SeGMA: Semi-Supervised Gaussian Mixture Autoencoder

We propose a semi-supervised generative model, SeGMA, which learns joint probability distribution of data and their classes is implemented in typical Wasserstein autoencoder framework. choose mixture Gaussians as target latent space, provides natural splitting into clusters. To connect Gaussian components with correct classes, we use small amount labeled classifier induced by the distribution. SeGMA optimized efficiently due to Cramer-Wold distance maximum mean discrepancy penalty, yields closed-form expression for spherical and, thus, obviates need sampling. While preserves all properties its predecessors achieves at least good performance on standard benchmark sets, it presents additional features: 1) interpolation between any pair points space produces realistically looking samples; 2) combining property disentangling class style information, able perform continuous transfer from one another; 3) possible change intensity characteristics point moving representation away specific components.