Gan-based Enhanced Deep Subspace Clustering Networks

In recent years,subspace clustering has been investigated in image content compress and representation learning,in image processing,image/video segmentation in computer vision and gene expression profile clustering,in bio-informatics.Compared to traditional clustering analysis,subspace clustering,by searching low-dimensional subspaces to explore structures in data,not only reduces the time complexity and hardware requirements,but also helps to alleviate negative effects of high-dimensional noise,thereby obtaining better performances.However,it robustness is now restricted by the size of input,and the lack of original data distribution learning brings the inconsistency from high-dimensional to latent space.In this paper,we propose two GAN-based enhanced deep subspace clustering approaches: Deep Subspace Clustering via Dual Adversarial Generative Networks(DSCDAG),and SelfSupervised Deep Subspace Clustering with Adversarial Generative Networks(S^2DSC-AG).In DSC-DAG,distributions of both the inputs and corresponding latent representations are learnt via adversarial training simultaneously.Besides,there are two kinds of synthetic representations to facilitate fine-tuning of the encoder: combinations of latent representations with random combination coeffificients,and representations of real-like inputs derived from noise variables.In S^2DSC-AG,a self-supervised information learning module substitutes for the adversarial learning in latent space,since both modules play the same role in learning discriminative latent representations.We analyze connections between these methods and demonstrate their equivalences.We conduct extensive experiments on multiple real-world data sets against state-of-the-art subspace clustering methods in terms of the accuracy,normalized mutual information and purity,where the robustness and effectiveness are showed.Visualization results demonstrate convergence and interpretability of our proposed methods.