I don't have experience with the Amazon AMIs for deep learning, but I'll give the tip that you can use spot instances with limited lifespans (x instance at $y price for at most z hours) to reduce cost and the risk you forget to shut something down.
Doing spot requests is a great way to keep the costs down if you just want to rent a GPU for a while. But bear in mind that GPU instance types tend to spike in price all-at-once when somebody brings a huge compute job online, which will exceed your spot request's max bid with no warning and immediately terminate your instance. So make sure whatever you're doing can be interrupted!