A technique that combines features extracted at different levels of abstraction from multiple encoders into a unified representation.