseline and ultra-short baseline systems. Nonetheless, in ROSB target localization methods, the trajectory of the tracking vehicle near the localized target plays an important role in obtaining the best accuracy of the predicted target position. Here, we investigate a Reinforcement Learning (RL) approach to find the optimal path that an autonomous vehicle should follow in order to increase and optimize the overall accuracy of the predicted target localization, while reducing time and power consumption. To accomplish this objective, different experimental tests have been designed using state-of-the-art deep RL algorithms. Our study also compares the results obtained with the analytical Fisher information matrix approach used in previous studies. The results revealed that the policy learned by the RL agent outperforms trajectories based on these analytical solutions, e.g. the median predicted error at the beginning of the target’s localisation is 17% less. These findings suggest that using deep RL for localizing acoustic targets could be successfully applied to in-water applications that include tracking of acoustically tagged marine animals by autonomous underwater vehicles. This is envisioned as a first necessary step to validate the use of RL to tackle such problems, which could be used later on in a more complex scenarios, This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 893089. This work also received financial support from the Spanish Ministerio de Econom´ıa y Competitividad (SASES: RTI2018-095112-B-I00; BITER-ECO: PID2020-114732RBC31). This work acknowledges the ‘Severo Ochoa Centre of Excellence’ accreditation (CEX2019-000928-S), and from the Generalitat de Catalunya ”Sistemas de Adquisición Remota de datos y Tratamiento de la Informaci´on en el Medio Marino (SARTI-MAR)” 2017 SGR 376

In this study, we present a platform-portable deep reinforcement learning method that has been used as a path-planning system to localize underwater objects with autonomous vehicles. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 893089. This work also received financial support from the Spanish Ministerio de Economía y Competitividad (BITERECO: PID2020-114732RBC31). This work acknowledges the ’Severo Ochoa Centre of Excellence’ accreditation (CEX2019-000928-S).

Segmentation of 3D images to identify cells and their molecular outputs can be difficult and tedious. Machine learning algorithms provide a promising alternative to manual analysis as emerging 3D image processing technology can save considerable time. For those unfamiliar with machine learning or 3D image analysis, the rapid advancement of the field can make navigating the newest software options confusing. In this paper, two open-source machine learning algorithms, Cellpose and Stardist, are compared in their application on a 3D light sheet dataset counting fluorescently stained proliferative cell nuclei. The effects of image tiling and background subtraction are shown through image analysis pipelines for both algorithms. Based on our analysis, the relative ease of use of Cellpose and the absence of need to train a model leaves it a strong option for 3D cell segmentation despite relatively longer processing times. When Cellpose's pretrained model yields results that are not of sufficient quality, or the analysis of a large dataset is required, Stardist may be more appropriate. Despite the time it takes to train the model, Stardist can create a model specialized to the users' dataset that can be iteratively improved until predictions are satisfactory with far lower processing time relative to other methods. This project has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 841047 and NSF CMMI #1727518.

Unmanned aerial vehicles (UAV), also known as drones have been used for a variety of reasons and the commercial drone market growth is expected to reach remarkable levels in the near future. However, some drone users can mistakenly or intentionally fly into flight paths at major airports, flying too close to commercial aircraft or invading people’s privacy. In order to prevent these unwanted events, counter-drone technology is needed to eliminate threats from drones and hopefully they can be integrated into the skies safely. There are various counter-drone methods available in the industry. However, a counter-drone system supported by an artificial intelligence (AI) method can be an efficient way to fight against drones instead of human intervention. In this paper, a deep reinforcement learning (DRL) method has been proposed to counter a drone in a 3D space by using another drone. In a 2D space it is already shown that the deep reinforcement learning method is an effective way to counter a drone. However, countering a drone in a 3D space with another drone is a very challenging task considering the time required to train and avoid obstacles at the same time. A Deep Q-Network (DQN) algorithm with dueling network architecture and prioritized experience replay is presented to catch another drone in the environment provided by an Airsim simulator. The models have been trained and tested with different scenarios to analyze the learning progress of the drone. Experiences from previous training are also transferred before starting a new training by pre-processing the previous experiences and eliminating those considered as bad experiences. The results show that the best models are obtained with transfer learning and the drone learning progress has been increased dramatically. Additionally, an algorithm which combines imitation learning and reinforcement learning is implemented to catch the target drone. In this algorithm, called deep q-learning from demonstrations (DQfD), expert demonstrations data and self-generated data by the agent are sampled and the agent continues learning without overwriting the demonstration data. The main advantage of this algorithm is to accelerate the learning process even if there is a small amount of demonstration data. This work was funded partially by the AGAUR under grant 2020PANDE00141, the Ministry of Science and Innovation of Spain under grant PID2020-116377RB-C21 and the SESAR Joint Undertaking (JU) project CORUS-XUAM, under grant SESAR-VLD2 101017682. The JU receives support from the European Union’s Horizon 2020 research and innovation program and SESAR JU members other than the Union.

