Why it matters

Polysemanticity motivates SAE. Understanding shapes interp challenges.

Advertisement

The architecture

Neuron activates for multiple concepts.

Hard to name / interpret.

Related to superposition.

PolysemanticityNeuron activationsacross inputsMultiple conceptsco-fireHinders interpmeaningful features hiddenMotivates SAE + dictionary learning to decompose into monosemantic features
Polysemanticity.
Advertisement

How it works end to end

Elhage et al. superposition explanation.

Compression under limited dim.

SAE aims to decompose.