FIG. 1.

scVAE with the semisupervised extension. This is the design for both MOVAE and our proposed system SISUA. The implementation of ${ℒ}_y\left(y,f_{\theta }\left(y\right)\right)$ [Eq. (4)] is a major difference between SISUA and MOVAE. SISUA leverages probabilistic embedding to regulate the amount of information backpropagated from the supervised objectives, which is discussed in the Section 2.3. MOVAE, multioutput variational autoencoder; scVAE, single-cell variational autoencoder; SISUA, SemI-SUpervised generative Autoencoder.