Search

TinyML tiene como objetivo la implementación de aplicaciones de aprendizaje automático en dispositivos de poco tamaño y baja potencia, como los microcontroladores. Normalmente los dispositivos periféricos necesitan estar conectados a centro de datos para poder ejecutar aplicaciones de aprendizaje automático. Sin embargo, este método no es posible en muchos escenarios, como en la falta de conectividad. Este estudio investiga las herramientas y técnicas utilizadas en TinyML, las limitaciones en la utilización de dispositivos de baja potencia y la viabilidad de implementar aplicaciones avanzadas de aprendizaje automático en microcontroladores. Se desarrollaron tres programas para comprobar la viabilidad de implementar aplicaciones de aprendizaje automático en microcontroladores. El primero, una aplicación capaz de reconocer un conjunto de palabras claves. El segundo, un programa capaz de entrenar un modelo de red neuronal en el microcontrolador siguiendo un enfoque de aprendizaje en línea. Y el tercero, un programa de aprendizaje federado capaz de entrenar un único modelo global con la agregación de modelos locales entrenados en múltiples microcontroladores. Los resultados muestran un óptimo rendimiento de las tres aplicaciones una vez desplegadas en los microcontroladores. El desarrollo de aplicaciones básicas de TinyML resulta sencillo una vez entendidos el proceso de aprendizaje automático. Sin embargo, el desarrollo de aplicaciones avanzadas es muy complejo, ya que requiere un profundo conocimiento tanto del aprendizaje automático como de los sistemas embebidos. Estos resultados demuestran la viabilidad de implementar con éxito aplicaciones avanzadas de aprendizaje automático en microcontroladores, y por lo tanto, desvelan un futuro brillante para TinyML. TinyML aims to implement machine learning (ML) applications on small, and low-powered devices like microcontrollers. Typically, edge devices need to be connected to data centers in order to run ML applications. However, this approach is not possible in many scenarios, such as lack of connectivity. This project investigates the tools and techniques used in TinyML, the constraints of using low-powered devices, and the feasibility of implementing advanced machine learning applications on microcontrollers. To test the feasibility of implementing ML applications on microcontrollers, three TinyML programs were developed. The first, a basic keyword spotting application able to recognize a set of words. The second, a program for training a neural network model on a microcontroller following an online learning approach. And the third, a federated learning program able to train a single global model with the aggregation of local models trained on multiple microcontrollers. The results show optimal performance in all three applications once deployed on microcontrollers. The development of basic TinyML applications is straightforward when the machine learning pipeline is understood. However, the development of advanced applications turned out to be very complex, as it requires a deep understanding of both machine learning and embedded systems. These results prove the feasibility of successfully implementing advanced ML applications on microcontrollers, and thus, unveil a bright future for TinyML.

1r Premi Domènec Valero 2021 Aquest treball s’emmarca en l’adopció de la Intel·ligència Artificial en l’entorn industrial i es pretén mostrar les xarxes neuronals artificials com a mitjà per fer front a possibles mancances o limitacions actuals i l’automatització de processos que no aportin valor afegit. Per l’elaboració d’aquest treball s’ha realitzat un profund estudi i recerca, en concret, en les xarxes neuronals artificials com a l’algoritme empleat en Aprenentatge Automàtic. Simulen el comportament de les neurones biològiques i són capaces d’aprendre a trobar els patrons que resideixen en el conjunt de dades d’entrada i elaborar-ne un model. Es va detectar el procés de control de qualitat com un procés crític en el sector industrial per garantir la qualitat del producte i augmentar el rendiment productiu total. Per aquest motiu, s’ha implementat un sistema intel·ligent basat en el disseny i entrenament d’un algoritme amb xarxes neuronals convolucionals que ha sigut capaç d’aprendre correctament a classificar imatges de peces defectuoses i no defectuoses resultants d’un procés productiu. L’algoritme s’ha desenvolupat amb les llibreries TensorFlow i Keras d’Aprenentatge Automàtic. Per tal d’aconseguir un model de classificació òptim, es va realitzar un anàlisi de resultats obtinguts per diferents valors d’hiperparàmetres. Es va obtenir una precisió del 97.5% per la classificació de les imatges de validació per la qual cosa s’ha conclòs que les xarxes neuronals artificials poden presentar-se com una alternativa rendible a la programació convencional per la seva flexibilitat i adaptació a entorns amb molta variabilitat. This project refers to the adoption of Artificial Intelligence in the industrial environment and aims to show the artificial neural networks as a means to deal with possible weaknesses or current limitations and the automation of processes that do not provide added value. To elaborate this project, a deep study and research has been carried out, in particular, in artificial neural networks as the algorithm used in Machine Learning. They simulate the behaviour of biological neurons and are able to learn to find the patterns that reside in the input dataset and develop a model. The quality inspection process was detected as a critical process in the industrial sector to ensure product quality and increase overall production efficiency and performance. For this reason, an intelligent system has been implemented based on the design and training of an algorithm with convolutional neural networks that has been able to correctly learn to classify images of defective and non-defective parts resulting from a production process. The algorithm was developed with the TensorFlow and Keras machine learning libraries. In order to achieve an optimal classification model, an analysis of results obtained by different hyperparameter values was performed. An accuracy of 97.5% was obtained for the classification of validation images so it has been concluded that artificial neural networks can be a cost-effective alternative to conventional programming due to their flexibility and adaptation to environments with a lot of variability. Award-winning

En la última década han sido grandes y numerosos los avances realizados en el campo de la inteligencia artificial y el aprendizaje automático. Estos avances fueron impulsados inicialmente gracias al descubrimiento de que la capacidad para el cálculo matricial que tienen las GPUs podía aplicarse en el entrenamiento de modelos de aprendizaje profundo, de forma que modelos que hubieran tardado semanas en entrenarse a principios de la década ahora se pueden entrenar en cuestión de horas, a lo que se añadió la revolución que supone el cloud computing, que permite trasladar la ejecución de los modelos de aplicaciones de aprendizaje automático a grandes centros de computación. Esta nueva realidad ha permitido la creación de modelos de aprendizaje automático cada vez más grandes y complejos, capaces de aportar soluciones a problemas que podrían resultar impensables en el pasado. Aun así, en la segunda mitad de la última década se han detectado situaciones donde sería muy beneficioso poder ejecutar modelos de aprendizaje automático en dispositivos de potencia computacional reducida sin tener que recurrir al uso de los centros de computación antes mencionados, de forma que el paradigma cambia de intentar únicamente conseguir que los modelos tengan el mayor grado de acierto a intentar maximizar la eficacia de los modelos a la vez que se minimizan los recursos necesarios para ejecutarlos. TinyML es la máxima expresión de este cambio de paradigma. TinyML es un campo de rápido crecimiento dentro de las tecnologías de aprendizaje automático, que incluye algoritmos, hardware y software capaces de realizar análisis de datos provenientes de sensores presentes en un microcontrolador, en el mismo dispositivo que tiene los sensores, con un consumo energético extremadamente reducido. Este tipo de aplicaciones permiten utilizar técnicas de aprendizaje automático en gran número de nuevas situaciones aun por explorar, pero debido a que se trata de una tecnología de reciente popularización la información relativa a la misma es muy limitada. Este trabajo tiene el objetivo que documentar y analizar el proceso de creación de aplicaciones TinyML y aportar información sobre que decisiones se pueden tomar durante su desarrollo y sus consecuencias, además de evaluar la eficacia real de este tipo de aplicaciones una vez desplegadas. La información resultante de esta investigación servirá para que nuevos desarrolladores interesados en esta tecnología puedan aprender a usarla de forma adecuada y dispongan de datos suficientes para tomar decisiones informadas a lo largo del desarrollo de este tipo de aplicaciones. In the last decade there have been great and numerous advances made in the field of artificial intelligence and machine learning. These advances were initially driven by the discovery that GPUs' matrix computing capabilities could be applied in deep learning model training, so models that would have taken weeks to train at the beginning of the decade can now be trained in a matter of hours, and the revolution brought by cloud computing, which makes it possible to transfer the execution of machine learning application models to large computing centers. This new reality has allowed the creation of increasingly large and complex machine learning models, capable of providing solutions to problems that might have been unthinkable in the past. Even so, at the second half of the last decade, people started to notice situations where it would be very beneficial to be able to run machine learning models on devices with reduced computational power without having to resort to the use of the aforementioned computing centers. This changed the paradigm from only trying to get models with the highest accuracy, to instead trying to maximize the accuracy of the models while minimizing the resources required to run them. TinyML is the ultimate expression of this paradigm shift. TinyML is a rapidly growing field within machine learning technologies, which includes algorithms, hardware, and software capable of performing data analysis from sensors present in a microcontroller, in the same device that has said sensors, with extremely low energy consumption. This type of application allows the use of machine learning techniques for many new situations yet to be explored, but due to the fact that it is a recently popularized technology, the information related to it is very limited. The objective of this project is to document and analyze the process of creating TinyML applications and provide information on what decisions can be made during their development and the consequences of those decisions, in addition to evaluating the real effectiveness of this type of applications once deployed. The information resulting from this research will shall be used by developers interested in this technology, so that they can learn to use it properly and have enough data to make informed decisions throughout the development of this type of applications.

