Search

[EN] Although Long-Short Term Memory (LSTM) networks and deep Transformers are now extensively used in offline ASR, it is unclear how best offline systems can be adapted to work with them under the streaming setup. After gaining considerable experience on this regard in recent years, in this paper we show how an optimized, low-latency streaming decoder can be built in which bidirectional LSTM acoustic models, together with general interpolated language models, can be nicely integrated with minimal performance degradation. In brief, our streaming decoder consists of a one-pass, real-time search engine relying on a limited-duration window sliding over time and a number of ad hoc acoustic and language model pruning techniques. Extensive empirical assessment is provided on truly streaming tasks derived from the well-known LibriSpeech and TED talks datasets, as well as from TV shows on a main Spanish broadcasting station.

[EN] Although Long-Short Term Memory (LSTM) networks and deep Transformers are now extensively used in offline ASR, it is unclear how best offline systems can be adapted to work with them under the streaming setup. After gaining considerable experience on this regard in recent years, in this paper we show how an optimized, low-latency streaming decoder can be built in which bidirectional LSTM acoustic models, together with general interpolated language models, can be nicely integrated with minimal performance degradation. In brief, our streaming decoder consists of a one-pass, real-time search engine relying on a limited-duration window sliding over time and a number of ad hoc acoustic and language model pruning techniques. Extensive empirical assessment is provided on truly streaming tasks derived from the well-known LibriSpeech and TED talks datasets, as well as from TV shows on a main Spanish broadcasting station.

[EN] The cascade approach to Speech Translation (ST) is based on a pipeline that concatenates an Automatic Speech Recognition (ASR) system followed by a Machine Translation (MT) system. Nowadays, state-of-the-art ST systems are populated with deep neural networks that are conceived to work in an offline setup in which the audio input to be translated is fully available in advance. However, a streaming setup defines a completely different picture, in which an unbounded audio input gradually becomes available and at the same time the translation needs to be generated under real-time constraints. In this work, we present a state-of-the-art streaming ST system in which neural-based models integrated in the ASR and MT components are carefully adapted in terms of their training and decoding procedures in order to run under a streaming setup. In addition, a direct segmentation model that adapts the continuous ASR output to the capacity of simultaneous MT systems trained at the sentence level is introduced to guarantee low latency while preserving the translation quality of the complete ST system. The resulting ST system is thoroughly evaluated on the real-life streaming Europarl-ST benchmark to gauge the trade-off between quality and latency for each component individually as well as for the complete ST system.

[EN] The cascade approach to Speech Translation (ST) is based on a pipeline that concatenates an Automatic Speech Recognition (ASR) system followed by a Machine Translation (MT) system. Nowadays, state-of-the-art ST systems are populated with deep neural networks that are conceived to work in an offline setup in which the audio input to be translated is fully available in advance. However, a streaming setup defines a completely different picture, in which an unbounded audio input gradually becomes available and at the same time the translation needs to be generated under real-time constraints. In this work, we present a state-of-the-art streaming ST system in which neural-based models integrated in the ASR and MT components are carefully adapted in terms of their training and decoding procedures in order to run under a streaming setup. In addition, a direct segmentation model that adapts the continuous ASR output to the capacity of simultaneous MT systems trained at the sentence level is introduced to guarantee low latency while preserving the translation quality of the complete ST system. The resulting ST system is thoroughly evaluated on the real-life streaming Europarl-ST benchmark to gauge the trade-off between quality and latency for each component individually as well as for the complete ST system.

