Search

Amb els ràpids avenços tecnològics, tant les persones com les empreses renoven els seus dispositius molt sovint. Avui dia, molts d'aquests dispositius són prematurament reciclats. Però, malgrat ser obsolets, encara tenen valor i poden convertir-se en dispositius de segona mà en comptes de ser directament reciclats, el que estalvia més recursos. Existeixen organitzacions que reben centenars o milers de dispositius i han de decidir per cadascun si cal reciclar-lo o guardar-lo per a donar-li un altre ús. La decisió requereix saber el valor del dispositiu i el cost que tindria guardar-lo fins que algú el compri. A més, pel fet que els dispositius són configurables, també cal tenir en compte si és bona idea extreure components d'altres dispositius per a reparacions o millorares a d'altres, el que augmenta la complexitat del problema. Aquest projecte proporciona una eina per a estimar el preu que té un dispositiu en el mercat i facilitar la presa de decisions. El projecte també desenvolupa una aplicació web que integra l'eina i facilita el seu ús., Due to the rapid evolution of hardware, individuals and organizations renew their devices frequently. Nowadays, most replaced devices are prematurely recycled. However, those devices, despite being obsolete, still hold value and can have a second owner, which saves more resources than recycling. Refurbishers receive hundreds or thousands of devices from organizations and need to make the decision of either recycling the device or storing it so it can be used by other organizations. To make a wise choice, the value of the device and the cost of storing it should be taken into account. Moreover, since devices can be configured, the decision could also involve taking components from one device to upgrade or repair another, which makes the problem more complex. This project aims at providing a solution to this problem by building a tool for predicting the price of second-hand devices based on machine learning so refurbishers can estimate the price a buyer would pay for the device. The project also aims at developing a web application that integrates the tool and eases its use.

Autoencoders are essential in the field of machine learning because of the wide range of applications and distinctive talents they have. The ability of autoencoders to learn condensed and effective representations of complicated input data is one of the main factors in their significance. Autoencoders offer effective data compression by encoding the input data into a lower-dimensional latent space, which is useful in situations with constrained storage or bandwidth. Autoencoders are also frequently employed for unsupervised learning tasks like data generation, dimensionality reduction, and anomaly detection. Without relying on explicit labels or supervision, they enable us to find underlying patterns and structures in the data. Overall, the versatility and utility of autoencoders make them a fundamental tool in the machine learning toolbox, empowering researchers and practitioners to tackle a wide range of problems across diverse domains. Generative models, such as autoencoders, play a fundamental role in machine learning by enabling the creation of new, synthetic data that closely resembles the original input distribution. These models have revolutionised various domains, including image generation, text synthesis, and music composition, among many others. By capturing the underlying patterns and structures of the training data, generative models provide a powerful framework for creative applications, data augmentation, and simulation studies. This project aimed to explore the capabilities and applications of autoencoders, a type of neural network architecture, in the field of machine learning. The main focus of the project was to refactor legacy code used for image identification and transform it into an autoencoder capable of generating MNIST images. Through extensive experimentation and analysis, the project demonstrated the effectiveness of autoencoders in learning representations of input data and generating high-quality synthetic images. The findings of this

En el camp del Machine Learning, és molt comú tenir la necessitat d'obtenir conjunts de dades amb característiques específiques per poder desenvolupar models de predicció i provar-los exhaustivament. Actualment, l'obtenció d'aquests conjunts pot ser una tasca complicada a causa de la manca d'eines que permetin obtenir dades ajustades al que es necessita. Els únics recursos disponibles en l'actualitat són els grans repositoris amb conjunts de dades existents i la creació manual de conjunts de dades. A partir d'aquesta manca de recursos, aquest projecte està enfocat en la creació d'un algorisme que permeti als usuaris generar conjunts de dades complets i amb diferents característiques de manera ràpida i eficaç. Per a això, l'algorisme permet la utilització de diversos tipus de variables, diversos mètodes per generar conjunts diferents, l'addició de valors perduts i l'addició de dificultat sobre la variable objectiu. Els resultats experimentals obtinguts durant el desenvolupament del projecte mostren que les tècniques utilitzades funcionen correctament i s'obtenen conjunts de dades útils per observar les característiques i el comportament d'alguns models de Machine Learning ja creats. Per concloure cal destacar que l'algorisme suposa una nova eina que pot servir com a base d'altres projectes i que és susceptible de noves millores que permetin afegir més característiques interessants als conjunts i que, per tant, li permetin arribar a més usuaris., In the field of Machine Learning, it is very common to have the need to obtain datasets with specific characteristics in order to develop prediction models and test them. Currently, obtaining such datasets can be a challenging task due to the lack of tools that allow for obtaining data tailored to the required specifications. The only available resources at present are large repositories with existing datasets and the manual creation of datasets. Addressing this lack of resources, this project is focused on the creation of an algorithm that enables users to quickly and efficiently generate complete datasets with various characteristics. To achieve this, the algorithm allows for the use of different types of variables, multiple methods for generating distinct datasets, the addition of missing values, and the introduction of difficulty regarding the target variable. The experimental results obtained during the project's development demonstrate that the employed techniques function correctly, yielding datasets that are useful for examining the characteristics and behavior of pre-existing Machine Learning models. In conclusion, it is worth noting that this algorithm represents a novel tool that can serve as a foundation for other projects. Furthermore, it is open to further improvements that could introduce more interesting features to the datasets and consequently expand its accessibility to a wider range of users.

En el camp de les simulacions virtuals de xarxes basades en grafs, s'utilitzen els que es diu GNN (Graph Neural Networks) en conjunt amb NDT (Network Digital Twin), que serveixen per replicar xarxes reals. El gran problema de la digitalització d'aquestes xarxes és el cost en temps i recursos, i la poca flexibilitat que té la rèplica digital per modificar-se a temps real. Recentment, solucions basades en Aprenentatge Autònom han anat guanyant rellevància degut a la seva rapidesa. El departament del BNN (Barcelona Neural Networking Center) ha creat una eina anomenada RouteNet-Fermi que es nodreix d'aquesta digitalització per analitzar xarxes virtuals i treure informació sobre el seu rendiment. RouteNet-Fermi és, a hores d'ara, una de les eines més avançades d'aquest àmbit. Tot i això, la interacció amb l'eina és complicada i per un usuari no familiaritzat amb aquests formats pot ser complicat. Es proposa el desenvolupament d'una eina que es relacioni amb RouteNet-Fermi, pero que a diferència d'altres ja existents, faciliti la configuració a temps real. Per dur a terme aquest projecte s'optarà per implementar un model web basat en un apartat Front-end i un altre Back-end, que mitjançant la interacció de l'usuari a través de la web, s'anirà modificant una xarxa virtual, per posteriorment interaccionar amb una API que processarà les dades fent ús de RouteNet-Fermi. S'ha demostrat que aquest projecte redueix de forma molt significativa el temps necessari per simular i replicar xarxes reals com a virtuals, ademés de facilitar la visualització de les dades resultants. El projecte té un potencial de creixement per futures necessitats elevat i permet la personalització adequada per cada usuari de l'eina., In the field of virtual simulations of graph-based networks, the so-called GNN (Graph Neural Networks) are used together with NDT (Network Digital Twin), which serve to replicate real networks. The big problem with the digitization of these networks is the cost in time and resources, and the little flexibility that the digital replica has to modify itself in real time. Recently, solutions based on Autonomous Learning have been gaining relevance due to their speed. The BNN (Barcelona Neural Networking Center) department has created a tool called RouteNet-Fermi that feeds on this digitization to analyze virtual networks and extract information about their performance. RouteNet-Fermi is, by now, one of the most advanced tools in this area. However, the interaction with the tool is complicated and for a user not familiar with these formats it can be complicated to use. It is proposed to develop a tool that is related to RouteNet-Fermi, but that, unlike others that already exist, facilitates real-time configuration. To carry out this project, it will be decided to implement a web model based on a Front-end section and another Back-end section, which through user interaction through the web, a virtual network will be modified, to later interact with an API that will process the data using RouteNet-Fermi. This project has been shown to significantly reduce the time needed to simulate and replicate real and virtual networks, in addition to facilitating the visualization of the resulting data. The project has a high growth potential for future needs and allows the appropriate customization for each user of the tool.

There is an increasing demand to enhance infrastructure asset management within the drinking water sector. A key factor for achieving this is improving the accuracy of pipe failure prediction models. Machine learning-based models have emerged as a powerful tool in enhancing the predictive capabilities of water distribution network models. Extensive research has been conducted to explore the role of explanatory variables in optimizing model outputs. However, the underlying mechanisms of incorporating explanatory variable data into the models still need to be better understood. This review aims to expand our understanding of explanatory variables and their relationship with existing models through a comprehensive investigation of the explanatory variables employed in models over the past 15 years. The review underscores the importance of obtaining a substantial and reliable dataset directly from Water Utilities databases. Only with a sizeable dataset containing high-quality data can we better understand how all the variables interact, a crucial prerequisite before assessing the performance of pipe failure rate prediction models., EF-O acknowledges the financial support provided by the “Agencia de Gestió d’Ajust Universitaris I de Recerca” (https:// agaur. gencat. cat/ en/) through the Industrial Doctorate Plan of the Secretariat for Universities and Research of the Department of Business and Knowledge of the Government of Catalonia, under the Grant DI 093-2021. Additionally, EF-O appreciates the economic support received from the Water Utility Aigües de Barcelona, Empresa Metropolitana de Gestió del Cicle Integral de l'Aigua., Peer Reviewed, Postprint (published version)

