method
active
method:downstream-client-feature-analysisDownstream Client Feature Analysis
Examining downstream neurons that rely on a given feature to verify its functional role
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.
- Proposed objective decomposing activations into predictable and novel components.
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Authors argue features are model properties because logit effects and ablations are consistent with feature interpretations
- Dual interpretation of features: in addition to responding to inputs, features also act to increase probability of specific output tokens
- Method of optimizing input to cause a neuron to fire maximally, used to characterize what a neuron detects; establishes causal link
- Model recalls known upstream PR solution.
- Phenomenon where a feature in a small SAE splits into multiple finer features in a larger SAE.
- Frequency-weighted spectral value πkλkj determining which features are retained.