En este Trabajo Final de Máster se pretende desplegar y aprovisionar de forma desatendida y distribuida un conjunto de nodos que lleven a cabo el entrenamiento de un modelo de deep learning (DL). Para este fin, se utilizará el framework de DL Tensorflow, junto con la herramienta Infrastructure Manager que permite el despliegue de infraestructuras complejas de forma programática. Se requiere que la infraestructura contemple: la gestión de los datos de entrenamiento, el entrenamiento del modelo con estos datos y la persistencia del modelo entrenado resultante. Se utilizarán plataformas de Cloud público como Amazon Web Services y, en la medida de lo posible, dispositivos de cómputo específico como GPGPUs., In this Master¿s thesis we focus on deploying and provisioning in an unattended and distributed way a set of nodes that conduct the training of a deep learning model. To this end, the deep learning framework TensorFlow will be used, as well as the Infrastructure Manager that allows us to deploy complex infrastructures programmatically. The nal infrastructure should consider: data handling, model training using this data, and the persistence of the trained model. For this purpose, it will be used public Cloud platforms such as Amazon Web Services (AWS) and General-Purpose Computing on Graphics Processing Units (GPGPU).

En este Trabajo Final de Máster se pretende desplegar y aprovisionar de forma desatendida y distribuida un conjunto de nodos que lleven a cabo el entrenamiento de un modelo de deep learning (DL). Para este fin, se utilizará el framework de DL Tensorflow, junto con la herramienta Infrastructure Manager que permite el despliegue de infraestructuras complejas de forma programática. Se requiere que la infraestructura contemple: la gestión de los datos de entrenamiento, el entrenamiento del modelo con estos datos y la persistencia del modelo entrenado resultante. Se utilizarán plataformas de Cloud público como Amazon Web Services y, en la medida de lo posible, dispositivos de cómputo específico como GPGPUs., In this Master¿s thesis we focus on deploying and provisioning in an unattended and distributed way a set of nodes that conduct the training of a deep learning model. To this end, the deep learning framework TensorFlow will be used, as well as the Infrastructure Manager that allows us to deploy complex infrastructures programmatically. The nal infrastructure should consider: data handling, model training using this data, and the persistence of the trained model. For this purpose, it will be used public Cloud platforms such as Amazon Web Services (AWS) and General-Purpose Computing on Graphics Processing Units (GPGPU).

En este Trabajo Final de Máster se pretende desplegar y aprovisionar de forma desatendida y distribuida un conjunto de nodos que lleven a cabo el entrenamiento de un modelo de deep learning (DL). Para este fin, se utilizará el framework de DL Tensorflow, junto con la herramienta Infrastructure Manager que permite el despliegue de infraestructuras complejas de forma programática. Se requiere que la infraestructura contemple: la gestión de los datos de entrenamiento, el entrenamiento del modelo con estos datos y la persistencia del modelo entrenado resultante. Se utilizarán plataformas de Cloud público como Amazon Web Services y, en la medida de lo posible, dispositivos de cómputo específico como GPGPUs., In this Master¿s thesis we focus on deploying and provisioning in an unattended and distributed way a set of nodes that conduct the training of a deep learning model. To this end, the deep learning framework TensorFlow will be used, as well as the Infrastructure Manager that allows us to deploy complex infrastructures programmatically. The nal infrastructure should consider: data handling, model training using this data, and the persistence of the trained model. For this purpose, it will be used public Cloud platforms such as Amazon Web Services (AWS) and General-Purpose Computing on Graphics Processing Units (GPGPU).

[EN] Nowadays, there is an increasing demand for machine learning techniques which can deal with problems where the instances are produced as a stream or in real time. In these scenarios, online learning is able to learn a model from data that comes continuously. The adaptability, efficiency and scalability of online learning techniques have been gaining interest last years with the increasing amount of data generated every day. In this paper, we propose a novel binary classification approach based on nonlinear mapping functions under an online learning framework. The non-convex optimization problem that arises is split into three different convex problems that are solved by means of Passive-Aggressive Online Learning. We evaluate both the adaptability and generalization of our model through several experiments comparing with the state of the art techniques. We improve significantly the results in several datasets widely used previously by the online learning community. (C) 2018 Elsevier Ltd. All rights reserved.