A cost-effective, non-invasive, and easy-to-use tool is presented that uses the 6-axis inertial sensor of the smartphone or a specific wearable sensor, boosted by machine learning, to support early differential diagnosis of Parkinson’s disease and Essential Tremor. A dedicated web server helps extract the kinematic indexes from the recorded signals, implement the machine learning models and return the resulting classification to the App. Thus, clinicians can use this App as a support tool in the clinic, contributing to performing motor evaluations in the uncertain and undecided stages of the diseases and promoting appropriate, fast, and timely therapeutic responses., This work was supported by Serra Húnter Programme (Generalitat de Catalunya), Peer Reviewed, Postprint (published version)

El present treball de fi de grau se centra en l’exploració i l’aplicació de diversos algorismes d’aprenentatge automàtic per a la resolució de tres problemes específics. Per a això, s’estudien dos casos d’aprenentatge supervisat i un altre d’aprenentatge no supervisat. Dins de l’àmbit de l’aprenentatge supervisat, es desenvolupa un problema de classificació binària per a predir la baixa dels empleats en una companyia i un problema de regressió per tal de pronosticar la demanda energètica actual. S’apliquen diferents algorismes amb l’objectiu de trobar aquells que proporcionen els millors resultats per a les taques plantejades. Així mateix, es resol un problema de recomanació de pel·lícules d’aprenentatge no supervisat basat en contingut i popularitat i s’analitzen els resultats mitjançant les recomanacions proporcionades per a una pel·lícula en concret. Per a la resolució dels tres problemes, se seleccionen les bases de dades adequades i es fa un treball previ d’exploració, preparació i neteja de dades per a obtenir uns resultats òptims., El presente trabajo de fin de grado se centra en la exploración y la aplicación de diversos algoritmos de aprendizaje automático para la resolución de tres problemas específicos. Para ello, se estudian dos casos de aprendizaje supervisado y otro de aprendizaje no supervisado. Dentro del ámbito del aprendizaje supervisado, se desarrolla un problema de clasificación binaria para predecir la baja de los empleados en una compañía y un problema de regresión para pronosticar la demanda energética actual. Se aplican diferentes algoritmos con el objetivo de encontrar aquellos que proporcionan los mejores resultados para las tareas planteadas. Asimismo, se resuelve un problema de recomendación de películas de aprendizaje no supervisado basado en contenido y popularidad y se analizan los resultados mediante las recomendaciones proporcionadas para una película en concreto. Para la resolución de los tres problemas, se seleccionan las bases de datos adecuadas y se realiza un trabajo previo de exploración, preparación y limpieza de datos para obtener unos resultados óptimos., The present undergraduate thesis focuses on the exploration and application of various machine learning algorithms to solve three specific problems. To do so, two cases of supervised learning and one case of unsupervised learning are studied. Within the scope of supervised learning, a binary classification problem is developed to predict employee attrition in a company, as well as a regression problem to forecast current energy demand. Different algorithms are applied with the objective of finding those that provide the best results for the given tasks. Additionally, an unsupervised learning problem of movie recommendations based on content and popularity is solved, and the results are analyzed through the recommendations provided for a specific movie. To solve the three problems, appropriate databases are selected, and preliminary work is done to explore, prepare, and clean the data in order to obtain optimal results.

This works proposes location estimation in 5G systems using Reconfigurable Intelligent Surfaces (RIS) to optimize positioning accuracy where LOS is obstructed. Machine Learning techniques are considered for the optimization of the RIS phases. The primary focus is on developing a RIS-aided localization algorithm to improve the accuracy of user positioning using compressed sensing techniques. The integration of a RIS assists the base station in obtaining more accurate location by mitigating the problems caused by blockages of the LoS. The present study suggests a phase optimization method that minimizes the Cramer- Rao Lower Bound (CRLB) in order to find the optimal RIS phase configuration. However, the computation of the CRLB for different scenarios has a high computational cost. To address this challenge, Machine Learning (ML) techniques have been applied to develop a network that can calculate the phases that minimize the CRLB based solely on the estimated position of the user. Moreover, in order to provide a comprehensive comparison of the results, the optimization of the RIS phases based on SNR criteria has also been considered., Este trabajo propone la estimación de ubicación en sistemas 5G utilizando Superficies Inteligentes Reconfigurables (RIS) para optimizar la precisión de posicionamiento cuando la línea de visión (LOS) está obstruida. Se consideran técnicas de Machine Learning para la optimización de las fases de la RIS. El enfoque principal se centra en desarrollar un algoritmo de localización asistido por RIS para mejorar la precisión del posicionamiento del usuario utilizando técnicas de compresión de señales (CS). La integración de una RIS ayuda a la estación base a obtener una ubicación más precisa al mitigar los problemas causados por los bloqueos de la línea de visión. El presente estudio sugiere un método de optimización de fases que minimiza el Límite Inferior de Cramer-Rao (CRLB) para encontrar la configuración óptima de fases de RIS. Sin embargo, el cálculo del CRLB para diferentes escenarios tiene un alto costo computacional. Para abordar este desafío, se han aplicado técnicas de Aprendizaje Automático para desarrollar una red neuronal que pueda calcular las fases que minimizan el CRLB basándose únicamente en la posición estimada del usuario. Además, con el fin de proporcionar una comparación exhaustiva de los resultados, también se ha considerado la optimización de las fases de RIS basada en criterios de Relación Señal-Ruido (SNR)., Aquest treball proposa l'estimació de la ubicació en sistemes 5G utilitzant Superfícies Intel·ligents Reconfigurables (RIS) per optimitzar la precisió de posicionament quan la línia de visió (LOS) està obstruïda. Es consideren tècniques de Machine Learning per a l'optimització de les fases de la RIS. L'enfocament principal es centra en desenvolupar un algoritme de localització assistit per una RIS per millorar la precisió del posicionament de l'usuari utilitzant tècniques de compressió de senyal (CS). La integració d'una RIS ajuda a l'estació base a obtenir una ubicació més precisa mitigant els problemes causats per bloqueigs de la LOS. L'estudi actual suggereix un mètode d'optimització de fases que minimitza el Límit Inferior de Cramer-Rao (CRLB) per trobar la configuració òptima de fases de la RIS. No obstant això, el càlcul del CRLB per a diferents escenaris té un cost computacional elevat. Per abordar aquest repte, s'han aplicat tècniques de Machine Learning (ML) per desenvolupar una xarxa neuronal que pugui calcular les fases que minimitzen el CRLB basant-se únicament en la posició estimada de l'usuari. A més, per proporcionar una comparació exhaustiva dels resultats, també s'ha tingut en compte l'optimització de les fases de RIS basada en criteris de Relació Senyal-Soroll (SNR).

Els models climàtics estimen la dinàmica dels fluids a la Terra resolent equacions contínues en quadrícules finites. Els processos físics per sota de la resolució dels models climàtics tenen efectes estadístics a les escales resoltes. Per tenir-los en compte, els models incorporen parametritzacions. El desenvolupament de parametritzacions efectives és crucial per produir prediccions climàtiques precises a llarg termini. En aquest treball, abordem aquest problema utilitzant l'scattering transform per estudiar i modelar el problema de clausura en el domini d'scattering, i desenvolupem un model generatiu i un mètode de superresolució per a camps de turbulència., Los modelos climáticos estiman la dinámica de los fluidos en la Tierra resolviendo ecuaciones en cuadrículas finitas. Los procesos físicos por debajo de la resolución de los modelos climáticos tienen efectos estadísticos en las escalas resueltas. Para tenerlos en cuenta, los modelos incorporan parametrizaciones. El desarrollo de parametrizaciones efectivas es crucial para producir predicciones climáticas precisas a largo plazo. En este trabajo, abordamos este problema utilizando la scattering transform para estudiar y modelar el problema de clausura en el dominio de scattering, y desarrollamos un modelo generativo y un método de superresolución para campos de turbulencia., Climate models estimate the dynamics of the fluids on Earth by solving equations on finite grids. The physical processes below the resolution of climate models have statistical effects on the resolved scales. To account for them, models incorporate subgrid parametrizations. Devising effective parametrizations is crucial when producing accurate, long-term climate predictions. In this work, we address this challenge by using the scattering transform to study and model the closure problem in the scattering domain and we develop a generative model and a super-resolution method for turbulence fields., Outgoing

Air quality monitoring sensor networks focusing on air pollution measure pollutants that are regulated by the authorities, such as CO, NO2, NO, SO2, O3, and particulate matter (PM10, PM2.5). However, there are other pollutants, such as black carbon (BC), which are not regulated, have a major impact on health, and are rarely measured. One solution is to use proxies, which consist of creating a mathematical model that infers the measurement of the pollutant from indirect measurements of other pollutants. In this paper, we propose a robust machine learning proxy (RMLP) framework for estimating BC based on nonlinear machine learning methods, calibrating the low-cost sensors (LCSs), and adding robustness against noise and data missing in the LCS. We show the impact of LCS data aggregation, denoising and missing imputation on BC estimation, and how the concentrations estimated by the BC proxy approximate the values obtained by a reference instrument with an accurate BC sensor., This work is supported by projects H2020 FIRE-RES, PID 2019-107910RB-I00, CEX 2018-000794-S, 2017 SGR41, 2021 SGR01059, and the support of Sec. d’Universitats i Recerca de la Generalitat de Catalunya i del Fons Social Europeu., Peer Reviewed, Postprint (author's final draft)

Recent research has shown the possibility to achieve microconfined turbulence utilizing supercritical fluids under high-pressure transcritical conditions. This has generated increasing interest in the hybrid nature of the fluid thermophysical properties near the pseudo-boiling region. However, obtaining thermophysical experimental data of microfluidic transcritical systems is challenging due to the difficulties associated with high-pressure experiments, and the focus on velocity-related data acquisition of traditional experimental techniques, such as particle image velocimetry and particle tracking velocimetry. To overcome these challenges, this work introduces a novel framework based on a physics-constrained neural network to recover thermophysical information from velocity data at high-pressure turbulent transcritical conditions. By incorporating thermophysical constraints and boundary conditions into the network architecture and the loss function, the proposed model surpasses non-physically informed approaches in terms of accuracy, satisfaction of thermophysical constraints and preservation of bulk quantities., This work is funded by the European Union (ERC, SCRAMBLE, 101040379)., Peer Reviewed, Postprint (published version)

