Multi-graph Regularized Deep Auto-Encoders for Multi-view Image Representation.

Deep auto-encoders combined with the manifold construction have attracted much attention as they can preserve local manifolds when the encoding function is learned. This paper proposes a novel framework which is named multi-graph regularized deep auto-encoders (MGDAE). Different from the previous work of graph regularized auto-encoders, our proposed framework incorporates multiple manifolds to well preserve multiple localities in multi-view datasets. With this framework, the low multi-view dimensional features can be obtained due to diverse graph constructions, which vary smoothly along the geodesics of multi-view data manifolds and facilitate the deep auto-encoders. Extensive experimental results on COIL20 and CIFAR-10 for image classification demonstrate our proposed framework outperforms other auto-encoders proposed in deep learning literature, such as AE and LAE.