Sparse autoencoders capture concept manifolds globally via compact feature groups or locally via selective tiling but typically recover them suboptimally in a fragmented dilution regime.
An R2 near 1 at n=k i indicates compact subspace capture
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Do Sparse Autoencoders Capture Concept Manifolds?
Sparse autoencoders capture concept manifolds globally via compact feature groups or locally via selective tiling but typically recover them suboptimally in a fragmented dilution regime.