© 2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. In 5G and beyond radios, the increased bandwidth, the fast-changing waveform scenarios, and the operation of large array multiple-input multiple-output (MIMO) transmitter architectures have challenged both the polynomial and the artificial neural network (ANN) MIMO adaptive digital predistortion (DPD) schemes. This article proposes training data selection methods and dimensionality reduction techniques that can be combined to enable relevant reductions of the DPD training time and the implementation complexity for MIMO transmitter architectures. In this work, the combination of an efficient uncorrelated equation selection (UES) mechanism together with orthogonal least squares (OLS) is proposed to reduce the training data length and the number of basis functions at every behavioral modeling matrix in the polynomial MIMO DPD scheme. For ANN MIMO DPD architectures, applying UES and principal component analysis (PCA) is proposed to reduce the input dataset length and features, respectively. The UES-OLS and the UES-PCA techniques are experimentally validated for a 2×2 MIMO test setup with strong power amplifier (PA) input and output crosstalk. This work was supported in part by the MCIN/AEI/10.13039/501100011033 under Project PID2020-113832RB-C22 and Project PID2020-113832RB-C21; and in part by the European Union-NextGenerationEU through the Spanish Recovery, Transformation and Resilience Plan, under Project TSI-063000-2021-121 (MINECO UNICO Programme).

The growth of data and its storage is becoming more and more important every day. However, occasionally this information is gathered but never used, or perhaps it is improperly gathered, making the extraction of the insides difficult. As a result, while beginning any project, choosing the analysis method is just as crucial as choosing the design of the data collection strategy. Most of the time, we only focus on the analysis of the data and do not consider how it was gathered or whether the fields were actually valuable or just added noise to what we were searching for. For this reason, a trustworthy data set has been chosen for this project. The data came from a telecom company, which, like other modern businesses, collects a lot of data. However, in this case, the data was published on the machine learning web competition Kaggle, where participants competed to build the best model to predict consumer behaviour. One of the key considerations in optimizing any organization's income is preventing customer churn. It happens when customers quit utilizing a company's goods or services, and is also referred to as customer attrition. The main goal of this master's thesis is to analyse a Churn database and categorise the clients in order to determine whether they are likely to leave the company. To do this, two machine learning techniques will be used in the current document. Extreme Gradient Boosting and Random Forest. In order to achieve high performance, the Random Forest (RF) method creates a large number of low-performance models and combines them. In this case, the lower-performance method is called Decision Tree, so it will be explained in more detail in the following document. Similar work is done by eXtreme Gradient Boosting (XGB), although it builds new models based on earlier findings. Both are quite effective predictor models, even with unbalanced data, as will be demonstrated in the next document. This adds another level of complexity that the algorithms must overcome to execute effectively. Different performance indicators will be provided and examined in order to determine which one is the greatest indicator to choose the best model during the process of determining the best model. Sensitivity, Specificity, Precision, F1 Score, and Geometric Mean are a few of the markers that are listed. Additionally, their trends for the various parameter values of the examined models will be shown and analysed. The strong performance of these machine learning algorithms will once more be supported in this thesis. the affirmation of the significance and practical use of these methodologies, as in the case of this project, to comprehend processes and behaviours. All fields can benefit from the information gleaned, and a successful application will undoubtedly yield financial rewards. The two machine learning applied algorithms' default and best models are finally shown, and their advantages and disadvantages will be evaluated while taking into account the many scenarios that exist. This thesis will demonstrate the good performance of both models, with XGB significantly outperforming RF. It will also demonstrate that while XGB performs better on precision and RF has better results on sensitivity

Data anonymization is an important tool for protecting personal privacy and prevent identification of individuals. δ-DOCA is a semantic anonymization algorithm that protects datasets by modifying them such that they are similar to the original dataset in a semantic sense, while ensuring that the original dataset cannot be re-identified. This thesis proposes a framework that, using Machine Learning techniques, evaluates to what extent the protected data reveals information about the original data and the impact of this protection on the quality of the data. This is done by training a set of models with the protected data and comparing their performance to the performance of models trained with the original data. Results show that the protection deteriorates the quality of the data, which is reflected in the performance of the models, and that there is a trade-off between privacy and accuracy.

