Generative Adversarial Networks

Please enable javascript in your browser.

Memarzadeh, Amir Reza | 2021

731 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 54341 (02)
University: Sharif University of Technology
Department: Mathematical Sciences
Advisor(s): Haji Mirsadeghi, Mir Omid
Abstract:
In this thesis we try to understand one of the most important subfield of deep learning, the generative adversarial networks. In this framework the goal is to reach a generator that generates samples from a target distribution. The target distribution is usually su- per high dimensional and we only have sample access to it. primarily , this distribution was used to be for set of Images (e.g. images of celebrity faces) and GANs performed well in this setting. In this framework two models work simultaneously: a generator tries to generate realistic samples from the target distribution and a discriminator or critic tries to distinguish real samples from generated (fake) samples or more precisely the critic tries to tell the distance between the real and generated samples. One of the known ways for the critic model is to use a relatively new notion of distance between distributions called maximum mean discrepancy(MMD). It is based on kernel methods and RKHS theory. In our study we tried to identify models of this type and show its usefulness. At the end some results are demonstrated
Keywords:
Generative Adversarial Networks ; Maximum Mean Discrepency ; Kernel Methods ; Reproducing Kernel Hilbert Space ; Kernel Learning