Some theoretical properties of GANs (Q2196234): Difference between revisions

The Generative Adversarial Networks, GANs, are generative algorithms implemented by a system of two neural networks competing to each other in a zero-sum game framework. GANs have been introduced by \textit{I. Goodfellow} et al., [``Generative adversarial nets'', Adv. Neural Inf. Proc. Sys. 27, 2672--2680 (2014)]. The aim of the present paper is to present a theoretical investigation of the original GANs algorithm. In the second section some optimality properties of the generative process in terms of Jensen-Shannon divergence are analyzed. The third section is devoted to approximation properties and one shows that the Jensen-Shannon divergence is a suitable criterion. Statistical properties are extensively discussed in the fourth section. Large sample properties of the parameter describing the empirically selected generator are studied. Conclusions and further research are presented in the fifth section and the proofs of results are contained in the section six.