Online learning and generalization of parts-based image representations by Non-Negative Sparse Autoencoders