Non-Legitime traffic in terms of automated internet bot traffic is a long-standing problem causing a huge economic impact and lack of trust in companies and administrations worldwide. For years, Artificial Intelligence and especially Machine Learning have been a key players fighting and helping the stakeholder to analyse and detect fraud instances automatically. However, it does not exist a reliable ground truth public dataset to evaluate and compare the proposed methodologies in the literature. Throughout this thesis, it is developed a public dataset consisting of legitimate and fraudulent web mouse movements extracted from real bot engines. In addition, it is evaluated using two Machine Learning models based on Decisions Tree classifier called LightGBM whilst the second one is based on Recurrent Neural Networks outperforming the accuracy El tráfico no legítimo en términos de tráfico automatizado de bots de Internet es un problema que se perpetua durante el tiempo. Este tráfico provoca un gran impacto económico y desconfianza en empresas y administraciones en todo el mundo. Durante años, la Inteligencia Artificial y especialmente el Machine Learning han sido un jugador clave para lucha y ayuda a la parte interesada a analizar y detectar instancias de fraude automáticamente. Sin embargo, no existe un conjunto de datos públicos fiables y verídicos para evaluar y comparar las metodologías propuestas en la literatura. A lo largo de esta tesis, se ha desarrolado un conjunto de datos público que consiste en movimientos legítimos y fraudulentos de ratón. Además, se evalúa mediante dos modelos Machine Learning basados en un clasificador Decisions Tree llamado LightGBM y el segundo se basa en Redes Neuronales Recurrentes. El tràfic no legítim en termes de trànsit automatitzat de bots d'Internet és un problema que es perpetua durant el temps. Aquest tràfic provoca un gran impacte econòmic i desconfiança en empreses i administracions en tot el món. Durant anys, la Intel·ligència Artificial i especialment el Machine Learning han estat un jugador clau per a la lluita i l'ajuda a la part interessada a analitzar i detectar instàncies de frau automàticament. Però, no existeix un conjunt de dades públiques fiables i verídiques per a avaluar i comparar les metodologies propostes en la literatura. Alllarg d'aquesta tesi, s'ha desenvolupat un conjunt de dades públics que consisteix en moviments legítims i fraudulents de ratolí. A més, s'avalua mitjançant dos models de Machine Learning basats en un classificador Decisions Tree anomenat LightGBM i el segon es basa en Xarxes Neuronals Recurrents.

Having an increasing number of patients in the emergency department constitutes an obstacle to the admissions process and hinders the emergency department (ED)’s ability to deal with the continuously arriving demand for new admissions. In addition, forecasting is an important aid in many areas of hospital management, including elective surgery scheduling, bed management, and staff resourcing. Therefore, this paper aims to develop a precise prediction model for admissions in the Integral Healthcare System for Public Use in Catalonia. These models assist in reducing overcrowding in emergency rooms and improve the quality of care offered to patients. Data from 60 EDs were analyzed to determine the likelihood of hospital admission based on information readily available at the time of arrival in the ED. The first part of the study targeted the obtention of models with high accuracy and area under the curve (AUC), while the second part targeted the obtention of models with a sensitivity higher than 0.975 and analyzed the possible benefits that could come from the application of such models. From the 3,189,204 ED visits included in the study, 11.02% ended in admission to the hospital. The gradient boosting machine method was used to predict a binary outcome of either admission or discharge. This research was funded by Ministerio de Ciencia e Innovación Torres Quevedo grant number PTQ2021-012147. Peer Reviewed Objectius de Desenvolupament Sostenible::3 - Salut i Benestar Objectius de Desenvolupament Sostenible::3 - Salut i Benestar::3.8 - Assolir la cobertura sanitària universal, en particular la protecció contra els riscos financers, l’accés a serveis de salut essencials de qualitat i l’accés a medicaments i vacunes segurs, eficaços, assequibles i de qualitat per a totes les persones

