> To avoid getting stuck in local minima, we make sure we use the proper learning rate This sentence is extremely misleading. It implies that gradient descent

> To avoid getting stuck in local minima, we make sure we use the proper learning rate This sentence is extremely misleading. It implies that gradient descent always reaches a global minimum if the "proper" learning rate is used. Guarantees of that form are only available for restricted families of function such as Lipschitz-smooth convex functions.