concept
active
concept:l0-norm-of-feature-activationsL0 Norm of Feature Activations
Average number of nonzero feature entries per input; primary measure of activation sparsity in the autoencoder
Neighborhood — ranked by edge-count
Concepts (1)
concept
- Feature Sparsityassociated_withProperty that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
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.
- Measure of features activated by clean vs adversarial inputs.
- Normalizing steering coefficient by local residual-stream norm to ensure comparability across checkpoints
- Clamping a feature's value to zero to measure its causal effect on model output.
- Removes superposition at cost of loss.
- Sparsity penalty weighting by sum of per-layer decoder norms, enabling comparability with per-layer SAEs.
- Time-dependent correlation structure in activations.
- Rotation-invariance argument.
- Assumption that small anchor changes can produce sharp performance shifts when conditions are favorable.