Thesis ESTRATEGIAS DIVIDE Y VENCERÁS PARA ENTRENAMIENTO DE REDES NEURONALES EN PRESENCIA DE MÚLTIPLES CLASES
Date
2019-10
Authors
MOLINA BARRA, GABRIEL ANDRES
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Actualmente las redes neuronales artificiales se han convertido en una de las metodologías más utilizadas para abordar problemas de aprendizaje en inteligencia artificial, lo que impulsa a crear modelos y arquitecturas cada vez más complejas para resolver distintos problemas. En esta memoria proponemos un nuevo método para el entrenamiento de redes neuronales en problemas de clasificación con múltiples categorías. En estos problemas, una red implementa una función de decisión que se utiliza para determinar la clase a la que pertenece un determinado dato de entrada (imagen, texto, etc). Cuando se tienen muchas categorías posibles, la metodología convencional consiste en aprender una distribución de probabilidad sobre todas las clases simultáneamente, usando una función objetivo que penaliza errores en todas las clases. Pero la efectividad de dichas redes no es simplemente observar el promedio del accuracy o la función de pérdida, si no que más bien se tiene que observar los resultados obtenidos por cada clase, ya que no se puede tener un accuracy alto en un par de clases y uno muy bajo en el resto de ellas. La idea que exploraremos en esta memoria consiste en descomponer este problema de aprendizaje en varias partes cuyas soluciones, combinadas apropiadamente, den una mejor solución a algunos problemas donde se ha superado incluso la habilidad humana para reconocer imágenes, como por ejemplo la clasificación de imágenes en el dataset CIFAR10 donde se ha alcanzado un accuracy del 90.82 %.
Currently artificial neural networks have become one of the most used methodologies to address learning problems in artificial intelligence, which drives to create increasingly complex models and architectures to solve different problems. In this report we propose a new method for the training of neural networks in classification problems with multiple categories. In these problems, a network implements a decision function that is used to determine the class to which a certain input data belongs (image, text, etc). When there are many possible categories, the conventional methodology consists of learning a probability distribution over all classes simultaneously, using an objective function that penalizes errors in all classes. But the result of the effectiveness of such networks is not simply to observe the average accuracy or loss function, but rather to observe the results obtained by each class, since you can not have a high accuracy in a pair of classes and one very low in the rest of them. The idea that we will explore in this memory is to break down this learning problem into several parts whose solutions, properly combined, give a better solution to some problems where even the human ability to recognize images has been overcome, such as the classification of images in the CIFAR10 dataset where an accuracy of 90.82 % has been reached.
Currently artificial neural networks have become one of the most used methodologies to address learning problems in artificial intelligence, which drives to create increasingly complex models and architectures to solve different problems. In this report we propose a new method for the training of neural networks in classification problems with multiple categories. In these problems, a network implements a decision function that is used to determine the class to which a certain input data belongs (image, text, etc). When there are many possible categories, the conventional methodology consists of learning a probability distribution over all classes simultaneously, using an objective function that penalizes errors in all classes. But the result of the effectiveness of such networks is not simply to observe the average accuracy or loss function, but rather to observe the results obtained by each class, since you can not have a high accuracy in a pair of classes and one very low in the rest of them. The idea that we will explore in this memory is to break down this learning problem into several parts whose solutions, properly combined, give a better solution to some problems where even the human ability to recognize images has been overcome, such as the classification of images in the CIFAR10 dataset where an accuracy of 90.82 % has been reached.
Description
Keywords
REDES NEURALES (Ciencia de la Computación) , INTELIGENCIA ARTIFICIAL , ALGORITMOS COMPUTACIONALES