SOGN: novel generative model using SOM.

Generative models such as variational autoencoder (VAE) and generative adversarial network (GAN) have been widely applied to many areas including image synthesis and voice generation. However, they have some problems that VAE makes blur images and GAN is difficult to learn due to mode collapsing. A novel generative model is proposed using a self-organising map (SOM) termed a selforganising generative network (SOGN). In the SOGN, training images are first mapped to SOM and then the output space of SOM is transformed into 2D vector spaces. These vector values are used as latent vectors to train the generative network such as artificial neural networks or convolutional neural networks. Experimental results with MNIST and CIFAR-10 datasets showed that their generative model was easy to train without mode collapsing and made more clean images than VAE. It was also confirmed that the manifold is well-observed without generating by the average effect of multiple images. [ABSTRACT FROM AUTHOR]