Stacked capsule autoencoder based generative adversarial network

Abstract

Convolutional neural network based generative adversarial networks have become the dominant generative model in the field of generative deep learning. But limitations of convolutional neural networks affect generative adversarial networks also, since most of the current generative adversarial networks are based on convolutional neural networks. The main limitation of convolutional neural networks is that they are invariant. In other words, convolutional neural networks can’t preserve spatial information of features in an image. In contrast, capsule networks gained attention in recent years due to their equivariant architecture which preserves spatial information.

Stacked capsule autoencoder is a type of capsule networks that is able to overcome the limitations that convolutional neural networks suffer from. Stacked capsule autoencoder is an equivariant model which preserves spatial, relational, geometrical information between parts and objects in an image. So in this research we implemented a generative adversarial network which uses stacked capsule autoencoder as the discriminator of it, by replacing the conventional convolutional neural network discriminator.

Then we evaluated the implementation of stacked capsule autoencoder based generative adversarial network using MNIST images. As the qualitative evaluation we observed the visual quality of generated images. Quality and diversity of the generated images are acceptable. Then we evaluated our model quantitatively using inception score for MNIST. Findings of this research show that, the stacked capsule autoencoder can be used as the discriminator of a generative adversarial network instead a convolutional neural network and its performances are plausible.