What is the primary risk of a high learning rate in gradient descent?