As I understand, this is only about this side experiment with ImageNet data which uses logistic regression on those neurons in some cryptic way; I was trying to comprehend the core work (faces) before that.