Learning Deep Generative Models with Doubly Stochastic MCMC
This addresses efficient inference for deep generative models, which is important for practitioners in machine learning, though it appears incremental as an extension of existing MCMC techniques.
The paper tackles approximate Bayesian inference for deep generative models by introducing doubly stochastic gradient MCMC, which uses mini-batch data and a neural adaptive importance sampler. The method outperforms state-of-the-art competitors in tasks like density estimation, data generation, and missing data imputation.
We present doubly stochastic gradient MCMC, a simple and generic method for (approximate) Bayesian inference of deep generative models (DGMs) in a collapsed continuous parameter space. At each MCMC sampling step, the algorithm randomly draws a mini-batch of data samples to estimate the gradient of log-posterior and further estimates the intractable expectation over hidden variables via a neural adaptive importance sampler, where the proposal distribution is parameterized by a deep neural network and learnt jointly. We demonstrate the effectiveness on learning various DGMs in a wide range of tasks, including density estimation, data generation and missing data imputation. Our method outperforms many state-of-the-art competitors.