[EN] Nowadays, the research on computer vision and machine learning is in its best moment. The computational capacity and communications currently available in any device, have risen new challenges. Among them, the task of human or object recognition on images and video are impulsed by the best universities and technological companies. Concretely, human activity recognition in videos has a direct application in many environments: security systems, interaction analysis, illness identification, etc. For this reason, this project proposes a prospective study about the task of THUMOS competition on computer vision. In this task, it is required to classify videos by activity, among a set of 101 activities, belonging to 5 different kinds: Human-Human interaction, Human-Object interaction, sports, body-motion, and playing musical instruments. This thesis proposes, applied to this task for the first time, a model based on artificial neural networks that uses improved Dense Trajectories as a feature extraction technique. This thesis will analize the current state-of-the-art, and it will perform experiments in order to obtain the best model for this task, and afterwards, these experiments will be compared with the results provided by the approaches on the top ten of the THUMOS classification, [ES] Actualmente, la investigación en el campo de la visión por computador y el aprendizaje automático se encuentra en su mejor momento. La capacidad de cómputo y de comunicación disponible hoy en dia en cualquier dispositivo ha despertado nuevos desafios. Entre ellos, las tareas de reconocimiento de personas o elementos dentro de imágenes o vídeos, se encuentran impulsadas por las mejores universidades y empresas tecnológicas. Concretamente, el reconocimiento de la actividad llevada a cabo por personas dentro de los vídeos, comprende una tarea que tiene aplicabilidad directa en numerosos entornos: sistemas de seguridad, análisis de la interacción, identificación de enfermedades, etc. Por ello, en este proyecto se propone un estudio prospectivo sobre la tarea planteada en la competición de visión por computador THUMOS. En esta tarea, se requiere la clasificación de vídeos por actividad, de entre un conjunto de 101 actividades, pertenecientes a 5 diferentes grupos: interacción humano-humano, interacción humanoobjeto, deportes, movimientos corporales y personas tocando diversos instrumentos. En este trabajo se plantea un modelo basado en redes neuronales artificiales, que se aplica por primera vez a esta tarea, utilizando la técnica del estado del arte improved Dense Trajectories para la extracción de características. Se analizará, además, el estado de la cuestión hasta el momento, y se llevará a cabo la experimentación con el objetivo de obtener el mejor modelo, para posteriormente comparar los resultados con los obtenidos en las aproximaciones que conforman el top-ten de la clasificación.

[EN] Nowadays, the research on computer vision and machine learning is in its best moment. The computational capacity and communications currently available in any device, have risen new challenges. Among them, the task of human or object recognition on images and video are impulsed by the best universities and technological companies. Concretely, human activity recognition in videos has a direct application in many environments: security systems, interaction analysis, illness identification, etc. For this reason, this project proposes a prospective study about the task of THUMOS competition on computer vision. In this task, it is required to classify videos by activity, among a set of 101 activities, belonging to 5 different kinds: Human-Human interaction, Human-Object interaction, sports, body-motion, and playing musical instruments. This thesis proposes, applied to this task for the first time, a model based on artificial neural networks that uses improved Dense Trajectories as a feature extraction technique. This thesis will analize the current state-of-the-art, and it will perform experiments in order to obtain the best model for this task, and afterwards, these experiments will be compared with the results provided by the approaches on the top ten of the THUMOS classification, [ES] Actualmente, la investigación en el campo de la visión por computador y el aprendizaje automático se encuentra en su mejor momento. La capacidad de cómputo y de comunicación disponible hoy en dia en cualquier dispositivo ha despertado nuevos desafios. Entre ellos, las tareas de reconocimiento de personas o elementos dentro de imágenes o vídeos, se encuentran impulsadas por las mejores universidades y empresas tecnológicas. Concretamente, el reconocimiento de la actividad llevada a cabo por personas dentro de los vídeos, comprende una tarea que tiene aplicabilidad directa en numerosos entornos: sistemas de seguridad, análisis de la interacción, identificación de enfermedades, etc. Por ello, en este proyecto se propone un estudio prospectivo sobre la tarea planteada en la competición de visión por computador THUMOS. En esta tarea, se requiere la clasificación de vídeos por actividad, de entre un conjunto de 101 actividades, pertenecientes a 5 diferentes grupos: interacción humano-humano, interacción humanoobjeto, deportes, movimientos corporales y personas tocando diversos instrumentos. En este trabajo se plantea un modelo basado en redes neuronales artificiales, que se aplica por primera vez a esta tarea, utilizando la técnica del estado del arte improved Dense Trajectories para la extracción de características. Se analizará, además, el estado de la cuestión hasta el momento, y se llevará a cabo la experimentación con el objetivo de obtener el mejor modelo, para posteriormente comparar los resultados con los obtenidos en las aproximaciones que conforman el top-ten de la clasificación.

