Tesis (Arquitectura y Tecnología de Computadores)

URI permanente para esta colecciónhttps://hdl.handle.net/11441/11295

Examinar

Mostrando 1 - 20 de 36

Study, Design and Implementation of Neuromorphic Systems through a Spiking Boolean Computing Paradigm
(2025-01-24) Ayuso Martínez, Álvaro; Jiménez Moreno, Gabriel; Domínguez Morales, Juan Pedro; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
In recent years, advances in transistor integration within digital computers have enabled them to be reduced to near-atomic scales, pushing this technology to its physical and thermal limits. This trend, which has also significantly increased production costs, reinforces the belief that Moore's law is going to become obsolete in the coming years. However, although doubts may arise about the possibility of further improving the efficiency of digital computers, these disappear when considering the brain, which is the most powerful and efficient system known. It is not based on transistors but on neurons and achieves high performance with minimal power consumption, both characteristics emerging mainly from the massive parallelism inherent to the nervous system. Inspired by the principles of neuromorphic engineering, this work proposes replacing transistors in digital circuits with neurons to harness these benefits. By abstracting neuronal function, it is possible to apply Boolean algebra to the design of Spiking Neural Networks under specific conditions, in a similar way to how it is applied to the design of digital circuits. Thus, this work lays the foundation for spiking Boolean computation through the spiking implementation of basic logic gates, providing a systematic approach for designing these networks, which could be valuable for researchers in the field. It also explores the development of complex spiking blocks for specialized applications, in which the development of the spiking computer is highlighted, and presents an extensive set of experiments whose results demonstrate their correct functionality mainly on two different neuromorphic platforms, SpiNNaker and Dynap-SE. The final implementations have been shown to behave as expected in challenging environments and under conditions comparable to those found in biology.
On the Optimization of geolocation Problems in wireless sensor Networks and connected Objects
(2024-12-13) Mani, Rahma; Ríos Navarro, José Antonio; Liouane, Noureddine; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
As localization represents the main core of various wireless sensor network applications including the Internet of Things (IoT), Industry 4.0, e-health, e-agriculture, etc..., several localization algorithms have been suggested in our research work. In fact, obtaining precise information about the location of the sensor nodes is crucial for making the data collected useful and meaningful. In our studies, we consider the type of range-free sensor network exploiting the wireless sensor connectivity. Range-free localization is a vital aspect of wireless sensor networks, aimed at estimating the positions of sensor nodes without requiring precise distance measurements or specialized hardware. The main advantages of connectivity-based algorithms for localizing multi-hop sensor networks are their simplicity, low energy consumption and acceptable accuracy. Firstly, We proposed a fuzzy-ELM learning machine to provide precise geolocation in WSNs. This method for training WSNs combines two techniques within the field of artificial intelligence: the Extreme Learning Machine (ELM) combined with the concept of Fuzzy-Logic. Fuzzy logic concepts are used for minimizing the environnemental incertainty by using a set of localization rules and ELM machine learning procedure is used as fastest localization machine learning. Subsequently, we replaced a large number of GPS-equipped anchors with a single mobile anchor. We employed a regularized bounding box to predict the positions of the unknown nodes using regularization factors and we used the Kalman filter algorithm to smooth and improve the accuracy of the sensor nodes to be localized. Moreover, the use of a reconfigurable FPGA board allows for real-time processing of sensory data, enabling us to make fast and accurate decisions in dynamic environments. In effect, the performance of FPGA hardware implementation presents a new achievement in localization due to its easy testing and fast implementation. Furthermore, we have put a great deal of effort into research to extend the 2D positioning algorithms in WSNs to 3D that reflects reality and the most practical applications. We assessed the perfor-mance of our algorithms using exhaustive experiments by evaluating parameters such as localization accuracy whilst changing other simulation characteristics such as the effect of communication range, number of nodes and the mobile anchor node trajectory on several isotropic and anisotropic topologies. Our results show that our approaches can efficiently locate unknown nodes with good latency and high accuracy.
Developing digital and analog spiking neural Networks for interacting with biological Tissues in the Case Study of temporal lobe Epilepsy
(2024-10-25) Ahmadi Farsani, Javad; Serrano Gotarredona, María Teresa; Linares Barranco, Alejandro; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
The HERMES (Hybrid Enhanced Regenerative Medicine Systems) project aims to pioneer an innovative approach to treating temporal lobe epilepsy by leveraging the symbiotic integration of tissue engineering, neuromorphic engineering, and artificial intelligence (AI). Within this project, my role focused on the neuromorphic engineering aspect, specifically designing a spiking neural network (SNN) using CMOS integrated circuits (ICs) with memristors to demonstrate its potential for pattern recognition tasks. To facilitate the implementation of the SNN, memristors were chosen as synaptic components due to their emerging properties and potential benefits in memory calculation applications. A collaborative effort with researchers from Politecnico di Milano (POLIMI) provided access to memristor chips, enabling the incorporation of these devices into the SNN design. Additionally, collaboration with biologists from the Italian Institute of Technology (IIT) enriched the project by providing access to a microelectrode array (MEA) system capable of recording local field potentials from brain tissues, particularly, rodent brain slices. This real-time interaction with brain tissues through the MEA system presented opportunities to validate the functionality of the SNN in a physiological context. The implementation of the SNN began with simulation and modeling on a Digital Signal Processor (DSP) platform, specifically the DSK6455 board, which replicated the functionality of the DSP embedded within the MEA system. This initial validation step ensured the viability of the SNN design before proceeding to the costly and time-intensive process of ASIC chip design. Experimental results demonstrated the capability of the SNN to learn and adapt using Spike-Timing-Dependent Plasticity (STDP) algorithms, laying the foundation for further hybrid experiments conducted with the IIT team. Subsequent stages of the project involved the design and fabrication of ASIC chips and PCBs to create a fully custom hardware SNN, culminating in pattern recognition experiments conducted using FPGA boards controlling the ASICs and memristor chips. As a result of the thesis, the DSP was employed for modelling, evaluation, and testing of the SNN and its constituent building blocks, intended for eventual design on an ASIC. Additionally, using the DSP, a closed-loop real-time biohybrid in vitro system was developed to interact with an actual rodent brain slice for seizure detection. The designs initially implemented on the DSP were transitioned to ASIC utilizing a 180nm technology. In the pre-synaptic part, a compact field potential to spike converter was engineered, capable of extracting spikes from overshoots, undershoots, or both. Meanwhile, in the post-synaptic part, a current-attenuator was employed to diminish current prior to its input into the neuron circuit, resulting in a significantly compact neuron circuit design. Experimental findings indicated that the SNN shows notable power efficiency when compared with state of the arts.
