Extending the Adversarial Loss Function of Soft-Intro VAE for Stronger Disentaglement

Information

• Publication Type: Student Project
• Workgroup(s)/Project(s):
  • Visualization Group
• Date: 2023
• Date (Start): March 2022
• Date (End): March 2023
• Matrikelnummer: 01529399
• First Supervisor:
  • Hubert Ramsauer
  • Manuela Waldner

Abstract

Disentanglement is hard to achieve in unsupervised representation learning. It can be negatively affected by trying to improve the reconstruction quality of the generated output. To try to alleviate these problems, this project combines two approaches that improve disentanglement and reconstruction quality, specifically β-TCVAE[1] and Soft-Intro-VAE[2]. The hypothesis was that a model that uses a combined loss function of both approaches can retain the positive aspects of both. The results did not confirm the hypothesis and showed no improvement in disentanglement metrics and worse reconstruction results compared to Soft-Intro-VAE[2]. The code for this project is available on GitHub.

Additional Files and Images

Weblinks

• GitHub

BibTeX

@studentproject{matt-2023-vae,
  title =      "Extending the Adversarial Loss Function of Soft-Intro VAE
               for Stronger Disentaglement",
  author =     "Matthias Matt",
  year =       "2023",
  abstract =   "Disentanglement is hard to achieve in unsupervised
               representation learning. It can be negatively affected by
               trying to improve the reconstruction quality of the
               generated output. To try to alleviate these problems, this
               project combines two approaches that improve disentanglement
               and reconstruction quality, specifically β-TCVAE[1] and
               Soft-Intro-VAE[2]. The hypothesis was that a model that uses
               a combined loss function of both approaches can retain the
               positive aspects of both. The results did not confirm the
               hypothesis and showed no improvement in disentanglement
               metrics and worse reconstruction results compared to
               Soft-Intro-VAE[2]. The code for this project is available on
               GitHub.",
  month =      feb,
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2023/matt-2023-vae/",
}