claim
active
claim:the-spd-decomposition-of-w-out-is-more-accurate-than-apd-s-whose-error-was-due-to-a-poorly-chosen-top-kThe SPD decomposition of W_out is more accurate than APD's, whose error was due to a poorly chosen top-k
Author interpretation.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Contradicts APD's 100 components for W_out.
- Neuron contributions of SPD subcomponents closely match target model across all 100 featuressupportsFigure 7.
- SPD decomposes W_in of compressed computation model into 100 subcomponents, one per input featuresupportsNeuron contributions match target.
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.
- Central claim.
- SPD finds 100 W_in components and a single rank-50 W_out component spanning both layers in the 2-layer modelfinding0.779Multi-layer mechanisms.
- Robustness.
- SPD finds 102 W_in components and a rank-51 W_out component spanning all three layers in the 3-layer modelfinding0.759APD struggled beyond two layers.
- Source of robustness.
- SPD avoids shrinkage.
- Ground truth mechanisms identified.
- Core assessment of SDL.