Computational Neuromorphic Architectures for Modeling the Hippocampus Formation applied to Robotic Navigation
(2024-07-01) Casanueva Morato, Daniel; Jiménez Moreno, Gabriel; Domínguez Morales, Juan Pedro; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
The brain is the most powerful machine that exists, capable of efficiently solving complex problems and far surpassing the capabilities of current systems. In recent decades, neuromorphic engineering has been responsible for the study, design and implementation of hardware and software systems that mimic the behavior, structure and functioning of the brain to achieve such superior capabilities. Within computing systems, memory is a critical component that limits the evolution of these systems by becoming a bottleneck in the flow of information. Additionally, despite the significant growth in computing, the robotics field has seen less significant evolution. Within the brain, the hippocampus stands out for its participation in episodic memory; it can learn and store a large amount of information through association from different brain sensory nuclei, while also being able to recall it based on different fragments of itself. Therefore, this work focuses on the study, design and implementation of neuromorphic memory systems bio-inspired by the hippocampus. A variety of models are proposed to explore different functionalities and paradigms observed in the neuromorphic domain (biological plausibility, analog or digital technology and simulation or emulation). These models, which are capable of learning, forgetting and recalling spiking information, have been developed using Spiking Neural Networks and implemented on various special-purpose hardware platforms for such type of networks. Furthermore, these models have been integrated into robotic platforms for learning, mapping and navigating environments and trajectories. These are the first implementations on specialpurpose hardware platforms for Spiking Neural Networks of fully functional memory models bio-inspired by the hippocampus, paving the way for the development of future, more complex neuromorphic systems.
Aplicaciones de la inteligencia artificial en la optimización de la salud y el rendimiento en equipos deportivos: un enfoque en la cardiología deportiva
(2024-06-24) Muñoz-Macho, Adolfo Antonio; Domínguez Morales, Manuel Jesús; Sevillano Ramos, José Luis; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
El enfoque en la protección de la salud y la optimización del rendimiento deportivo de los atletas dentro de los equipos deportivos profesionales ha cobrado una importancia creciente en los últimos años. En este contexto, la inteligencia artificial (IA) emerge como una herramienta de alto interés, capaz de transformar la manera en que se monitorean, analizan y mejoran tanto el rendimiento físico como el bienestar y salud de los jugadores. Esta tesis se centra en la aplicación de tecnologías de IA para el análisis y la mejora del rendimiento y la salud en equipos deportivos profesionales, abarcando desde la detección temprana de posibles lesiones o enfermedades hasta la gestión de datos de entrenamientos y partidos para maximizar la efectividad y eficiencia de cada jugador. En su desarrollo concreto, se abordan las posibles mejoras en el ámbito de la cardiología deportiva aportando nuevas herramientas y visiones innovadoras y de interés. La tesis se ha elaborado como compendio de publicaciones, donde como primera investigación se ha explorado el estado del arte de las aplicaciones actuales de la IA en equipos deportivos profesionales con la intención de desvelar las aplicaciones prácticas realizadas y descubrir los posibles vacíos de conocimiento. Se ha explorado más a fondo las aplicaciones orientadas a salud y prevención y manejo de lesiones y también orientado al rendimiento con datos de posicionamiento global deportivo, test de bienestar y otros medios de valoración del rendimiento. Dentro de este estudio, se observó la evaluación de la efectividad de los sistemas en función del tipo de técnica utilizada, revelando como la IA puede ayudar a mejoras significativas en la capacidad de entender los procesos que ayudan a la mejora de rendimiento físico de los jugadores y en una reducción en la incidencia de lesiones y otras incidencias. Como segunda aportación se ha generado el primer dataset público de electrocardiogramas de deportistas profesionales que será de utilidad para la comunidad científica. En tercer lugar, se presenta un estudio de caso en el que se aplican varias técnicas de IA en un equipo de fútbol profesional, demostrando cómo la tecnología puede utilizarse para ayudar en el seguimiento de datos cardiológicos a través del electrocardiograma, prevenir una falta de diagnóstico, mejorar la clasificación de riesgo y, en última instancia, elevar el nivel de seguimiento de la salud y rendimiento del equipo. Como últimas aportaciones se presenta un estudio desarrollando una herramienta de visualización de ECGs en un capítulo de libro producto de una presentación a un congreso internacional y una evolución de esta iniciativa hacia la simulación y la posibilidad de modificación de datos con fines educativos en un quinto artículo que está enviado y en proceso de revisión. Esta tesis contribuye al campo de la IA aplicada al deporte, ofreciendo una perspectiva integral sobre cómo las tecnologías de este tipo pueden ser empleadas de manera efectiva para fomentar el rendimiento deportivo y la salud de los atletas en el ámbito de los deportes de equipo. A través de la construcción y análisis de modelos predictivos avanzados, se establece un marco para la aplicación futura de estas tecnologías, no solo en el fútbol, sino también en una amplia gama de disciplinas deportivas, abriendo nuevas vías para la investigación y la práctica en la intersección entre la inteligencia artificial, el deporte y la medicina deportiva.