El aprendizaje automático (Machine Learning ML), aparece por primera vez alrededor de los años cincuenta, pero en los últimos años se ha vivido una gran expansión en este campo. Cada día el ML se ve más involucrado en la vida de las personas con la finalidad de mejorarla, es evidente que todavía queda mucho por descubrir acerca de este campo. Este trabajo de final de grado es una aportación a un proyecto de Machine Learning Function Orchestrator (MLFO), el trabajo consiste en la implementación de una base de datos centralizada para un orquestador de ML, lo que nos permite almacenar toda la información de una manera eficiente. Para almacenar esta información es necesario analizar los datos recibidos según el formato, organizándolos en tablas de la manera más estructurada posible. Finalmente, una vez almacenada esta información, utilizaremos un algoritmo de optimización para transmitir al MLFO el mejor mapeo de la infraestructura, reduciendo la utilización de los recursos, para asignar las diferentes funcionalidades de ML a cada data center. L’aprenentatge automàtic (Machine Learning ML), apareix per primer cop al voltant dels anys cinquanta, però és als últims anys on s’ha viscut una gran expansió dins d’aquest camp. Cada dia el ML es veu més involucrat en la vida de les persones amb la finalitat de millorar-la, és evident que encara hi ha molt per descobrir al voltant d’aquest camp. Aquest treball de fi de grau és una aportació a un projecte de Machine Learning Function Orchestrator (MLFO), el treball consisteix en la implementació d’una base de dades centralitzada per a un orquestrador de ML, permetent-nos emmagatzemar tota la informació d’una manera eficient. Per emmagatzemar aquesta informació és necessari analitzar totes les dades rebudes segons el format, organitzant-les en taules de la forma més estructurada possible. Finalment, un cop hem emmagatzemat aquesta informació, utilitzarem un algoritme d’ptimització per transmetre al MLFO el millor mapeig de la infraestructura, reduint la utilització dels recursos, per assignar les diferents funcionalitats de ML a cada data center. Machine Learning (ML) first appears around the 1950s, but it is in recent years when this field has grown exponentially. Every day machine learning is more involved in people’s lives in order to improve it, but it is clear that there is still much to discover about this field. This final degree thesis is a contribution to a Machine Learning Function Orchestration (MLFO) project, this thesis consists of the implementation of a centralized data base for a ML orchestrator, and with that implementation it is possible to store data in an efficient way. To load that amount of data is necessary to analyse whole information and store in different tables in a structured way. Finally, once this information is stored, an optimization algorithm will be used to transmit to the MLFO the best mapping of the infrastructure, reducing the use of resources, to assign the ML functions into data centres.

Symbolic Regression (SR) is a Machine Learning method that seeks for a mathematical expression to find the best model given a dataset. This remains a really hard task as, unlike classic Machine Learning, both model parameters and model structure are unknown. Typically, algorithms for Symbolic Regression have been based on classical heuristic search algorithms as Genetic Programming, Simulated Annealing or Bayesian approaches. Nevertheless, in the last years a new movement trying to tackle the problem using Neural Networks has arisen. The results obtained show a promising research pathway but the usage of these networks for SR is still under-explored. In this thesis, we propose a data-dependent Transformer Model trained in two phases (supervised phase and Reinforcement Learning phase) to improve the state-of-the-art in Symbolic Regression. The performance of our model is compared to its direct predecessor, NeSymRes and, gplearn, a widely used algorithm based on Genetic Programming. The results obtained demonstrated that we have succeeded in building a model that is able to create high accurate and low complex equations given a dataset (X, Y) as an input. Our model outperforms the compared algorithms in terms of accuracy, recoverability rate, complexity and execution time. Most importantly, we help to settle the usage of Neural Networks for SR as a promising and interesting future line of research.

In recent years there has been a major increase in malicious activities on the Internet, which caused major disruptions and sever economic and privacy implications. For this reason, the research group with which this Thesis will be developed published the paper "Unveiling the potential of Graph Neural Networks for robust Intrusion Detection". In this paper, the authors propose "a novel Graph Neural Network (GNN) model that uses a non-standard message-passing architecture specially designed to process and learn from host-connection graphs". In, the training and evaluation of the model has been done in its entirety with the well- known CIC-IDS2017, thus further evaluation and testing is needed. To this end, the model has to be modified and each dataset has to be adequately preprocessed for the evaluation of the model, and if necessary include it in the training has to be modified and each dataset has to be adequately preprocessed for the evaluation of the model, and if necessary include it in the training.

