JumpReLU SAE

Creator
Creator
Seonglae Cho
Created
Created
2025 Mar 17 11:6
Editor
Edited
Edited
2025 May 20 14:43
Refs
Refs

Generalized Gated SAE by learning threshold by zeroing out pre-activations

+ L0 loss
JumpReLU(x,t)={x,if x < exp(t)0,otherwiseJumpReLU(x, t) = \begin{cases} x, & \text{if x < exp(t)}\\ 0, & \text{otherwise} \end{cases}

JumpReLU SAE with
Unit step function

Does this mean they efficiently implemented the gating mechanism using JumpReLU activation?
σ(z)=JumpReLUθ(z)=zH(zθ)\sigma(z) = JumpReLU_\theta(z) = z \odot H(z - \theta)
notion image
notion image
 
 
google jumprelu preliminary
gemma scope jumprelu
differentiable pre-act-loss
 
 
 

Recommendations