eh2022-hidden-nonlinearity-essential-for-superposition

Summary

The non-linear activation in a hidden layer is what lets a superposed representation actually perform useful computation; without it, the shared neurons can only do simple linear mixing and the whole mechanism for packing multiple features into the same space collapses. In practical terms, any system that relies on superposition must treat that non-linearity as load-bearing, not optional, or the approach stops working.