LawrenceC comments on Sparse Coding, for Mechanistic Interpretability and Activation Engineering

LawrenceC 23 Sep 2023 19:19 UTC
4 points
0
We train such an autoencoder to convergence, driving towards an $L^{0}$
This is a typo right? IT should say L^1
- David Udell 23 Sep 2023 19:20 UTC
  4 points
  0
  Parent
  No, towards an $L^{0}$ value. $L^{1}$ is the training proxy for that, though.
  - LawrenceC 23 Sep 2023 19:21 UTC
    4 points
    0
    Parent
    Oh, okay, makes sense.