claim
active
claim:mlp-layers-relay-many-features-without-nonlinear-modification-but-along-a-different-axis-explaining-driftMLP layers relay many features without nonlinear modification but along a different axis, explaining drift
Interpretation of pre/post MLP correlation.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Findings (2)
finding
- Feature decoder directions are more stable than chance but drift substantially across layers even where norm is strongassociated_withFeatures not merely relayed via residual connections.
- May explain drift: MLPs relay features along different axes.
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.
- Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters
- Self-repair effect.
- Fig. 10.
- From drift findings.
- Zeroing MLP layers to localize task neurons.
- Feed-forward neural network with hidden layers, capable of representing non-linearly separable functions.
- Argues that intervening control layers decompose the genotype-phenotype mapping into two easier problems.
- Structural finding showing modular organization within the sparse neuron set