Malicious websites are currently one of the most severe threats to internet users. Traditional methods that detect malicious websites, such as blacklists, do not update frequently, and they cannot detect new attackers. To address this need, we propose a system that can be inserted into a firewall or Google Chrome extension. Starting with a dataset that contains both malicious and benign websites, we classify it by extracting numerous features. Furthermore, the features are parsed, analyzed, and preprocessed to create easily separable and categorized data. With the preprocessed data, the study proposes a Deep Learning (DL) model to classify each sample. The model consists of a Feed-Forward Neural Network (FFNN). We evaluate different combinations of neurons in the model and computes an in-depth study of the best performing network. Our results show up to 99.88% of detection of malicious websites and 2.61% of false hits (i.e. malicious websites classified as benign). Additionally, the system was tested with 10000 unseen websites, and the false hit rate decreased to 1.026%. To approach the system to end-users, an interface is created to insert the suspicious URL and return the prediction output. The interface response time gives a mean response time of 2.53 seconds. This value is beneath the 4-second limit where the user's start to lose attention. Overall, the proposed system can correctly classify both previously seen and unseen data, protecting thus against new attackers. Additionally, the interface creates a space where the user can quickly consult the URL maliciousness. Hoy en día, las páginas web maliciosas son una de las mayores amenazas que pueden tener los usuarios de internet. Los métodos tradicionales para detectar páginas web maliciosas, como las listas negras, no se actualizan con la rapidez necesaria para poder identificar con precisión nuevos tipos de ataques en la web. Para solucionar esta necesidad proponemos un sistema que se puede añadir en un cortafuegos o en una extensión de Google Chrome. Empezando con un conjunto de datos (dataset) que contiene páginas web maliciosas y benignas, los clasificamos extrayendo numerosos atributos. Además, los atributos se formatean, se analizan y se procesan previamente para crear datos fácilmente separables y clasificables. El modelo consiste en una red neuronal prealimentada (FFNN). Evaluamos distintas combinaciones de neuronas en el modelo y realizamos un estudio en profundidad de la red con mejor rendimiento. Nuestros resultados muestran una precisión de detección de páginas web maliciosas de hasta el 99,88% y, un 2,61% de detección de falso positivo (es decir, sitios web malignos clasificados como benignos). El sistema ha sido probado con 10.000 páginas webs no vistas previamente y la detección de falso positivo disminuye hasta el 1,026%. Para hacer más accesible el sistema a los usuarios finales, se desarrolla una interfaz gráfica que al insertar una URL sospechosa devuelve el resultado obtenido de esta. También se estudia el tiempo de respuesta de la interfaz. Este proporciona un valor medio de 2,53 segundos, que es un valor inferior al límite de 4 segundos en que el usuario empieza a perder la atención. En general el sistema propuesto es capaz de clasificar correctamente los datos tanto vistos con anterioridad como los que no. Por tanto, protege de nuevos ataques. Aparte, la interfaz crea un espacio donde el usuario puede consultar rápidamente la malicia de la URL. Les webs malicioses en aquest moment són una de les amenaces més grans que poden tenir els usuaris d'internet. Els mètodes tradicionals que detecten pàgines web malicioses, com les llistes negres, no s'actualitzen amb la rapidesa necessària per poder detectar amb precisió nous atacs a la web. Per a solucionar aquesta necessitat proposem un sistema que es pot afegir a un tallafoc o una extensió de Google Chrome. Començant amb un conjunt de dades (dataset) que conté llocs web malicioses i benignes, els classifiquem traient nombrosos atributs. Addicionalment, els atributs es formaten, s'analitzen i es processen prèviament per crear dades fàcilment separables i classificables. Amb les dades processades, l'estudi proposa un model d'Aprenentatge Profund (DL) per a classificar cada mostra. El model consisteix en una xarxa neuronal directa (FFNN). Avaluem diferents combinacions de neurones al model i realitzem un estudi en profunditat de la xarxa amb millor rendiment. Els nostres resultats mostren una precisió de detecció de pàgines web malicioses de fins al 99,88% i el 2,61% de detecció de fals positiu (és a dir, llocs web malignes classificats com a benignes). A part, el sistema ha estat provat amb 10.000 llocs webs no vistos prèviament i la detecció de fals positiu disminueix fins a l'1,026%. Per fer més accessible el sistema als usuaris finals, es desenvolupa una interfície que a l'inserir una URL sospitosa retorna el resultat obtingut d'aquesta. El temps de resposta d'aquesta interfície dóna un valor mitjà de 2,53 segons, que és un valor inferior al límit de 4 segons en què l'usuari comença a perdre l'atenció. En general, el sistema proposat és capaç de classificar correctament les dades vistes amb anterioritat i les que no. Per tant, protegeix contra nous atacants. A més, la interfície crea un espai on l'usuari pot consultar ràpidament la malícia de la URL.

Actualmente el método de detección de fragilidad de un adulto mayor es una prueba larga que implica varias horas de dedicación por parte tanto del paciente como del psicólogo y medico geriatra encargado. Las Pruebas médicas/psicológicas más conocidos para determinar la fragilidad son el de VIG y el FRIED quienes de hecho no necesariamente arrojan los mismos resultados. Dentro de estas pruebas una de las variables a utilizar es la máxima fuerza de apriete con la mano dominante de 3 intentos, cada uno con una duración aproximada de 6 segundos. Además de acuerdo con el testimonio de uno de los médicos entrevistados en antecedentes a esta investigación “Solo con apretarles la mano se cuál es su condición”. Basados en que de por si todo apunta a que la fuerza de apriete es importante para determinar si un paciente se encuentra o no en estado de fragilidad, se propuso la siguiente pregunta: ¿Si el apriete máximo nos da mucha información, podría toda la señal recogida durante la acción de apriete darnos aún más información? Este Trabajo de investigación persigue relacionar los resultados de la prueba de VIG de 83 pacientes (DDBB facilitada por el grupo de investigación CETpD-UPC) con las muestras (entre 1 y 3 por paciente) de las señales de apriete de todos los pacientes y así diseñar un clasificador que permite determinar el estado de fragilidad de un paciente solo con la señal del apriete de su mano dominante. La investigación arroja como resultado que el clasificador obtenido, creado a base de redes neuronales superficiales (SNN), consigue un desempeño de hasta 97% de predictibilidad bajo ciertas condiciones, tomando en cuenta que se utilizaron métodos de prevención de sobreajuste, diferentes técnicas de agrupación de data (se resalta que la caracterización de la señal de apriete fue basada en los métodos de caracterización de señal de sistemas de control industrial), el uso del algoritmo SMOTE para la creación sintética de observaciones con el objetivo balancear el set de entrenamiento y métodos de validación cruzada a 2 niveles, primero entre las SNN que componen al predictor FrailDetectionNET y segundo a nivel de las muestras por paciente para logran identificar si el diagnóstico del paciente es correcto, todo esto realizado dentro de un entorno MATLAB. Currently, the method to detect frailty in an elderly adult is a long test that involves several hours of dedication on the part of the patient as well as the psychologist and geriatrician in charge. The best known medical / psychological tests to determine frailty are the VIG and the FRIED, which in fact do not necessarily give the same results. Within these tests, one of the variables to be used is the maximum Hand Grip Strength with the dominant hand of 3 attempts, each one lasting approximately 6 seconds. In addition, according to the testimony of one of the doctors interviewed in the background of this investigation "I only know what their condition is by squeezing their hands. Since everything points out to the fact that the hand grip strength is important in figuring out whether a patient is in a state of fragility or not, the following question was proposed: If the maximum hand grip strength gives us a lot of information, could all the signal collected during the hand grip strength action give us even more information? This research work seeks to relate the results of the VIG test of 83 patients (DDBB provided by the CETpD-UPC research group) with the samples (between 1 and 3 per patient) of the hand grip strength signals of all patients and thus design a classifier that allows to determine the state of frailty of a patient only with the signal of the squeeze of his dominant hand. The research shows that the obtained classifier, created based on shallow neural networks (SNN), achieves a performance of up to 97% accuracy under certain conditions, taking into account that the following tools were used: overfitting prevention methods, different grouping techniques of data (it is highlighted that the characterization of the tightening signal was based on the signal characterization methods of industrial control systems), the use of the SMOTE algorithm (to create synthetic observations in order to balance the training set) and methods of cross-validation at 2 levels, first between the SNNs that make up the FrailDetectionNET predictor and second at the level of samples per patient to identify if the patient's diagnosis is correct; all this done within a MATLAB environment.

In recent years, second generation artificial neural networks (ANNs) have revolutionized the field of machine learning. These networks are mainly trained using supervised learning algorithms, e.g. back-propagation, and their neurons have a static activation that represents the spike rate of a biological neuron. However, physiological evidence suggests that biological neurons also take into account spike timing to encode information, and that their learning mechanisms are unsupervised. These difference motivated the exploration of spiking neural networks (SNNs) as third generation ANNs, that employ spiking neurons as computational units and are endowed with bio-inspired Hebbian learning mechanisms, e.g. spike timing dependent plasticity (STDP). In parallel, a growing interest to extend neural networks to machine learning tasks involving non-Euclidean data has led to the development of graph neural networks (GNNs). However, the research on SNNs is very recent and a third generation GNN implementation is still lacking. In this work, we develop a new learning rule for SNNs that we call GraphSTDP, which represents a first step towards the implementation of a third generation GNN.

