A NOVEL INFORMATION-THEORETIC EXPLANATION OF GENERATIVE ADVERSARIAL NETWORKS

Chen, Haochen

Permalink

https://hdl.handle.net/2142/114984 Copy

Description

Title

A NOVEL INFORMATION-THEORETIC EXPLANATION OF GENERATIVE ADVERSARIAL NETWORKS

Author(s)

Chen, Haochen

Issue Date

2022-05

Keyword(s)

Generative adversarial networks; channel capacity; unsupervised learning

Abstract

Generative Adversarial Network (GAN) is widely used for instance generation in multiple fields. It models a zero-sum game between a generator network and a discriminator network while using theJensen-Shannon divergence between the target distribution and the generated one as the optimization goal. We propose an alternative explanation of such a framework using a binary-input channel.We also show that our proposed goal is equivalent to the original Jensen-Shannon divergence whileenforcing a tighter performance bound. We then introduce an additional trainable parameter πwith only few modification on the vanilla implementation of GAN. We further experiment with ourproposed variant of GAN on MNIST and CIFAR-10, producing ∼ 10 less on the Fr´echet inceptiondistance and ∼ 0.5 more on the inception score.

Type of Resource

text

Language

en

Handle URL

https://hdl.handle.net/2142/114984

Owning Collections