Year: 2018
Type: dissertation
Abstract: Generative models have been one of the major research fields in unsupervised deep \nlearning during the last years. They are achieving promising results in learning the distribution \nof multidimensional variables as well as in finding meaningful hidden representations \nin data. \nThe aim of this thesis is to gain a sound understanding of generative models through a \nprofound study of one of the most promising and widely used generative models family, \nthe variational autoencoders. In particular, the performance of the standard variational \nautoencoder (known as VAE) and the Gaussian Mixture variational autoencoder (called \nGMVAE) is assessed. First, the mathematical and probabilistic basis of both models is \npresented. Then, the models are implemented in Python using the Tensorflow framework. \nThe source code is freely available and documented in a personal GitHub repository created \nfor this thesis. Later, the performance of the implemented models is appraised in \nterms of generative capabilities and interpretability of the hidden representation of the \ninputs. Two real datasets are used during the experiments, the MNIST and "Frey faces". \nResults show the models implemented work correctly, and they also show the GMVAE \noutweighs the performance of the standard VAE, as expected.
Author: Sánchez Martín, Pablo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7062466804