Dropout Induced Noise for Co-Creative GAN Systems

TL;DR

This paper demonstrates how Dropout can be used in Generative Adversarial Networks to generate multiple different outputs to one input, as an alternative to latent space exploration.

Abstract

This paper demonstrates how Dropout can be used in Generative Adversarial Networks to generate multiple different outputs to one input. This method is thought as an alternative to latent space exploration, especially if constraints in the input should be preserved, like in A-to-B translation tasks.

Figures (5)

• Figure 1: Model trained with a Dropout rate of 0.8. Images generated with Dropout rates ranging from 0 to 0.8. Generation was repeated three times to show variety in output. Especially with higher Dropout rates, generated images differ a lot.
• Figure 2: Model trained with no Dropout. Images generated with Dropout rates ranging from 0 to 0.8. Generation was repeated three times to show variety in output. Especially with higher Dropout rates, generated images look distorted or broken.
• Figure 3: Model trained with a Dropout rate of 0.8. Images generated with Dropout rates ranging from 0 to 0.8 and no scaling ($p_{scale}=0$). Generation was repeated three times to show variety in output.
• Figure 4: Model trained with a Dropout rate of 0.8. Images generated with Dropout rates ranging from 0 to 0.8 and no scaling. The first row shows a model with Dropout applied to all hidden layers in generation, the bottom row has Dropout only applied to the first hidden layer.
• Figure 5: Model trained on facade label masks. Left images shows input, right images show variety of output if Dropout rate of 0.5 is used for generation.