Redes neuronales convolucionales

Abstract

Artificial neural networks are a class of machine learning algorithms involved in many of the most spectacular applications in Artificial Intelligence. Deep neural networks or multilayer networks produce the best empirical results in the classification of images or texts. From a theoretical point of view, understanding the reasons for the success of these algorithms is still a pending issue. In addition to the convergence problems of the learning algorithms, the enormous overparameterization of many types of neural networks makes, perhaps, very likely that the classification rules obtained following this method suffer from overfitting. The goal of this project is to study the design of neural networks adapted to image analysis, without abundant parametrization, but oriented to take advantage of the special structure of this type of data. The gain of this approach in terms of control of the overfit will be studied and applied to the classification of some appropriate image dataset.