λEnter Calculus CompanionRegistered manuscripts, kept distinct

Manuscript Psy01-A@1 · Psy01-A.tex · section · 94477fa3

Logit perturbation

Stability, hard-mask discontinuity, and schedule dependence › Logit perturbation · lines 1140–1152

Back to manuscript

Consider two networks $θ, θ^{'}$ , evaluated on the same input and the same hidden Gaussian vectors. Define

a_{ℓ} = max {\norm B_{ℓ}_{op}, \norm B_{ℓ}^{'}_{op}}, e_{ℓ} = \norm B_{ℓ} - B_{ℓ}^{'}_{op}, β_{ℓ} = \norm b_{ℓ} - b_{ℓ}^{'}_{2} .

Displayed formulas in this passage

unlabeled · line 1144
$a_{ℓ} = max {\norm B_{ℓ}_{op}, \norm B_{ℓ}^{'}_{op}}, e_{ℓ} = \norm B_{ℓ} - B_{ℓ}^{'}_{op}, β_{ℓ} = \norm b_{ℓ} - b_{ℓ}^{'}_{2} .$

Nearby in this section