claim
active
claim:gradients-flow-through-every-subcomponent-on-every-datapoint-in-spd-aiding-stability-unlike-apd-top-kGradients flow through every subcomponent on every datapoint in SPD, aiding stability unlike APD top-k
Source of robustness.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Four hypothesized sources.
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 splits W_out into 50 subcomponents forming a single rank-50 component in compressed computation modelfinding0.761Contradicts APD's 100 components for W_out.
- SPD decomposes W_in of compressed computation model into 100 subcomponents, one per input featurefinding0.760Neuron contributions match target.
- Motivation.
- Author interpretation.
- The magnitude of the normalized gradients (choice of αk) plays an important role in performance.claim0.753Insight about gradient normalization scaling.
- App. B.4.