Improving Generative Adversarial Networks with Bayesian techniques

E N, Supriya and P K, Srijith (2019) Improving Generative Adversarial Networks with Bayesian techniques. Masters thesis, Indian institute of technology Hyderabad.

Text
Mtech_Thesis_TD1512_2019.pdf
Restricted to Repository staff only until 15 July 2020.
Download (891kB) | Request a copy

Abstract

Generative adversarial networks(GAN) are popular Deep learning models that can implicitly learn rich distribution over complex high-dimensional data like image, speech etc. Recent success of GANs has enforced the need for better training methods, owing to their versatile applications. Well known challenges in training GANs include non-convergence, unstable gradients, mode collapse and sensitivity to hyperparameters. While generative models are used to generate data in small data scenarios, deep neural network based models are data-hungry, tend to overfit and model selection is a cumbersome process. In this work, we introduce probabilistic techniques to train GANs using recent advances in Bayesian techniques. As Bayesian models are known to perform well on small data and have implicit model selection we use these techniques to propose new training methods for GAN in order to alleviate the mode collapse and non convergence issues. Experiments are conducted using MNIST dataset and generation process is studied.

[error in script]

IITH Creators: