$$\hat x = W_d f(x) + b_d, \qquad f(x) = \mathrm{JumpReLU}_\theta\big(W_e x + b_e\big), \qquad \mathcal L(x) = \lVert x - \hat x \rVert_2^2 + \lambda \lVert f(x) \rVert_0$$

Equation 9 · AI Research

Probing, Sparse Autoencoders, Patching, and Steering: The Main Interpretability Methods, Compared

This equation states an equality: the expressions on both sides have the same value under the article’s assumptions.

The formal objective has evolved since the earliest versions, and the direction of that evolution is itself informative. A dictionary decoder reconstructs the activation x from a sparse code f(x) : $\hat x = W_d f(x) + b_d, \qquad f(x) = \mathrm{JumpReLU}_\theta\big(W_e x + b_e\big), \qquad \mathcal L(x) = \lVert x - \hat x \rVert_2^2 + \lambda \lVert f(x) \rVert_0$. Earlier versions of this objective penalised the code’s $\ell_1$ norm as a differentiable stand-in for sparsity, but an $\ell_1$ penalty also shrinks the magnitude of every active feature, distorting reconstruction in a way that has nothing to do with how many features are active. Rajamanoharan and colleagues introduced JumpReLU, a thresholded activation function with a learned per-feature cutoff $\theta$ trained through a straight-through gradient estimator, which lets the…

Equation guide → · Article →