Federated Learning permet que els protocols de Machine Learning utilitzin dades privades en dispositius locals per evitar potencials bretxes de seguretat, permetent l'ús de les teves dades sesne posar la seva seguretat en perill. S'esta estudiant Federated Learning per implementar-ho en areas més privades com hospitals, permetent el ús de les dades mediques privades sense compartir-les. El projecte té la intenció d'aprendre més sobre Federated Learning i extendre un prototip existent, implementant un sistema peer-to-peer i provar-lo. Un cop funcional, s'analitzaran els resultats obtinguts amb els dos sistemes (centralitzat i descentralitzat) i s'extreuran conclusions basades en aquests resultats.

Machine learning-based systems are now part of a wide array of real-world applications seamlessly embedded in the social realm. In the wake of this realisation, strict legal regulations for these systems are currently being developed, addressing some of the risks they may pose. This is the coming of age of the interpretability and explainability problems in machine learning-based data analysis, which can no longer be seen just as an academic research problem. In this tutorial, associated to ESANN 2021 special session on “Interpretable Models in Machine Learning and Explainable Artificial Intelligence”, we discuss explainable and interpretable machine learning as post-hoc and ante-hoc strategies to address these problems and highlight several aspects related to them, including their assessment. The contributions accepted for the session are then presented in this context

In a world in which the internet gives us access to any kind of information, there are still limitations when the source of such information is presented in another language. Online translators are far from perfect, this is why language machine translation is a trending topic in the field of deep learning. The purpose of this project is to use the Transformer architecture, developed by Google in 2017, in the context of Multilingual Machine Translation and to improve its results both in translation and a common intermediate representation. The Transformer model is focused on self-attention and composed by an encoder and decoder that rely on a common intermediate representation of the source language. For the purpose of raising the BLEU score that defines the quality of the translation and enhancing the common intermediate representation, we have introduced Part-Of-Speech tagging in the encoder of the model. We perform experiments with four languages (English, Spanish, French and German) both in Machine Translation and in Cross-lingual Natural Language Inference. Finally, we visualize the intermediate representation and make experiments to see how source embeddings codify gender information. Comparing a baseline model without tagging with the new POS tagged codes, the translation BLEU has decreased 0.50 points on average. In the case of NLI, the accuracies have also decreased 8% on average, showing that the POS tagged models do not improve the performances of these tasks. However, in the gender experiments of the encoder embeddings, the accuracy of the gender classification for professions has increased by 1.1%. En un mundo donde Internet nos da acceso a cualquier tipo de información, aún hay limitaciones cuando la fuente de esta se presenta en otro idioma. Los traductores en línea están lejos de ser perfectos y es por eso que la traducción automática entre idiomas es un tema de tendencia en el campo del aprendizaje profundo. El objetivo de este proyecto es utilizar la arquitectura del Transformer, desarrollada por Google en 2017, en el contexto de la traducción automática multilingüe. El modelo se centra en el concepto de la atención y está compuesto por un codificador y un decodificador basados en una representación intermedia del lenguaje de origen. Con el objetivo de aumentar la BLEU, puntuación que define la calidad de la traducción, y mejorar la representación intermedia, hemos introducido el etiquetado de categorías gramaticales (Part-Of-Speech) en el codificador del modelo. Hemos realizado experimentos con cuatro idiomas (inglés, español, francés y alemán) tanto en traducción automática como en Comprensión del Lenguaje Natural (CLN). Finalmente, hemos visualizado la representación intermedia y hemos hecho experimentos para ver cómo las representaciones de palabras en vectores codifican la información de género. Al comparar un modelo base sin etiquetado Part-Of-Speech con los nuevos códigos etiquetados, la BLEU de traducción ha disminuido 0.50 puntos de media. En el caso de la CLN, las precisiones también han disminuido un 8%, demostrando que los modelos etiquetados no mejoran el rendimiento de estas tareas. Sin embargo, la precisión de la clasificación de género para profesiones ha aumentado en un 1.1%. En un món on l'Internet ens dóna accés a qualsevol tipus d'informació, encara hi ha limitacions quan la font d'aquesta informació es presenta en un altre idioma. Els traductors en línia són lluny de ser perfectes i és per això que la traducció automàtica entre idiomes és un tema de tendència en el camp de l'aprenentatge profund. L'objectiu d'aquest projecte és utilitzar l'arquitectura del Transformer, desenvolupada per Google al 2017, en el context de la traducció automàtica multilingüe. El model es centra en el concepte de l'atenció i està compost per un codificador i un decodificador basats en una representació intermèdia del llenguatge d'origen. Amb l'objectiu d'augmentar la BLEU, puntuació que defineix la qualitat de la traducció, i millorar la representació intermèdia comuna, hem introduït l'etiquetatge de categories gramaticals (Part-Of-Speech) al codificador del model. Hem realitzat experiments amb quatre idiomes (anglès, castellà, francès i alemany) tant en traducció automàtica com en Comprensió del Llenguatge Natural (CLN). Finalment, hem visualitzat la representació intermèdia i hem fet experiments per veure com les representacions de paraules en vectors codifiquen la informació de gènere. En comparar un model base sense etiquetatge Part-Of-Speech amb els nous codis etiquetats, la BLEU de traducció ha disminuït 0.50 punts de mitjana. En el cas de la CLN, les precisions també han disminuït un 8%, demostrant que els models etiquetats no milloren el rendiment d'aquestes tasques. No obstant això, la precisió de la classificació de gènere per a professions ha augmentat en un 1.1%.

Deep neural networks have been a leading research topic within the machine learning field for the past few years. The introduction of graphical processing units (GPUs) and hardware ad- vances made possible the training of deep neural networks. Previously the training procedure was impossible due to the huge amount of training samples required. The new trained introduced architectures have outperformed the classical methods in different classification and regression problems. With the introduction of 5G technology, related to low-latency and online applica- tions, the research on decreasing the computational cost of deep learning architectures while maintaining state-of-art performance has gained huge interest. This thesis focuses on the use of Self Size-estimating Feedforward Network (SSFN), a feed- forward multilayer network. SSFN presents low complexity on the training procedure due to a random matrix instance used in its weights. Its weight matrices are trained using a layer-wise convex optimization approach (a supervised training) combined with a random matrix instance (an unsupervised training). The use of deterministic transforms is explored to replace random matrix instances on the SSFN weight matrices. The use of deterministic transforms automat- ically reduces the computational complexity, as its structure allows to compute them by fast algorithms. Several deterministic transforms such as discrete cosine transform, Hadamard trans- form and wavelet transform, among others, are investigated. To this end, two methods based on features statistical parameters are developed. The proposed methods are implemented on each layer to decide the deterministic transform to use. The effectiveness of the proposed approach is illustrated by SSFN for object classification tasks using several benchmark datasets. The results show a proper performance, similar to the original SSFN, and also consistency across the different datasets. Therefore, the possibility of introducing deterministic transformations in machine learning research is demonstrated.

Due to the increase in population in large urban areas, mobility in large cities has become increasingly problematic and difficult. With the desire to find solutions to this situation, the use of micro-mobility vehicles has become popular. In this sense the solution presented proposes to make use of Deep Learning and Computer Vision techniques in order to identify, automatically, possible defects where micro-mobility vehicles circulate, in order to alert authorities and other users and improve user safety. Three solutions were devised to determine which one best suited the needs of the project. The first two to test the two models separately and the third one with the intention of combining them, with ensemble learning techniques, to try to improve the results obtained. Debido al incremento de la población en los núcleos urbanos, la movilidad en las grandes ciudades se ha vuelto cada vez más problemática y difícil. Con el afán de encontrar soluciones a esta situación, se ha popularizado el uso de vehículos de micro-movilidad. En este sentido la solución que se presenta propone hacer uso de técnicas de Deep Learning y Computer Vision para identificar posibles defectos en las carreteras de forma automática y mejorar la seguridad de los usuarios. Se proponen tres soluciones para determinar cuál es el modelo que se adapta mejor a las necesidades del proyecto. Los dos primeros para probar los modelos por separado y el tercero con la intención de juntarlos, con técnicas de ensemble learning, para intentar mejorar los resultados obtenidos. Degut a l'increment de la població en els grans nuclis urbans, la mobilitat a les grans ciutats s'ha tornat cada cop més problemàtica i difícil. Amb l'afany de trobar solucions a aquesta situació, s'ha popularitzat l'ús de vehicles de micro-mobilitat. En aquest sentit la solució que es presenta proposa fer ús de tècniques de Deep Learning i Computer Vision per identificar possibles defectes a les carreteres de forma automàtica i millorar la seguretat dels usuaris. Es proposen tres solucions per determinar quin és el model que s'adapta millor a les necessitats del projecte. Els dos primers per provar dos models per separat i el tercer amb la intenció d'ajuntar-los, amb tècniques d'ensemble learning, per intentar millorar els resultats obtinguts.

