Want to understand neuron-level sparsity?

[TO BE WRITTEN]

• subject of a lot of mechinterp, rather less mathematical theory
• studies of ICA (as opposed to PCA)
• L1-norm over L2-norm
• some of Arthur’s work (deep nets induce some kind of L1-ish sparsity)
  • esp from small init!
• “condensation phenomenon”
• some of Dan Kunin’s recent work (AGF)
• Q: can you study neuron densities and avoid worrying about neuron sparsity?
• Q: do inf-width muP nets learn sparsity? (probably, right? so does that mean Anthropic’s superposition hyp is wrong?)