claim
active
claim:when-saes-can-optimally-tile-common-feature-manifolds-at-the-expense-of-discovering-rarer-featuresWhen β<α, SAEs can optimally tile common feature manifolds at the expense of discovering rarer features
Core concern.
Source paper
extracted_from(2025) · Michaud, Eric J. · Gorton, Liv · McGrath, Tom
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (6)
finding
- With radial variation, ReLU SAE manifolds plateau early near n≈2d_i latentsassociated_withcontradictsStrongest argument against pathology.
- Simulated SAE with 100M latents discovers only ~3M features when β≪αassociated_withsupportsFig 2 right.
- Higher-dimensional hyperspheres accommodate ≥10^4 latents without plateau; β≈0.05 for dim 6-8associated_withFig 5.
- Fig 6.
- Core result.
- Fig 1.
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.
- Theoretical framework answer.
- Justifies focusing on direction-based SAEs.
- Features act as local patch detectors.
- Core critique of sparse autoencoders: they break the geometric structure of representations, making it harder to see the big picture.
- Unsure.
- Future-work perspective.
- Top SAE features on the years manifold show localized smooth tuning curves with overlapping supportfinding0.788Reminiscent of population coding.
- Extension of mechanistic interpretability findings to the metacognitive domain