Towards Visual Explanations for Convolutional Neural Networks via Input Resampling