Internet of Things (IoT) has acquired useful and powerful advances thanks to the Machine Learning (ML) implementations. But the implementation of Machine Learning in IoT devices with data centers has some serious problems (data privacy, network bottleneck, etc). Tiny Machine Learning (TinyML) arose in order to have an independent edge device executing the ML program without the necessity of any data center. But there is still the need for high performance computers to train the ML model. But, can this situation improve? This project goes through TinyML and two TinyML techniques capable to train the ML model on-device (what we call TinyML On-Device Learning or TinyODL): TinyML with Online-Learning (TinyOL) and Federated Learning (FL). We study both techniques in a theoretical analysis and try to develop one TinyODL app. Internet of Things (IoT) ha obtingut uns forts avantatges molt usables gràcies a les implementacions del Machine Learning (ML). Però la implementació del Machine Learning en dispositius IoT utilitzant centres de dades porta una sèrie de problemes a tenir en compte (privacitat de les dades, el coll d'ampolla de la xarxa, etc.). Tiny Machine Learning (TinyML) va sorgir amb l'objectiu de tenir dispositious IoT independents executant el programa d'ML sense la necessitat d'un centre de dades. Però encara hi ha la necessitat de fer servir ordinadors d'alta potència per poder entrenar el model d'ML. Així i tot, es pot millorar aquesta situació? Aquest projecte estudia el TinyML i dues de les seves tècniques, del que anomenem TinyML On-Device Learning o TinyODL, capaces d'entrenar el model d'ML en el mateix dispositiu (on-device learning): TinyML with Online Learning (TinyOL) i Federated Learning (FL). S'estudien les dues tècniques des d'una anàlisi teòrica i provem de desenvolupar una aplicació TinyODL. Internet of Things (IoT) ha obtenido unas muy buenas y usables mejoras gracias a las implementaciones del Machine Learning (ML). Pero la implementación de Machine Learning en dispositivos IoT utilizando centros de datos conlleva una serie de problemas a tener en cuenta (privacidad de los datos, el cuello de botella de la red, etc.). Tiny Machine Learning (TinyML) surgió con el objetivo de tener dispotivios IoT independientes ejecutando el programa de ML sin la necesidad de un centro de datos. Pero aún existe la necesidad de usar ordenadores de alta potencia para poder entrenar el modelo de ML. Aún así, se puede mejorar esta situación? Este proyecto estudia el TinyML y dos de sus técnicas, de lo que llamamos TinyML On-Device Learning o TinyODL, capaces de entrenar el model de ML en el mismo dispotivio (on-device learning): TinyML with Online Learning (TinyOL) y Federated Learning (FL). Se estudian las dos técnicas desde un anáisis teórico y probamos de desarrollar una aplicación TinyODL.

Finding automatization opportunities for email interactions can have positive effects for several industries, especially in tasks such as reading, receiving, writing and responding emails, categorizing emails or even to prevent loss of productivity and financial loses by dealing with spam, or improve users' satisfaction; even improving automatic categorization systems can mitigate negative impacts on personal and organization performance. Furthermore, people who work in companies spend around 28 % of their time reading and answering emails. In this project we proposed a methodology based on NLP and Unsupervised Machine Learning to look for opportunities of automation arising from recurrent email patterns found in email texts. We intent to facilitate the linguistic analysis in order to retrieve interaction patterns that can trigger automation actions. We proposed CRISP-DM methodology that lays the groundwork for detection of automatization opportunities in tasks relates. We compared the unsupervised machine learning methods K-Means, DBSCAN, and HDBSCAN with four clustering metrics applied to the Enron e-mails dataset transformed into paragraph vectors and performed several experiments with Word Mover's Distance, Euclidean Distance, L2-Norm and Cosine Similarity. Although our process yielded limited results in the detection of email interactions, we found that DBSCAN combined with Euclidean Distance was the best method among all scores. This project also contributes to the parameterization literature of said clustering algorithms as well as showing which methods, distances and scores settings are relevant for unsupervised email mining.

