concept
active
concept:interference-weightsInterference Weights
Logit weight contributions from a feature that arise due to superposition with other features, not from the feature's own causal role
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Discussion.
Methods (2)
method
- Fisher effectivenessaboutSecond-order KL estimate of a weight's effect on outputs using Fisher metric.
- HelpfulnessaboutAverage change in loss when a weight is ablated.
Concepts (1)
concept
- Weight superpositionassociated_withVirtual weights between features forced into superposition by low-dimensional projections.
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.
- When non-orthogonal features cause logistic regression to identify a suboptimal probe direction
- Frequent-task updates overwrite rare-task features.
- Identifying interference weights is necessary but not sufficient for reading global circuitsclaim0.765Discussion.
- Novelty claim.
- Cost of non-orthogonal feature embedding.
- Interpretation of motivating example.
- Baseline MTL approach minimizing sum of task losses with equal weights; suffers from task balancing
- Concluding statement.