concept
active
concept:loss-landscape-curvature-hessianLoss landscape curvature (Hessian)
Central quantity linking memorization and generalization.
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.
- Sparsely active parameter-space subnetworks reconstructing loss gradients.
- Angular margin technique borrowed from face recognition and applied to the prototype contrast loss in SAE training
- Loss on randomly masked subcomponents matching target outputs, with layerwise auxiliary.
- Minimizes the geometric mean loss.
- LTPBR principle that natural geomorphic processes should replace engineered grading in restoration design.
- The combination of SOHMs and vascular tension that defines the brain's inference process.
- The natural erosion and deposition processes that low-tech designs aim to harness instead of grading (principle 8).
- Addressing disparity in loss magnitudes across tasks at the loss level