Theoretical question: can RN generalize to shape-color combinations not previously seen in training data?

[PabloMessina]

Say, if in the training data the RN saw many green circles, and many yellow triangles, but it never saw green triangles or yellow circles, would the RN perform well if asked a question about those shape-color combinations it never saw in the training data? In other words, can the RN learn the abstract concept of "color", the abstract concept of "shape" and generalize those concepts to understand new questions involving novel color-shape combinations?