concept
active
concept:constructive-abstractionConstructive Abstraction
Type of abstraction map where node information is computed from non-overlapping neuron sets
Neighborhood — ranked by edge-count
Concepts (3)
concept
- Constructive Causal Abstractionrelated_toFormal definition: H is a constructive abstraction of L under alignment Π when interchange interventions have equivalent effects at both levels.
- Distributed AbstractionextendsKey notion where alignment map ϕ maps neurons block-wise to latent variables before constructive abstraction
- Strong τ-AbstractionextendsNotion where all interventions on algorithm A are allowed; gives equivalence between algorithm and DNN
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 framework the paper uses alongside feature geometry to deepen mechanistic understanding of LMs
- The formal method used to establish that the identified circuit causally mediates the model's cyclic reasoning behavior
- Question prompting the use of Category for linear transformations.
- The emergent human capacity that Nicholson argues is enabled by loose parts in environments.
- Denotation function µ decomposes over operations so meaning of compound expressions follows from meanings of parts
- Programming technique to restructure a fine-grained Linda program for efficiency by replacing live data structures with passive ones and coarser-grain processes.
- Internal actions like commitments not necessarily expressed in output, but on whose fulfillment correctness depends.
- Graded notion of causal abstraction measured by IIA; when IIA is alpha < 100%, the model is alpha-on-average approximately abstract.