Cycles Improve Conditional Generators: Synthesis and Augmentation for Data Mining

Abstract

Conditional Generative Adversarial Networks (CGANs) are diversely utilized for data synthesis in applied sciences and natural image tasks. Conditional generative models extend upon data generation to account for labeled data by estimating joint distributions of samples and labels. We present a family of modified CGANs which demonstrate the inclusion of reconstructive cycles between prior and data spaces inspired by BiGAN and CycleGAN improves upon baselines for natural image synthesis with three primary contributions. The first is a study proposing three incremental architectures for conditional data generation which demonstrate improvement on baseline generation quality for a natural image data set across multiple generative metrics. The second is a novel approach to structure latent representations by learning a paired structured condition space and weakly structured variation space with desirable sampling and supervised learning properties. The third is a proposed utilization of conditional image synthesis for supervised learner data set augmentation as an alternative generation metric. Additional experiments demonstrate the successes of inducing cycles in conditional GANs for both image synthesis and image classification over comparable models with no additional tweaks or modifications. We release our source code, models, and experiments here: https://github.com/alexander-moore/Cycles-Improve-Conditional-Generators.