finding
active
finding:a-single-grassmannian-bsf-block-covers-all-curve-orientations-of-inceptionv1-as-one-connected-manifold-whereas-neurons-and-sae-atoms-tile-itA single Grassmannian BSF block covers all curve orientations of InceptionV1 as one connected manifold, whereas neurons and SAE atoms tile it.
Curve manifold recovery.
Source paper
extracted_from(2026) · Thomas Fel · Matthew Kowal · Mozes Jacobs · Dron Hazra +21
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Multiplicity of SAE curve features is an artifact of the featurizer, not a property of the network.supportsReinterpretation of curve detectors.
Hypotheses (1)
hypothesis
- A block-sparse code should hold the curve manifold together where a directional code fragments it.associated_withPrediction tested on InceptionV1.
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.
- MDL result at 20% distortion.
- InceptionV1 neuron 4e:55 responds to cat faces, fronts of cars, and cat legs as unrelated stimulifinding0.778Concrete example of polysemantic neuron demonstrating the challenge to the circuits agenda
- Rationale for manifold steering.
- Fig. 10.
- Recontextualizes Olah et al. 2020.
- Highlights the non-genetic control of large-scale anatomy.
- Demonstrates that meaningful algorithms can be read directly off floating-point weights in a neural network
- Central claim motivating DAS over prior methods.