Rethinking interpretation: Input-agnostic saliency mapping of deep visual classifiers