Feature Learning for Nonlinear Dimensionality Reduction toward Maximal Extraction of Hidden Patterns