A Deep Generative Approach to Conditional Sampling

We propose a deep generative approach to sampling from conditional distribution based on unified formulation of and generalized nonparametric regression function using the noise-outsourcing lemma. The proposed aims at learning generator, so that random sample target can be obtained by transforming drawn reference distribution. generator is estimated nonparametrically with neural networks matching appropriate joint distributions Kullback-Liebler divergence. An appealing aspect our method it allows either or both predictor response high-dimensional handle continuous discrete type predictors responses. show consistent in sense converges underlying under mild conditions. Our numerical experiments simulated benchmark image data validate demonstrate outperforms several existing density estimation methods. Supplementary materials for this article are available online.