Rapid development in the network infrastructure has resulted in sophisticated attacks which are hard to detect using typical network intrusion detection systems (NIDS). There is a strong need for efficient NIDS to detect these known attacks along with ever-emerging zero-day exploits. Existing NIDS are more focused on detecting known attacks using supervised machine learning approaches, achieving better performance for known attacks but poor detection of unknown attacks. Many NIDS have utilized the unsupervised approach, which results in better detection of unknown anomalies. In this paper, we proposed a Hybrid NIDS based on Semisupervised One-Class Support Vector Machine (OC-SVM) and Supervised Random Forest (RF) algorithms. This detection system has several stages. The First stage is based on OC-SVM, which filters benign and malicious traffic. The next stages use many parallel supervised models and an additional OC-SVM model to separate known and unknown attacks from malicious traffic. The previous process is done so that known attacks are classified by their type, and unknown attacks are detected. The proposed NIDS is tested on the standard public dataset CSE-CIC-IDS-2018. The evaluation results show that the system achieves a high accuracy, 99.45%, for detecting known attacks. Our proposed NIDS achieves an accuracy of 93.99% for unknown or zero-day attacks. The overall accuracy of the proposed NIDS is 95.95%. The system significantly improves the detection of known and unknown anomalies using a hybrid approach., This project has received funding from the European Union’s H2020 research and innovation programme under the grant agreement No. 952644, from the Spanish Ministry of Science and Innovation under contract PID2021-124463OB- I00, and from the Catalan Government under contract 2021 SGR 00326., Peer Reviewed, Postprint (author's final draft)

In vivo magnetic resonance spectroscopy (MRS) has two modalities, single-voxel (SV) and multivoxel (MV), in which one or more contiguous grids of SVs are acquired. Purpose: To test whether MV grids can be classified with models trained with SV. Methods: Retrospective study. Training dataset: Multicenter multiformat SV INTERPRET, 1.5T. Testing dataset: MV eTumour, 3T. Two classification tasks were completed: 3-class (meningioma vs. aggressive vs. normal) and 4-class (meningioma vs. low-grade glioma vs. aggressive vs. normal). Five different methods were tested for feature selection. The classification was implemented using linear discriminant analysis (LDA), random forest, and support vector machines. The evaluation was completed with balanced error rate (BER) and area under the curve (AUC) on both sets. The accuracy in class prediction was calculated by developing a solid tumor index (STI) and segmentation accuracy with the Dice score. Results: The best method was sequential forward feature selection combined with LDA, with AUCs = 0.95 (meningioma), 0.89 (aggressive), 0.82 (low-grade glioma), and 0.82 (normal). STI was 66% (4-class task) and 71% (3-class task) because two cases failed completely and two more had suboptimal STI as defined by us. Discussion: The reasons for failure in the classification of the MV test set were related to the presence of artifacts., H2020-EU.1.3.—EXCELLENT SCIENCE—Marie Skłodowska-Curie Actions, grant number H2020-MSCA-ITN-2018-813120. Proyectos de investigación en salud 2020, grant numbers PI20/00064 and PI20/00360. Spanish Ministerio de Economía y Competitividad SAF2014-52332-R. Centro de Investigación Biomédica en Red en Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN [http://www.ciber-bbn.es/en, accessed on 12 January 2023], CB06/01/0010), an initiative of the Instituto de Salud Carlos III (Spain) co-funded by the EU Fondo Europeo de Desarrollo Regional (FEDER). Spanish AEI PID2019-104551RB-I00 grant. Xartecsalut, 2018 XARDI 00,016 and 2021 XARDI 00021. eTUMOUR: FP6-2002-LIFESCIHEALTH- 503094. INTERPRET: FP5-IST-1999-10310., Peer Reviewed, Postprint (published version)

El paper "Differential Machine Learning", publicat el 2020 per Antoine Savine i Brian Huge obre una nova porta al càlcul de derivats financers a través d'un disseny d'una nova Xarxa Neuronal, nucli del Deep Learning, creant un vincle entreaquests camps. Aquesta tesi preten donar una comprensió matemàtica de tots els conceptes introduits al paper, complementant-lo i ampliant-lo. L'altra part de la tesis treballa en torn a la possibilitat d'estendre la "twin network" i els seus càlculs per incloure la resta de gregues, implementant les seves prediccions i comparant els resultados amb les fòrmules tancades originals obtingudes pel model de Black-Scholes. En conclusió, la tesis contribueix a la comprensió de la "twin network" i mostra algunes aplicacions de Deep Learning en Finances, concretament en el càlcul d'opcions europeas i les seves respectives gregues. Traducción realizada con la versión gratuita del traductor www.DeepL.com/Translator, El paper "Differential Machine Learning", publicado en 2020 por Antoine Savine y Brian Huge abre una nueva puerta al cálculo de derivados financieros a través de un diseño de una nueva Red Neuronal, núcleo del Deep Learning, creando un vínculo entre estos dos campos. Esta tesis pretende dar una comprensión matemática de todos los conceptos introducidos en el paper, complementándolo y ampliándolo. La otra parte de esta tesis trabaja en torno a la posibilidad de extender la "twin network" y sus cálculos para incluir el resto de griegas, implementando sus predicciones y comparando sus resultados con las fórmulas cerradas originales obtenidas por el modelo Black-Scholes. En conclusión, esta tesis contribuye a la comprensión de la "twin network" y muestra algunas aplicaciones de Deep Learning en Finanzas, concretamente en el cálculo de opciones europeas y sus respectivas griegas., The paper ”Differential Machine Learning”, published in 2020 by Antoine Savine and Brian Huge opens a new door to the calculation of financial derivatives through a design of a new Neural Network, the core of Deep Learning, creating a link between these two fields. This thesis aims to give a mathematical understanding of all the concepts introduced in the paper, complementing it and extending it. The other part of this thesis works around the possibility of extending the twin network and its calculations to include the rest of the Greeks, implementing its predictions and comparing its results to the original closed formulas obtained by the Black-Scholes model. In conclusion, this thesis contributes to the understanding of the twin network and showcases some applications of Deep Learning in Finance, specifically in the calculation of European call options and its respective Greeks.

El cáncer de pulmón es una enfermedad que se caracteriza por el crecimiento anormal de células malignas en los tejidos pulmonares. Es uno de los cánceres más comunes y está estrechamente relacionado con el consumo de tabaco, aunque también puede afectar a personas no fumadoras. Este tipo de cáncer puede tener graves consecuencias para la salud y su detección temprana, así como el acceso a tratamientos adecuados, son fundamentales para mejorar las perspectivas de supervivencia y calidad de vida de los pacientes. Por lo tanto, la detección del cáncer de pulmón es crucial para identificar la enfermedad en etapas tempranas y poder así mejorar las opciones de tratamiento. Este proyecto de fin de grado tiene como objetivo el desarrollo de una red neuronal convolucional (CNN) que a partir de imágenes histopatológicas de tumores pulmonares puede clasificar tumores benignos, carcinomas o adenocarcinomas. La prioridad de la red neuronal es maximizar el recall, para no brindar falsos negativos en los diagnósticos realizados, ya que este error podría comportar consecuencias graves para los pacientes. Además, se desarrollan diversos modelos de aprendizaje automático para clasificar a los pacientes según la presencia de cáncer en sus pulmones y también la gravedad de este. De esta manera, se puede realizar un análisis en profundidad de las variables que tienen mayor impacto con la presencia y gravedad de un cáncer, tomando como variables tanto posibles potenciadores de la presencia de cáncer como síntomas presentes en el paciente., Lung cancer is a disease characterised by the abnormal growth of malignant cells in lung tissues. It is one of the most common cancers and is closely linked to tobacco use, although it can also affect non-smokers. This type of cancer can have serious health consequences and its early detection, as well as access to appropriate treatments, are essential to improve patients' prospects of survival and quality of life. Therefore, screening for lung cancer is crucial to identify the disease in its early stages and improve treatment options. This thesis presents the development of a convolutional neural network (CNN) for the classification of histopathological images of lung tumours into benign tumours, carcinomas or adenocarcinomas. The priority of the neural network is to maximise recall in order to avoid false negatives in the diagnoses made, as this error could have serious consequences for patients. In addition, several machine learning models are developed to classify patients according to the potential presence of cancer in their lungs and also to classify the severity of the cancer. In this way, an in-depth analysis of the variables that have the greatest impact on the presence and severity of a cancer can be carried out, taking as variables both possible enhancers of the presence of cancer and symptoms.

Support Vector Machines (SVMs) represent a category of supervised machine learning algorithms that find extensive application in both classification and regression tasks. In these algorithms, kernel functions are responsible for measuring the similarity between input samples to generate models and perform predictions. In order for SVMs to tackle data analysis tasks involving mixed data, the implementation of a valid kernel function for this purpose is required. However, in the current literature, we hardly find any kernel function specifically designed to measure similarity between mixed data. In addition, there is a complete lack of significant examples where these kernels have been practically implemented. Another notable characteristic of SVMs is their remarkable efficacy in addressing high-dimensional problems. However, they can become inefficient when dealing with large volumes of data. In this project, we propose the formulation of a kernel function capable of accurately capturing the similarity between samples of mixed data. We also present an SVM algorithm based on Bagging techniques that enables efficient analysis of large volumes of data. Additionally, we implement both proposals in an updated version of the successful SVM library LIBSVM. Moreover, we evaluate their effectiveness, robustness and efficiency, obtaining promising results.

