Manifold Regularized Stacked Autoencoder for Feature Learning

Stacked auto encoders enjoy their popularization with the prosperity of deep learning in recent years. However, relative studies seldom exploit the intrinsic information buried in the interrelations between the samples with respect to deep networks. Regarding this, the manifold regularization is introduced to analyze the neighborhood of each training sample, which leads to a manifold regularized stacked auto encoder hierarchical framework with deep multilayer substructures. A series of experiments are conducted upon MNIST and Yale Faces using locally linear embedding as the manifold regularization module. The results show that neighborhood analysis should be combined with stacked auto encoders to achieve some notable promotions of their performances.