finding
active
finding:grassmannian-bsf-with-b-1-describes-dinov3-activations-in-fewer-bits-than-the-b-1-sae-at-every-width-and-sparsity-minimum-at-b-between-2-and-4Grassmannian BSF with b>1 describes DINOv3 activations in fewer bits than the b=1 SAE at every width and sparsity, minimum at b between 2 and 4.
MDL result at 20% distortion.
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
- Support-term savings.
Datasets (1)
dataset
- DINOv3aboutMain vision model whose activations are decomposed.
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.
- Fig. 10.
- Curve manifold recovery.
- Block and Grassmannian featurizers keep utilization ≈0.6–0.66 at k=16, while Group Lasso falls to ≈0.17.finding0.777Per-feature occupancy.
- On DINOv3, classification and segmentation tolerate R²≈0.8, while depth needs R²≈0.9 to retain 95% accuracy.finding0.775Task-dependent distortion floors.
- Featurizer-specific optima.
- Concept dimensionality.
- Interpretation of stable rank.
- Fig. 5.