x

LESSWRONG
LW

ZachMaas

Message

Independent mech interp research, working on SAE feature geometry + dependence structure. Background in computational biology and transcriptomics.

8

1

3y

ZachMaas

Subscribe

Message

Independent mech interp research, working on SAE feature geometry + dependence structure. Background in computational biology and transcriptomics.

8

1

3y

Preliminary Results on Building Graphs from SAEs

TLDR: * I use nodewise LASSO to build approximate sparse conditional dependence graphs over SAE features, with resampling and null controls * Initial experiments produce graphs with small standalone modules that are stable under resampling * These modules frequently correspond to coherent-looking linguistic features, and are only weakly aligned with...

Mar 279