Estudio y aplicación de la teoría psicofisiológica de la emoción para la detección y clasificación del estado afectivo del usuario mediante técnicas de Machine Learning para el apoyo en procesos de rehabilitación
(2022-10-14) Muñoz Saavedra, Luis; Domínguez Morales, Manuel Jesús; Miró Amarante, María Lourdes; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
En este trabajo se realiza una contribución a los sistemas automáticos de detección y clasificación de emociones, con el fin de ayudar a la adherencia en los procesos de rehabilitación. Para ello se sigue un proceso analítico en el que, en primera instancia, se realiza un estudio profundo de la teoría emocional y su evolución a lo largo de los años, así como su incidencia sobre la fisiología de la persona (psicofisiología). Seguidamente, se demuestra esta incidencia mediante la elaboración de un clasificador del estado emocional basado en redes neuronales profundas, utilizando un conjunto de datos públicos. Una vez demostrada la teoría, se diseña e implementa un dispositivo portable capaz de recolectar información en tiempo real de diversas variables fisiológicas del paciente. Con dicho dispositivo, se elabora un protocolo propio de evaluación basado en la exposición de los participantes a diversos estímulos audiovisuales, aprobado por un comité ético, con el objetivo de recolectar un conjunto de datos de participantes sobre los cuales se estudia la teoría psicofisiológica en profundidad. Para ello se sigue un proceso de optimización y evaluación hasta obtener el clasificador con la combinación de variables fisiológicas que mejor determina el estado emocional del usuario. Los resultados demuestran que la combinación de sensores formada por la respuesta galvánica de la piel y la señal de hemoglobina oxigenada en sangre conforma el conjunto de variables fisiológicas que mejor representa el estado emocional. Estos resultados demuestran la viabilidad de hacer uso de este tipo de sistemas en procesos de rehabilitación, proporcionando información útil al profesional sanitario para adaptar los ejercicios y la intensidad de los mismos dependiendo del paciente.
Neuromorphic auditory computing: towards a digital, event-based implementation of the hearing sense for robotics
(2022-09-23) Gutiérrez Galán, Daniel; Linares Barranco, Alejandro; Jiménez Fernández, Ángel Francisco; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
In this work, it is intended to advance on the development of the neuromorphic audio processing systems in robots through the implementation of an open-source neuromorphic cochlea, event-based models of primary auditory nuclei, and their potential use for real-time robotics applications. First, the main gaps when working with neuromorphic cochleae were identified. Among them, the accessibility and usability of such sensors can be considered as a critical aspect. Silicon cochleae could not be as flexible as desired for some applications. However, FPGA-based sensors can be considered as an alternative for fast prototyping and proof-of-concept applications. Therefore, a software tool was implemented for generating open-source, user-configurable Neuromorphic Auditory Sensor models that can be deployed in any FPGA, removing the aforementioned barriers for the neuromorphic research community. Next, the biological principles of the animals' auditory system were studied with the aim of continuing the development of the Neuromorphic Auditory Sensor. More specifically, the principles of binaural hearing were deeply studied for implementing event-based models to perform real-time sound source localization tasks. Two different approaches were followed to extract inter-aural time differences from event-based auditory signals. On the one hand, a digital, event-based design of the Jeffress model was implemented. On the other hand, a novel digital implementation of the Time Difference Encoder model was designed and implemented on FPGA. Finally, three different robotic platforms were used for evaluating the performance of the proposed real-time neuromorphic audio processing architectures. An audio-guided central pattern generator was used to control a hexapod robot in real-time using spiking neural networks on SpiNNaker. Then, a sensory integration application was implemented combining sound source localization and obstacle avoidance for autonomous robots navigation. Lastly, the Neuromorphic Auditory Sensor was integrated within the iCub robotic platform, being the first time that an event-based cochlea is used in a humanoid robot. Then, the conclusions obtained are presented and new features and improvements are proposed for future works.
Deep Learning-based Computer-Aided Diagnosis systems: a contribution to prostate cancer detection in histopathological images
(2021-07-20) Durán López, Lourdes; Linares Barranco, Alejandro; Vicente Díaz, Saturnino; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
In this work, novel computer-aided diagnosis systems for medical image analysis focusing on prostate cancer are proposed and implemented. First, the histopathology of prostate cancer was studied, along with the Gleason Grading System, which measures the aggressiveness of a tumor through different patterns with the purpose of driving therapies dealing with this disease. Furthermore, a study of Deep Learning techniques, particularly focusing on neural networks applied to medical image analysis, was conducted. Based on these studies, a Deep Learning-based system to detect malignant regions in gigapixel-size whole-slide prostate cancer tissue images was proposed and developed, which is able to report spatial information of the malignant areas. This solution was evaluated in terms of performance and execution time, obtaining promising results when compared to other state-of-the-art methods. Since the implemented system locates malignant regions within the image without providing a global class, a customWide & Deep network was developed to report a slide-level label per image. The proposed system provides a fast screening method for analyzing histopathological images. Next, a neural network was proposed to assign a specific Gleason pattern to the malignant areas of the tissue. Finally, with the purpose of developing a global computeraided diagnosis system for prostate cancer detection and classification, the three aforementioned subsystems were combined, allowing a complete analysis of histopathological images by reporting whether the sample is normal or malignant, and, in the last case, a heatmap of the malignant areas with their corresponding Gleason pattern. The studied algorithms were also used for other medical image analysis tasks. The performance of these systems were evaluated, discussing the obtained results, presenting conclusions and proposing improvements for future works.
