framework
active
framework:block-sparse-featurizers-bsfBlock-Sparse Featurizers (BSF)
Featurizers whose atoms are blocks of directions rather than single directions.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Block norm for presence; PCA coordinates for intrinsic geometry.
Frameworks (4)
framework
- Grassmannian BSFextendsTied encoder/decoder with orthonormal charts and learned scalar gamma.
- Group Lasso BSFextendsLinear encoder with block soft-threshold and group lasso penalty.
- Vanilla BSFextendsFree encoder/decoder with block top-k projection.
- Structured sparsityimplementsClassical literature on sparsity priors over admissible supports.
Artifacts (1)
artifact
- Released code.
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.
- Discussion of cortical analogy.
- Block and Grassmannian featurizers keep utilization ≈0.6–0.66 at k=16, while Group Lasso falls to ≈0.17.finding0.755Per-feature occupancy.
- BSF codes improve on the basic probe for classification and approach it for segmentation and depth.finding0.753Linear probing.
- If BSFs capture concepts, moving within a block should manipulate behavior according to concept geometry.hypothesis0.752Motivates diffusion steering.
- BSF codes surpass the SAE in best single-concept F1 on ImageNet-1k at every sparsity, gap widening with b.finding0.747Class selectivity.
- Foundational empirical result enabling all downstream analysis
- MDL result at 20% distortion.
- The primary novel framework introduced in the paper for learning facet-level personality control vectors