finding
active
finding:with-radial-variation-relu-sae-manifolds-plateau-early-near-n-2d-i-latentsWith radial variation, ReLU SAE manifolds plateau early near n≈2d_i latents
Strongest argument against pathology.
Source paper
extracted_from(2025) · Michaud, Eric J. · Gorton, Liv · McGrath, Tom
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (4)
claim
- When β<α, SAEs can optimally tile common feature manifolds at the expense of discovering rarer featuresassociated_withcontradictsCore concern.
- Appendix A.1.
- Synthetic experiments.
- Explains plateau.
Questions (1)
question
- Left open.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Fig 6.
- SAEs do not achieve compact capture; they are in dilution.
- Features act as local patch detectors.
- Higher-dimensional hyperspheres accommodate ≥10^4 latents without plateau; β≈0.05 for dim 6-8finding0.765Fig 5.
- Inception-v1 SAE shows meaningful fraction of latents with very high nearest-neighbor cosinefinding0.748Fig 8; may be duplication.
- Subspace capture on synthetic manifolds peaks near k=4 and degrades at lower and higher sparsityfinding0.747Sparsity sweet spot in synthetic benchmark (R^128, c=512).
- Top SAE features on the years manifold show localized smooth tuning curves with overlapping supportfinding0.747Reminiscent of population coding.
- Evidence in real models.