Vanishing and Exploding Gradients

Gradients can shrink toward zero or grow too large in deep networks, disrupting training.