- Keywords: Langevin dynamics, amortized inference, latent variable model, deep generative model
- Abstract: Markov chain Monte Carlo (MCMC), such as Langevin dynamics, is valid for approximating intractable distributions. However, its usage is limited in the context of deep latent variable models since it is not scalable to data size owing to its datapoint-wise iterations and slow convergence. This paper proposes the amortized Langevin dynamics (ALD), wherein datapoint-wise MCMC iterations are entirely replaced with updates of an inference model that maps observations into latent variables. Since it no longer depends on datapoint-wise iterations, ALD enables scalable inference from large-scale datasets. Despite its efficiency, it retains the excellent property of MCMC; we prove that ALD has the target posterior as a stationary distribution with a mild assumption. Furthermore, ALD can be extended to sampling from an unconditional distribution such as an energy-based model, enabling more flexible generative modeling by applying it to the prior distribution of the latent variable. Based on ALD, we construct a new deep latent variable model named the Langevin autoencoder (LAE). LAE uses ALD for autoencoder-like posterior inference and sampling from the latent space EBM. Using toy datasets, we empirically validate that ALD can properly obtain samples from target distributions in both conditional and unconditional cases, and ALD converges significantly faster than traditional LD. We also evaluate LAE on the image generation task using three datasets (SVHN, CIFAR-10, and CelebA-HQ). Not only can LAE be trained faster than non-amortized MCMC methods, but LAE can also generate better samples in terms of the Fréchet Inception Distance (FID) compared to AVI-based methods, such as the variational autoencoder.
- One-sentence Summary: We introduce amortization method of Langevin dynamics for learning deep latent variable models.