Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders

Faheem Moolla; Terence L. van Zyl; Hairong Wang

doi:10.1007/978-3-031-49002-6_20

Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders

Faheem Moolla
, Terence L. van Zyl
, Hairong Wang

Institute for Intelligent Systems

University of the Witwatersrand

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Semi-supervised learning is a deep learning paradigm that has shown significant value for general machine learning and generative modelling. To date, Generative Adversarial Networks (GANs) still suffer from challenges related to mode collapse and other sources of instability. Further, little research has been done to investigate how incorporating semi-supervised learning (using SGAN) and pre-training (using VAE) into GAN training might alleviate some of these challenges. To this end, this study proposes SSGAN, a combination of VAE and SGAN, to tackle some of these challenges. Our extensive qualitative and quantitative analysis shows that the proposed approach significantly improves the stability of GAN training and the quality of generated images. Further, the results indicate that this can be done with relatively few additional labelled examples. In conclusion, continued research and exploring foundation models and other semi- and self-supervised learning mechanisms will likely lead to further improvements.

Original language	English
Title of host publication	Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings
Editors	Anban Pillay, Edgar Jembere, Aurona J. Gerber
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	300-314
Number of pages	15
ISBN (Print)	9783031490019
DOIs	https://doi.org/10.1007/978-3-031-49002-6_20
Publication status	Published - 2023
Event	4th Southern African Conference for Artificial Intelligence Research, SACAIR 2023 - Muldersdrift, South Africa Duration: 4 Dec 2023 → 8 Dec 2023

Publication series

Name	Communications in Computer and Information Science
Volume	1976 CCIS
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	4th Southern African Conference for Artificial Intelligence Research, SACAIR 2023
Country/Territory	South Africa
City	Muldersdrift
Period	4/12/23 → 8/12/23

Keywords

GAN
Semi-supervised GAN
VAE

ASJC Scopus subject areas

General Computer Science
General Mathematics

Access to Document

10.1007/978-3-031-49002-6_20

Cite this

Moolla, F., van Zyl, T. L., & Wang, H. (2023). Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders. In A. Pillay, E. Jembere, & A. J. Gerber (Eds.), Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings (pp. 300-314). (Communications in Computer and Information Science; Vol. 1976 CCIS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-49002-6_20

Moolla, Faheem ; van Zyl, Terence L. ; Wang, Hairong. / Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders. Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings. editor / Anban Pillay ; Edgar Jembere ; Aurona J. Gerber. Springer Science and Business Media Deutschland GmbH, 2023. pp. 300-314 (Communications in Computer and Information Science).

@inproceedings{8d29739df58749b7b9915dd2ac146a4d,

title = "Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders",

abstract = "Semi-supervised learning is a deep learning paradigm that has shown significant value for general machine learning and generative modelling. To date, Generative Adversarial Networks (GANs) still suffer from challenges related to mode collapse and other sources of instability. Further, little research has been done to investigate how incorporating semi-supervised learning (using SGAN) and pre-training (using VAE) into GAN training might alleviate some of these challenges. To this end, this study proposes SSGAN, a combination of VAE and SGAN, to tackle some of these challenges. Our extensive qualitative and quantitative analysis shows that the proposed approach significantly improves the stability of GAN training and the quality of generated images. Further, the results indicate that this can be done with relatively few additional labelled examples. In conclusion, continued research and exploring foundation models and other semi- and self-supervised learning mechanisms will likely lead to further improvements.",

keywords = "GAN, Semi-supervised GAN, VAE",

author = "Faheem Moolla and \{van Zyl\}, \{Terence L.\} and Hairong Wang",

note = "Publisher Copyright: {\textcopyright} 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.; 4th Southern African Conference for Artificial Intelligence Research, SACAIR 2023 ; Conference date: 04-12-2023 Through 08-12-2023",

year = "2023",

doi = "10.1007/978-3-031-49002-6\_20",

language = "English",

isbn = "9783031490019",

series = "Communications in Computer and Information Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "300--314",

