It happens when gradients become too small for early layers to learn.

Deep networks can suffer from tiny gradients during backpropagation. Better activations, normalization, residual connections, and initialization help.