As the demand for Network Function Virtualization accelerates, service providers are expected to advance the way they manage and orchestrate their network services to offer lower latency services to their future users. Modern services require complex data flows between Virtual Network Functions, placed in separate network domains, risking an increase in latency that compromises the offered latency constraints. This shift requires high levels of automation to deal with the scale and load of future networks. In this paper, we formulate the Service Function Chaining (SFC) placement problem and then we tackle it by introducing SCHEMA, a Distributed Reinforcement Learning (RL) algorithm that performs complex SFC orchestration for low latency services. We combine multiple RL agents with a Bidding Mechanism to enable scalability on multi-domain networks. Finally, we use a simulation model to evaluate SCHEMA, and we demonstrate its ability to obtain a 60.54% reduction of average service latency when compared to a centralised RL solution.

In the last decades, the interest in predicting tropospheric ozone levels (O₃) has increased due to its detrimental effect on population health and vegetation. Although certain factors such as solar radiation are well known to have an influence on ozone levels, the effect of other variables is less clear. In this study, several regression models based on the Random Forest (RF) algorithm are generated to predict the daily maximum hourly ozone concentration level (1hO₃) and the daily maximum 8-hours average ozone concentration level (8hO₃) one day ahead in Barcelona, using air quality data, meteorological data and time variables as inputs. Two versions of the model are considered: taking information from the whole year and focusing only on summer months (May to September). In addition, classification models are created, based on thresholds inspired by current regulations for both outputs. RF regression models capture the time variation of tropospheric ozone through the year and they generate accurate estimations with acceptable deviation between the observations and predictions. In general, the categorical models of 1hO₃ show suitable and lower error rates than 8hO₃. However, the categories, which gather the most of the tropospheric ozone values have high accuracy and the categories with few values inside them have low accuracy. Consequently, these categorical models are not useful as a tool to alert the population about a specific ozone event. The analysis of RF models shows that the tropospheric ozone level (1hO₃ or 8hO₃ according to the model) of the previous day to the prediction has the strongest association to the output. The importance of other inputs varies between the models considered; while solar radiation and day of the year are the main variables after O₃ for the whole year models, relative humidity, average dew-point deficit and weekday are also relevant in the summer models.

En aquest treball volem investigar com la Intel·ligència Artificial pot ajudar a resoldre problemes del món real com ara fer exercici físic, minimitzant el nombre de lesions i millorant la tècnica d’una persona principiant. D’altra banda, aconseguir que la solució obtinguda pugui ser usable per tothom. L’objectiu és desenvolupar un software mitjançant tècniques d'Aprenentatge Automàtic (com ara la Visió per Computador) per detectar la correcta execució d’un exercici físic, com per exemple l’esquat. Realitzarem el procés de crear i entrenar el model (algoritme basat en aprenentatge automàtic que ha après a solucionar un problema específic) perquè aprengui a detectar si un esquat està tècnicament ben fet o no, i en cas de no estar bé, informi dels errors detectats i de com es podria millorar la tècnica. Per poder crear aquest model també crearem un conjunt de dades propi, és a dir, un conjunt d’exemples que serveixen perquè el model aprengui. En el nostre cas, aquest conjunt de dades contindrà vídeos que hauran gravat diverses persones fent esquats correctament i d’altres incorrectament. Per últim, testejarem aquest model amb un conjunt de vídeos no usats per entrenar el model, fins aconseguir un model amb un encert alt en la detecció i que, per tant, sigui una bona solució per al problema que plantegem. També crearem una API REST (Application Program Interface) que englobi la solució obtinguda i tot el preprocessament de les dades necessari amb l’objectiu de poder comunicar-nos amb el model i fer-lo accessible per qualsevol aplicació que es connecti a l’API. Finalment, dissenyarem i implementarem una petita aplicació web/mòbil que es connecti a l’API i processi la resposta obtinguda per mostrar-la a l’usuari de manera entenedora. Fent-ho llegible, aconseguim que l’usuari pugui fer ús dels serveis que ofereix el nostre sistema d’intel·ligència artificial. L’aplicació està pensada per ser accessible des de qualsevol dispositiu (escriptori o mòbil). En aquesta memòria detallem tot el procés que hem dut a terme i com hem dissenyat, implementat i testejat tots els components del servei. In this project, we want to investigate how Artificial Intelligence can help solve real-world problems, such as minimising the number of injuries and improving the technique of a beginner, and how this solution can be used by everyone. The aim is to develop software with Machine Learning techniques (such as Computer Vision) to detect the correct execution of physical exercises such as squat. We will carry out the process of creating and training a model (an algorithm based on automatic learning that has learned to solve a specific problem) to detect whether a squat is well done or not, and if not, the different kinds of errors. In order to create this model we also create a dataset, which is a set of examples that serve to help the model learn and which contains recorded videos of people doing correct and incorrect squats. Finally, we will test this model with other videos to find a model that has a high accuracy and therefore is a good solution. We will also create a REST API (Application Program Interface) that contains the obtained solution and all the necessary data preprocessing in order to be able to communicate with the model and make it accessible to any application that connects to the API. In the end, we will design and implement a small web/mobile application that connects to the API and process the response obtained to show it to the user in an understandable way. By making it human readable, we get the user to use the services offered by our artificial intelligence system. The application is designed to be accessible from any device (computer or mobile). This report details the whole process we have carried out and how we have designed, implemented and tested all the components of the service

Convolutional Neural Networks (ConvNets) at present achieve remarkable performance in image classification tasks. However, current ConvNets cannot guarantee the capabilities of the mammalian visual systems such as invariance to contrast and illumination changes. Some ideas to overcome the illumination and contrast variations usually have to be tuned manually and tend to fail when tested with other types of data degradation. In this context, we present a new bio-inspired {entry} layer, M6, which detects low-level geometric features (lines, edges, and orientations) which are similar to patterns detected by the V1 visual cortex. This new trainable layer is capable of coping with image classification even with large contrast variations. The explanation for this behavior is the monogenic signal geometry, which represents each pixel value in a 3D space using quaternions, a fact that confers a degree of explainability to the networks. We compare M6 with a conventional convolutional layer (C) and a deterministic quaternion local phase layer (Q9). The experimental setup {is designed to evaluate the robustness} of our M6 enriched ConvNet model and includes three architectures, four datasets, three types of contrast degradation (including non-uniform haze degradations). The numerical results reveal that the models with M6 are the most robust in front of any kind of contrast variations. This amounts to a significant enhancement of the C models, which usually have reasonably good performance only when the same training and test degradation are used, except for the case of maximum degradation. Moreover, the Structural Similarity Index Measure (SSIM) is used to analyze and explain the robustness effect of the M6 feature maps under any kind of contrast degradations., Comment: "For associated code, see https://gitlab.com/monogenic-layer-m6/monogenic-layer-trainablebio-inspired-cnnlayerforcontrastinvariance"

Recent years have seen the vast potential of Graph Neural Networks (GNN) in many fields where data is structured as graphs (e.g., chemistry, logistics). However, implementing a GNN prototype is still a cumbersome task that requires strong skills in neural network programming. This poses an important barrier to researchers and practitioners that want to apply GNN to their specific problems but do not have the needed Machine Learning expertise. In this paper, we present IGNNITION, a novel open-source framework for fast prototyping of GNNs. This framework is built on top of TensorFlow, and offers an intuitive high-level abstraction that allows the user to define its GNN model via a YAML file, being completely oblivious to the tensor-wise operations made internally by the model. At the same time, IGNNITION offers great flexibility to build any GNN-based architecture. To showcase its versatility, we implement two state-of-the-art GNN models applied to the field of computer networks, which differ considerably from well-known standard GNN architectures. Our evaluation results show that the GNNs produced by IGNNITION are equivalent in performance to implementations directly coded in TensorFlow. This work has received funding from the European Union’s Horizon 2020 research and innovation programme within the framework of the NGI-POINTER Project funded under grant agreement No. 871528. This paper reflects only the authors’ view; the European Commission is not responsible for any use that may be made of the information it contains. This work was also supported by the Spanish MINECO under contract TEC2017-90034-C2-1-R (ALLIANCE) and the Catalan Institution for Research and Advanced Studies (ICREA).

In this master thesis we develop a model that surpasses previous studies to be able to detect cyberbullying and other disorders that are a common behaviour in teenagers. We analyze short sentences in social media with new techniques that haven?t been studied in depth in language processing in order to be able to detect these problems. Deep learning is nowadays the common approach for text analysis. However, struggling with dataset size is one of the most common problems. It is not optimal to dedicate thousands of hours to label data by humans every time we want to create a new model. Different techniques have been used over the years to solve or at least minimize this problem, for instance transfer learning or self-learning. One of the most known ways to solve this is by data augmentation. In this thesis we make use of active learning and self-training to address having restrictions of labelled data. We have used data that has not been labeled to improve the performance of our models. The architecture of the model is composed of a Bert model plus a linear layer that projects the Bert sentence embedding into the number of classes we want to detect. We take advantage of this already functional model to label new data that we will use afterwards to create our final model. Using noise techniques we modify the data so the final model has to predict less structured data and learn from difficult scenarios. Thanks to this technique we were able to improve the results in some of the classes, for instance the F-score modified increases by 7% for substance abuse (drugs, alcohol, etc) and 3% in disorders (anxiety, depression and distress) while keeping the performance of the other classes.

