Please use this identifier to cite or link to this item: https://hdl.handle.net/10356/147096
Title: Adaptive generative adversarial network (GAN) for small datasets
Authors: Liu, Chang
Keywords: Engineering::Computer science and engineering::Data
Engineering::Computer science and engineering::Software::Programming techniques
Issue Date: 2021
Publisher: Nanyang Technological University
Source: Liu, C. (2021). Adaptive generative adversarial network (GAN) for small datasets. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/147096
Abstract: This paper starts from the basic mathematic knowledge in the deep learning network, then introduces some helpful and important infrastructure networks, it will also show programming tools and pods to help us build up our network. After that, this paper introduces the basic principle of GAN and analyzes the relevant classical GAN model. For small data sets, the method of using complex distribution such as Gaussian mixture models is proposed to enhance the simple sampling noise and a new model is created on this base which is called DeLiGAN. The main idea of the DeLiGAN is to increase the modeling capability of the prior distribution rather than increasing the depth of the model and reparametrize potential space into a mixed Gaussian model. By comparing the training results of different GAN models, it is concluded that DeLiGAN does perform better on the small data sets. This model provides a training model under a small data set, which can help us better train the neural network and improve training efficiency
URI: https://hdl.handle.net/10356/147096
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:EEE Theses

Files in This Item:
File Description SizeFormat 
Final Dissertation_G1901185A.pdf
  Restricted Access
1.6 MBAdobe PDFView/Open

Page view(s)

237
Updated on Feb 5, 2023

Download(s)

13
Updated on Feb 5, 2023

Google ScholarTM

Check

Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.