Vanishing Gradients

Vanishing gradients make early layers learn very slowly in deep networks.

What problem does the 'vanishing gradient' issue cause during training?