Health Recommender Systems for Behavior Change: Exploring their Potential for Smoking Cessation
(2022-03-24) Hors Fraile, Santiago; Fernández-Luque, Luis; Vries, Hein de; Schneider, Francine; Civit Balcells, Antón; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
Smoking has several harmful effects on our health and affects our organs, leading to the incidence of many life-threatening diseases. Furthermore, it is one of the most preventable causes of death. Despite its detrimental effect on our health, quitting smoking is challenging due to the tobacco addictive chemicals and humans’ psychological dependency on it. Nonetheless, there are different approaches to support people willing to stop smoking. One method is eHealth computer tailoring, which helps personalize feedback given to smokers based on psychological models of behavioral change based on pre-defined if-then-else rules. These methods showed to generate positive results in terms of high abstinence rates and cost-effectiveness. However, new innovative solutions are available to improve the eHealth methods for smoking cessation further. One of those methods is related to recommender systems technology. Recommender systems are AI algorithms that can select the most relevant item (such as a piece of text, book, movie, or product) from a set of items for each user. Depending on the type of recommender system, relevance is determined considering different methods and variables. A commonly used method for calculating relevance is the “collective intelligence” approach. This approach uses algorithms to generate a user profile for each user (e.g., using demographic variables) and calculate how relevant a specific item is based on the given relevance of that item for users with similar user profiles.These systems can learn from user feedback over time in that the users rate the relevance of the recommended items, which helps train the system for making future recommendations. For decades, the scientific community has explored the relevance of these systems in other fields such as leisure (movie recommendations on Netflix) and e-commerce (product recommendations on Amazon). Due to their potential and proven effectiveness in other fields but limited application in the healthcare sector, which began onlya few years ago, studying how these systems can be applied for smoking cessation is crucial.In Chapter 2 of this dissertation, we have conducted a scoping review to assess the existing knowledge and research gaps using recommender systems in healthcare,also known as health recommender systems (HRSs). We assessed their technical and healthcare aspects through this review. Based on its results, we then generated a new taxonomy for these types of systems. Next, we provided a detailed description of a health recommender system (HRS) design process with collective intelligence grounded in behavioral science for smoking cessation using the I-Change model as an example. In Chapter 3, we explained all the steps and the system design, including algorithm components, messages creation, and user interface design, to help interested stakeholders better understand such systems, which would provide inspiration and a basis for future studies. Furthermore, we performed an assessment study to test the created HRS using collective intelligence in a real-world setting with a follow-up period of six months. The control condition was a simpler version of the created HRS in this assessment, except for the collective intelligence component. In Chapter 4, we reported the protocolof this study and analyzed the actual results regarding the appreciation, engagement, dropouts, and smoking abstinence generated by the system (Chapter 5). Chapter 1 provides a general introduction to the problem associated with smoking cessation. First, it introduces different existing support approaches, focusing on the ones related to behavioral change and their application in computer-tailored interventions. Then, it presents the recommender system technology and its different types that exist as an option for facilitating computer-tailored interventions. Further, it highlights the appreciation and engagement metrics, which are the factors that complement abstinence for intervention success. Chapter 2 contains a scoping review that provides an analysis of the state-of-the-art HRS, identifying the research gaps and the elements that should be improved when applying this technology to the healthcare sector. From this study, we identified that the collaborative filtering technique was the most-used information filtering method. However, it was also observed that there is a lack of applying behavioral change theories and factors in HRS studies. Furthermore, these studies neither implemented the principles of tailoring nor assessed their (cost)-effectiveness. Therefore, a taxonomy was proposed to facilitate consistent classification and better comprehension of these systems. This taxonomy included the domain of the study (e.g., the type of population, country, therapeutic area), the methodology and procedures of the study (the duration, number of users, outcomes), health behavior change factors (e.g., self-efficacy, social influence, attitudes), and the technical aspects required to understand the algorithm (e.g., recommendation technology, profile generation techniques).Chapter 3 provides a multidisciplinary and comprehensive description of the design process of an HRS for supporting smoking cessation that uses collective intelligence in combination with the I-Change behavioral change model. This detailed description contributed to help reveal the process of how an HRS can be built to support behavioral change interventions. This process had not been disclosed in detail before, and this lack of transparency can act as a barrier for behavioral change researchers in using HSR technology. The new system was built based on a previous HRS that utilized a mobile app to support smokers trying to stay abstinent by sending them motivational messages. First, we identified the areas that needed improvements based on the app’s usage data. Then, we implemented relevant changes to our new system design (e.g., increasing the granularity of the possible user feedback from three options to five options). Our final mobile app was supposed to be more streamlined and usable thanthe first version. The generated HRS was a hybrid algorithm with a knowledge-based step and a collaborative-filtering step in cascade. It used 58 variables to compute the similarity formula for choosing recommendations; from the total, 47 were related to the determinants of the I-Change model. Altogether, 331 motivational messages were created, and ten different health communication methods were considered for their design. Chapter 4 explains the protocol to be followed to assess the system created in Chapter 3. This protocol included the description of a clinical pilot and a public pilot. We used the latter one to analyze the HRS in this dissertation. Chapter 5 presents, discusses and reflects on the results obtained from the public pilot. The public pilot was a double-blinded experiment. Those smokers who can read English or Mandarin and download a mobile app from the Internet were eligible to participate. After creating their account and answering questions relevant to their user profile (e.g., name, age, gender, level of addiction, and motivation to quit), they can set a quitting day to start receiving personalized motivational text messages via the mobile app. Smokers were randomly allocated to the group where such messages were generated by the new HRS, which was described in Chapter 3, or to the group associated with a simpler version of the algorithm, without collective intelligence (using only the