concept
active
concept:interference-between-featuresInterference between features
Cost of non-orthogonal feature embedding.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Thinkers (1)
thinker
- Saxe et al.studiesLearning dynamics of linear networks used for mathematical understanding.
Concepts (2)
concept
- Feature Interferencerelated_toWhen non-orthogonal features cause logistic regression to identify a suboptimal probe direction
- Superpositionassociated_withPhenomenon where models represent more features than dimensions via almost-orthogonal directions.
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.
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Extra feature representing unexplained error to make isomorphism exact.
- Phenomenon where a feature in a small SAE splits into multiple finer features in a larger SAE.
- Patterns of which features activate together across tokens; preserved by covariance pooling but lost in mean pooling.
- Features with similar decoder magnitude across both models.
- Problems suggesting sparsity breaks down as interpretability proxy.
- Phenomenon where sparsity yields gerrymandered latents for hierarchical concepts.