Aligning a model's sensitivity structure to match the statistical structure of task-irrelevant variations in data.