Master's Thesis
Generative Deep Clustering with the Hierarchical and Relativistic Wasserstein Autoencoder GAN
— 2021
Key information
Authors:
Supervisors:
Published in
December 2, 2021
Abstract
Clustering is one of the most important tasks in the unsupervised learning field, and traditional methods (like k-means) struggle when dealing with high-dimensional data. For this reason, many researchers have successfully attempted to use deep learning techniques to map the data into a lower dimensional, cluster friendly space. The process usually consists of using an autoencoder-based structure to find a meaningful latent space and then using traditional clustering algorithms on it. In this dissertation, I developed a novel approach to clustering by using a generative hierarchical autoencoder containing multiple latent groups (the vanilla autoencoder has just one) that are regularized by latent relativistic discriminators in an adversarial manner. These latent groups are then clustered by the density-based clustering algorithm HDBSCAN, which automatically estimates the number of clusters. The resulting model is named Hierarchical and Relativistic Wasserstein Autoencoder GAN (HRWAEGAN) and this work will show that utilizing multiple latent spaces does not offer any advantages when it comes to the task of clustering image datasets.
Publication details
Authors in the community:
Gustavo Augusto Toscano Morais
ist197152
Supervisors of this institution:
Pedro Filipe Zeferino Tomás
ist146645
Fields of Science and Technology (FOS)
electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering
Publication language (ISO code)
eng - English
Rights type:
Embargo lifted
Date available:
October 14, 2022
Institution name
Instituto Superior Técnico