In recent years, the interest in Artificial Intelligence (AI) has experienced a significant growth, which has contributed to the emergence of new research directions such as Explainable Artificial Intelligence (XAI). The ability to apply AI approaches to solve various problems in many industrial areas has been mainly achieved by increasing model complexity and the use of various black-box models that lack transparency. In particular, deep neural networks are great at dealing with problems that are too difficult for classic machine learning methods, but it is often a big challenge to answer the question why a neural network made such a decision and not another. The answer to this question is extremely important to ensure that ML models are reliable and can be held liable for the decision-making process. Over a relatively short period of time a plethora of methods to tackle this problem have been proposed, but mainly in the area of computer vision and natural language processing. Few publications have been published so far in the context of explainability in time series. This Thesis aims to provide a comprehensive literature review of the research in XAI for time series data as well as to achieve and evaluate local explainability for a model in time series forecasting problem. The solution involved framing a time series forecasting task as a Remaining Useful Life (RUL) prognosis for turbofan engines. We trained two Bi-LSTM models, with and without attention layer, on the C-MAPSS data set. The local explainability was achieved using two post-hoc explainability techniques SHAP and LIME as well as extracting and interpreting the attention weights. The results of explanations were compared and evaluated. We applied the evaluation metric which incorporates the temporal dimension of the data. The obtained results indicate that LIME technique outperforms other methods in terms of the fidelity of local explanations. Moreover, we demonstrated the potential of attention mechanisms to make a deep learning model for time series forecasting task more interpretable. The approach presented in this work can be easily applied to any time series forecasting or classification scenario in which we want to achieve model interpretability and evaluation of generated explanations.

L'objectiu d'aquest projecte és facilitar el desenvolupament de projecte sobre Machine Learning per a que pugui ser entenedor per a tothom relacionat amb aquest. Per a realitzar això, s'ha desenvolupat una plantilla que conté un decorador introduïnt MLFlow per a la monitorització d'aquests projectes, així com un seguit de mecanismes que fa que executar-lo estigui a l'abast de tothom. Finalment, s'han duit a terme un seguit de proves amb el projecte inicial juntament amb la plantilla final generada. The objective of this purpose is to facilitate the development of projects related to Machine Learning so that they can be understood by everyone related to it. To do this, a template has been developed that contains a decorator introducing MLFlow for monitoring these projects, as well as a series of mechanisms that allow anyone to run it. Finally, some tests have been carried out with the initial project together with the final generated template.

G-Protein Coupled Receptors (GPCRs) are a big family of eukaryotic cell transmembrane proteins, responsible for numerous biological processes. From a practical viewpoint around 34\% of the drugs approved by the US Food and Drug Administration target these receptors. They can be analyzed from their simulated molecular dynamics, including the prediction of their behavior in the presence of drugs. In this paper, the capability of Long Short-Term Memory Networks (LSTMs) are evaluated to learn and predict the molecular dynamic trajectories of a receptor. Several models were trained with the 3D position of the amino acids of the receptor considering different transformations on the position of the amino acid, such as their centers of mass, the geometric centers and the position of the $\alpha$--carbon for each amino acid. The error of the prediction of the position was evaluated by the mean average error (MAE) and root-mean-square deviation (RMSD). The LSTM models show a robust performance, with results comparable to the state-of-the-art in non-dynamic 3D predictions. The best MAE and RMSD values were found for the mass center of the amino acids with 0.078 {\AA} and 0.156 {\AA} respectively. This work shows the potential of LSTM to predict the molecular dynamics of GPRCs., Comment: G-Protein Coupled Receptors, LSTM, Molecular Dynamics

Cities are becoming data-driven, re-engineering their processes to adapt to dynamically changing needs. A.I. brings new capabilities, effectively enlarging the space of policy interventions that can be explored and applied. Therefore, new tools are needed to augment our capacity to traverse this space and find adequate policy interventions. Digital twins are revealing themselves as powerful tools for policy experimentation and exploration, allowing faster and more complete explorations while avoiding costly interventions. However, they face some problems, among them data availability and model scalability. We introduce a digital twin framework based on an A.I. and a synthetic data model on NO2 pollution as a proof-of-concept, showing that this approach is feasible for policy evaluation and (autonomous) intervention and solves the problems of data scarcity and model scalability while enabling city level Open Innovation.

G protein-coupled receptors are a large super-family of cell membrane proteins that play an important physiological role as transmitters of extra-cellular signals. Signal transmission through the cell membrane depends on the conformational changes of the transmembrane region of the receptor and the investigation of the dynamics in these regions is therefore key. Molecular Dynamics (MD) simulations can provide information of the receptor conformational states at the atom level and machine learning (ML) methods can be useful for the analysis of these data. In this paper, Recurrent Neural Networks (RNNs) are used to evaluate whether the MD can be modeled focusing on the different regions of the receptor (intra-cellular, extra-cellular and each transmembrane regions (TM)). The best results, as measured by root-mean-square deviation (RMSD), are 0.1228 Å for TM4 of the 2rh1 (inactive state) and 0.1325 Å for TM4 of the 3p0g (active state), which are comparable to the state-of-the-art in non-dynamic 3-D predictions, showing the potential of the proposed approach. This work is funded by Spanish PID2019-104551RB-I00 research project and by the PRE2020-092428 Ph.D. training program, through the Ministry Science and Innovation.