knowledge-based step), selected and sent these messages. A total of 371 participants were eligible to be part of the study analysis. Smokers were followed up for six months, starting from their quitting day, and were asked weekly about their smoking abstinence through a voluntary question in the app.Moreover, we measured their message appreciation and engagement. The attributes (factors) considered as possible indicators of differences in the study outcomes included the motivation to quit, nicotine dependence, age, gender, and completion of the extended user profile questionnaire. They were studied as potential covariates in the statistical analysis. No statistically significant differences were found neither for the analysis on available data of the 7D-PP abstinence averaging the abstinence reports across the study nor for the penalized imputation analysisof both the 7D-PP abstinence averaging the abstinence reports across the study and the 7D-PP considering only the last available abstinence report. However, the analysis on available data for the 7D-PP considering only the last available abstinence reportshowed lower abstinence rates in the HRS using collective intelligence. Also, the results showed that the HRS using collective intelligence did not have statistically significant differences for message appreciation, number of rated messages, and number of quitting attempts. However, the collective intelligence algorithm performed worse regarding the number of abstinence reports and active days. The sub-group analysis showed that the completion of the extended user profile did significantly impact the engagement of the participants reducing the number of dropouts in both groups and increasing the number of quitting attempts in participants who received messages selected with the collective intelligence. Finally, Chapter 6 provides a general discussion of the main findings and conclusions of all the studies presented in this dissertation (from chapters 2–5). It also contains the main methodological considerations for this dissertation, such as the strengths and limitations, risks, reflections for practice, and the impact of this thesis on the scientific community. In conclusion, the studies presented in this dissertation showed that although HRSs are gaining traction in the healthcare sector, they are still novel, with underreported details and suboptimal application, as they do not take advantage of the behavioral change theories. However, we have shown that they can be used as an alternative approach to traditional tailoring for behavioral interventions by embedding behavioral science in the design of theseemergent systems. We compared the HRSs with and without collective intelligence technology for a trial for smoking cessation, measuring their performance in real-life conditions. The results showed that despite showing some positive results in terms of engagement –number of quitting attempts -when completing the extended user profile, the HRS using collective intelligence did not manage to improve smoking behavior, appreciation, and engagement compared to the other HRS. In addition, some of the engagement and abstinence metrics led to worse results. Furthermore, although we achieved better smoking cessation outcomes than quitting cold turkey or with brief clinician advice, our HRS did not improve the abstinence rates achieved by other approaches in smoking cessation, such as traditional computer tailoring. Further, it is still unclear why the theoretical potential of collective intelligence did not provide the expected benefits in our study. Therefore, future research is needed to find out how HRS-based interventions, using or not using the collective intelligence technology, can be improved to achieve better outcomes in terms of behavioral change.
Memristor Based Event Driven Neuromorphic Nano-CMOS Processor
(2021-02-19) Mohan, Charanraj; Linares Barranco, Bernabé; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
‘Neuromorphic engineering’ has been showing significant developments in recent days. The word ‘neuromorphic’ was first coined by Caver Mead, which is morphing biological brain on-chip [1]. The main idea is to use the sub-threshold currents of transistors and mimic the biophysical properties that the neurons have. These brain-inspired neuromorphic computing systems have attracted research interest since they are alternate to classical von Neumann [2], computer architectures mainly because of the co-existence of memory and processing units. The renowned neuromorphic chips in the last few decades are Neurogrid [3], Truenorth [4], BrainScaleS [5], and SpiNNaker [6]. Memristors are the fourth fundamental passive-bipolar device, that links charge and flux non-linearly. When Chua coined the word ‘Memristor’ in the late 70s, there was no hint of the existence of the device [7]. Later when the physical existence of the device was shown by HP Labs, it sparked a new wave of enthusiasm among the neuromorphic community [8]. Properties such as non-volatile storage, nano-size existence, non-abrupt switching transition, continuously distributed resistance states, and repeatable behavior convinced the neuromorphic researcher to realize memristors as favorable synaptic elements for neuromorphic systems. In this scenario, the research activities carried out in this doctoral dissertation demonstrates a neuromorphic processing chip for event-driven learning, using memristors as synapses, which are integrated monolithically above the CMOS layers. Although memristors emerged as a potential synapse to solve the density challenge, scalability remains an important bottleneck. Neuromorphic systems should be made more scalable to realize large networks. To contribute to this, we focus on significant challenges in memristor-based neuromorphic hardware. They are- 1) Implementing an on-chip three-stage bulk-based calibration scheme for memristive crossbars and using its low-power inference for recognizing patterns using template matching, programming, and learning. 2) Designing a new current attenuator that is used for efficient crossbar read-outs with a scale-down factor of about 104. The thesis also demonstrates- characterization of three different memristors on various test-benches such as- ArC One Instrument, a full-custom test-PCB, and using probe station with semiconductor parameter analyzer.
A Contribution to Deep Learning based Medical Image Diagnosis Aids
(2020-11-13) Civit Masot, Javier; Vicente Díaz, Saturnino; Domínguez Morales, Manuel Jesús; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
In this work, an in-depth study about the use of Deep Learning techniques to support healthcare professionals for the recognition of pathologies using medical images is carried out. Most of the research presented in this work is focused on the detection of glaucoma using images of the eye fundus; however, in order to demonstrate the feasibility of the processing systems implemented in this work, other types of images are used (in this case, X-ray images) to detect another completely different pathology, such as the detection of patients with COVID-19. Thus, in this work the classic detection techniques for these pathologies are studied, an in-depth study of the techniques based on Deep Learning is carried out, several treatment models are implemented with specific pre-processing stages adapted to the problem itself; and, finally, these systems are tested using large databases in order to demonstrate the feasibility of the those classification systems. The results obtained demonstrate that Deep Learning techniques can be used as a diagnosis aid of those diseases that require medical images analysis. In this way, the human workload required for these tasks is greatly reduced.
