hypothesis
active
hypothesis:superposition-hypothesis-networks-represent-more-features-than-neurons-as-almost-orthogonal-directionsSuperposition hypothesis: networks represent more features than neurons as almost-orthogonal directions
Central hypothesis.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Findings (1)
finding
- Superposition occurs in ReLU output toy models: more features than dimensions represented when features are sparseassociated_withsupportsCore demonstration.
Claims (1)
claim
- Hypothesis about real networks.
Frameworks (1)
framework
- ReLU output toy modelsupportsx'=ReLU(W^TWx+b), main model demonstrating superposition.
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.
- Superposition hypothesis: neural networks represent more features than dimensions using almost-orthogonal directions.hypothesis0.930Explanation for why dictionary learning can recover many more features than dimensions.
- Interpretation of the cars-in-superposition circuit finding as an intentional representational strategy
- Linear representation hypothesis: neural networks represent meaningful concepts as directions in their activation spaces.hypothesis0.824Foundation for interpreting features as linear directions.
- Explanation for neuron interpretability.
- Open question.
- Supported by abs model.
- Speculative extension of universality to neuroscience, with high-low frequency detectors as a candidate prediction
- superposition is deeply linked to the ability to identify and enumerate over all features in a modelquote0.797Safety rationale.