Supplementary materials for: Synthesizing the preferred inputs for neurons in neural networks via deep generator networks

While the visualizations of many neurons appear to be great-looking, showing canonical images of a class (e.g. a table lamp with shade and base, Fig. 1); many others do not (e.g. dog visualizations often show blobs of texture instead of a dog standing with 4 legs, Fig. S18). We investigate whether this is a shortcoming of our method or these non-canonical visualizations are actually reflecting some property of the neurons.