[EN] Nowadays, the research on computer vision and machine learning is in its best moment. The computational capacity and communications currently available in any device, have risen new challenges. Among them, the task of human or object recognition on images and video are impulsed by the best universities and technological companies. Concretely, human activity recognition in videos has a direct application in many environments: security systems, interaction analysis, illness identification, etc. For this reason, this project proposes a prospective study about the task of THUMOS competition on computer vision. In this task, it is required to classify videos by activity, among a set of 101 activities, belonging to 5 different kinds: Human-Human interaction, Human-Object interaction, sports, body-motion, and playing musical instruments. This thesis proposes, applied to this task for the first time, a model based on artificial neural networks that uses improved Dense Trajectories as a feature extraction technique. This thesis will analize the current state-of-the-art, and it will perform experiments in order to obtain the best model for this task, and afterwards, these experiments will be compared with the results provided by the approaches on the top ten of the THUMOS classification, [ES] Actualmente, la investigación en el campo de la visión por computador y el aprendizaje automático se encuentra en su mejor momento. La capacidad de cómputo y de comunicación disponible hoy en dia en cualquier dispositivo ha despertado nuevos desafios. Entre ellos, las tareas de reconocimiento de personas o elementos dentro de imágenes o vídeos, se encuentran impulsadas por las mejores universidades y empresas tecnológicas. Concretamente, el reconocimiento de la actividad llevada a cabo por personas dentro de los vídeos, comprende una tarea que tiene aplicabilidad directa en numerosos entornos: sistemas de seguridad, análisis de la interacción, identificación de enfermedades, etc. Por ello, en este proyecto se propone un estudio prospectivo sobre la tarea planteada en la competición de visión por computador THUMOS. En esta tarea, se requiere la clasificación de vídeos por actividad, de entre un conjunto de 101 actividades, pertenecientes a 5 diferentes grupos: interacción humano-humano, interacción humanoobjeto, deportes, movimientos corporales y personas tocando diversos instrumentos. En este trabajo se plantea un modelo basado en redes neuronales artificiales, que se aplica por primera vez a esta tarea, utilizando la técnica del estado del arte improved Dense Trajectories para la extracción de características. Se analizará, además, el estado de la cuestión hasta el momento, y se llevará a cabo la experimentación con el objetivo de obtener el mejor modelo, para posteriormente comparar los resultados con los obtenidos en las aproximaciones que conforman el top-ten de la clasificación.

