It's not a Lottery, it's a Race: Understanding How Gradient Descent Adapts the Network's Capacity to the Task