Estudio e integración en sistemas empotrados de algoritmos de Aprendizaje Supervisado basados en Redes Neuronales Artificiales para el análisis de perturbaciones y eventos asociados a la marcha
(2020-11-13) Luna Perejón, Francisco; Civit Balcells, Antón; Domínguez Morales, Manuel Jesús; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
En este trabajo se contribuye al avance en el análisis de aspectos asociados a la marcha para la identificación de anomalías en la misma y de eventos de peligro, mediante el uso de algoritmos de Aprendizaje Supervisado basados en Redes Neuronales Artificiales. Se pretende estudiar la viabilidad de integración de estos algoritmos complejos en sistemas empotrados con recursos limitados y que puedan ejecutarse en tiempo real, posibilitando la creación de dispositivos de monitorización portátiles con una autonomía adecuada a su propósito. Se han realizado una serie de estudios que contribuyen en soluciones para dos problemas particulares del ámbito. Por una parte, la precisión y rendimiento de Redes Neuronales Feedforward para el análisis biomecánico de la pisada del usuario. Por otra parte, la efectividad de Redes Neuronales Recurrentes con Compuertas para la detección de actividades de vida diaria, caídas y eventos de riesgo durante la marcha del individuo. Los resultados obtenidos abren la posibilidad de nuevas vías de investigación. Se exponen propuestas de mejora que quedarán indicadas como trabajos futuros, y cuya finalidad es la creación de un sistema integrado de análisis de la marcha.
Neuromorphic deep convolutional neural network learning systems for FPGA in real time
(2019-12-13) Tapiador Morales, Ricardo; Jiménez Fernández, Ángel Francisco; Linares Barranco, Alejandro; Jiménez Moreno, Gabriel; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
Deep Learning algorithms have become one of the best approaches for pattern recognition in several fields, including computer vision, speech recognition, natural language processing, and audio recognition, among others. In image vision, convolutional neural networks stand out, due to their relatively simple supervised training and their efficiency extracting features from a scene. Nowadays, there exist several implementations of convolutional neural networks accelerators that manage to perform these networks in real time. However, the number of operations and power consumption of these implementations can be reduced using a different processing paradigm as neuromorphic engineering. Neuromorphic engineering field studies the behavior of biological and inner systems of the human neural processing with the purpose of design analog, digital or mixed-signal systems to solve problems inspired in how human brain performs complex tasks, replicating the behavior and properties of biological neurons. Neuromorphic engineering tries to give an answer to how our brain is capable to learn and perform complex tasks with high efficiency under the paradigm of spike-based computation. This thesis explores both frame-based and spike-based processing paradigms for the development of hardware architectures for visual pattern recognition based on convolutional neural networks. In this work, two FPGA implementations of convolutional neural networks accelerator architectures for frame-based using OpenCL and SoC technologies are presented. Followed by a novel neuromorphic convolution processor for spike-based processing paradigm, which implements the same behaviour of leaky integrate-and-fire neuron model. Furthermore, it reads the data in rows being able to perform multiple layers in the same chip. Finally, a novel FPGA implementation of Hierarchy of Time Surfaces algorithm and a new memory model for spike-based systems are proposed.
Network traffic characterisation, analysis, modelling and simulation for networked virtual environments
(2019-11-18) Font Calvo, Juan Luis; Sevillano Ramos, José Luis; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
Networked virtual environment (NVE) refers to a distributed software system where a simulation, also known as virtual world, is shared over a data network between several users that can interact with each other and the simulation in real-time. NVE systems are omnipresent in the present globally interconnected world, from entertainment industry, where they are one of the foundations for many video games, to pervasive games that focus on e-learning, e-training or social studies. From this relevance derives the interest in better understanding the nature and internal dynamics of the network tra c that vertebrates these systems, useful in elds such as network infrastructure optimisation or the study of Quality of Service and Quality of Experience related to NVE-based services. The goal of the present work is to deepen into this understanding of NVE network tra c by helping to build network tra c models that accurately describe it and can be used as foundations for tools to assist in some of the research elds enumerated before. First contribution of the present work is a formal characterisation for NVE systems, which provides a tool to determine which systems can be considered as NVE. Based on this characterisation it has been possible to identify numerous systems, such as several video games, that qualify as NVE and have an important associated literature focused on network tra c analysis. The next contribution has been the study of this existing literature from a NVE perspective and the proposal of an analysis pipeline, a structured collection of processes and techniques to de ne microscale network models for NVE tra c. This analysis pipeline has been tested and validated against a study case focused on Open Wonderland (OWL), a framework to build NVE systems of di erent purpose. The analysis pipeline helped to de ned network models from experimental OWL tra c and assessed on their accuracy from a statistical perspective. The last contribution has been the design and implementation of simulation tools based on the above OWL models and the network simulation framework ns-3. The purpose of these simulations was to con rm the validity of the OWL models and the analysis pipeline, as well as providing potential tools to support studies related to NVE network tra c. As a result of this nal contribution, it has been proposed to exploit the parallelisation potential of these simulations through High Throughput Computing techniques and tools, aimed to coordinate massively parallel computing workloads over distributed resources.