Deep Learning Convolutional Neural Networks have been successfully used in many applications. Its versatility lies in reducing the number of parameters to train while maintaining or improving the feature representation capabilities offered by other architectures. Due to its success, Convolutional Networks have become the architecture of choice for image and video processing applications. The application of Convolutional Networks to wind time series is still limited, being an area with high potential for developing new approaches. This paper explores several deep learning models and applies them to wind time series for multi-step forecasting. The time series used for the experimentation are multidimensional time-stamped multi-variate meteorological data. We use a large dataset of wind data from the National Renewable Laboratory with 126,692 wind sites, requiring the use of High Performance Computing. The experimentation results show how Convolutional Networks are a valid approach for wind time series forecasting.

The advent of next-generation sequencing technologies allowed relative quantification of microbiome communities and their spatial and temporal variation. In recent years, supervised learning (i.e., prediction of a phenotype of interest) from taxonomic abundances has become increasingly common in the microbiome field. However, a gap exists between supervised and classical unsupervised analyses, based on computing ecological dissimilarities for visualization or clustering. Despite this, both approaches face common challenges, like the compositional nature of next-generation sequencing data or the integration of the spatial and temporal dimensions. Here we propose a kernel framework to place on a common ground the unsupervised and supervised microbiome analyses, including the retrieval of microbial signatures (taxa importances). We define two compositional kernels (Aitchison-RBF and compositional linear) and discuss how to transform non-compositional beta-dissimilarity measures into kernels. Spatial data is integrated with multiple kernel learning, while longitudinal data is evaluated by specific kernels. We illustrate our framework through a single point soil dataset, a human dataset with a spatial component, and a previously unpublished longitudinal dataset concerning pig production. The proposed framework and the case studies are freely available in the kernInt package at https://github.com/elies-ramon/kernInt., This work was funded by projects PID2019-108829RB-I00, AGL2016–78709-R, and AGL2017–88849-R awarded by the Spanish Ministry of Economy and Competitiveness. ER has funding from a FI-AGAUR Ph.D. studentship grant, with the support of the Secretaria d’Universitats i Recerca de la Generalitat de Catalunya and the European Social Fund. YR-C is recipient of a Ramon y Cajal post-doctoral fellowship (RYC2019-027244-I) from the Spanish Ministry of Science and Innovation. We acknowledge further financial support from the Spanish Ministry of Economy and Competitiveness through the “Severo Ochoa Programme for Centres of Excellence in R&D” 2016–2019 (SEV-2015-0533),” and from the EU through the BFU2016–77236-P (MINECO/AEI/FEDER, EU).

Real-time velocity estimation is a core task in autonomous driving, which is carried out based on available raw sensors such as wheel odometry and motor currents. When the system dynamics and observations can be modeled together as a fully known linear Gaussian state-space (SS) model, the celebrated Kalman filter (KF) is a low complex- ity optimal solution. However, both linearity of the underlying SS model and accurate knowledge of it are often not encountered in practice. This work proposes to estimate the velocity using a hybrid data-driven (DD) implementation of the KF for non-linear systems, coined KalmanNet. KalmanNet integrates a compact recurrent neural network in the flow of the classical KF, retaining low computational complexity, high data ef- ficiency, and interpretability, while enabling operation in non-linear SS models with partial information. We apply KalmanNet on pilatus driverless, an autonomous racing car as part of the Formula Student (FS) Driverless competition. Our results demon- strate the ability of KalmanNet to outperform a state-of-the-art implementation of the KF that uses a postulated SS model, while being applicable on the vehicle control unit used by the car. La estimación de la velocidad en tiempo real es una tarea fundamental en la conducción autónoma, que se lleva a cabo a partir de los sensores disponibles, como la odometría de las ruedas y las corrientes del motor. Cuando la dinámica del sistema y las observaciones pueden modelarse conjuntamente como un modelo lineal gaussiano de espacio de estados (SS) totalmente conocido, el célebre filtro de Kalman (KF) es una solución óptima de baja complejidad. Sin embargo, tanto la linealidad del modelo SS subyacente como el conocimiento preciso del mismo no suelen darse en la práctica. Este trabajo propone estimar la velocidad utilizando una implementación híbrida basada en datos (DD) y en el KF para sistemas no lineales, denominada KalmanNet. KalmanNet integra una red neuronal recurrente compacta en el flujo del KF clásico, conservando una baja complejidad computacional, alta efi ciencia e interpretabilidad, al tiempo que permite operar en modelos de SS no lineales con información parcial. Aplicamos KalmanNet en pilatus driverless, un coche de carreras autónomo que forma parte de la competición Formula Student (FS) Driverless. Nuestros resultados demuestran la capacidad de KalmanNet para superar una implementación de vanguardia del KF que utiliza un modelo de SS postulado, a la vez que es aplicable en la unidad de control del vehículo utilizada por el coche. L'estimació de la velocitat en temps real és una tasca fonamental en la conducció autònoma, que es porta a terme a partir dels sensors disponibles, com la odometría de les rodes i els corrents de l'motor. Quan la dinàmica de sistema i les observacions es poden modelar conjuntament com un model lineal gaussià d'espai d'estats (SS) totalment conegut, el cèlebre filtre de Kalman (KF) és una solució òptima de baixa complexitat. No obstant això, tant la linealitat de el model SS subjacent com el coneixement precís de la mateixa no solen donar-se en la pràctica. Aquest treball proposa estimar la velocitat fent servir un sistema híbrid basat en dades (DD) i el Kalman Filter per a sistemes no lineals, anomenada KalmanNet. KalmanNet integra una xarxa neuronal recurrent compacta en el flux de l'KF clàssic, conservant una baixa complexitat computacional, alta eficiència i interpretabilitat, al mateix temps que permet operar en models de SS no lineals amb informació parcial. Apliquem KalmanNet en Pilatus driverless, un cotxe de carreres autònom que forma part de la competició Formula Student (FS) driverless. Els nostres resultats demostren la capacitat d'KalmanNet per superar una implementació d'avantguarda del KF que utilitza un model de SS postulat, alhora que és aplicable a la unitat de control de el vehicle utilitzat pel cotxe.

Este proyecto se centra en la evaluación de la precisión de tres implementaciones distintas para un sistema de posicionamiento en interiores (IPS) que utiliza tecnología Bluetooth Low Energy (BLE). Cada una de las tres implementaciones se diferencian en la técnica utilizada para realizar el posicionamiento. Estas técnicas son: el modelo Path Loss Model (PLM) y los algoritmos de aprendizaje automático Support Vector Regressor (SVR) y Multi-Layer Perceptron (MLP). Para conseguir una mayor generalización de los resultados los experimentos se han realizado en dos escenarios distintos: una habitación y una aula. Finalmente, se ha implementado una aplicación que utiliza la técnica seleccionada y que realiza el posicionamiento a tiempo real de un teléfono móvil. This project is focused on the evaluation of the accuracy of three different implementations for an indoor positioning system (IPS) which uses Bluetooth Low Energy (BLE) technology. Each of the three implementations differs on the technology used to perform the positioning. These techniques are the Path Loss Model (PLM) and the Support Vector Regressor (SVR) and Multi-Layer Perceptron (MLP) machine learning algorithms. To achieve a greater generalization of the results, the experiments have been performed in two different scenarios: a room and a classroom. Finally, an application which uses the selected technique and performs the real-time location of a smartphone has been implemented. Aquest projecte se centra en l’evaluació de la precisió de tres implementacions diferents per un sistema de posicionament en interiors (IPS) que utilitza tecnologia Bluetooth Low Energy (BLE). Cadascuna de les tres implementacions es diferencien en la tècnica utilitzada per a realitzar el posicionament. Aquestes tècniques són: el model Path Loss Model (PLM) i els algorismes d’aprenentatge automàtic Support Vector Regressor (SVR) i Multi-Layer Perceptron (MLP). Per aconseguir una major generalització dels resultats els experiments s’han realitzat en dos escenaris diferents: una habitació y una aula. Finalment, s’ha implementat una aplicació que utilitza la tècnica seleccionada i que realitza el posicionament a temps real d’un telèfon mòbil.

Emotion analysis of messages using machine learning techniques is a difficult and cumbersome task requiring a major effort to obtain reliable results. This challenge is even more pronounced when the target language is not English, but Spanish. To overcome this challenge, this paper describes how UPC Team applied sentiment analysis on social media messages (in particular, on Twitter) written in Spanish and, related to events that took place in April 2019 from different domains. To this aim, we present a machine learning model based on BERT and describe the results obtained to reach an accuracy of 65% approx. and the 12th position in the ranking, for this second edition of the contest for emotion detection of Spanish tweets EmoEva-lEs@IberLEF2021.

L'objectiu principal d'aquesta tesi és dissenyar i implementar un Sistema Basat en el Coneixement capaç d'aprendre mitjançant la interacció contínua amb l'usuari. El component central del sistema és un Knowledge Graph que gestiona el coneixement del sistema en tot moment. El Knowledge Graph s'inicialitza amb l'anàlisi de documents del domini i després, el sistema va aprenent nous conceptes i relacions mitjançant la interacció amb l'usuari, ampliant la representació amb les noves experiències. The main goal of this thesis is to design and implement a Knowledge-Based System capable of learning through continuous interaction with the user. The central component of the system is a Knowledge Graph that manages system knowledge at all times. The Knowledge Graph is initialized with the analysis of the documents in the domain, and then the system learns new concepts and relationships through interaction with the user, expanding the representation with new experiences.

