PROJECT TITLE :

Segmented Generative Networks Data Generation in the Uniform Probability Space

ABSTRACT:

Recent developments in the field of generative networks have demonstrated that it is possible for deep neural networks to generate data that is analogous to that of the real world. In order to circumvent the unsolvability of the posterior distribution, a few implicit probabilistic models have been developed. These models use a stochastic process to directly generate data. The ability to model data does, however, require an in-depth knowledge and understanding of the statistical dependence of the data, which can be preserved and studied in appropriate latent spaces. In this article, we present a segmented generation process that is accomplished through linear and nonlinear manipulations in a same-dimensional latent space, which is the destination to which data are projected. We develop a segmented approach for the generation of dependent data by utilizing the concept of copula. This approach was inspired by the well-known stochastic method that is used to generate correlated data. The generation process is broken up into two separate frames: the first one embeds the information about the covariance or the copula in the uniform probability space, and the second one embeds the information about the marginal distribution in the sample domain. An empirical method to sample directly from implicit copulas is also provided by the proposed network structure, which is known as a segmented generative network (SGN). We evaluate the presented method in three different application scenarios in order to demonstrate its generalizability. These scenarios include a toy example, handwritten digits, and face image generation.