question
active
question:how-could-ripples-in-feature-manifolds-affect-sae-scalinghow could ripples in feature manifolds affect SAE scaling?
Unsure.
Source paper
extracted_from(2025) · Michaud, Eric J. · Gorton, Liv · McGrath, Tom
Neighborhood — ranked by edge-count
Papers (1)
paper
- Understanding sparse autoencoder scaling in the presence of feature manifoldsassociated_withmentions
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.
- Core critique of sparse autoencoders: they break the geometric structure of representations, making it harder to see the big picture.
- Core concern.
- Special tokens have very large magnitude, so their SAE features also have large magnitude and sit far from other pointshypothesis0.777First explanation for isolated features.
- Theoretical framework answer.
- Top SAE features on the years manifold show localized smooth tuning curves with overlapping supportfinding0.775Reminiscent of population coding.
- Features act as local patch detectors.
- Related to point-based simplicial limit.
- Link to prior work.