Variational autoencoders with triplet loss for representation learning

Learning low dimensional meaningful representations of data is an important task for classification, visualization and compression. Using autoencoders for representation learning is a successful application of deep learning. Recently, variational autoencoders have also been developed. These are more advantageous than autoencoders since these are generative and have a compact form in the latent space. In order to improve the clustering performance of variational autoencoders in the latent space, the use of variational autoencoders with triplet loss is proposed in this study.