A study of traits that affect learnability in GANs

TL;DR

This paper investigates how different traits of datasets influence the learnability of GANs through empirical experiments on synthetic data, aiming to understand and improve training success.

Contribution

It provides empirical insights into the traits affecting GAN learnability, using parameterized synthetic datasets to analyze factors influencing training outcomes.

Findings

Certain dataset traits significantly impact GAN training success

Empirical relationships between dataset properties and learnability are identified

Guidelines for designing datasets to improve GAN training are suggested

Abstract

Generative Adversarial Networks GANs are algorithmic architectures that use two neural networks, pitting one against the opposite so as to come up with new, synthetic instances of data that can pass for real data. Training a GAN is a challenging problem which requires us to apply advanced techniques like hyperparameter tuning, architecture engineering etc. Many different losses, regularization and normalization schemes, network architectures have been proposed to solve this challenging problem for different types of datasets. It becomes necessary to understand the experimental observations and deduce a simple theory for it. In this paper, we perform empirical experiments using parameterized synthetic datasets to probe what traits affect learnability.