RemixNet: Generative Adversarial Networks for Mixing Multiple Inputs

Generative Adversarial Networks (GANs) are a class of Artificial Neural Network architectures for generating realistic data distributions from a small number of latent variables. The latent variables are typically drawn IID from a Normal distribution and fed as input. Empirically, it has been shown that GANs are able to produce realistic images from as few as 100 latent variables. While numerous studies have focused on modifying the original GAN architecture, few have explored the structure of the input latent variables. In this study we propose placing structure on the input for aiding GANs to converge faster, lead insight into the structure of the underlying data distribution, and produce potentially more creative outputs. To achieve a GAN with such constraints, we call our proposed architecture RemixNet. In RemixNet the latent variables are a result of multiple draws from our existing dataset and "remixed" into a new distribution. This idea of combining existing data into new data is inspired from the artistic and musical practice of collage making and remixes. For example, an artist may take an existing recording of music, modify each with a filter and combine to form a new composition. We show several experiments on the MNIST dataset to gain some intuition on various RemixNet architectures. We also compare empirically RemixNet with GAN and show that the RemixNet architecture converges significantly faster to realistic images.