Oftalmological images have shown promising results in diagnostic imaging in recent years. In this work, 2D OCTA cuts as well as 3D OCTA images and the radiomic features extracted therefrom will be analyzed using machine learning techniques to support diagnostic decision making in diabetes mellitus and diabetic retinopathy. For this purpose, a cross-sectional analysis of these medical images collected in a large scale prospective OCTA trial cohort and their respective radiomic features is performed and evaluated with classification models such as Logistic Regression, LDA, Linear SVC and RBF SVC for the respective classification problem (presence of DM and DR). 808 eyes from 404 patients were studied, of which 660 eyes belonged to diabetes mellitus patients and 148 eyes from healthy control eyes. The diagnosis of DM resulted in an AUC value of almost 0.8 with 2D OCTA cuts and 0.7 with 3D OCTA images by only considering the radiomic features. With the addition of clinical data, these values improve slightly. For the diagnosis of DR, 461 eyes with DM but without DR and 199 with DR were available. Again, with the addition of clinical data, these values improve slightly. AUC values of just below 0.7 for 2D cuts and slightly lower values for 3D images were obtained. Similar to before, the results improve after adding the clinical data. Now, however, percentally stronger than before. Radiomic techniques applied to OCTA images have shown accurate classification of DM and DR patients. The combination of retinal images and clinical data superior performance than induvidual tests only.

Deep learning (DL) is attracting considerable interest as it currently achieves remarkable performance in many branches of science and technology. However, current DL cannot guarantee capabilities of the mammalian visual systems such as lighting changes. This paper proposes a deterministic entry layer capable of classifying images even with low-contrast conditions. We achieve this through an improved version of the quaternion monogenic wavelets. We have simulated the atmospheric degradation of the CIFAR-10 and the Dogs and Cats datasets to generate realistic contrast degradations of the images. The most important result is that the accuracy gained by using our layer is substantially more robust to illumination changes than nets without such a layer. The authors would like to thank to CONACYT and Barcelona supercomputing Center. Sebastián Salazar-Colores (CVU 477758) would like to thank CONACYT (Consejo Nacional de Ciencia y Tecnología) for the financial support of his PhD studies under Scholarship 285651. Ulises Moya and Ulises Cortés are member of the Sistema Nacional de Investigadores CONACyT.

This project implements Reinforcement Learning algorithms to machine-to-machine (M2M) communication, because the number of devices that need to connect to each other is constantly growing, ways to mitigate the overload in the network have been investigated, one of them is the main objective of this work, which is based on controlling traffic congestion to provide resources to the maximum number of devices that want to access the medium. Reinforcement Learning techniques are used to activate and control the ACB factor of access to the medium, which regulates the percentage of machines that can access among all those that present data to be sent in a given communications channel. Obtaining as a result an excellent success rate of MTCD with eNB by implementing Q-Learning and Deep Q-Learning algorithms. Este proyecto implementa algoritmos de Reinforcement Learning a la comunicación de máquinas con máquinas (M2M). Debido a que el número de dispositivos que necesitan conectarse entre sí no para de crecer, se han investigado las formas de mitigar la sobrecarga en la red, una de ellas es el objetivo principal de este trabajo, que se basa en controlar la congestión del tráfico para proporcionar recursos al máximo número de dispositivos que quieran acceder al medio. Mediante técnicas de Reinforcement Learning se activa y controla el factor ACB de acceso al medio, que regula el porcentaje de máquinas que pueden acceder entre todas las que presentan datos a enviar en un canal de comunicaciones determinado. Obteniendo como resultado una excelente tasa de éxito de los MTCD con eNB mediante la implementación de algoritmos de Q-Learning y Deep Q-Learning. Aquest projecte implementa algorismes de Reinforcement Learning a la comunicació de màquines amb màquines (M2M), pel fet que el nombre de dispositius que necessiten connectar-se entre si no para de créixer, s'han investigat les maneres de mitigar la sobrecàrrega en la xarxa, una d'elles és l'objectiu principal d'aquest treball, que es basa en controlar la congestió del trànsit per a proporcionar recursos al màxim nombre de dispositius que vulguin accedir al mitjà. Mitjançant tècniques de Reinforcement Learning s'activa i controla el factor ACB d'accés al mitjà, que regula el percentatge de màquines que poden accedir entre totes les que presenten dades a enviar en un canal de comunicacions determinat. Obtenint com a resultat una excel·lent taxa d'èxit dels MTCD amb eNB mitjançant la implementació d'algorismes de Q-Learning i Deep Q-Learning.

Artificial intelligence (AI)-driven zero-touch massive network slicing is envisioned to be a disruptive technology in beyond 5G (B5G)/6G, where tenancy would be extended to the final consumer in the form of advanced digital use-cases. In this paper, we propose a novel model-free deep reinforcement learning (DRL) framework, called collaborative statistical Actor-Critic (CS-AC) that enables a scalable and farsighted slice performance management in a 6G-like RAN scenario that is built upon mobile edge computing (MEC) and massive multiple-input multiple-output (mMIMO). In this intent, the proposed CS-AC targets the optimization of the latency cost under a long-term statistical service-level agreement (SLA). In particular, we consider the Q-th delay percentile SLA metric and enforce some slice-specific preset constraints on it. Moreover, to implement distributed learners, we propose a developed variant of soft Actor-Critic (SAC) with less hyperparameter sensitivity. Finally, we present numerical results to showcase the gain of the adopted approach on our built OpenAI-based network slicing environment and verify the performance in terms of latency, SLA Q-th percentile, and time efficiency. To the best of our knowledge, this is the first work that studies the feasibility of an AI-driven approach for massive network slicing under statistical SLA. This work has been supported in part by the research projects MonB5G (871780), ZEROTO6G, AGAUR (2017-SGR-891), and PROGRESSUS (876868).

Working paper We apply the two-step machine-learning method proposed by Claveria et al. (2021) to generate country-specific sentiment indicators that provide estimates of year-on-year GDP growth rates. In the first step, by means of genetic programming, business and consumer expectations are evolved to derive sentiment indicators for 19 European economies. In the second step, the sentiment indicators are iteratively re-computed and combined each period to forecast yearly growth rates. To assess the performance of the proposed approach, we have designed two out-of-sample experiments: a nowcasting exercise in which we recursively generate estimates of GDP at the end of each quarter using the latest survey data available, and an iterative forecasting exercise for different forecast horizons We found that forecasts generated with the sentiment indicators outperform those obtained with time series models. These results show the potential of the methodology as a predictive tool.

The COVID-19 pandemic has already caused more than 150,000,000 cases worldwide. In Spain this has lead to a massive and simultaneous saturation of all sanitary regions. Coherently, the quick and consistent understanding of the COVID-19 disease requires of the combined analysis of thousands of medical records generated by dozens of different institutions. In the context of the publicly funded CIBERES-UCI-COVID project, we have gathered, cleaned and preprocessed data from heterogeneous sources - more than 30 hospitals, with different data entry systems - in order to produce a unified database, of more than 6.000 patients, that is used in several clinical studies being carried by different multidisciplinary groups. In this paper, we identify the complexities we encountered, the solutions we applied, and we summarise the statistical and machine learning techniques we have applied for the studies. © 2021 The authors and IOS Press.

Pre-training Reinforcement Learning (RL) agents in a task-agnostic manner has shown promising results. However, previous works still struggle to learn and discover meaningful skills in high-dimensional state-spaces. We approach the problem by leveraging unsupervised skill discovery and self-supervised learning of state representations. In our work, we learn a compact latent representation by making use of variational or contrastive techniques. We demonstrate that both allow learning a set of basic navigation skills by maximizing an information theoretic objective. We assess our method in Minecraft 3D maps with different complexities. Our results show that representations and conditioned policies learned from pixels are enough for toy examples, but do not scale to realistic and complex maps. We also explore alternative rewards and input observations to overcome these limitations. The code is publicly available at https://github.com/juanjo3ns/mineRL.

