finding
active
finding:k-fac-eigenvector-bands-show-sorted-activation-strength-top10-10-25-25-50-bottom50-unlike-svdK-FAC eigenvector bands show sorted activation strength top10%>10-25%>25-50%>bottom50%, unlike SVD
Supports curvature basis interpretability.
Source paper
extracted_from(2025) · Merullo, Jack · Vatsavaya, Srihita · Bushnaq, Lucius · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Section 4.
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.
- Vision disentanglement.
- Disentanglement evidence.
- Basis for editing.
- If curvature measures memorization, eigenvector bands will activate differently on memorized vs clean datahypothesis0.768Section 4 hypothesis.
- K-FAC suppresses the broadest range of unseen memorized text without requiring labeled forget dataclaim0.753Practical advantage over BSN.
- Weight components ordered high to low curvature.
- Better transfer to unseen memorized data.
- Regularization effect.