claim
active
claim:neural-networks-represent-features-as-directions-in-activation-spaceNeural networks represent features as directions in activation space
Linear representation hypothesis supported by empirical and theoretical arguments.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Frameworks (1)
framework
- Linear Representation Hypothesisassociated_withThe hypothesis that models internalize concepts as approximately linear directions in representation space; used to interpret MDS injection behavior
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.
- Linear representation hypothesis: neural networks represent meaningful concepts as directions in their activation spaces.hypothesis0.860Foundation for interpreting features as linear directions.
- How important is the geometry of activation space for explaining neural network behavior?question0.841Open question A.1.1a.5.
- Does the geometric structure of activation space causally shape neural network behavior?question0.834Central research question driving the work.
- Extends convergence argument to brain-machine alignment
- First of three speculative claims forming the foundation of the circuits research agenda
- The paper's central thesis statement, presented prominently after the abstract
- Superposition hypothesis: neural networks represent more features than dimensions using almost-orthogonal directions.hypothesis0.814Explanation for why dictionary learning can recover many more features than dimensions.
- Closing statement.