As the traffic demand in Mobile Communication Systems is constantly increasing, the new 5G NR standards need to adapt its Self-Organizing Network functions. In order to cope with its complexity, Artificial Intelligence (AI) and Machine Learning (ML) based techniques are being used. These techniques need appropriate datasets based on real-life data to be trained and evaluated because wireless networks are hard to simulate through mathematical models. This project aims to provide datasets useful for the testing of AI and ML algorithms based on data from the Campus Nord mobile network. Moreover, the second purpose of the project is to determine whether connectivity sharing techniques could be beneficial in order to improve the network performance for all its users. All the data has been gathered during several measurement campaigns using a terminal with an Android operating system. Two Mobile Network Tools have been used: Qualipoc Android, for the data measurement, and Romes Replay for the later visualization and exportation of the parameters needed. Finally, all the data processing, plotting and tabulation have been done using Matlab. By the end of the project, two datasets, one for the LTE and another for 5G NR have been created. These datasets along with several coverage and base station identifying maps provide the necessary information to characterize the mobile network in Campus Nord. In addition, it has been determined that connectivity sharing techniques are viable. It has been proved that the sharing user does not lose service quality and the service received by the second user is only limited by the capabilities of the connection with the first one. Debido a que la demanda en los Sistemas de Comunicaciones Móbiles está en constante crecimiento, los nuevos estándares de 5G necesitan adaptar sus funciones de auto-organización de la red. Con el fin de ocuparse de su complejidad, son necesarias técnicas basadas en Inteligencia Artificial (AI) y Machine lerning (ML). Estas técnicas, necesitan conjuntos de datos apropiados basados en datos reales para ser entrenadas y evaluadas, ya que las redes móviles son difíciles de simular mediante modelos matemáticos. Este proyecto está orientado a proporcionar conjuntos de medidas útiles para el testeo de algoritmos de AI y ML basados en datos extraídos de la red móvil del Campus Nord. Además, el segundo propósito del proyecto es determinar si las técnicas de conexión compartida podrían ser beneficioises para mejorar las prestaciones de la red para todos sus usuarios. Todos los datos han sido obtenidos durante varias campañas de medidas utilizando un terminal con sistema operativo Android. Dos herramientas de red móvil (Mobile Network Tools, MNT) han sido utilizadas: Qualipoc Android, para las medidas de datos, y Romes Replay, para la posterior visualización y exportación de los parámetros necesarios. Finalmente, todo el procesamiento de datos, graficación y tabulación han sido hechos mediante Matlab. Al final del proyecto, dos conjuntos de medidas, uno para LTE y el otro para 5G han sido creados. Estos conjuntos de medidas, junto con varios mapas de cobertura y de identificación de estaciones base, proporcionan la información necesaria para caracterizar la red móvil en el Campus Nord. Adicionalmente, ha sido determinado que las técnicas de conexión compartida son viables. Ha sido demostrado que el usuario que comparte conexión no pierde calidad de servicio y que el servicio recibido por el segundo usuario está únicamente limitado por las capacidades de la conexión con el primero. Degut a que la demanda en els Sistemes de Comunicacions Mòbils està en constant creixement, els nous estàndards de 5G necessiten adaptar les seves funcions d'auto-organització de la xarxa. Per tal d'ocupar-se de la seva compexitat, són necessàries tècniques basades en Intel·ligència Artificial (AI) i Machine Lerning (ML). Aquestes tècniques, necessiten conjunts de dades apropiats basats en dades reals per tal de ser entrenades i avaluades, ja que les xarxes mòbils són difícils de simular mitjançant models matemàtics. Aquest projecte està orientat a proporcionar conjunts de mesures útils pel testeig d'algoritmes d'AI i ML basats en dades extretes de la xarxa mòbil del Campus Nord. A més, el segon propòsit del projecte és determinar si les tècniques de connexió compartida podrien ser beneficioses per tal de millorar les prestacions de la xarxa per a tots els seus usuaris. Totes les dades han estat obtingudes durant diverses campanyes de mesures utilitzant un terminal amb sistema operatiu Android. Dues eines de xarxa mòbil (Mobile Network Tools, MNT) han estat emprades: Qualipoc Android, per les mesures de dades, i Romes Replay, per la posterior visualització i exportació dels paràmetres necessaris. Finalment, tot el processament de dades, graficació i tabulació han estat fets mitjançant Matlab. Al final del projecte, dos conjunts de mesures, un per LTE i l'altre per 5G han estat creats. Aquests conjunts de mesures, juntament amb diversos mapes de cobertura i d'identificació d'estacions base, proporcionen la informació necessària per caracteritzar la xarxa mòbil al Campus Nord. Addicionalment, ha estat determinat que les tècniques de connexió compartida són viables. Ha estat demostrat que l'usuari que comparteix connexió no perd qualitat de servei i que el servei rebut pel segon usuari està únicament limitat per les capacitats de la connexió amb el primer.

Recent progress in machine learning frameworks makes it now possible to run an inference with sophisticated machine learning models on tiny microcontrollers. Model training, however, is typically done separately on powerful computers. There, the training process has abundant CPU and memory resources to process the stored datasets. In this work, we explore a different approach: training the model directly on the microcontroller. We implement this approach for a keyword spotting task. Then, we extend the training process using federated learning among microcontrollers. Our experiments with model training show an overall trend of decreasing loss with the increase of training epochs. This work was partially funded by the Spanish Government under contracts PID2019-106774RB-C21, PCI2019-111850-2 (DiPET CHIST-ERA), PCI2019-111851-2 (LeadingEdge CHIST-ERA), and the Generalitat de Catalunya as Consolidated Research Group 2017- SGR-990.

In this study we combine the results of two independent analyses to position Spanish regions according to both the characteristics of the time series of international tourist arrivals and the accuracy of predictions of arrivals at the regional level. We apply a seasonal-trend decomposition procedure based on non-parametric regression to isolate the different components of the series and calculate the main time series features. Predictions are generated with several machine learning models in a recursive multi-step-ahead forecasting experiment. Finally, we summarize all the information from the two previous experiments using categorical principal component analysis. By overlapping the distribution of the regions and the component loadings of each variable along both dimensions, we observe that entropy and dispersion show an inverse relation with forecast accuracy, but the interactions between the rest of the features and accuracy are heavily dependent on the forecast horizon. On this evidence, we conclude that in order to increase forecast accuracy of tourist arrivals at the regional level, model selection should be region-specific and based on the forecast horizon. This research was supported by by the projects ECO2016-75805-R and TEC2015-69266-P from the Spanish Ministry of Economy and Competitiveness. We would like to thank the Editor and two anonymous referees for their useful comments and suggestions.

Glioblastoma is the most frequent aggressive primary brain tumor amongst human adults. Its standard treatment involves chemotherapy, for which the drug temozolomide is a common choice. These are heterogeneous and variable tumors which might benefit from personalized, data-based therapy strategies, and for which there is room for improvement in therapy response follow-up, investigated with preclinical models. This study addresses a preclinical question that involves distinguishing between treated and control (untreated) mice bearing glioblastoma, using machine learning techniques, from magnetic resonance-based data in two modalities: MRI and MRSI. It aims to go beyond the comparison of methods for such discrimination to provide an analytical pipeline that could be used in subsequent human studies. This analytical pipeline is meant to be a usable and interpretable tool for the radiology expert in the hope that such interpretation helps revealing new insights about the problem itself. For that, we propose coupling source extraction-based and radiomics-based data transformations with feature selection. Special attention is paid to the generation of radiologist-friendly visual nosological representations of the analyzed tumors. L.M.N., A.V., C.A., A.P.C., M.J.L. and A.S. received EC funding from the ATTRACT project, under Grant Agreement 777222. A.V. and E.R. acknowledge funding from Spanish MINECO TIN2016-79576-R research project. A.S and M.J.L-C acknowledge funding from Spanish MINECO RTI2018-098682-B-I00.

Amb l’arribada de la quarta revolució industrial, la gran quantitat de dades disponibles y el poder de computació i emmagatzematge dels dispositius, el Machine Learning (ML) ha assumit gran protagonisme a nivell industrial. El ML permet abordar reptes de fabricació, com la millora de la qualitat dels productes i la optimització dels processos industrials. L’objectiu principal d’aquest projecte és destacar la important eina que suposa el ML a la industria i als processos de fabricació. També, pretén conscienciar de la necessitat d’afegir al currículum acadèmic d’un enginyer nocions bàsiques sobre aquesta innovadora eina. Per una banda, es mostren i defineixen els tres tipus d’aprenentatge automàtic: supervisat, no supervisat i per reforç. Aquesta primera part va acompanyada de la presentació d’alguns dels algoritmes més utilitzat i els elements principals per al correcte funcionament d’aquests. Per una altra banda, es recullen i sintetitzen 45 casos on s’estan aplicant tècniques de ML en l’àmbit de la enginyeria mecànica. Aquests casos es classifiquen en funció de l’objectiu principal de l’algoritme entre: diagnòstic i classificació de fallides, predicció i pronòstic de defectes i optimització de sistemes. Un cop analitzats els estudis plantejats, s’observa que les tècniques de ML s’estan aplicant amb èxit i que aquesta utilització d’algoritmes de ML proporciona grans resultats de precisió i fiabilitat que garanteixen la productivitat i eficiència dels processos. Con la llegada de la cuarta revolución industrial, la gran cantidad de datos disponibles y el alto poder de computación y almacenamiento de los dispositivos, el Machine Learning (ML) ha cobrado un gran protagonismo a nivel industrial. El ML permite abordar desafíos de fabricación, tales como la mejora de la calidad de los productos y la optimización de los procesos industriales. El objetivo principal de este proyecto es destacar la importante herramienta que supone el ML en la industria y en los procesos de fabricación. También, pretende concienciar de la necesidad de añadir en el currículo académico de un ingeniero nociones básicas sobre esta novedosa herramienta. Por un lado, se muestran y se definen los tres tipos de aprendizaje automático: supervisado, no supervisado y por refuerzo. Esta primera parte va acompañada de la presentación de algunos de los algoritmos más utilizados y los principales elementos para el correcto entrenamiento de estos. Por otro lado, se recogen y sintetizan 45 casos donde se están aplicando técnicas de ML en el ámbito de la ingeniería mecánica. Estos casos se clasifican dependiendo del objetivo principal del algoritmo entre: diagnóstico y clasificación de fallos, predicción y pronóstico de defectos y optimización de sistemas. Tras el análisis de los estudios de caso recogidos, se observa que se están aplicando con éxito las diferentes técnicas de ML. Este uso de algoritmos proporciona grandes resultados de precisión y fiabilidad que garantizan la productividad y eficiencia de los procesos. With the arrival of the fourth industrial revolution; large amounts of data, higher computing power and larger storage devices have become available. Machine Learning (ML) has taken on a major role at industrial level. ML enables you to tackle manufacturing challenges, such as improving product quality and optimizing industrial processes. The main objective of this project is to highlight the importance of ML as a tool in the industry and manufacturing processes. This study also aims to raise awareness of the need for adding basic notions of ML into the engineering curriculum. On one hand, the study defines three types of ML: supervised, unsupervised and reinforcement. This first part is accompanied by the presentation of some of the most used algorithms and the main elements for their correct training. On the other hand, 45 cases of ML techniques applied in the field of mechanical engineering are collected and synthesized. These cases are classified, depending on the main objective of the algorithm between: diagnosis and classification of failures, prediction and prognosis of defects and optimization of systems. After analysing the collected case studies, it was observed that the different ML techniques are being applied successfully. The use of these algorithms provides great results of precision and reliability that guarantee the productivity and efficiency of the processes.

