Folded Hamiltonian Monte Carlo for Bayesian Generative Adversarial Networks

Narges Pourshahrokhi, Samaneh Kouchaki, Yunpeng Li, Payam M. Barnaghi

Abstract: Generative Adversarial Networks (GANs) can learn complex distributions over images, audio, and data that are difficult to model. We deploy a Bayesian formulation for unsupervised and semi-supervised GAN learning. We propose Folded Hamiltonian Monte Carlo (F-HMC) within this framework to marginalise the weights of the generators and discriminators. The resulting approach improves the performance by having suitable entropy in generated candidates for generator and discriminators' weights. Our proposed model efficiently approximates the high dimensional data due to its parallel composition, increases the accuracy of generated samples and generates interpretable and diverse candidate samples. We have presented the analytical formulation as well as the mathematical proof of the F-HMC. The performance of our model in terms of autocorrelation of generated samples on converging to a high dimensional multi-modal dataset exhibits the effectiveness of the proposed solution. Experimental results on high-dimensional synthetic multi-modal data and natural image benchmarks, including CIFAR-10, SVHN and ImageNet, show that F-HMC outperforms the state-of-the-art methods in terms of test error rates, runtimes per epoch, inception score and Frechet Inception Distance scores.