Iterative Deep Subspace Clustering

Lei Zhou, Shuai Wang, Xiao Bai, Jun Zhou, Edwin R. Hancock

Abstract: Recently, deep learning has been widely used for subspace clustering problem due to the excellent feature extraction ability of deep neural network. Most of the existing methods are built upon the auto-encoder networks. In this paper, we propose an iterative framework for unsupervised deep subspace clustering. In our method, we first cluster the given data to update the subspace ids, and then update the representation parameters of a Convolutional Neural Network (CNN) with the clustering result. By iterating the two steps, we can obtain not only a good representation for the given data, but also more precise subspace clustering result. Experiments on both synthetic and real-world data show that our method outperforms the state-of-the-art on subspace clustering accuracy.