Report with tips and tools for PhD students who need to start working with tools like OMNeT++, SUMO; OpenStreetMaps. We collect here a list of tips to help them to speed up the learning process. Also, we introduce the researcher to machine learning techniques focusing on neural networks. We give tools, references and useful examples to start with. Finally, we also comment how to use latex editors easily. Nowadays, network simulators are frequently used by researchers to work with different kind of communication networks. In this context, the objective of this work is to introduce the user to the main used simulation tools for wireless networks, in particular vehicular ad hoc networks (VANETs). In this report, we gather some useful tools and tips for researchers, including installation and operation of OMNET++, SUMO and VEINS. We think this information will be useful for new users and will save them a lot of time and effort at the beginning of their research works. For this, we have used a widely known distribution of Linux (Ubuntu 16.X). We describe how to integrate simulators OMNET++, VEINS, and SUMO to test vehicular communication protocols and services in an urban environment. This tutorial includes a virtual machine as additional material to help researchers. Our final objective is to assist researchers with their tasks associated to the development of novel proposals that require to perform simulations to show the benefits of their approaches.

In recent years, research has focused on generating mechanisms to assess the levels of subjects&rsquo, cognitive workload when performing various activities that demand high concentration levels, such as driving a vehicle. These mechanisms have implemented several tools for analyzing the cognitive workload, and electroencephalographic (EEG) signals have been most frequently used due to their high precision. However, one of the main challenges in implementing the EEG signals is finding appropriate information for identifying cognitive states. Here, we present a new feature selection model for pattern recognition using information from EEG signals based on machine learning techniques called GALoRIS. GALoRIS combines Genetic Algorithms and Logistic Regression to create a new fitness function that identifies and selects the critical EEG features that contribute to recognizing high and low cognitive workloads and structures a new dataset capable of optimizing the model&rsquo, s predictive process. We found that GALoRIS identifies data related to high and low cognitive workloads of subjects while driving a vehicle using information extracted from multiple EEG signals, reducing the original dataset by more than 50% and maximizing the model&rsquo, s predictive capacity, achieving a precision rate greater than 90%.

Skin lesions are a severe disease globally. Early detection of melanoma in dermatoscopy im-ages significantly increases the survival rate. However, the accurate recognition of melanomais extremely challenging due to the following reasons: low contrast between lesions and skin,appearances of artifacts, etc. Hence, reliable automatic detection of skin tumors is very usefulto increase the accuracy and efficiency of dermatologists. In this MSc thesis we explore the useof contextual information to improve the performance of Deep Learning classifiers in the areaof skin lesion image classification. This information was obtained by means of patient embed-dings, attention, and a hand-crafted contextual sampler. The proposed methods were evaluatedon the ISIC 2020 dataset. Experimental results show that this techniques yield better results butthey are still not significant enough. It is necessary to make future research on other ways toaggregate this information.

A challenge to solve when analyzing multimorbidity patterns in elderly people is the management of a high number of characteristics associated with each patient. The main variables to study multimorbidity are diseases, however other variables should be considered to better classify the people included in each pattern. Age, sex, social class and medication are frequently used in the typing of each multimorbidity pattern. Subsequently the cardinality of the set of features that characterize a patient is very high and normally, the set is compressed to obtain a patient vector of new variables whose dimension is noticeably smaller than that of the initial set. To minimize the loss of information by compression, traditionally Principal Component Analysis (PCA) based projection techniques have been used, which although they are generally a good option, the projection is linear, which somehow reduces its flexibility and limits the performance. As an alternative to the PCA based techniques, in this paper, it is proposed to use autoencoders, and it is shown the improvement in the obtained multimorbidity patterns from the compressed database, when the registered data on about a million patients (5 years' follow-up) are processed. This work demonstrates that autoencoders retain a larger amount of information in each pattern and results are more consistent with clinical experience than other approaches frequently found in the literature.Clinical relevance- From an epidemiological perspective, the contribution is relevant, since it allows for a more precise analysis of multimorbidity patterns, leading to better approaches to patient health strategies. Work supported by Carlos III Institute of Health, Spanish Ministry of Economy and Competitiveness and ERDF (UE) funds under Grant PI16/00639, and in part by Catalan Government funds under Grants 2017 SGR 578 AGAUR and Strategic Plan for Research in Health PERIS grant number SLT002/16/00058

El projecte de fi de grau “Machine Learning a partir de dades clíniques” és un estudi del món Machine Learning i les diverses tècniques que ofereix, i com aplicar aquestes tècniques sobre dades de pacients reals. L’objectiu principal del projecte és crear una solució informàtica basada en el Machine Learning que permet predir les dades d’insulina i de medicació adients a un pacient diabètic, basant-se tant en les dades del propi pacient com en les de tots els pacients dels que se’n conserven dades. Per tal de dur a terme aquest objectiu, s’ha fet un estudi exhaustiu de les principals tècniques utilitzades actualment, i s’han desenvolupat tres mòduls de Python que ofereixen una solució al problema inicial, cadascun dels quals oferint solucions diverses. En la memòria d’aquest projecte es mostra informació bàsica d’estadística i de l’àmbit del Machine Learning, camps estretament lligats. A més, es mostra la implementació informàtica de la solució proposada i es fa un anàlisi minuciós dels resultats obtinguts. The Bachelor Thesis “Machine Learning on medical data” is a study on Machine Learning and the various techniques it offers, and how these techniques can be applied on data from real patients. The main goal of the project is to create a Machine Learning based solution that allows to predict the dose of insulin and the right medication for a diabetic patient. These predictions are both based on the data from the actual patient and from all the previous ones. In order to fulfill the project’s objectives, a deep study on the main techniques that are currently in use has been carried out. Furthermore, three Python modules that offer solutions to the initial problem have been developed, each of those offering a new perspective on the solution. The thesis contains basic information on statistics and Machine Learning techniques, which are highly related. In addition to that, the thesis shows the implementation of the given solution, and a thorough analysis of the results.

This paper presents a study showing the benefits of the EfficientNet models compared with heavier Convolutional Neural Networks (CNNs) in the Document Classification task, essential problem in the digitalization process of institutions. We show in the RVL-CDIP dataset that we can improve previous results with a much lighter model and present its transfer learning capabilities on a smaller in-domain dataset such as Tobacco3482. Moreover, we present an ensemble pipeline which is able to boost solely image input by combining image model predictions with the ones generated by BERT model on extracted text by OCR. We also show that the batch size can be effectively increased without hindering its accuracy so that the training process can be sped up by parallelizing throughout multiple GPUs, decreasing the computational time needed. Lastly, we expose the training performance differences between PyTorch and Tensorflow Deep Learning frameworks. This work was partially supported by the Spanish Ministry of Science and Innovation and the European Regional Development Fund under contract TIN2015- 65316-P, by the BSC-CNS Severo Ochoa program SEV-2015-0493, and grant 2017-SGR-1414 by Generalitat de Catalunya and by the research agreement CaixaBank-BSC 2016-2021.

Trabajo presentado en el CVPR Workshop on Text and Documents in the Deep Learning Era (CVPRW), celebrado online del 14 al 19 de junio de 2020, Text Spotting in the wild consists of detecting and recognizing text appearing in images (e.g. signboards, traffic signals or brands in clothing or objects). This is a challenging problem due to the complexity of the context where texts appear (uneven backgrounds, shading, occlusions, perspective distortions, etc.). Only a few approaches try to exploit the relation between text and its surrounding environment to better recognize text in the scene. In this paper, we propose a visual context dataset 1 for Text Spotting in the wild, where the publicly available dataset COCO-text [40] has been extended with information about the scene (such as objects and places appearing in the image) to enable researchers to include semantic relations between texts and scene in their Text Spotting systems, and to offer a common framework for such approaches. For each text in an image, we extract three kinds of context information: objects in the scene, image location label and a textual image description (caption). We use state-of-the-art out-of-the-box available tools to extract this additional information. Since this information has textual form, it can be used to leverage text similarity or semantic relation methods into Text Spotting systems, either as a post-processing or in an end-to-end training strategy., This work is supported by the KASP Scholarship Program and by the he Spanish government under projects HuMoUR TIN2017-90086-R and Mar´ıa de Maeztu Seal of Excellence MDM-2016-0656.