concept
active
concept:shrinkage-of-learned-parametersShrinkage of learned parameters
APD showed ML2R≈0.9; SPD avoids it.
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.
- Systematic underestimation of non-zero feature activations due to L1 sparsity penalty.
- Authors argue features are model properties because logit effects and ablations are consistent with feature interpretations
- Inference of parameters encoding contingencies of the world (e.g., likelihood matrix A) at slower timescale than perception.
- Motivating question.
- Distillation of why learning generalises.
- Loss a smaller model never reaches even asymptotically.
- Loss gains a smaller model recovers with infinite data.
- Scaling laws for dictionary learning are unknown and needed to assess feasibility on frontier models