Describir: A Manifold Learning Perspective on Representation Learning: Learning Decoder and Representations without an Encoder