En els darrers anys, el sector del pàdel ha crescut d'una manera exponencial. Cada vegada hi ha més practicants professionals i no professionals d'aquest esport. Concretament, dues de cada nou persones que practiquen esport a Espanya, són practicants de pàdel. Cada vegada hi ha més marques i empreses que s'incorporen en aquest sector i, això fa que el mercat creixi cada vegada més. En aquest treball de fi de grau (TFG), el que es farà és millorar i desenvolupar noves funcionalitats per professionalitzar una plataforma especialitzada en pàdel. La plataforma està basada en la recopilació i anàlisi de dades del mercat del pàdel. En un futur, es vol evolucionar la plataforma cap a l'extracció i predicció de dades futures. Per tant, el present treball busca el desenvolupament d'un petit spin-off d'un predictor de vendes amb l'ajuda de les dades actuals de la plataforma i de l'aprenentatge automàtic. Com a solució del problema proposat, es proporcionaran diferents models típics de machine learning, cadascun amb els seus paràmetres i variables, amb els que es podrà predir els valors de vendes futures. Com a desenllaç, es donaran els resultats de l'avaluació dels models i els valors de vendes predits. Els models seran escollits en un procés selectiu en el qual es provaran moltes variables i paràmetres per a donar les millors solucions. S'analitzaran els resultats i s'escolliran els models amb millors puntuacions tenint en compte diversos escenaris. Després de tot el treball fet, es conclou que el desenvolupament d'aquest sistema de predicció pot ajudar a les empreses del sector a anticipar-se com més aviat millor als possibles esdeveniments futurs per ajustar els marges de fabricació, construcció i venda dels seus productes., In recent years, the padel sector has grown exponentially. There are more and more professional and non-professional practitioners of this sport. Specifically, two out of nine people who practice sports in Spain are padel practitioners. There are more and more brands and companies that have been incorporated in this sector, and this makes the market grow more and more. In this Final Degree Project, what will be done is to improve and develop new features to professionalize a platform specialized in padel tennis. The platform is based on the collection and analysis of data from the padel market. In the future, we want to evolve the platform towards the extraction and prediction of future data. Therefore, the present work seeks the development of a small spin-off of a sales predictor with the help of the platform's current data and machine learning. As a solution to the proposed problem, different machine learning models will be provided, each with its own parameters and variables, with which future sales values can be predicted. As a conclusion, the results of the evaluation of the models and predicted sales values will be given. The models will be chosen in a selective process in which many variables and parameters will be tested to give the best solutions. The results will be analyzed, and the models with the best scores will be chosen taking into account various scenarios. After all the work done, it is concluded that the development of this prediction system can help companies in the sector to anticipate as early as possible to possible future events in order to adjust the margins of manufacturing, construction and sale of their products.

In on-device training of machine learning models on microcontrollers a neural network is trained on the device. A specific approach for collaborative on-device training is federated learning. In this paper, we propose embedded federated learning on microcontroller boards using the communication capacity of a LoRa mesh network. We apply a dual board design: The machine learning application that contains a neural network is trained for a keyword spotting task on the Arduino Portenta H7. For the networking of the federated learning process, the Portenta is connected to a TTGO LORA32 board that operates as a router within a LoRa mesh network. We experiment the federated learning application on the LoRa mesh network and analyze the network, system, and application level performance. The results from our experimentation suggest the feasibility of the proposed system and exemplify an implementation of a distributed application with re-trainable compute nodes, interconnected over LoRa, entirely deployed at the tiny edge., This work was supported by the Spanish Government under contracts PID2019-106774RB-C21, PCI2019-111851-2 (LeadingEdge CHIST-ERA), PCI2019-111850-2 (DiPET CHIST-ERA)., Peer Reviewed, Postprint (published version)

Knowledge bases are centralized repositories used for developing knowledge-oriented information systems. They are essential for adaptive, specialized knowledge in dialogue systems, supporting up-to-date domain-specific discussions with users. However, designing large-scale knowledge bases presents multiple challenges in data collection and knowledge exploitation. Knowledge graphs provide various research opportunities by integrating decentralized data and generating advanced knowledge. Our contribution presents a knowledge base in the form of a knowledge graph for extended knowledge generation for mobile apps and features extracted from related natural language documents. Our work encompasses the knowledge graph completion and deductive and inductive knowledge requests. We evaluated the effectiveness and performance of these knowledge strategies, which can be used as on-demand knowledge requests used by third-party software systems., With the support from the Secretariat for Universities and Research of the Ministry of Business and Knowledge of the Government of Catalonia and the European Social Fund. This paper has been funded by the Spanish Ministerio de Ciencia e Innovación under project / funding scheme PID2020-117191RB-I00 / AEI/10.13039/501100011033., Peer Reviewed, Postprint (author's final draft)

Current trends in networking propose the use of Machine Learning (ML) for a wide variety of network optimization tasks. As such, many efforts have been made to produce ML-based solutions for Traffic Engineering (TE), which is a fundamental problem in Internet Service Provider (ISP) networks. Nowadays, state-of-the-art TE optimizers rely on traditional optimization techniques, such as Local search, Constraint Programming, or Linear programming. In this paper, we present MAGNNETO, a distributed ML-based framework that leverages Multi-Agent Reinforcement Learning and Graph Neural Networks for distributed TE optimization. MAGNNETO deploys a set of agents across the network that learn and communicate in a distributed fashion via message exchanges between neighboring agents. Particularly, we apply this framework to optimize link weights in Open Shortest Path First (OSPF), with the goal of minimizing network congestion. In our evaluation, we compare MAGNNETO against several state-of-the-art TE optimizers in more than 75 topologies (up to 153 nodes and 354 links), including realistic traffic loads. Our experimental results show that, thanks to its distributed nature, MAGNNETO achieves comparable performance to state-of-the-art TE optimizers with significantly lower execution times. Moreover, our ML-based solution demonstrates a strong generalization capability to successfully operate in new networks unseen during training., This publication is part of the Spanish I+D+i project TRAINER-A (ref. PID2020-118011GBC21), funded by MCIN/AEI/10.13039/501100011033. This work is also partially funded by the Catalan Institution for Research and Advanced Studies (ICREA), the Secretariat for Universities and Research of the Ministry of Business and Knowledge of the Government of Catalonia, and the European Social Fund., Peer Reviewed, Postprint (author's final draft)

L'ús de bateries de segona vida és una de les estratègies més prometedores per a resoldre el problema del reciclatge de piles en el futur. L'objectiu d'aquesta recerca és aportar solucions pràctiques de garbellat, reagrupació i reutilització donant un nou ús a les piles de liti retirades. Donar una segona oportunitat a aquestes bateries retirades no sols té beneficis ecològics i mediambientals, sinó que també és rendible des del punt de vista econòmic. Per a aquest projecte, s'ha aprofitat el conjunt de dades que va registrar la NASA per a observar el comportament de les bateries elèctriques d'ió-liti al llarg de la seva vida per a realitzar un estudi que permet predir la vida útil d'una bateria. Mitjançant diferents tècniques d'aprenentatge automàtic, principalment SVM, s'ha aconseguit entrenar un model que pugui predir aquesta característica. D'aquesta manera es pretén aportar en un camp on la sostenibilitat i l'economia circular són cada vegada més importants. En el futur, es podrien implementar estratègies basades en l'ús de bateries de segona vida en diferents sectors, des de l'emmagatzematge d'energia fins a la mobilitat elèctrica. Realitzant aquest projecte es pretén continuar pel camí de la recerca a partir de tècniques de ML i IA per a contribuir a una economia més sostenible i responsable., The use of second life batteries is one of the most promising strategies to solve the problem of battery recycling in the future. The objective of this research is to provide practical solutions of screening, regrouping and reuse by giving a new use to retired lithium batteries. Giving a second chance to these retired batteries not only has ecological and environmental benefits, but is also cost-effective from an economic point of view. For this project, the data set recorded by NASA to observe the behavior of lithium-ion electric batteries over their lifetime has been used to conduct a study to predict the lifetime of a battery. Using different machine learning techniques, mainly SVM, it has been possible to train a model that can predict this characteristic. In this way, it aims to contribute to a field where sustainability and circular economy are becoming increasingly important. In the future, strategies based on the use of second life batteries could be implemented in different sectors, from energy storage to electric mobility. This project aims to continue on the path of research based on ML and AI techniques to contribute to a more sustainable and responsible economy.

Energy distribution flow through the general grid has traditionally been hierarchical and unidirectional from generation agents to customers. However, contemporary changes in the conception of energy uses have led to the creation of self-established smart energy communities, where members actively participate in a federation that allows them to generate, buy, store, consume, supply and trade their own energy autonomously. This novel paradigm presents several challenges that can be modelled and solved by making use of multiple computing and informatics domains, for instance, to forecast day-ahead energy needs, to optimally plan energy usage and trading, or to register energy transactions among members securely and transparently. With the goals of moving towards a democratized self-governance model, decarbonizing the economy, and accelerating the energy transition at the user level, a new comprehensive architecture is devised as a multi-level framework that integrates optimization models, machine learning tools and smart grid resources for managing communities constituted by electric vehicle smart charging stations. This framework is implemented and thoroughly evaluated.

