A model architecture where multiple tasks share the same learned representations but have task-specific output layers.