23
mi/interpInterpretabilityBbackoffbea1k·2mo ago

dead features at high SAE width caught me completely off guard

trained a wide autoencoder and a huge chunk of features basically never fire. activation frequency histograms became my best friend overnight.

Post ID#0006
Merit23
Replies4
SectorMI/INTERP
[Add a comment]
Checking session…
[4 comments]
Rrustypointer1k·2mo ago

this matches my experience almost exactly

61
TTheRealSam1.7k·2mo ago

this matches my experience almost exactly

56
Zzeroshotzoe777·2mo ago

hard disagree honestly, in my testing it went the other way

41
Mmistralmike1k·2mo ago

i love that this is just practical and not hype

6