Blow-up under super-critical correlated initialization
Prove that a ResNet with correlated initialization blows up at initialization under super-critical scaling \(\lambda_L\) larger than \(L^{-H}\ell(L)^{q/2}\), under suitable non-degeneracy or lower-bound assumptions on the activation function.
References
The conjectured blow-up for super-critical scaling is more difficult to prove.
— Correlated initialization of deep residual networks
(2609.03589 - Benning et al., 3 Sep 2026) in Remark 2.7, “Super-critical scaling”