Comparison of Maximum Likelihood and GAN-based training of Real NVPs
This work addresses training evaluation for generative models, but appears incremental as it builds on existing GAN and one-shot learning methods.
The paper compared maximum likelihood and Wasserstein GAN training for a generator, showing that an independent critic helps detect overfitting and proposing a novel fast-learning critic based on one-shot learning ideas.
We train a generator by maximum likelihood and we also train the same generator architecture by Wasserstein GAN. We then compare the generated samples, exact log-probability densities and approximate Wasserstein distances. We show that an independent critic trained to approximate Wasserstein distance between the validation set and the generator distribution helps detect overfitting. Finally, we use ideas from the one-shot learning literature to develop a novel fast learning critic.