We derive well-known learning rules from an objective that casts learning rules as policies for navigating uncertain loss landscapes