* It is computationally efficient compared to stochastic gradient descent.
* It improves generalization by finding flat minima.
* It improves convergence by using mini-batches. We can approximate the gradient of the entire training set, which might help to avoid local minima.
