representations are embedding-aware); and mapping features into latent space using a multilayer perceptron. g, Using the continuity over the manifold, the network is trained using unsupervised ...