finding
active
finding:spd-uses-20-subcomponents-on-tms5-2-but-only-5-learn-columns-others-negligible-normSPD uses 20 subcomponents on TMS5-2 but only 5 learn columns; others negligible norm
Figure 1.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Ground truth mechanisms identified.
- SPD splits W_out into 50 subcomponents forming a single rank-50 component in compressed computation modelfinding0.799Contradicts APD's 100 components for W_out.
- SPD finds 100 W_in components and a single rank-50 W_out component spanning both layers in the 2-layer modelfinding0.787Multi-layer mechanisms.
- SPD decomposes W_in of compressed computation model into 100 subcomponents, one per input featurefinding0.777Neuron contributions match target.
- SPD finds 102 W_in components and a rank-51 W_out component spanning all three layers in the 3-layer modelfinding0.775APD struggled beyond two layers.
- Ground-truth columns recovered with negligible shrinkage.
- Recovery with hidden identity.
- Neuron contributions of SPD subcomponents closely match target model across all 100 featuresfinding0.749Figure 7.