concept
active
concept:polysemanticityPolysemanticity
Neurons that respond to multiple unrelated concepts, limiting interpretability.
Neighborhood — ranked by edge-count
Papers (4)
paper
Frameworks (1)
framework
- Superposition Hypothesisassociated_withCore theoretical framework: neural networks represent more features than neurons by encoding features as directions in superposition
Claims (1)
claim
- Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters
Concepts (6)
concept
- monosemanticityassociated_withrelated_toInterpretability property where a latent feature represents a single semantic concept; benchmarked across architectures.
- Feature Densityassociated_withFraction of training tokens on which a given feature has nonzero activation; used as proxy metric for autoencoder quality
- Virtual Weightsassociated_withImplicit weights directly connecting any pair of layers computed by multiplying output weights of one layer with input weights of another through the residual stream
- Model-exclusive featuresassociated_withFeatures writing predominantly to one model.
- Privileged Basisassociated_withA property of activations where neural network features align with basis dimensions due to sparse activation functions; absent in the residual stream but present in tokens, attention patterns, and MLP activations
- Neuron Doctrineassociated_withNeuroscience view of neurons as unit; analog to early interpretability.
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.
- A head can implement multiple behaviors via high-rank output without superposition.
- A neuron that responds to multiple unrelated inputs, posing a major challenge for circuit-level interpretation
- Most pragmatically important question.
- Major prediction checked against InceptionV1.
- Features that correspond to a single semantic concept and are effective for steering behavior.
- Quantitative assessment of feature quality using clinical concepts across models.
- Neurons responding to a single feature.
- Property of developmental systems where functions are encapsulated in modules with simple triggers, enhancing evolvability.