Selective Sampling and Mixture Models in Generative Adversarial Networks

TL;DR

This paper introduces a multi-generator GAN framework that models mixture distributions, enabling selective sampling from individual components, with demonstrated feasibility on simple models and the MNIST dataset.

Contribution

It presents a novel multi-generator extension to GANs that captures mixture components and allows selective sampling, enhancing interpretability and diversity.

Findings

Feasibility shown analytically and experimentally.

Generators learn distinct mixture components.

Selective sampling from individual components is possible.

Abstract

In this paper, we propose a multi-generator extension to the adversarial training framework, in which the objective of each generator is to represent a unique component of a target mixture distribution. In the training phase, the generators cooperate to represent, as a mixture, the target distribution while maintaining distinct manifolds. As opposed to traditional generative models, inference from a particular generator after training resembles selective sampling from a unique component in the target distribution. We demonstrate the feasibility of the proposed architecture both analytically and with basic Multi-Layer Perceptron (MLP) models trained on the MNIST dataset.