concept
active
concept:feature-interferenceFeature Interference
When non-orthogonal features cause logistic regression to identify a suboptimal probe direction
Neighborhood — ranked by edge-count
Papers (2)
paper
Frameworks (1)
framework
- Core theoretical framework: neural networks represent more features than neurons by encoding features as directions in superposition
Concepts (3)
concept
- Interference between featuresrelated_toCost of non-orthogonal feature embedding.
- Adversarial examplesassociated_withLinked to superposition interference.
- Maximum Margin Separatorassociated_withThe direction logistic regression converges to on linearly separable data; shown to be suboptimal for identifying truth direction
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.
- Frequent-task updates overwrite rare-task features.
- Logit weight contributions from a feature that arise due to superposition with other features, not from the feature's own causal role
- Extra feature representing unexplained error to make isomorphism exact.
- Synthesizing inputs maximizing component activation.
- Phenomenon where sparsity yields gerrymandered latents for hierarchical concepts.
- Interpretable concept representations hypothesized as basic units of computation.
- Asymmetric transfer after fine-tuning: high-density bases (B10) are more robust.
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work