I believe it is the case that deep NNs at the highest/deepest levels have interpretable feature-specific cells as well. Not sure what they think the difference is.