editor = "Anban Pillay and Edgar Jembere and \{J. Gerber\}, Aurona",

booktitle = "Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings",

address = "Germany",

}

Moolla, F, van Zyl, TL & Wang, H 2023, Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders. in A Pillay, E Jembere & A J. Gerber (eds), Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings. Communications in Computer and Information Science, vol. 1976 CCIS, Springer Science and Business Media Deutschland GmbH, pp. 300-314, 4th Southern African Conference for Artificial Intelligence Research, SACAIR 2023, Muldersdrift, South Africa, 4/12/23. https://doi.org/10.1007/978-3-031-49002-6_20

Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders. / Moolla, Faheem; van Zyl, Terence L.; Wang, Hairong.
Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings. ed. / Anban Pillay; Edgar Jembere; Aurona J. Gerber. Springer Science and Business Media Deutschland GmbH, 2023. p. 300-314 (Communications in Computer and Information Science; Vol. 1976 CCIS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders

AU - Moolla, Faheem

AU - van Zyl, Terence L.

AU - Wang, Hairong

PY - 2023

Y1 - 2023

N2 - Semi-supervised learning is a deep learning paradigm that has shown significant value for general machine learning and generative modelling. To date, Generative Adversarial Networks (GANs) still suffer from challenges related to mode collapse and other sources of instability. Further, little research has been done to investigate how incorporating semi-supervised learning (using SGAN) and pre-training (using VAE) into GAN training might alleviate some of these challenges. To this end, this study proposes SSGAN, a combination of VAE and SGAN, to tackle some of these challenges. Our extensive qualitative and quantitative analysis shows that the proposed approach significantly improves the stability of GAN training and the quality of generated images. Further, the results indicate that this can be done with relatively few additional labelled examples. In conclusion, continued research and exploring foundation models and other semi- and self-supervised learning mechanisms will likely lead to further improvements.

AB - Semi-supervised learning is a deep learning paradigm that has shown significant value for general machine learning and generative modelling. To date, Generative Adversarial Networks (GANs) still suffer from challenges related to mode collapse and other sources of instability. Further, little research has been done to investigate how incorporating semi-supervised learning (using SGAN) and pre-training (using VAE) into GAN training might alleviate some of these challenges. To this end, this study proposes SSGAN, a combination of VAE and SGAN, to tackle some of these challenges. Our extensive qualitative and quantitative analysis shows that the proposed approach significantly improves the stability of GAN training and the quality of generated images. Further, the results indicate that this can be done with relatively few additional labelled examples. In conclusion, continued research and exploring foundation models and other semi- and self-supervised learning mechanisms will likely lead to further improvements.

KW - GAN

KW - Semi-supervised GAN

KW - VAE

UR - https://www.scopus.com/pages/publications/85180529297

U2 - 10.1007/978-3-031-49002-6_20

DO - 10.1007/978-3-031-49002-6_20

M3 - Conference contribution

AN - SCOPUS:85180529297

SN - 9783031490019

T3 - Communications in Computer and Information Science

SP - 300

EP - 314

BT - Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings

A2 - Pillay, Anban

A2 - Jembere, Edgar

A2 - J. Gerber, Aurona

PB - Springer Science and Business Media Deutschland GmbH

T2 - 4th Southern African Conference for Artificial Intelligence Research, SACAIR 2023

Y2 - 4 December 2023 through 8 December 2023

ER -

Moolla F, van Zyl TL, Wang H. Improving Semi-supervised Learning in Generative Adversarial Networks Using Variational AutoEncoders. In Pillay A, Jembere E, J. Gerber A, editors, Artificial Intelligence Research - 4th Southern African Conference, SACAIR 2023, Proceedings. Springer Science and Business Media Deutschland GmbH. 2023. p. 300-314. (Communications in Computer and Information Science). doi: 10.1007/978-3-031-49002-6_20