Neural networks and symbolic learning techniques can be seen as operating at different levels of abstraction. Our work focuses on understanding differences between their capabilities, and on combining their strengths.