Increasing model capacity during training by adding computational loops or depth, rather than using fixed architecture.