Machine Learning requires an enormous amount of mathematical computation per second. Several architectures have been proposed to match the computation requirements and improve the calculation efficiency. Among these, the Systolic Array accelerators show promising results. These accelerators are composed of several Processing Elements (PEs), arranged on multiple symmetrical lines, which include a Multiply-And-Accumulate module. Specific multi-precision multipliers are increasingly popular since they can execute different precision multiplications and they can be integrated into Systolic Array accelerators. In this work, a multi-precision multiplier is proposed. The objective of the design is to build up a multiplier formed by combining smaller and equal multipliers. Depending on the number of small multipliers that are used, the size of the final multiplier changes. The mathematical foundation is based on decomposing the operands as the addition of several numbers and then applying the distributive and associative properties of mathematics. Different designs have been implemented with small multipliers able to execute 4x4 and 8x8 multiplications. A total of six multipliers have been implemented: three multipliers are made with 4x4 small multipliers and their output size is 16,32,64. The other three multipliers are realized with 8x8 small multipliers and their output size is 32,64,128. The multiplier benefits of hardware re-utilization and shows promising results in terms of area and power consumption. Less energy is consumed and less surface area is required compared to previously developed solutions. Compared to other structures performing similar functions, reusing multipliers can save about 30% of the area with a penalty of only 8% of the delay.

The simultaneous measurement of soil water content and water table levels is of great agronomic and hydrological interest. Not only does soil moisture represent the water available for plant growth but also water table levels can affect crop productivity. Furthermore, monitoring soil saturation and water table levels is essential for an early warning of extreme rainfall situations. However, the measurement of these parameters employing commercial instruments has certain disadvantages, with a high cost of purchase and maintenance. In addition, the handling of commercial devices makes it difficult to adapt them to the specific requirements of farmers or decision-makers. Open-source IoT hardware platforms are emerging as an attractive alternative to developing flexible and low-cost devices. This paper describes the design of a datalogger device based on open-source hardware platforms to register water table levels and soil moisture data for agronomic applications. The paper begins by describing energy-saving and wireless transmission techniques. Then, it summarizes the linear calibration of the phreatimeter sensor obtained with laboratory and field data. Finally, it shows how non-linear machine-learning techniques improve predictions over classical tools for the moisture sensor (SKU: SEN0193). This work is supported by Universidad Nacional del Litoral, research grants CAID 50520190100249LI and CAAT 03012020, and the National Spanish funding PID2019-107910RB-I00, by regional project 2017SGR-990, and with the support of Secretaria d’Universitats i Recerca de la Generalitat de Catalunya i del Fons Social Europeu.

Deep learning and high-performance computing have augmented and speed-up the scope of video-based vehicles' massive counting. The automatic vehicle counts result from the detection and tracking of the vehicles in certain areas or Regions of Interest (ROI). In this paper, we propose a technique to create a counting area with different traffic-flow directions based on YOLO and DBSCAN You Only Look Once version five (YOLOv5) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN). We compare the performance of the method against the manually counted ground truth. The proposed method showed that it is possible to generate the ROIs (counting areas) according to the traffic flow using deep learning techniques with relatively good accuracy (less than 5 % error). These results are promising but we need to explore the limits of this method with more street-view configurations, time and other detection and tracking algorithms, and in an HPC environment.

Emotional and physical health are strongly connected and should be taken care of simultaneously to ensure completely healthy persons. A person’s emotional health can be determined by detecting emotional states from various physiological measurements (EDA, RB, EEG, etc.). Affective Computing has become the field of interest, which uses software and hardware to detect emotional states. In the IoT era, wearable sensor-based real-time multi-modal emotion state classification has become one of the hottest topics. In such setting, a data stream is generated from wearable-sensor devices, data accessibility is restricted to those devices only and usually a high data generation rate should be processed to achieve real-time emotion state responses. Additionally, protecting the users’ data privacy makes the processing of such data even more challenging. Traditional classifiers have limitations to achieve high accuracy of emotional state detection under demanding requirements of decentralized data and protecting users’ privacy of sensitive information as such classifiers need to see all data. Here comes the federated learning, whose main idea is to create a global classifier without accessing the users’ local data. Therefore, we have developed a federated learning framework for real-time emotion state classification using multi-modal physiological data streams from wearable sensors, called Fed-ReMECS. The main findings of our Fed-ReMECS framework are the development of an efficient and scalable real-time emotion classification system from distributed multimodal physiological data streams, where the global classifier is built without accessing (privacy protection) the users’ data in an IoT environment. The experimental study is conducted using the popularly used multi-modal benchmark DEAP dataset for emotion classification. The results show the effectiveness of our developed approach in terms of accuracy, efficiency, scalability and users’ data privacy protection. Arijit Nandi is a fellow of Eurecat’s “Vicente López” PhD grant program. This study has been partially funded by ACCIÓ Spain (Pla d’Actuació de Centres Tecnològics 2021) under the project TutorIA.