[EN] Nowadays, the research on computer vision and machine learning is in its best moment. The computational capacity and communications currently available in any device, have risen new challenges. Among them, the task of human or object recognition on images and video are impulsed by the best universities and technological companies. Concretely, human activity recognition in videos has a direct application in many environments: security systems, interaction analysis, illness identification, etc. For this reason, this project proposes a prospective study about the task of THUMOS competition on computer vision. In this task, it is required to classify videos by activity, among a set of 101 activities, belonging to 5 different kinds: Human-Human interaction, Human-Object interaction, sports, body-motion, and playing musical instruments. This thesis proposes, applied to this task for the first time, a model based on artificial neural networks that uses improved Dense Trajectories as a feature extraction technique. This thesis will analize the current state-of-the-art, and it will perform experiments in order to obtain the best model for this task, and afterwards, these experiments will be compared with the results provided by the approaches on the top ten of the THUMOS classification, [ES] Actualmente, la investigación en el campo de la visión por computador y el aprendizaje automático se encuentra en su mejor momento. La capacidad de cómputo y de comunicación disponible hoy en dia en cualquier dispositivo ha despertado nuevos desafios. Entre ellos, las tareas de reconocimiento de personas o elementos dentro de imágenes o vídeos, se encuentran impulsadas por las mejores universidades y empresas tecnológicas. Concretamente, el reconocimiento de la actividad llevada a cabo por personas dentro de los vídeos, comprende una tarea que tiene aplicabilidad directa en numerosos entornos: sistemas de seguridad, análisis de la interacción, identificación de enfermedades, etc. Por ello, en este proyecto se propone un estudio prospectivo sobre la tarea planteada en la competición de visión por computador THUMOS. En esta tarea, se requiere la clasificación de vídeos por actividad, de entre un conjunto de 101 actividades, pertenecientes a 5 diferentes grupos: interacción humano-humano, interacción humanoobjeto, deportes, movimientos corporales y personas tocando diversos instrumentos. En este trabajo se plantea un modelo basado en redes neuronales artificiales, que se aplica por primera vez a esta tarea, utilizando la técnica del estado del arte improved Dense Trajectories para la extracción de características. Se analizará, además, el estado de la cuestión hasta el momento, y se llevará a cabo la experimentación con el objetivo de obtener el mejor modelo, para posteriormente comparar los resultados con los obtenidos en las aproximaciones que conforman el top-ten de la clasificación.

[ES] La actual industria del entretenimiento exige la generación de contenidos de diversa índole, tanto para ofrecer al jugador nuevas experiencias, como para sorprender al espectador con nuevos planteamientos. Dentro de estos contenidos, la generación de escenarios donde se desarrolla la acción es uno de los mayores retos a los que se enfrentan los sistemas gráficos interactivos. En concreto, los escenarios urbanos, plantean una amplia problemática y una gran complejidad. Plasmar qué distribución de carreteras establecer, o qué estructura y tamaño tendrán los edificios de una forma automática, es un problema de estudio desde hace décadas. Por ello, en este trabajo, se plantea un marco teórico y práctico donde se ha desarrollado una librería de generación de entornos urbanos sintéticos, así como su interacción en 3D bajo la tecnología del navegador, todo ello mediante WebGL y three.js., [EN] The current state of the industry of entertainment, requires generating contents of diverse kind, in order to offer new experiences to the player and surprise the spectator with new approaches. Regarding the contents, one of the major challenges for the interactive graphics systems is the generation of scenarios where action occurs. These scenarios, specifically the urban ones, contemplate wide difficulties and a huge complexity. Thus, since decades, one important case of study is shaping and set up the road distribution or reflect the size or which shape will have the buildings in an automatic way. For that reason, a theoretical and practical framework to deal with this problem is proposed in this project. Thus, a library of generation of urban synthetic environment has been developed, allowing the interaction in 3D under the technology of the browser, all by means of WebGL and three.js.

USER’S MANUAL. Version 2.1.0. January 2015, Magentix2 is an agent platform for open Multiagent Systems. Its main objective is to bring agent technology to real domains: business, industry, logistics, e-commerce, health-care, etc. Magentix2 platform is proposed as a continuation of the first Magentix platform. The final goal is to extend the functionalities of Magentix, providing new services and tools to allow the secure and optimized management of open Multiagent Systems. Nowadays, Magentix2 provides support at three levels: - Organization level, technologies and techniques related to agent societies. - Interaction level, technologies and techniques related to communications between agents. - Agent level, technologies and techniques related to individual agents (such as reasoningand learning). Thus, Magentix2 platform uses technologies with the necessary capacity to cope with the dynamism of the system topology and with flexible interactions, which are both natural consequences of the distributed and autonomous nature of its components. In this sense, the platform has been extended in order to support flexible interaction protocols and conversations, indirect communication and interactions among agent organizations. Moreover, other important aspects cover by the Magentix2 project are the security issues.