The master's degree thesis is composed of theoretical and practical parts. The theoretical part describes the basics of predictive data analytics and machine learning algorithms for classification such as Logistic Regression, Decision Tree, Random Forest, SVM, and KNN. We also describe different evaluation metrics such as Recall, Precision, Accuracy, F1 Score, Cohen's Kappa, Hamming Loss, and Jaccard Index that are used to measure the performance of these algorithms. Additionally, we record the time taken for the training and prediction processes to provide insights into algorithm scalability. The key part master's thesis is the practical part that compares these algorithms with a self-implemented tool that shows results for different evaluation metrics on seven datasets. First, we describe the implementation of an application for testing where we measure evaluation metrics scores. We tested these algorithms on all seven datasets using Python libraries such as scikit-learn. Finally, we analyze the results obtained and provide final conclusions.

The rise in cyber-attacks and cyber-crime is causing more and more organizations and individuals to consider the correct implementation of their security systems. The consequences of a security breach can be devastating, ranging from loss of public confidence to bankruptcy. Traditional techniques for detecting and stopping malware rely on building a database of known signatures using known samples of malware. However, these techniques are not very effective at detecting zero-day exploits because there are no samples in their malware signature databases. The limitation of not being able to detect zero-day exploits leaves organisations vulnerable to new and evolving malware threats. To address this challenge, this thesis proposes a novel approach to malware detection using machine learning techniques. The proposed approach creates a user profile that trains a machine learning model using only normal user behaviour data, and detects malware by identifying deviations from this profile. In this way, the proposed approach can detect zero-day malware and other previously unknown threats without having a specific database of malware signatures. The proposed approach is evaluated using real-world datasets, and different machine learning algorithms are compared to evaluate their performance in detecting unknown threats. The results show that the proposed approach is effective in detecting malware, achieving high accuracy and low false positive rates. This thesis contributes to the field of malware detection by providing a new perspective and approach that complements existing methods, and has the potential to improve the overall security of organisations and individuals in the face of evolving cybersecurity threats.

Ultra-high-density single-atom catalysts (UHD-SACs) present unique opportunities for harnessing cooperative effects between neighboring metal centers. However, the lack of tools to establish correlations between the density, type, and arrangement of the isolated metal atoms with the support surface properties hinders efforts to engineer advanced material architectures. Here, we precisely describe the metal center organization in various mono- and multimetallic UHD-SACs based on nitrogen-doped carbon (NC) supports by coupling transmission electron microscopy with tailored machine-learning methods (released as a user-friendly web app) and density functional theory simulations. Our approach quantifies the non-negligible presence of multimers with increasing atom density, characterizes the size and shape of these low-nuclearity clusters, and identifies surface atom density criteria to ensure isolation. Further, it provides previously inaccessible experimental insights into coordination site arrangements in the NC host, uncovering a repulsive interaction that influences the disordered distribution of metal centers in UHD-SACs. This observation holds in multimetallic systems, where chemically-specific analysis quantifies the degree of intermixing. These fundamental insights into the materials chemistry of single-atom catalysts are crucial for designing catalytic systems with superior reactivity., This publication was created as part of NCCR Catalysis (grant number 180544), a National Centre of Competence in Research funded by the Swiss National Science Foundation. A. R.-F. acknowledges funding from the Generalitat de Catalunya and the European Union under Grant 2023 FI-3 00027. N.L. acknowledges support from the Ministerio de Ciencia e Innovación, ref. no. RTI2018-101394-B-100, and the Severo Ochoa Grant, MCIN/AEI/10.13039/501100011033-CEX2019-000925-S. The authors thank BSC-RES for generously providing computational resources., Peer Reviewed, Postprint (published version)

El auge de los sistemas de machine learning (ML), la mejora de sus capacidades y el mayor tamaño de los sistemas, ha incrementado el impacto medioambiental de los modelos ML. Sin embargo, la información sobre cómo se mide, comunica y evalúa la huella de carbono de los modelos de ML es escasa. Este proyecto, basado en un análisis de 1.417 modelos de ML y conjuntos de datos asociados en Hugging Face, el repositorio más popular para modelos de ML preentrenados, tiene como objetivo proporcionar una solución integrada para comprender, informar y optimizar la eficiencia de carbono de los modelos de ML. Además, implementamos una aplicación web que genera etiquetas de eficiencia energética para modelos de ML y permite visualizar sus emisiones de carbono. Con menos del 1% de los modelos en Hugging Face proporcionando información sobre las emisiones de carbono, el proyecto subraya la necesidad de mejorar las prácticas de reporte energético y la promoción del desarrollo de modelos eficientes en carbono dentro de la comunidad Hugging Face. Para abordar esta cuestión, ofrecemos una herramienta web que produce etiquetas de eficiencia energética para modelos de ML, una contribución que fomenta la transparencia y el desarrollo de modelos sostenibles dentro de la comunidad de ML. Permite la creación de etiquetas energéticas, al tiempo que proporciona valiosas visualizaciones de los datos de emisiones de carbono. Esta solución integrada constituye un paso importante hacia prácticas de IA más sostenibles medioambientalmente., The rise of machine learning (ML) systems has increased their environmental impact due to the enhanced capabilities and larger model sizes. However, information about how the carbon footprint of ML models is measured, reported, and evaluated remains scarce and scattered. Aims: This project, based on an analysis of 1,417 ML models and associated datasets on Hugging Face, the most popular repository for pretrained ML models, aims to provide an integrated solution for understanding, reporting, and optimizing the carbon efficiency of ML models. Moreover, we implement a web-based application that generates energy efficiency labels for ML models and visualizes their carbon emissions. With less than 1% of models on Hugging Face currently reporting carbon emissions, the project underscores the need for improved energy reporting practices and the promotion of carbon-efficient model development within the Hugging Face community. To address this, we offer a web-based tool that produces energy efficiency labels for ML models, a contribution that encourages transparency and sustainable model development within the ML community. It enables the creation of the energy labels, while also providing valuable visualizations of carbon emissions data. This integrated solution serves as an important step towards more environmentally sustainable AI practices.

Factors such as road traffic, challenging ambient temperature conditions, and extended periods of driving have detrimental effects on the physical and mental well-being of a driver. These factors can alter the stress levels, thereby diminishing his or her capacity to make effective decisions when faced with hazardous situations on the road. In this regard, this study presents a novel approach utilizing a multiple-input multiple-output (MIMO) radar system to accurately assess driver stress levels by measuring both physiological signals and driving behavior. The proposed method, assisted by a machine learning technique, provides comprehensive insights to classify the stress level of a driver into three states: drowsiness, awakeness, and anxiety. The MIMO radar system captures the elongation distance and velocity of six specific regions of the frontal torso of the driver in an advanced driving simulator based on virtual reality. This allows for the extraction of vital physiological parameters such as heart rate, respiratory rhythm, and breathing patterns over time, as well as the identification of changes in driving style determined by variations in their relative position in the seat and control of the steering wheel. Then, a fully connected neural network (FCNN) model is trained with the acquired data, and its performance is evaluated with volunteers submitted to different driving situations that induce stress in the driver. The findings show an accuracy in drowsiness detection of 90%, awakeness of 96% and anxiety of 85%., This work was supported by the Spanish “Comision Interministerial de Ciencia y Tecnologia” (CICYT) under projects PID2019-107885GB-C31, PID2022-136869NB-C31 and MDM2016-O6OO; Catalan Research Group 2017 SGR 219; and ”Secretaría Nacional de Educación Superior, Ciencia, Tecnología e Innovación” (SENESCYT) from the Ecuadorian government., Peer Reviewed, Postprint (author's final draft)

La digitalització d'éssers humans en entorns 3D ha estat durant dècades objecte d'estudi en la visió per computador i els gràfics digitals, però és encara un problema obert. Avui dia, cap tecnologia és capaç de digitalitzar persones amb una qualitat i dinamisme excel·lent, i que pugui ser utilitzada en motors 3D, com ara un casc de realitat virtual o un telèfon mòbil, en temps real. En aquesta tesi, intentem contribuir a aquest problema explorant com combinar els dos mètodes més usats en els últims anys: \textit{neural radiance fields} i models paramètrics 3D. Intentem dissenyar un model capaç de crear avatars digitals i animables de cares humans a velocitats raonables. La nostra feina se centra principalment a crear un model d'aprenentatge automàtic capaç de generar un avatar facial a partir d'una col·lecció d'imatges i càmeres, però també desenvolupem una eina per integrar l'obtenció d'aquestes dades, de manera que podem provar el nostre mètode en dades reals. A més, també implementem una llibreria per generar dades sintètiques, per tal de controlar els errors que podrien sorgir quan s'obtenen dades reals, per exemple problemes amb la calibració de les càmeres, i facilitar el desenvolupament d'altres projectes relacionats amb humans., La digitalización de seres humanos en entornos 3D ha sido durante décadas objeto de estudio en la visión por computador y los gráficos digitales, pero es aún un problema abierto. Actualmente, ninguna tecnología es capaz de digitalizar personas con una cualidad y dinamismo excelente, y que pueda ser usada en motores 3D, como por ejemplo un casco de realidad virtual o un teléfono móvil, en tiempo real. En esta tesis, intentamos contribuir a este problema explorando como combinar los dos métodos más usados en los últimos años: \textit{neural radiance fields} y modelos paramétricos 3D. Intentamos diseñar un modelo capaz de crear avatares digitales y animables de caras humanas a velocidades razonables. Nuestro trabajo se centra principalmente en crear un modelo de aprendizaje automático capaz de generar un avatar facial a partir de una colección de imágenes y cámaras, pero también desarrollamos una herramienta para obtener estos datos, de manera que podemos probar nuestro método en datos reales. Además, también implementamos una librería para generar datos sintéticos, para poder controlar los errores que pueden surgir al obtener datos reales, como problemas con la calibración de las cámaras, y facilitar el desarrollo de otros proyectos relacionados con humanos., Digitalizing humans in 3D environments has been a subject of study in computer vision and computer graphics for decades, but it still remains an open problem. No current technology can digitalize humans with excellent quality and dynamism that can be used in 3D engines, such as in a virtual reality headset or a mobile phone, at real-time speeds. In this thesis, we aim to contribute to this problem by exploring how to combine the two most commonly used approaches in recent years: neural radiance fields and parametric 3D meshes. We attempt to design a model capable of creating digital, animatable avatars of human faces at reasonable speeds. Our work focuses mostly on creating a machine learning model capable of generating a facial avatar from a set of images and camera poses, but we will also build a pipeline to integrate all steps of obtaining such data, allowing us to demonstrate our method in real-world data. Additionally, we implement a framework to generate synthetic data, in order to alleviate the errors in obtaining real-data, such as problems with camera calibration, and facilitate the development of other human-related projects., Outgoing