Supervised machine learning is widely used in biology and deserves more scrutiny. We present a set of community-wide recommendations (DOME) aiming to help establish standards of supervised machine learning validation in biology. Formulated as questions, the DOME recommendations improve the assessment and reproducibility of papers when included as supplementary material. The work of the Machine Learning Focus Group was funded by ELIXIR, the research infrastructure for life-science data. IW was funded by the A*STAR Career Development Award (project no. C210112057) from the Agency for Science, Technology and Research (A*STAR), Singapore. D.F. was supported by Estonian Research Council grants (PRG1095, PSG59 and ERA-NET TRANSCAN-2 (BioEndoCar)); Project No 2014-2020.4.01.16-0271, ELIXIR and the European Regional Development Fund through EXCITE Center of Excellence. S.C.E.T. has received funding from the European Union’s Horizon 2020 research and innovation programme under Marie Skłodowska-Curie Grant agreements No. 778247 and No. 823886, and Italian Ministry of University and Research PRIN 2017 grant 2017483NH8. Peer Reviewed "Article signat per 8 autors més 28 autors/es de l' ELIXIR Machine Learning Focus Group: Emidio Capriotti, Rita Casadio, Salvador Capella-Gutierrez, Davide Cirillo, Alessio Del Conte, Alexandros C. Dimopoulos, Victoria Dominguez Del Angel, Joaquin Dopazo, Piero Fariselli, José Maria Fernández, Florian Huber, Anna Kreshuk, Tom Lenaerts, Pier Luigi Martelli, Arcadi Navarro, Pilib Ó Broin, Janet Piñero, Damiano Piovesan, Martin Reczko, Francesco Ronzano, Venkata Satagopam, Castrense Savojardo, Vojtech Spiwok, Marco Antonio Tangaro, Giacomo Tartari, David Salgado, Alfonso Valencia & Federico Zambelli"

Este documento introduce al lector en el campo de la visión por computador, explicando sus fundamentos. Para ello, además de exponer su definición, se estudian los elementos que componen un sistema que aplica esta tecnología, y se describe en qué consiste el proceso que realiza. También se presentan las funciones que desarrollan estos sistemas y algunas de las aplicaciones en las que se utilizan actualmente. Se ha desarrollado una aplicación de visión artificial para implementar en la prótesis de la extremidad superior del cuerpo humano desarrollada por el proyecto universitario ARM2u. Para ello, se han comparado diversas placas de programación, actualmente en el mercado, mediante sus características técnicas, estableciendo que la mejor opción según nuestras necesidades es la Spresense de Sony. La aplicación desarrollada se basa en un modelo de inteligencia artificial, también llamado algoritmo, que incorpora redes neuronales convolucionales. El proceso seguido se ha realizado mediante el aplicativo web Edge Impulse, que permite efectuar todo el proceso de diseño y entrenamiento de dicho algoritmo de visión por computador. Para conseguir un modelo optimizado y con gran precisión se modifican, a base de ensayo y error, tanto el tiempo de entrenamiento como los parámetros internos del mismo. La técnica utilizada para el entrenamiento de los modelos ha sido transfer learning, comprobando sus grandes potencialidades. Para el caso de clasificación de imágenes también se ha estudiado la posibilidad de entrenar completamente un algoritmo con red neuronal personalizada. Los modelos conseguidos han alcanzado una precisión aceptable, superior al 85%, tanto el de clasificación de imágenes como el de detección de objetos. Se ha comprobado la influencia que ejerce el fondo de las imágenes en la respuesta de los diferentes modelos Objectius de Desenvolupament Sostenible::9 - Indústria, Innovació i Infraestructura