Learning A Task-Specific Deep Architecture For Clustering

While deep networks show to be highly effective in extensive applications, few efforts have been spent on studying its potential in clustering. In this paper, we argue that the successful domain expertise of sparse coding in clustering is still valuable, and can be combined with the key ingredients of deep learning. A novel feed-forward architecture, named TAG-LISTA, is constructed from graph-regularized sparse coding. It is then trained with task-specific loss functions from end to end. The inner connections of the proposed network to sparse coding leads to more effective training. Moreover, by introducing auxiliary clustering tasks to the hierarchy of intermediate features, we present DTAG-LISTA and obtain a further performance boost. We demonstrate extensive experiments on several benchmark datasets, under a wide variety of settings. The results verify that the proposed model performs significantly outperforms the generic architectures of the same parameter capacity, and also gains remarkable margins over several state-of-the-art methods.