Dissertation/Thesis Abstract

Unsupervised Generative Capsule Neural Networks
by Serowik, Justin J., Master's, Stevens Institute of Technology, 2018, 29; 13420362
Abstract (Summary)

Capsule Neural Networks (CapsNets) have recently emerged as a new neural network architecture aimed at improving on the weaknesses of Convolutional Neural Networks (CNNs). CapsNets use “capsules” to learn advanced structures of features in objects and possess equivariance. They have already shown that they can perform better on classification tasks than state-of-the-art CNNs. Utilizing these strengths, we can apply CapsNets to other tasks other than classification such as image generation. Image generation is already possible from the masked output of high level capsules in a supervised model, but in this paper I propose a method which uses CapsNets in a Generative Adversarial Network (GAN) architecture to generate images in an unsupervised learning setting. These Unsupervised Generative CapsNets (UGCNs) outperform similarly sized Fully-Connected Dense layers in terms of image reconstruction quality and Fréchet Inception Distance on the MNIST and CIFAR-10 data sets.

Indexing (document details)
Advisor: Man, Hong
Commitee: Wang, Rensheng
School: Stevens Institute of Technology
Department: Computer Engineering
School Location: United States -- New Jersey
Source: MAI 58/04M(E), Masters Abstracts International
Subjects: Computer Engineering, Artificial intelligence, Computer science
Keywords: Capsule neural networks, Deep learning, Generative adversarial networks, Image generation, Unsupervised capsule networks, Unsupervised learning
Publication Number: 13420362
ISBN: 978-0-438-83728-7
Copyright © 2021 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy