Clustering is one of the research hotspots of deep learning. Recently, deep generative models provide a new way to achieve clustering. However, from all the proposed models, the performance of the clustering variants based on the Variational Autoencoder(VAE) are superior than that based on the Generative Adversarial Network (GAN), which is mainly because the former allows the data to be multi-mode in latent space but the latter does not do so, making the boundaries of different classes obscure and difficult to distinguish. In this paper, we propose a new GAN-based clustering model named Auto-clustering Output Layer Generative Adversarial Network(ACOL-GAN), which replaces the normal distribution that standard GAN relied on with Gaussian mixture distribution generated by sampling networks and adopts the Auto-clustering Output Layer(ACOL) as the output layer in discriminator. Due to Graph-based Activity Regularization(GAR) terms, softmax nodes of parent-classes are specialized as the competition between each other occurs during training. The experimental results show that ACOL-GAN is superior to most unsupervised clustering algorithms on MNIST, USPS and Fashion-MNIST datasets.
The ACOL-GAN algorithm was compared to the most advanced clustering methods. As shown in Table 1, the ACOL-GAN clustering algorithm is better than or equal to other algorithms in most cases, and the performance on the Fashion-MNIST dataset is the best in the recorded algorithm clustering results.
The latent representation of the MNIST data set is visualized by t-SNE[5]. ACOL-GAN can quickly capture the data distribution of the MNIST dataset in the latent space and make the actual data to fit the specified Gaussian mixture for fast clustering. And only 5 epochs are needed to achieve different boundaries in the latent space.
The number of clusters required for the ACOL-GAN model is set 7 and 14 on the MNIST and Fashion-MNIST datasets. In the results of the cluster number set to 7 categories, similar pictures such as the combination of 0 and 6 in MNIST dataset and the combination of the sandals and the sneakers in Fashion-MNIST dataset are grouped together. The result that the number of clusters is set to 14 shows that even in the same class, some differences such as the tilt angle and the shape in the same category can be divided into a plurality of classes. Thus, by observing the experimental results, it can be found that ACOL-GAN can effectively accomplish the task even if the number of clusters is not the number of categories in the real world.