Batch gradient descent computes the gradient using all training samples. What is its main disadvantage?