concept
active
concept:circuit-simplification-via-persistent-featuresCircuit simplification via persistent features
Removing duplicate persistent features to shrink circuit depth.
Neighborhood — ranked by edge-count
Methods (1)
method
- Sparse CrosscoderimplementsSparse dictionary learning variant that reads and writes to multiple layers or models, yielding shared features.
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.
- Motivating example with 13 vs 3 features.
- Second of three speculative claims asserting that subgraphs of neural networks are tractable and meaningful objects of study
- The key novel property of DiffLogic CA — logic gate networks that are recurrent both spatially and temporally
- Open question raised by the 5-gate checkerboard solution and its generalization properties
- Third of three speculative claims asserting that learned features are not model-specific but represent universal solutions to learning problems
- A goal in mechanistic interpretability to identify sparse computational subgraphs; VPD promotes sparse parameter circuits.
- Hypothesis raised by the 5-gate checkerboard solution and its grid-size invariance
- Hypothesis that discretization finds minimum-size circuits equivalent to minimal algorithmic descriptions of patterns