Deep learning has transformed AI, yet its methodology in sequential decision-making settings remains largely bottom-up and ad hoc: architectures and losses are chosen by intuition and tuned against benchmarks, without a crisp account of what should be learned. This tutorial takes a top-down view instead, grounded in the language of classical planning, and asks two questions in turn. Part I (Hector Geffner) asks what it takes to learn the hidden lifted action model of a domain - predicates and action schemas - from action or state-action traces, covering symbolic algorithms alongside recent deep-learning counterparts. Part II (Blai Bonet) asks what it takes to learn a single general policy that solves every instance of a planning domain at once in the symbolic and neural settings.