Deep Learning for Generative Adversarial Networks and Change Detection

Abstract

The generalization capability of deep neural networks has led to an increase in its utilization for complex tasks across a wide array of applications, ranging from image classification, computer vision, cybersecurity, and healthcare. In this thesis, we look at applications of deep learning based techniques for change detection and generative modeling. The goal of this thesis is two-fold: (a) to provide quantitative measures for evaluating the performance of generative models, and; (b) to develop unsupervised algorithms to detect changes in time series data. Generative Adversarial Networks (GANs) are a popular framework that train two neural networks in an adversarial manner to generate synthetic samples that follow the distribution of input data. While the performance of GANs has been found to be better than other generative models in terms of the quality of the samples, they often suffer from the problem of mode collapse, i.e., synthetic samples tend to lack the diversity present in original data. Many approaches have been proposed to alleviate this phenomenon in GANs. The first contribution of this thesis are quantitative metrics that capture the extent of mode collapse, as well as the sample quality. The second contribution of this thesis is to devise an unsupervised algorithm for change detection. The proposed approach leverages deep learning based methods to estimate likelihood ratios between samples from two distributions. Subsequently, this methodology is used to devise an unsupervised change detection statistic. We also provide generalization of this framework to detect multiple changes, and for the online setting. We validate the performance of our approach using both synthetic and real-world datasets.