claim
active
claim:a-penalty-intrinsic-to-a-feature-must-be-blind-to-basis-choice-leaving-only-the-l2-norm-the-atomic-unit-is-the-subspaceA penalty intrinsic to a feature must be blind to basis choice, leaving only the l2 norm; the atomic unit is the subspace.
Rotation-invariance argument.
Source paper
extracted_from(2026) · Thomas Fel · Matthew Kowal · Mozes Jacobs · Dron Hazra +21
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Main theoretical claim, Prop. 1.
- ReLU/JumpReLU restrict blocks to a cone and fail to recover superpositions a signed code resolves.extendsJustifies no coordinate-wise nonlinearity.
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.
- Main proposal.
- Quantitative relationship between concept frequency and feature presence.
- Dismissal of earlier criteria as too narrow.
- L1-of-norms is preferable for loss comparability and exposing layer- or model-specific featuresclaim0.758All experiments used L1-of-norms.
- Load-bearing interpretive claim about the layer-specificity of Burger et al.'s finding.
- Ethics must be based on empirically-determined cognitive properties (goals, preferences, concerns) rather than parochial markers.
- can we use the feature basis to detect when fine-tuning a model increases the likelihood of undesirable behaviors?question0.754Question about practical safety application of feature monitoring.
- Supported by the finding that non-trivial rotations are required to find aligned representations.