How do CNNs propagate gradients through very deep architectures without collapse?

Gradient flow is preserved using residual connections, normalization, and careful initialization to prevent vanishing or exploding gradients.

In deep CNNs, gradients are propagated backward using chain rule. Without architectural support, gradients shrink exponentially. Residual connections provide alternate gradient paths, batch normalization stabilizes distributions, and initialization methods like He initialization maintain variance. Together, these techniques ensure stable optimization across deep layers.