Molecular classification of medulloblastoma is critical for the treatment of this brain tumor. Array-based DNA methylation profiling has emerged as a powerful approach for brain tumor classification. However, this technology is currently not widely available. We present a machine-learning decision support system (DSS) that enables the classification of the principal molecular groups—WNT, SHH, and non-WNT/non-SHH—directly from quantitative PCR (qPCR) data. We propose a framework where the developed DSS appears as a user-friendly web-application—EpiGe-App—that enables automated interpretation of qPCR methylation data and subsequent molecular group prediction. The basis of our classification strategy is a previously validated six-cytosine signature with subgroup-specific methylation profiles. This reduced set of markers enabled us to develop a methyl-genotyping assay capable of determining the methylation status of cytosines using qPCR instruments. This study provides a comprehensive approach for rapid classification of clinically relevant medulloblastoma groups, using readily accessible equipment and an easy-to-use web-application., The study was supported by Associations of Parents and Families of Children with Cancer and by funding of the Spanish Ministry of for Science, Innovation and University (grant PI20/00519; PI CL) and the Foundation La Marató TV3 (grant 201921-30; PI CL). We acknowledge the multidisciplinary team who helped in the molecular analyses and care of patients, and the BioBank Hospital Sant Joan de Déu of the Spanish BioBank Network for sample procurement. We also acknowledge Marta Fortuny for communication strategy advice and Eduard Puig for legal assistance and data protection regulations. Authors acknowledge the SJD Fundraising Team., Peer Reviewed, Article signat per 23 autors/es: Soledad Gómez-González, Joshua Llano, Marta Garcia, Alicia Garrido-Garcia, Mariona Suñol, Isadora Lemos, Sara Perez-Jaume, Noelia Salvador, Nagore Gene-Olaciregui, Raquel Arnau Galán, Vicente Santa-María, Marta Perez-Somarriba, Alicia Castañeda, José Hinojosa, Ursula Winter, Francisco Barbosa Moreira, Fabiana Lubieniecki, Valeria Vazquez, Jaume Mora, Ofelia Cruz, Andrés Morales La Madrid, Alexandre Perera, Cinzia Lavarino., Postprint (published version)

Individuals with ADHD face significant challenges in their daily lives due to difficulties with attention, hyperactivity, and impulsivity. These challenges are especially pronounced in the workplace or educational settings, where the ability to sustain attention and manage time effectively is crucial for success. Virtual reality (VR) software has emerged as a promising tool for improving productivity in individuals with ADHD. However, the effectiveness of such software depends on the identification of potential distractions and timely intervention. The proposed computer screenshot classification approach addresses this need by providing a means for identifying and analyzing potential distractions within VR software. By integrating Convolutional Neural Networks (CNNs), Optical Character Recognition (OCR), and Natural Language Processing (NLP), the proposed approach can accurately classify screenshots and extract features, facilitating the identification of distractions and enabling timely intervention to minimize their impact on productivity. The implications of this research are significant, as ADHD affects a substantial portion of the population and has a significant impact on productivity and quality of life. By providing a novel approach for studying, detecting, and enhancing productivity, this research has the potential to improve outcomes for individuals with ADHD and increase the efficiency and effectiveness of workplaces and educational settings. Moreover, the proposed approach holds promise for wider applicability to other productivity studies involving computer users, where the classification of screenshots and feature extraction play a crucial role in discerning behavioral patterns., Les persones amb TDAH s’enfronten a reptes importants en la seva vida diària a causa de les dificultats d’atenció, hiperactivitat i impulsivitat. Aquests reptes són especialment pronunciats al lloc de treball o en entorns educatius, on la capacitat de mantenir l’atenció i gestionar el temps de manera eficaç és crucial per a l’èxit. El software de realitat virtual (RV) s’ha revelat com a eina prometedora per millorar la productivitat de les persones amb TDAH. Tanmateix, l’eficàcia del software esmentat depèn de la identificació de distraccions potencials i de la intervenció oportuna. L’enfocament de classificació de captures de pantalla d’ordinador proposat aborda aquesta necessitat proporcionant un mitjà per identificar i analitzar les distraccions potencials dins del programari de RV. Mitjançant la integració de xarxes neuronals convolucionals (CNN), el reconeixement òptic de caràcters (OCR) i el processament del llenguatge natural (NLP), l’enfocament proposat pot classificar amb precisió les captures de pantalla i extreure’n característiques, facilitant la identificació de les distraccions i permetent una intervenció oportuna per minimitzar-ne l’impacte en la productivitat. Les implicacions d’aquesta investigació són importants, ja que el TDAH afecta una part substancial de la població i té un impacte significatiu a la productivitat i la qualitat de vida. En proporcionar un enfocament nou per estudiar, detectar i millorar la productivitat, aquesta investigació té el potencial de millorar els resultats per a les persones amb TDAH i augmentar l’eficiència i l’eficàcia dels llocs de treball i els entorns educatius. A més, l’enfocament proposat promet una aplicabilitat més gran a altres estudis de productivitat en què participin usuaris d’ordinadors, en què la classificació de captures de pantalla i l’extracció de característiques tenen un paper crucial a l’hora de discernir patrons de comportament.

Online machine learning (ML) inference services provide users with an interactive way to request for predictions in realtime. To meet the notable computational requirements of such services, they are increasingly being deployed in the Cloud. In this context, the efficient provisioning and optimization of ML inference services in the Cloud is critical to achieve the required performance and meet the dynamic queries by end-users. Existing provisioning solutions focus on framework parameter tuning and infrastructure resources scaling, without considering deployments based on containerization technologies. The latter promises reproducibility and portability features for ML inferences services. There is limited knowledge about the impact of distinct deployment schemes at the container-level on the performance of online ML inference services, particularly on how to exploit multi-container deployments and its relation with processor and memory affinity. In light of this, in this paper we investigate experimentally the containerization of ML inference services and analyze the performance of multi-container deployments that partition the threads belonging to an online learning application into multiple containers in each node. This paper shares the findings and lessons learned from conducting realistic client patterns on an image classification model across numerous deployment configurations, especially including the impact of container granularity and its potential to exploit processor and memory affinity. Our results indicate that fine-grained multi-container deployments and affinity are useful for improving performance (both throughput and latency). In particular, our experiments on single-node and four-node clusters show up to 69% and 87% performance improvement compared to the single-container deployment, respectively., This work was partially supported by Lenovo as part of Lenovo-BSC collaboration agreement, by the Spanish Government under contract PID2019-107255GB-C22, and by the Generalitat de Catalunya under contract 2021-SGR-00478 and under grant 2020 FI-B 00257., Peer Reviewed, Postprint (author's final draft)

Super Resolution (SR) techniques are powerful digital manipulation tools that have significantly impacted various industries due to their ability to enhance the resolution of lower quality images and videos. Yet, the real-world adaptation of SR models poses numerous challenges, which blind SR models aim to overcome by emulating complex real-world degradations. In this thesis, we investigate these SR techniques, with a particular focus on comparing the performance of blind models to their non-blind counterparts under various conditions. Despite recent progress, the proliferation of SR techniques raises concerns about their potential misuse. These methods can easily manipulate real digital content and create misrepresentations, which highlights the need for robust SR detection mechanisms. In our study, we analyze the limitations of current SR detection techniques and propose a new detection system that exhibits higher performance in discerning real and upscaled videos. Moreover, we conduct several experiments to gain insights into the strengths and weaknesses of the detection models, providing a better understanding of their behavior and limitations. Particularly, we target 4K videos, which are rapidly becoming the standard resolution in various fields such as streaming services, gaming, and content creation. As part of our research, we have created and utilized a unique dataset in 4K resolution, specifically designed to facilitate the investigation of SR techniques and their detection.

