framework
active
framework:theorem-3-features-are-learned-in-order-of-utilityTheorem 3: Features are Learned in Order of Utility
Width-N minimizer spans top-N eigenspace of M.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (2)
concept
- Feature utilityusesFrequency-weighted spectral value πkλkj determining which features are retained.
- Saddle-to-saddle dynamicsanalogous_toDynamics consistent with the static ordering rule.
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.
- Width buys features in order of utility, so only larger models learn infrequent or complex tasksclaim0.805Interpretation of Theorem 3.
- Learning dynamics.
- Second of three speculative claims asserting that subgraphs of neural networks are tractable and meaningful objects of study
- Authors argue features are model properties because logit effects and ablations are consistent with feature interpretations
- Common-task gradient bounded by residual.
- Discrete case.
- Quantitative relationship between concept frequency and feature presence.
- Baseline.