Cool to see this upvoted here, though I will say this is fairly deep down the weeds — if you’re new to the topic of analyzing trained neural networks neuron by neuron, I’d suggest starting with our introductory article: https://distill.pub/2020/circuits/zoom-in/
The "Universality" section is especially interesting. They found these high-low frequency neurons in a number of different network architectures, and at around the same network depth in each case. I wonder how this insight could be useful. Can you train a network layer by layer and then connect the pre-trained layers? Like maybe train a "high-low frequency detector" layer on generated inputs of high-low frequency images, and then somehow put it 1/3 of the way into a network, maybe one whose other layers have been trained similarly to recognize other features.