Els mètodes tradicionals que intenten resoldre tasques a la vida real amb robots mitjançant l'ús de la IA, tenen molts inconvenients. L'aprenentatge per imitació dona bons resultats a l'hora de resoldre tasques molt específiques, però necessita una gran quantitat de demostracions d'experts per poder superar el problema de propagació d'errors. Els mètodes centrats en l'aprenentatge per reforç també presenten un problema fonamental, necessiten una funció de recompensa curosament dissenyada, que pot ser arbitràriament complicada d'especificar, en especial quan tractem amb imatges, i ha de ser densa per a poder resoldre les tasques més complicades. En aquest treball, explorem com els robots poden aprendre en el món real mitjançant aprenentatge autosupervisat a través d'un procés d'anotacions retrospectives. En particular, per aconseguir-ho utilitzem opinions humanes en forma de comparacions binàries, per tal de guiar l'exploració. Aquest tipus d'anotacions es poden recollir asíncronament i de manera molt escalable mitjançant el crowdsourcing. També ensenyem com el nostre mètode és robust als errors en les anotacions i com en necessita molt menys que altres algoritmes. Finalment, ensenyem com podem modificar el nostre mètode perquè pugui ser executat contínuament, sense reinicis. En aquest estudi mostrem com el nostre algoritme supera a altres mètodes clàssics i com pot ser aplicat en el món real., Los métodos tradicionales que tratan de resolver tareas en el mundo real con robots entrenados mediante IA, tienen muchos inconvenientes. El aprendizaje por imitación da buenos resultados a la hora de resolver tareas muy específicas, pero necesita gran cantidad de demostraciones de expertos para poder lidiar con la propagación de errores. Enfoques basados en aprendizaje por refuerzo también presentan un problema fundamental, necesitan una función de recompensa cuidadosamente diseñada, la cual es específica a una tarea en concreto, puede ser arbitrariamente difícil de especificar, especialmente cuando trabajamos con imágenes, y necesita ser densa para permitir resolver las tareas más complicadas. En este trabajo, exploramos como los robots pueden aprender en el mundo real usando aprendizaje autosupervisado a través de un proceso de anotaciones retrospectivas. En particular, para lograrlo usamos opiniones humanas, en forma de comparaciones binarias, para guiar la exploración. Este tipo de anotaciones se pueden recoger asíncronamente y de manera muy escalable usando crowdsourcing. También mostramos como nuestro método es robusto a los errores en las anotaciones y como necesita mucha menos supervisión humana que otros algoritmos. Finalmente, mostramos cómo podemos modificar nuestro algoritmo para que pueda ser ejecutado continuamente, sin reseteos. En este estudio, mostramos como nuestro algoritmo supera a otros métodos clásicos y como puede ser aplicado en la vida real., Traditional methods that aim to solve tasks involving robots in the real world leveraging AI, suffer from many drawbacks: Behaviour Cloning shows great results at solving specific tasks, but it requires a considerably large amount of expert demonstrations to deal with compounding errors. Approaches based on Reinforcement Learning also present a fundamental problem, namely reward shaping, which is task-specific, can be non-trivial to specify, especially when dealing with images, and needs to be dense in order to allow us to solve hard tasks. In this work, we explore how robots can learn in the real world by using self-supervised learning via hindsight relabeling. In particular, to do so, we leverage human feedback, in the form of binary comparisons, to guide exploration. This type of feedback can be collected asynchronously and is easily scalable via crowd-sourcing. We also show that our method is robust to noisy labels and requires less feedback than other algorithms. Finally, we show how we can adapt our algorithm to make it run continuously without resets. In this work, we show how our algorithm outperforms all baselines in some simulation benchmarks and how it can also be used in the real world., Outgoing

En aquesta memòria es presenten totes les anàlisis resultants del treball dut a terme en el meu treball final de grau en el marc de la col·laboració entre l'inLab FIB i KaptureIO, per millorar l'eficiència i la qualitat en línies de producció de les Indústries 4.0. Actualment, es realitza un control de qualitat a partir de la combinació de dos mètodes ben diferenciats. Per començar, s'executa un control de toleràncies en les quals, a partir de la mesura de múltiples punts en el vehicle, es detecten possibles errates en el procés productiu. Tot seguit, si es detecta alguna errata, es du a terme una parada en la línia i s'elabora una anàlisi manual, el qual permetrà concloure les modificacions a dur a terme. En aquest projecte s'han analitzat les diferents variables que intervenen en el control de qualitat, buscant desenvolupar algorismes capaços d'anticipar-se a les possibles futures errates. A partir de la implementació d'algorismes d'aprenentatge automàtic s'ha assolit l'anticipació, virtualment, del 71.4% d'errates. Així doncs, els resultats d'aquest projecte suposen un pas endavant en la millora de l'eficiència i la qualitat en les línies de producció, gràcies al fet que es detecten amb anticipació les possibles errates o desviacions en la fabricació dels vehicles. Finalment, gràcies als resultats assolits, l'empresa vol continuar col·laborant amb nosaltres amb objectius més ambiciosos., This report presents all the analyzes resulting from the work carried out in my final degree work in the framework of the collaboration between the inLab FIB and KaptureIO, to improve the efficiency and quality in production lines of Industries 4.0. Currently, quality control is carried out using a combination of two distinct ways. To begin, a tolerance control is performed in which probable mistakes in the manufacturing process are recognized based on measurements of several points on the vehicle. If a mistake is identified, the line is stopped and a manual analysis is performed, allowing the alterations to be carried out to be completed. In this study, the various variables involved in quality control were examined in order to develop algorithms capable of anticipating potential future faults. Based on the application of machine learning techniques, a virtual prediction of 71.4% of mistakes has been obtained. As a result, the outcomes of this study constitute a step forward in improving efficiency and quality in production lines, because potential errors or deviations in car manufacturing are spotted in advance. Finally, as a result of the achievements obtained, the company want to continue partnering with us on more ambitious aims.

Aquesta tesi explora el paper de MLOps per a proporcionar eficiència i productivitat en el desplegament, monitoratge i gestió de models d'aprenentatge automàtic en entorns de producció. L'informe compta amb una part teòrica que pretén aportar detalls sobre el funcionament de les pràctiques de MLOps i les tecnologies relacionades. Això serveix de complement per a la part pràctica, la qual es considera la principal contribució. Consisteix en el desenvolupament d'un framework que pretén recollir algunes de les principals funcionalitats de MLOps. Això ajuda a complir l'objectiu principal de demostrar la seva utilitat i comprendre per què MLOps és cada vegada més important. El desenvolupament del framework implica l'ús de Python, Docker, Streamlit i Airflow, cadascun necessari per a proporcionar diferents funcionalitats de MLOps.

Aquest treball es centra en el desenvolupament d'algoritmes de Machine Learning per a la detecció i classificació de falles en una màquina a partir de dades generades pels seus sensors. Aquestes mesures dels sensors (un tacòmetre i sis acceleròmetres) han estat analitzades i transformades en els indicadors més adients per tal de que els models de Machine Learning tinguin més facilitat per trobar patrons en les dades i determinar l’estat de la màquina. S’ha realitzat un algoritme per a la detecció de falles, utilitzant com a model la distribució normal multivariada, i dos algoritmes per a la seva classificació segons el tipus de falla de què es tracti, mitjançant una xarxa neuronal i un Random Forest, respectivament. El fet de poder identificar falles abans de que causin complicacions importants pot conduir a una reducció en els costos de manteniment i a un allargament de la vida útil de les màquines. Per aquest motiu, s’han desenvolupat algoritmes de Machine Learning que duguin a terme aquesta tasca i s’ha obtingut una precisió molt elevada. A més, s’han dut a terme diverses metodologies per evitar l’overfitting en els algoritmes (que els models s’ajustin massa a les dades d’entrenament i no funcionin tan positivament per a dades no vistes prèviament).

La industria alimentaria actual es insostenible y altamente perjudicial para el medio ambiente. Dado que un cambio en el modelo de consumo de la población podría incentivar a dicha industria a evolucionar y mejorar, apoyar actividades como la avicultura de autoconsumo es importante. El objetivo de este proyecto es dar apoyo a estos avicultores proporcionando información precisa basada en conocimiento experto y la experiencia de otros usuarios para facilitar la selección de las razas que más se adecúan a su situación y necesidades. Para ello se han recopilado datos de usuarios y de distintas razas de gallina, a partir de los cuales se ha construido un sistema recomendador basado en técnicas de aprendizaje automático., The current food industry is unsustainable and highly prejudicial for the environment. Given that a change in the consumption model of the population might incentivate said industry to evolve and improve, giving support to activities such as self-consumption poultry farming is important. The aim of this project is to support these farmers by providing accurate information based on expert knowledge and other users' experience to help them select the breeds that best match their situation and needs. In order to do so, a recommender system based on machine learning techniques has been built using collected data from users and on different chicken breeds.

The primary objective of this research was to create a predictive model capable of identifying socio-demographic and clinical characteristics potentially associated with the risk of experiencing gender-based violence. Additionally, the aim was to predict the likelihood of suffering gender-based violence using machine learning techniques by evaluating data from the 2019 macro-survey on violence against women conducted by the Ministry of Equality of the Government of Spain. To achieve these goals, the study employed two feature selection methods: Recursive Feature Elimination (RFE) and LASSO Regularized Regression. Furthermore, two predictive models, namely Support Vector Machines (SVM) and Random Forest (RF), were utilized to analyze the data and make predictions.

El següent projecte de recerca té com a objectiu facilitar l’accés i l’ús d’imatges mèdiques, en específic imatges histopatològiques, als científics, enginyers i professionals de la salut mitjançant la generació d’imatges sintètiques. Aquestes imatges sintètiques, en tractar-se d’imatges considerades anònimes, compleixen amb les regulacions de protecció de privacitat de dades mèdiques. Aquest enfocament permet crear un conjunt de dades diverses i representatives que aborda els reptes de l’accés limitat i sensible a l’accés de dades mèdiques i proporcionar una eina per entrenar models d’aprenentatge automàtic que tinguin la capacitat de diagnosticar malalties, com per exemple, el càncer de mama, millorant la precisió dels models i la fiabilitat dels resultats., El siguiente proyecto de investigación tiene como objetivo facilitar el acceso y uso de imágenes médicas, en específico imágenes histopatológicas, a los científicos, ingenieros y profesionales de la salud mediante la generación de imágenes sintéticas. Estas imágenes sintéticas, al tratarse de imágenes consideradas anónimas, cumplen con las regulaciones de protección de privacidad de datos médicos. Este enfoque permite crear un conjunto de datos diversos y representativos que aborda los retos del acceso limitado y sensible al acceso de datos médicos y proporcionar una herramienta para entrenar modelos de aprendizaje automático que tengan la capacidad de diagnosticar enfermedades, como por ejemplo, el cáncer de mama, mejorando la precisión de los modelos y la confiabilidad de los resultados., The following research project aims to facilitate the access and use of medical images, specifically histopathological images, to scientists, engineers and health professionals through the generation of synthetic images. These synthetic images, as they are considered anonymous images, comply with the regulations for the protection of medical data privacy. This approach makes it possible to create a diverse and representative dataset that addresses the challenges of limited and sensitive access to medical data and provides a tool for training machine learning models that have the ability to diagnose diseases such as for example, breast cancer, improving the accuracy of the models and the reliability of the results.

