I imagined it like a Fourier transform. Distilling individual features from a combined signal shared between multiple neurons. The reason you need to do this is the superposition hypothesis: that multiple neurons are encoding more features than just the orthogonal vectors.