I worked on this paper, submitted to CVPR, while advised by Professor Song-Chun Zhu at the Center for Vision, Cognition, Learning, and Autonomy at UCLA.
The short-run Markov Chain Monte Carlo model can be considered a valid generative model just like a variational autoencoder (VAE) or a generative adversarial network (GAN). We use it to synthesize, interpolate, and reconstruct images. Image synthesis is performed by running Langevin dynamics from a uniform noise distribution, and the interpolation and reconstruction results rival or are qualitatively better than a VAE and GAN.
Please note that more information about this paper will be posted here after the final version is published to Arxiv. In the meantime, the images and content below are taken from a predecessor paper, “On Learning Non-Convergent Non-Persistent Short-Run MCMC Toward Energy-Based Model.”
Abstract: This paper studies a curious phenomenon in learning energy-based model (EBM) using MCMC. In each learning iteration, we generate synthesized examples by running a non-convergent, non-mixing, and non-persistent short-run MCMC toward the current model, always starting from the same initial distribution such as uniform noise distribution, and always running a fixed number of MCMC steps. After generating synthesized examples, we then update the model parameters according to the maximum likelihood learning gradient, as if the synthesized examples are fair samples from the current model. We treat this non-convergent short-run MCMC as a learned generator model or a flow model. We provide arguments for treating the learned non-convergent short-run MCMC as a valid model. We show that the learned short-run MCMC is capable of generating realistic images. More interestingly, unlike traditional EBM or MCMC, the learned short-run MCMC is capable of reconstructing observed images and interpolating between images, like generator or flow models.
A presentation I gave about the predecessor paper is here.