The main goal of this thesis is to design and develop an insurance recommender system using machine learning. The AI model will analyze the risks and opportunities within several determined cross selling prospects in order to offer products and options that are relevant. While using state of the art prediction algorithms and explainable AI, the aim is to adapt the offers based on what other similar people with similar portfolios have among other parallel strategies, for both our current and the prospective customer base.

Peer Reviewed, Postprint (published version)

This project starts by studying the finite element method for the steady state Navier-Stokes equations, afterwards it is implemented in Matlab and optimized via Vectorization achieving up to 5000x speed-up in some calculations. Then a reduced order model is studied to decrease the computational time of performing different simulations with slight modifications to the input parameters. Finally, the results obtained are compared against an already tested FEM code, Kratos Multiphysics, and against literature, and the performance of the developed solver for the equations is analyzed. It has been observed that the results obtained with the present work’s solver are almost equal to those made by the reference alternatives.

La telemedicina, una solució transformadora per connectar pacients i proveïdors de salut, especialment en àrees remotes i desateses, ha captat una atenció significativa pel seu potencial per aprofitar les tècniques de Processament de Llenguatge Natural (NLP, de l'anglès) en l'extracció automatitzada d'informació. Aquesta tesi se centra en desenvolupar un model de Reconeixement d'Entitats Nomenades (NER, de l'anglès) adaptat al llenguatge mèdic, amb l'objectiu de millorar l'eficiència i precisió dels serveis de telemedicina proporcionats per Abi Global Health. El projecte, conegut com a MermAId (Desenvolupament del Model de Reconeixement d'Entitats Mèdiques), va ser proposat pel Laboratori d'Intel·ligència Artificial (AI Lab) d'Abi Global Health en col·laboració amb la Universitat Politècnica de Catlalunya (UPC). El departament d'AI Lab és responsable de desenvolupar i gestionar diversos models de ML i IA per donar suport i millorar els serveis de l'empresa. El model final de MermAId és capaç d'identificar 27 classes d'entitats mèdiques en les oracions d'entrada. El model es basa en un model BERT preentrenat en un gran corpus de dades en anglès, que s'ha afinat amb preguntes de pacients i respostes de professionals de la salut recopilades durant les consultes mèdiques d'Abi. La tesi descriu els procediments de recopilació de dades, anotació i pre-processament, s'ha implicat la col·laboració amb experts mèdics per garantir la creació d'un conjunt de dades etiquetat d'alta qualitat. Posteriorment, s'han utilitzat tècniques avançades d'aprenentatge automàtic per entrenar el model NER, optimitzant el seu rendiment mitjançant experimentació iterativa i ajustament d'hiperparàmetres. El rendiment del model s'ha avaluat en un conjunt dades etiquetat per humans utilitzant mètriques personalitzades i restriccions de qualitat. El model resultant obté un rendiment gairebé de tecnologia d'avantguarda en comparació amb les solucions comercials existents i ha superat els algoritmes, Telemedicine, a transformative solution for connecting patients and healthcare providers, particularly in remote and underserved areas, has garnered significant attention for its potential to leverage natural language processing (NLP) techniques in automated information extraction. This thesis focuses on developing a Named Entity Recognition (NER) model tailored to medical language, with the goal of enhancing the efficiency and accuracy of telemedicine services provided by Abi Global Health. The project, known as MermAId (Medical Entity Recognition Model Development), was proposed by the Abi Global Health Artificial Intelligence Laboratory (AI Lab) in collaboration with Polytechnic University of Catalonia (UPC). The AI Lab department is responsible for developing and managing various ML and AI models to support and improve the company's services. The final MermAId model is capable of identifying 27 medical entities in text data. The model is based on a BERT model pretrained on a large corpus of English data, fine-tuned on patient questions and health professional responses collected during Abi's medical consultations. The thesis outlines the data collection, annotation, and preprocessing procedures, which entailed collaborating with medical experts to ensure the creation of a high-quality labeled dataset. Subsequently, advanced machine learning techniques were employed to train the NER model, optimizing its performance through iterative experimentation and hyperparameter tuning. The model's performance was assessed on a human-labeled set using customized evaluation metrics and quality restrictions. The resulting model achieved near state-of-the-art performance when compared to existing commercial solutions and outperformed the previous NER algorithms at Abi. MermAId will now serve as the main NER tool for all Abi products and services, ultimately contributing to improve telemedicine consultations worldwide.

La citometria de flux és una de les principals eines d'anàlisi cel·lular des de fa més de quaranta anys i és present tant en els laboratoris de recerca com en diagnòstic clínic. Les poblacions cel·lulars convencionalment es caracteritzen mitjançant el procediment de gating, el qual consisteix a utilitzar histogrames o gràfics de punts de manera seqüencial per tal d'anar delimitant manualment les poblacions. Aquest procediment, però, és lent, poc precís i no resulta adient per a les noves màquines de citometria capaces de treballar amb més de 30 paràmetres per cèl·lula. En els darrers anys noves tècniques computacionals han estat desenvolupades per gestionar les massives quantitats de dades generades, no obstant això, tot i que es pot trobar una gran varietat d'algoritmes a l'abast, internacionalment en el dia d'avui, no hi ha cap procediment estàndard a l'hora de fer servir un algoritme o altre que asseguri obtenir els resultats òptims. L'objectiu principal d'aquest projecte és dur a terme un estudi comparatiu entre diversos algoritmes d'anàlisi hiperdimensional basats en aprenentatge automàtic per tal de proporcionar als citometristes un punt de partida on explorar tècniques de visualització de dades generades per citometria de flux de manera reproduïble. Amb aquesta finalitat, s'ha implementat una metodologia estandarditzada per a l'avaluació del rendiment de les respectives eines d'anàlisi., La citometría de flujo es una de las principales herramientas de análisis celular desde hace más de cuarenta años y es presente tanto en los laboratorios de investigación como en diagnóstico clínico. Las poblaciones celulares convencionalmente se caracterizan mediante el procedimiento de gating, el cual consiste en utilizar histogramas o gráficos de puntos de manera secuencial para ir delimitando manualmente las poblaciones. Este procedimiento, pero, es lento, poco preciso y no resulta conveniente para las nuevas máquinas de citometría capaces de trabajar con más de 30 parámetros por célula. En los últimos años nuevas técnicas computacionales han sido desarrolladas para gestionar las masivas cantidades de datos generados, sin embargo, a pesar de que se puede encontrar una gran variedad de algoritmos al alcance, internacionalmente en el día de hoy, no hay ningún procedimiento estándar en la hora de usar un algoritmo u otro que asegure obtener los resultados óptimos. El objetivo principal de este proyecto es llevar a cabo un estudio comparativo entre varios algoritmos de análisis hiperdimensional basados en aprendizaje automático para proporcionar a los citometristas un punto de partida donde explorar técnicas de visualización de datos generados por citometría de flujo de manera reproducible. Con cuyo objeto, se ha implementado una metodología estandarizada para la evaluación del rendimiento de las respectivas herramientas de análisis., Flow cytometry is one of the main tools of cell analysis for over 40 years and is present in both research laboratories and clinical diagnosis. Cell populations are conventionally characterized by the gating procedure, which involves using sequential histograms or dot plots to manually delimit populations. This procedure, however, is slow, inaccurate and inappropriate for new cytometry machines capable of working with more than 30 parameters per cell. In recent years new computational techniques have been developed to handle the massive amounts of data generated, however, although a variety of algorithms can be found in reach, internationally today, there is no standard procedure for using an algorithm or other that ensures optimal results. The main objective of this project is to carry out a comparative study among various hyperdimensional analysis algorithms based on machine learning in order to provide cytometrists with a starting point where to explore data visualization techniques generated by flow cytometry in a reproducible way. To this end, a standardised methodology has been implemented for evaluating the performance of the respective analysis tools.

Huge efforts have been paid lastly to study the application of Machine Learning techniques to optical transport networks. Applications include Quality of Transmission (QoT) estimation, failure and anomaly detection, and network automation, just to mention a few. In this regard, the development of Optical Layer Digital Twins able to accurately model the optical layer, reproduce scenarios, and generate expected signals are of paramount importance. In this paper, we introduce two applications of Optical Layer Digital Twins namely, misconfiguration detection and QoT estimation. Illustrative results show the accuracy and usefulness of the proposed applications., The research leading to these results has received funding from the European Community through the MSCA MENTOR (G.A. 956713) and the HORIZON SEASON (G.A. 101096120) projects, the AEI through the IBON (PID2020-114135RB-I00) project, and by the ICREA institution., Peer Reviewed, Postprint (author's final draft)

Modern machine learning systems are increasingly trained on large amounts of data embedded in high-dimensional spaces. Often this is done without analyzing the structure of the dataset. In this work, we propose a framework to study the geometric structure of the data. We make use of our recently introduced non-negative kernel (NNK) regression graphs to estimate the point density, intrinsic dimension, and linearity of the data manifold (curvature). We further generalize the graph construction and geometric estimation to multiple scales by iteratively merging neighborhoods in the input data. Our experiments demonstrate the effectiveness of our proposed approach over other baselines in estimating the local geometry of the data manifolds on synthetic and real datasets., Our work was supported in part by DARPA grant (FA8750-19-2-1005) in the Learning with Less Labels (LwLL) program., Peer Reviewed, Postprint (author's final draft)