This paper proposed a joint framework for manifold learning, which simultaneously learns the low-dimensional embeddings of input data and a mapping function for out-of-sample extrapolation. This framework can be potentially applied to any existing manifold learning algorithm.