Evaluación heurística de la usabilidad en contextos específicos
(2019-09-19) Yáñez Gómez, Rosa; Sevillano Ramos, José Luis; Cascado Caballero, Daniel; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
La atención a la usabilidad de los sistemas de información, como requisito no funcional, es especialmente relevante al tratarse de un factor crítico que determina buena parte del volumen de desarrollos correctivos que van a necesitar los productos en su fase de mantenimiento y el grado de aceptación del usuario hacia dicho sistema, producto o programa. Para evaluar el grado de usabilidad de las interfaces y detectar precozmente el mayor número de problemas se diseñan planes de evaluación que seleccionarán y combinarán distintas herramientas teniendo en cuenta las características de la interfaz, los objetivos prioritarios de la evaluación y la fase de desarrollo del producto o prototipo objeto del análisis; así como cualesquiera otras circunstancias o condicionantes, sea el coste o el tiempo, la formación del personal, etc. La Evaluación Heurística es una herramienta de evaluación de la usabilidad basada en la intervención de expertos. Consiste en la comprobación de los elementos de una lista de heurísticas que, por su naturaleza, no pueden comprobarse automáticamente y requieren del criterio de un experto. Supone un método de bajo coste y buen rendimiento de detección de fallos que se adapta a distintos productos y objetivos, a las características de las distintas fases de desarrollo y, en definitiva, a muy distintos planes de evaluación. En la literatura coexisten distintas propuestas de listas heurísticas de carácter general que, en dominios específicos, deben adaptarse a las características del contexto para maximizar su eficacia. Un ejemplo es la heurística propuesta en esta Tesis, adaptada a la interacción con interfaces móviles. Estas listas adaptadas también se han multiplicado en la literatura atendiendo a distintos aspectos de los planes de evaluación. De este modo el volumen de propuestas a disposición de los evaluadores, para su selección y potencial combinación, resulta bastante inmanejable Tomando un contexto tan específico como los videojuegos —juegos a partir de ahora— , sin embargo, una revisión sistemática de la literatura permite observar que la lista generalista de Nielsen sigue siendo la más usada, incluso cuando resulta insuficiente a la hora de abarcar las características específicas y novedosas de este dominio de aplicación. Partiendo de las hipótesis de que el gran número de propuestas, con distintas características y objetivos, supone un obstáculo a la hora de que los evaluadores puedan seleccionar la más ajustada a su caso y de que incluso una selección ajustada desaprovecharía potenciales ventajas de otras propuestas, se ha desarrollado una herramienta denominada MUSE (Meta-heUristics uSability Evaluation tool) que es capaz de generar una lista heurística personalizada para un plan de evaluación específico tomando como base las distintas propuestas de la bibliografía y automatizando al máximo su selección y filtrado. Una evaluación preliminar de la herramienta aventura resultados prometedores en cuanto a su capacidad de mejorar la detección de vulneraciones de la usabilidad frente al uso de herramientas genéricas en el caso del contexto específico de los juegos y, particularmente, cuando los evaluadores involucrados no son expertos. La presente Tesis comprende, por tanto, tres trabajos principales: • La construcción y análisis de una heurística adaptada a un contexto específico: interfaces móviles. • Una revisión sistemática de la bibliografía que presenta el estado de las evaluaciones de usabilidad de juegos en el ámbito académico.• El desarrollo de la herramienta MUSE usando como base la concepción modular de las heurísticas y su manejo desde una metaheurística capaz de adaptarse, potencialmente, a nuevos contextos específicos y características concretas de los planes de evaluación, generando listas más eficaces.
Neuromorphic audio processing through real-time embedded spiking neural networks.
(2018-12-03) Domínguez Morales, Juan Pedro; Domínguez Morales, Manuel Jesús; Jiménez Fernández, Ángel Francisco; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
In this work novel speech recognition and audio processing systems based on a spiking artificial cochlea and neural networks are proposed and implemented. First, the biological behavior of the animal’s auditory system is analyzed and studied, along with the classical mechanisms of audio signal processing for sound classification, including Deep Learning techniques. Based on these studies, novel audio processing and automatic audio signal recognition systems are proposed, using a bio-inspired auditory sensor as input. A desktop software tool called NAVIS (Neuromorphic Auditory VIsualizer) for post-processing the information obtained from spiking cochleae was implemented, allowing to analyze these data for further research. Next, using a 4-chip SpiNNaker hardware platform and Spiking Neural Networks, a system is proposed for classifying different time-independent audio signals, making use of a Neuromorphic Auditory Sensor and frequency studies obtained with NAVIS. To prove the robustness and analyze the limitations of the system, the input audios were disturbed, simulating extreme noisy environments. Deep Learning mechanisms, particularly Convolutional Neural Networks, are trained and used to differentiate between healthy persons and pathological patients by detecting murmurs from heart recordings after integrating the spike information from the signals using a neuromorphic auditory sensor. Finally, a similar approach is used to train Spiking Convolutional Neural Networks for speech recognition tasks. A novel SCNN architecture for timedependent signals classification is proposed, using a buffered layer that adapts the information from a real-time input domain to a static domain. The system was deployed on a 48-chip SpiNNaker platform. Finally, the performance and efficiency of these systems were evaluated, obtaining conclusions and proposing improvements for future works.
