Abstract
Learning generative probabilistic models that can estimate the continuousdensity given a set of samples, and that can sample from that density, is oneof the fundamental challenges in unsupervised machine learning. In this paperwe introduce a new approach to obtain such models based on what we calldenoising density estimators (DDEs). A DDE is a scalar function, parameterizedby a neural network, that is efficiently trained to represent a kernel densityestimator of the data. Leveraging DDEs, our main contribution is to develop anovel approach to obtain generative models that sample from given densities. Weprove that our algorithms to obtain both DDEs and generative models areguaranteed to converge to the correct solutions. Advantages of our approachinclude that we do not require specific network architectures like innormalizing flows, ordinary differential equation solvers as in continuousnormalizing flows, nor do we require adversarial training as in generativeadversarial networks (GANs). Finally, we provide experimental results thatdemonstrate practical applications of our technique.