Joint Deep Dictionary Learning

Abstract

Inspired by the efficiency of dictionary learning and deep learning, we propose a novelty nonlinear feature representation method, namely Joint Deep Dictionary Learning (JDDL). JDDL embeds an ingenious design of network module which could achieve dictionary learning into common deep learning network structures. Specifically, JDDL learns the dictionary and sparse representation coefficients simultaneously in a low-dimensional latent space built upon deep auto-encoders. Hence, JDDL could produce compact and discriminative representation for the given data and even unseen data by incorporating dictionary learning and deep learning into a unified framework. Extensive experiments are conducted on four real-world data sets clustering to show that the proposed method could provide performance superior to many state of-the-art approaches.