New training technique would reveal the basis for machine-learning systems’ decisions.

System designed to label visual scenes according to type turns out to detect particular objects, too.