[EN] The exponential growth of social networks makes fingerprint let by users on the Internet a great source of information, with data about their preferences, needs, goals, profile and social environment. These data are distributed across di↵erent sources of information (social networks, blogs, databases, etc.) that may contain inconsistencies and their accuracy is uncertain. Paradoxically, this unprecedented availability of heterogeneous data has meant that users have more information available than they actually are able to process and understand to extract useful knowledge from it. Therefore, new tools that help users in their decision-making processes within the network (e.g. which friends to contact with or which products to consume) are needed. In this paper, we show how we have used a graph-based model to extract and model data and transform it in valuable knowledge to develop a persuasive social recommendation system1.

USER’S MANUAL. Version 2.1.0. January 2015, Magentix2 is an agent platform for open Multiagent Systems. Its main objective is to bring agent technology to real domains: business, industry, logistics, e-commerce, health-care, etc. Magentix2 platform is proposed as a continuation of the first Magentix platform. The final goal is to extend the functionalities of Magentix, providing new services and tools to allow the secure and optimized management of open Multiagent Systems. Nowadays, Magentix2 provides support at three levels: - Organization level, technologies and techniques related to agent societies. - Interaction level, technologies and techniques related to communications between agents. - Agent level, technologies and techniques related to individual agents (such as reasoningand learning). Thus, Magentix2 platform uses technologies with the necessary capacity to cope with the dynamism of the system topology and with flexible interactions, which are both natural consequences of the distributed and autonomous nature of its components. In this sense, the platform has been extended in order to support flexible interaction protocols and conversations, indirect communication and interactions among agent organizations. Moreover, other important aspects cover by the Magentix2 project are the security issues.

USER’S MANUAL. Version 2.1.0. January 2015, Magentix2 is an agent platform for open Multiagent Systems. Its main objective is to bring agent technology to real domains: business, industry, logistics, e-commerce, health-care, etc. Magentix2 platform is proposed as a continuation of the first Magentix platform. The final goal is to extend the functionalities of Magentix, providing new services and tools to allow the secure and optimized management of open Multiagent Systems. Nowadays, Magentix2 provides support at three levels: - Organization level, technologies and techniques related to agent societies. - Interaction level, technologies and techniques related to communications between agents. - Agent level, technologies and techniques related to individual agents (such as reasoningand learning). Thus, Magentix2 platform uses technologies with the necessary capacity to cope with the dynamism of the system topology and with flexible interactions, which are both natural consequences of the distributed and autonomous nature of its components. In this sense, the platform has been extended in order to support flexible interaction protocols and conversations, indirect communication and interactions among agent organizations. Moreover, other important aspects cover by the Magentix2 project are the security issues.

[EN] The exponential growth of social networks makes fingerprint let by users on the Internet a great source of information, with data about their preferences, needs, goals, profile and social environment. These data are distributed across di↵erent sources of information (social networks, blogs, databases, etc.) that may contain inconsistencies and their accuracy is uncertain. Paradoxically, this unprecedented availability of heterogeneous data has meant that users have more information available than they actually are able to process and understand to extract useful knowledge from it. Therefore, new tools that help users in their decision-making processes within the network (e.g. which friends to contact with or which products to consume) are needed. In this paper, we show how we have used a graph-based model to extract and model data and transform it in valuable knowledge to develop a persuasive social recommendation system1.

© ACM 2015 This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in ACM, In Proceedings of the 30th Annual ACM Symposium on Applied Computing (pp. 189-194).http://dx.doi.org/10.1145/2695664.2695732, Nowadays, social networks have an enormous impact in the society generating a lot of useful information to be employed in new social applications. In this paper, we show how we have used a graph-based model to extract and model data in order to develop a Social Recommendation System which recommends recipes in a social network.