Date of Award
Campus Access Thesis
Master of Science (MS)
Disentanglement is an active topic for generative modeling in the deep learning community. Three major research questions have been studied: 1) How to improve the performance of disentanglement with a limited number of labels in a dataset. 2) How latent codes are distributed in a geometric perspective. 3) Whether there exist clusters in the distribution of the latent code. To resolve these problems, this thesis focuses on: 1) Developing a plug-in generative model orientation beta-VAE which combines multiple generative models. Without decreasing the performance of the original beta-VAE, orientation beta-VAE learns the orientation factors with explicit instructions. We show the learning process can be done with a small sample size, and it has learned multi factors. 2) Showing the distribution of latent codes exists in clusters in geometry perspective, and it provides a rich property for analyzing the performance of disentanglement. 3) Traditional cluster algorithms were used to identify some clusters and assign labels for each cluster.
Li, Donghao, "Study of Disentangle Latent Representation from The Geometry Perspective" (2021). Graduate Masters Theses. 690.