Healthrecsys: Sistema recomendador para la salud
(2017-09-18) Sánchez Bocanegra, Carlos Luis; Sevillano Ramos, José Luis; Fernández Luque, Luis; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
Cuando se habla de información de salud en Internet, inmediatamente se piensa en información procedente de colecciones dirigidas por profesionales y por otras entidades que están en continua colaboración y creación de contenido. Esta es la fuente más ortodoxa y fiable. Pero el comportamiento de los pacientes y sus familiares en relación al cuidado de su enfermedad está en continua transformación desde el nacimiento de Internet. Con una actitud mucho más proactiva, los usuarios suelen realizar búsquedas de información relativa a su salud en Internet, encontrando otras fuentes distintas a las propias de los profesionales sanitarios. Estas fuentes alternativas suponen un enorme volumen de datos que pueden incluir información incorrecta o desfasada por el avance de la investigación científica. Incluso pueden encontrarse contenidos generados con objetivos distintos a la propia salud, como ocurre por ejemplo en comunidades pro-anorexia, antivacuna, etc. Según el caso, el acceso a estas fuentes de información alternativas puede resultar contraproducente o incluso perniciosa. Precisamente uno de los principales problemas que están surgiendo en la comunidad médica es gestionar la información de salud que se genera en Internet. Las nuevas vías de comunicación incorporan un cambio en el modelo asistencial tal y como lo vemos a día de hoy. Este nuevo cambio de visión obliga a los profesionales de la salud a incrementar las observaciones y a convertirse en guías que permitan el acceso a información tratada y validada. La labor de generación de contenidos de salud desde distintas fuentes, así como el control y selección de la información fiable disponible, exige la formalización y validación de los contenidos de salud en Internet. Por tanto, es necesario abordar esta labor desde un punto de vista formal y riguroso. Para ello, un concepto esencial es el de “ontología”, que puede definirse como una especificación de un vocabulario de representación que consta de las definiciones de clases, relaciones, funciones y otros objetos de un determinado dominio. Modelando los contenidos en base a ontologías se puede identificar los contenidos más adecuados para un determinado paciente en base a su enfermedad, sus preferencias o sus características. Esta Tesis se basa pues en los llamados Sistemas de Recomendación en Salud, en particular en los que se basan en el contenido, ítem o artículo con información de salud que busca el paciente o sus familiares. Se estudia la forma en que estos sistemas analizan el contenido que recomiendan, la forma en recuperarlo y tratarlo, teniendo en cuenta la heterogeneidad de fuentes y formatos. En particular, se hace uso de ontologías como diccionarios de salud con información adicional. A partir de técnicas basadas en el procesado del lenguaje natural, podremos interpretar en distintos niveles un documento, y automatizar su comprensión morfológica, sintáctica y semántica. En resumen, el uso de ontologías del área de salud, junto con un procesado del lenguaje natural, permitirá obtener una recomendación personalizada de contenidos de salud, que será evaluada experimentalmente. Dado el uso creciente de vídeos como vehículo de difusión de información sobre salud, esta Tesis se centra en el desarrollo de un sistema recomendador de vídeos en salud. Partiendo de YoutubeTM como plataforma más popular, se recopilan vídeos de salud de distintos Canales y Playlists validados por la comunidad sanitaria, y se estudian los metadatos asociados a cada vídeo. Se extraen los términos de salud de los metadatos del título y descripción del vídeo, y estos términos se filtran a partir de la terminología clínica SNOMED-CT, y a partir de ellos se recomiendan enlaces relacionados de MedlinePlus como fuente fiable de enlaces de salud. Para aumentar el número de recomendaciones, y mejorar su calidad, se plantea el uso de ontologías sanitarias (Bioontology) y técnicas de procesado del lenguaje natural (cTAKES). Posteriormente, estos elementos se ponen a prueba en aplicaciones webs que suministran información de salud válida relacionada con los vídeos visualizados, y finalmente se lleva a cabo una evaluación de la precisión de estas recomendaciones.
Evaluación y análisis de una aproximación a la fusión sensorial neuronal mediante el uso de sensores pulsantes de visión / audio y redes neuronales de convolución
(2017-07-19) Ríos Navarro, José Antonio; Linares Barranco, Alejandro; Jiménez Fernández, Ángel Francisco; Jiménez Moreno, Gabriel; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores
En este trabajo se pretende avanzar en el conocimiento y posibles implementaciones hardware de los mecanismos de Deep Learning, así como el uso de la fusión sensorial de forma eficiente utilizando dichos mecanismos. Para empezar, se realiza un análisis y estudio de los lenguajes de programación paralela actuales, así como de los mecanismos de Deep Learning para la fusión sensorial de visión y audio utilizando sensores neuromórficos para el uso en plataformas de FPGA. A partir de estos estudios, se proponen en primer lugar soluciones implementadas en OpenCL así como en hardware dedicado, descrito en systemverilog, para la aceleración de algoritmos de Deep Learning comenzando con el uso de un sensor de visión como entrada. Se analizan los resultados y se realiza una comparativa entre ellos. A continuación se añade un sensor de audio y se proponen mecanismos estadísticos clásicos, que sin ofrecer capacidad de aprendizaje, permiten integrar la información de ambos sensores, analizando los resultados obtenidos junto con sus limitaciones. Como colofón de este trabajo, para dotar al sistema de la capacidad de aprendizaje, se utilizan mecanismos de Deep Learning, en particular las CNN1, para fusionar la información audiovisual y entrenar el modelo para desarrollar una tarea específica. Al final se evalúa el rendimiento y eficiencia de dichos mecanismos obteniendo conclusiones y unas proposiciones de mejora que se dejarán indicadas para ser implementadas como trabajos futuros.
Categorizadores neuronales en VLSI
(1996) Serrano Gotarredona, María Teresa; Linares Barranco, Bernabé; Universidad de Sevilla. Departamento de Electrónica y Electromagnetismo
Análisis y evaluación del control de un robot móvil: aplicación a sillas de ruedas eléctricas
(1997) Díaz del Río, Fernando; Jiménez Moreno, Gabriel; Civit Balcells, Antón; Universidad de Sevilla. Departamento de Arquitectura y Tecnología de Computadores

Examinar

Envíos recientes