Deep Learning of Part-based Representation of Data Using Sparse Autoencoders with Nonnegativity Constraints