Image Representation Learning Using Graph Regularized Auto-Encoders

It is an important task to learn a representation for images which has low dimension and preserve the valuable information in original space. At the perspective of manifold, this is conduct by using a series of local invariant mapping. Inspired by the recent successes of deep architectures, we propose a local invariant deep nonlinear mapping algorithm, called graph regularized auto-encoder (GAE). The local invariant is achieved using a graph regularizer, which preserves the local Euclidean property from original space to the representation space, while the deep nonlinear mapping is based on an unsupervised trained deep auto-encoder. This provides an alternative option to current deep representation learning techniques with its competitive performance compared to these methods, as well as existing local invariant methods.