Generative Models for Data Synthesis - Archive ouverte HAL
Rapport (Rapport Contrat/Projet) Année : 2022

Generative Models for Data Synthesis

Kunwar Saaim
  • Fonction : Auteur
  • PersonId : 1209380
Supreeth Srinath
  • Fonction : Auteur
  • PersonId : 1209381
Shasha Fu
  • Fonction : Auteur
  • PersonId : 1209382

Résumé

Finding large quantities of high-quality data to train neural networks is one of the most challenging aspects for researchers since privacy restrictions and associated financial obligations make it difficult to gather the data. In recent years, Generative Adversarial Networks (GANs) have been extensively used in the generation of different types of datasets. Despite this, we cannot control the attributes that will be associated with a data sample generated by GANs. Combining GANs with Variational Autoencoders (VAEs) is an effective way of obtaining outputs which have the desired attributes. Our study examined the sampling capacity and quality of VAE and VAE-GAN for MNIST and Fashion MNIST datasets. Additionally, we introduce a new metric, Transformer Score (TS), to determine the quality of the generated data. It is based on a vision transformers network and demonstrates superiority over Inception Score.
Fichier principal
Vignette du fichier
Final_Report_MM811.pdf (1.88 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03911560 , version 1 (23-12-2022)

Identifiants

  • HAL Id : hal-03911560 , version 1

Citer

Kunwar Saaim, Supreeth Srinath, Shasha Fu. Generative Models for Data Synthesis. University of Alberta,Canada. 2022. ⟨hal-03911560⟩

Collections

LARA
89 Consultations
179 Téléchargements

Partager

More