Reproduce “Maximum-Likelihood Augmented Discrete Generative Adversarial Networks” on OpenResearch

About this paper

Tong Che, Yanran Li, Ruixiang Zhang, R Devon Hjelm, Wenjie Li, Yangqiu Song, Yoshua Bengio

Researchers from MILA and Hong Kong universities introduce Maximum-Likelihood Augmented Discrete Generative Adversarial Networks (MaliGAN), a framework that stabilizes GAN training for discrete data by formulating the generator's objective as a maximum-likelihood problem augmented by the discriminator's signal. The method achieves improved sample quality and lower perplexity on tasks like binarized MNIST, Chinese poem generation, and sentence-level language modeling compared to prior GAN and MLE approaches.

Key insights

• Reformulating the generator's objective to minimize KL divergence to an augmented target distribution, derived from the discriminator and a delayed generator, provides a more stable training signal.
• Normalizing the discriminator's relative score (D(x)/(1-D(x))) within a mini-batch effectively reduces variance and provides useful gradients even when samples are initially poor.
• Combining maximum likelihood training for initial tokens with adversarial training for subsequent tokens, alongside Monte Carlo Tree Search for granular rewards and single-sample renormalization, significantly stabilizes discrete GAN training.

Summary by alphaXiv.