Probing the Purview of Neural Networks via Gradient Analysis