Publicación: Neuroevolution of Bayesian Neural Networks
Cargando...
Fecha
2020-06-01
Autores
Editor/a
Director/a
Carmona Suárez, Enrique Javier
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingeniería Informática. Departamento de Inteligencia Artificial.
Resumen
Las Redes Neuronales Bayesianas han atraído mucha atención últimamente por parte de la comunidad investigadora porque muestran resultados prometedores en la gestión de la incertidumbre en modelos predictivos. Esta gestión es especialmente relevante en aplicaciones donde un falso positivo o negativo pueda tener un gran impacto como en el caso de los vehículos autónomos o en la diagnosis médica. Sin embargo, este tipo de redes todavía no alcanzan el nivel de las métricas de las Redes Neuronales tradicionales. Alguna de las causas es la complejidad añadida al entrenar redes que tienen parámetros extra para modelar la incertidumbre. En esta Tesina, diseñamos, implementamos y validamos un nuevo algoritmo llamado NEAT Bayesiano para evolucionar tanto la arquitectura como los pesos en Redes Neuronales Bayesianas "de propogación hacia delante". El algoritmo está basado en un algoritmo evolutivo previo llamado NEAT (por sus siglas en Inglés, Neuroevolución de topologías en constante aumento) que ha sido aplicado con éxito a Redes Neuronales más tradicionales. Las contribuciones principales de este trabajo son: (1) el diseño y validación del NEAT Bayesiano como una aproximación para resolver problemas de Aprendizaje Supervisado en problemas de media dimensionalidad al mismo tiempo que se proporciona una estimación de la incertidumbre para cada ejemplo; (2) la validación del NEAT Bayesiano como un algoritmo altamente robusto contra conjuntos de datos con muchos ejemplos mal etiquetados; y (3) el diseño e implementación del primer algoritmo basado en NEAT para redes "de propogación hacia delante" en una librería especializada de Aprendizaje Profundo.
Bayesian Neural Networks have been getting a lot of attention from the research community because they show promise on better handling uncertainty. This is specially relevant in applications where a false positive or negative can have a big impact such as in self-driving cars or medical diagnosis. However, they do not typically achieve the performance metrics of traditional Neural Networks as they still have many problems. Some of these problems are the issues that arise with the extra complexity of training a network that has extra parameters to model uncertainty. In this Master Thesis, we design, implement and validate a new algorithm called Bayesian-NEAT to both evolve architecture and parameters for feed-forward Bayesian Neural Networks. The algorithm is based on a previous evolutionary algorithm called NEAT (Neuroevolution of Augmenting Topologies) that is successfully applied to standard Neural Networks. The main contributions of this work are: (1) design and validation of Bayesian-NEAT as an approach for solving Supervised Learning problems of medium dimensionality and providing uncertainty estimates for individual predictions; (2) validate Bayesian-NEAT as highly robust algorithm against mislabeled data; and (3) design and implement first NEAT-based implementation of a feed-forward network in a Deep Learning framework.
Bayesian Neural Networks have been getting a lot of attention from the research community because they show promise on better handling uncertainty. This is specially relevant in applications where a false positive or negative can have a big impact such as in self-driving cars or medical diagnosis. However, they do not typically achieve the performance metrics of traditional Neural Networks as they still have many problems. Some of these problems are the issues that arise with the extra complexity of training a network that has extra parameters to model uncertainty. In this Master Thesis, we design, implement and validate a new algorithm called Bayesian-NEAT to both evolve architecture and parameters for feed-forward Bayesian Neural Networks. The algorithm is based on a previous evolutionary algorithm called NEAT (Neuroevolution of Augmenting Topologies) that is successfully applied to standard Neural Networks. The main contributions of this work are: (1) design and validation of Bayesian-NEAT as an approach for solving Supervised Learning problems of medium dimensionality and providing uncertainty estimates for individual predictions; (2) validate Bayesian-NEAT as highly robust algorithm against mislabeled data; and (3) design and implement first NEAT-based implementation of a feed-forward network in a Deep Learning framework.
Descripción
Categorías UNESCO
Palabras clave
búsqueda de Arquitecturas de Neuronas, Redes Neuronales Bayesianas, aprendizaje supervisado, NEAT, Neural Architecture search, Bayesian Neural Networks, supervised learning
Citación
Centro
Facultades y escuelas::E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial