activation-function-affects-polysemanticity
IN premise — summaries/2026/08/24/elhage-2022-toy-models-superposition-chunk-13.md
Created 2026-08-25T02:57:59+00:00
Kshitij Sachan at Redwood Research independently replicated the Elhage et al. 2022 toy model experiments and found that the choice of activation function (e.g., ReLU vs. alternatives) materially affects the degree of polysemanticity and shifts the phase boundary.
Summary
This observation means that polysemanticity, the tendency for a single neuron to carry multiple concepts at once, is not a fixed and unavoidable property of neural networks. It is partially shaped by a simple architectural choice like the activation function, which gives practitioners a concrete lever to control how tangled representations become and warns that results from one activation function do not automatically transfer to another.