Estudi de les xarxes neuronals convolucionals profundes mitjançant Caffe

Abstract

The deep learning techniques and the use of GPUs have made neural networks the leading option for solving some computational problems and it has been shown to produce the state-of-the-art results in many fields like computer vision, automatic speech recognition, natural language processing, and audio recognition. This final grade dissertation is divided into three parts: first, we will discuss theoretical concepts that allow us to understand how neural networks work. Second, we will focus in deep convolutional neural network to understand and learn how to build such networks that Caffe Framework use. Finally, the third part will be a compilation of all learned skills to make that a neural network will classify correctly the data set MNIST, and then we will change Caffe Framework files so it can read images with more channels and we will watch the results obtained. At last, we will match and improve the public state-of-the-art classification system of the data set Food-101. After all the work, our goals will be achieved: we will modify Caffe Framework and we will check that in the case of the MNIST we will improves the classification rate. But above all, the result that we want to emphasize is the to release of classification system for the Food-101 that will improve the accuracy from 56.40% to 74.6834% and finally we will propose ideas for improving this classification in the future.