concept
active
concept:feature-density

Feature Density

Fraction of training tokens on which a given feature has nonzero activation; used as proxy metric for autoencoder quality

Neighborhood — ranked by edge-count

Concepts (4)

concept
  • Polysemanticity
    associated_with
    Neurons that respond to multiple unrelated concepts, limiting interpretability.
  • Dead Neurons
    associated_with
    Autoencoder neurons that fail to activate across any datapoints during training; addressed via neuron resampling
  • Cluster of autoencoder features with extremely low activation density (~1e-7) that are generally not interpretable and appear to be training artifacts
  • Mode in feature density histogram around 1e-5 corresponding to interpretable features, contrasted with ultralow density cluster

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

  • Log-scale histogram of feature firing rates used as proxy for autoencoder quality during hyperparameter tuning
  • Approximate posterior probability distribution embodied in organism's internal states; organism's best guess about causes of sensations
  • The extracted set of sparse interpretable features from model embeddings via SAEs
  • Fraction of a dimension given to a feature.
  • Measure of superposition in toy models.
  • Decomposability, linearity, superposition, basis alignment.
  • Featuresconcept0.765
    Interpretable concept representations hypothesized as basic units of computation.
  • Density of Centersconcept0.760
    The degree to which centers are packed and overlapped, contributing to the life of the whole.