It's just a function of multi-targeting. When problem can be framed in a way that one of the outputs helps explain the other output I would call that component

It's just a function of multi-targeting. When problem can be framed in a way that one of the outputs helps explain the other output I would call that component to be an "explanation component". For one of my problems I know that the model would need to 'see' an orientation field in order to make its predictions, so instead of just asking for the predictions I also ask for an orientation field. The loss function has only a small weighting and is set against a very rough ground truth along with a total variation loss. This ends up producing a nicer orientation field than the ground truth.