Weights in the final layer of common visual models appear as horizontal bands. We investigate how and why.