ReGAN: RE[LAX|BAR|INFORCE] based Sequence Generation using GANs

TL;DR

This paper introduces ReGAN, a sequence generation method using GANs with advanced gradient estimators, achieving improved stability and quality in generating longer sequences compared to prior GAN-based approaches.

Contribution

It demonstrates the effectiveness of using efficient policy gradient estimators in GANs for sequence generation and provides a comparative analysis of recent unbiased gradient estimation techniques.

Findings

ReGAN achieves improved training stability.

Sequence quality improves with advanced gradient estimators.

The method performs well on synthetic datasets with varying sequence lengths.

Abstract

Generative Adversarial Networks (GANs) have seen steep ascension to the peak of ML research zeitgeist in recent years. Mostly catalyzed by its success in the domain of image generation, the technique has seen wide range of adoption in a variety of other problem domains. Although GANs have had a lot of success in producing more realistic images than other approaches, they have only seen limited use for text sequences. Generation of longer sequences compounds this problem. Most recently, SeqGAN (Yu et al., 2017) has shown improvements in adversarial evaluation and results with human evaluation compared to a MLE based trained baseline. The main contributions of this paper are three-fold: 1. We show results for sequence generation using a GAN architecture with efficient policy gradient estimators, 2. We attain improved training stability, and 3. We perform a comparative study of recent unbiased low variance gradient estimation techniques such as REBAR (Tucker et al., 2017), RELAX (Grathwohl et al., 2018) and REINFORCE (Williams, 1992). Using a simple grammar on synthetic datasets with varying length, we indicate the quality of sequences generated by the model.