Representation of PELICAN architecture, which is composed of three modules: the autoencoder, the contrastive, and the classification modules. The first module optimizes the autoencoder loss containing a sparsity parameter (see Eq. (10)). In the second module, the contrastive loss (see Eq. (11)) is optimized to bring the features with the same label together. Finally the third module performs the classification step optimizing a standard classification loss.