Figure 2.

The semi-supervised block, which aims to re-discover samples belonging to the known K classes, is composed of a data encoder model (e_θ), a data decoder model (d_ψ) and a predictor model (p_ϕ). The encoder maps the input data to a latent representation; the decoder reconstructs the initial data; a predictor learns to classify the known samples in the correct labeled classes. After the joint training of e_θ, d_ψ and p_ϕ, a collection of K One-Class Classifiers (OCCs) is trained using the latent embeddings to determine whether each of the unlabeled samples belongs to one of the known classes.