ACOL-GAN: Learning Clustering Generative Adversarial Networks through Graph-Based Activity Regularization

In recent years, deep learning has achieved great success in many fields. As the most basic machine learning task, clustering has also become one of the research hotspots. However, clustering performance based on Variational Autoencoder is generally better than that based on Generative Adversarial Network, which is mainly because the former implements multi-modal learning and there are obvious boundaries between different categories, while the latter does not. In this paper, we propose a new clustering model named ACOL-GAN, which replaces the normal distribution that standard GAN relied on with sampling networks and adopts the Auto-clustering Output Layer as the output layer in discriminator. Due to Graph-based Activity Regularization terms, softmax nodes of parent-classes are specialized as the competition between each other during training. The experimental results show that ACOL-GAN achieved the state-of-the-art performance for clustering tasks on MNIST USPS and Fashion-MNIST, with the highest accuracy on Fashion-MNIST.