We define our empathy learning problem as determining the extent by which a system can perceive user affect and intention as well as ambient context, construct models of these perceptions and of its interaction behavior with the user, and incrementally improve on its own its models in order to effectively provide empathic responses that change the ambient context. In concept, system selfimprovement can be viewed as changing its internal assumptions, programs or hardware. In a practical sense, we view this as rooting from a data-centric approach, i.e., the system learns its assumptions from recorded interaction data, and extending to growth-centric, i.e., such knowledge should be dynamically and continuously refined through subsequent interaction experiences as the system learns from new data. To demonstrate this, we return to the fundamental concept of affect modeling to show the data-centric nature of the problem and suggest how to move towards engaging the growth-centric. Lastly, given that an empathic system that is ambient intelligent has yet to be explored, and that most ambient intelligent systems are not affective let alone empathic, we submit for consideration our initial ideas on an empathic ambient intelligence in human-system interaction.