Deep Convex NMF for Image Clustering

Conventional matrix factorization methods fail to exploit useful information from rather complex data due to their single-layer structure. In this paper, we propose a novel deep convex non-negative matrix factorization method (DCNMF) to improve the ability of feature representation. In addition, the manifold and sparsity regularizers are imposed on each layer to discover the inherent structure of the data. For the formulated multi-layer objective, we develop an efficient iterative optimization algorithm, which can enhance the stability via layer-by-layer factorization and fine-tuning. We evaluate the proposed method by performing clustering experiments on face and handwritten character benchmark datasets; the results show that the proposed method obviously outperforms the conventional single-layer methods, and achieves the state-of-the-art performance.