The Deep Neural Networks (DNN) have become the main contributor in the field of machine learning (ML). Specifically in the computer vision (CV), there are applications like image and video classification, object detection and tracking, instance segmentation and visual question answering, image and video generation are some of the applications from many that DNNs have demonstrated magnificent progress. To achieve the best performance, the DNNs usually require a large number of labeled samples, and finding the optimal solution for such complex models with millions of parameters is a challenging task. It is known that, the data are not uniformly distributed on the sample space, rather they are residing on a low-dimensional manifold embedded in the ambient space. In this dissertation, we specifically investigate the effect of manifold assumption on various applications in computer vision. First we propose a novel loss sensitive adversarial learning (LSAL) paradigm in training GAN framework that is built upon the assumption that natural images are lying on a smooth manifold. It benefits from the geodesic of samples in addition to the distance of samples in the ambient space to differentiate between real and generated samples. It is also shown that the discriminator of a GAN model trained based on LSAL paradigm is also successful in semi-supervised classification of images when the number of labeled images are limited. Then we propose a novel Capsule projection Network (CapProNet) that models the manifold of data through the union of subspace capsules in the last layer of a CNN image classifier. The CapProNet idea has been further extended to the general framework of Subspace Capsule Network that not only does model the deformation of objects but also parts of objects through the hierarchy of sub- space capsules layers. We apply the subspace capsule network on the tasks of (semi-) supervised image classification and also high resolution image generation. Finally, we verify the reliability of DNN models by investigating the intrinsic properties of the models around the manifold of data to detect maliciously trained Trojan models.
| Identifer | oai:union.ndltd.org:ucf.edu/oai:stars.library.ucf.edu:etd2020-1674 |
| Date | 01 January 2021 |
| Creators | Edraki, Marzieh |
| Publisher | STARS |
| Source Sets | University of Central Florida |
| Language | English |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | Electronic Theses and Dissertations, 2020- |
Page generated in 0.0017 seconds