Contextual RNN-GANs for Abstract Reasoning Diagram Generation
arxiv.org
arxiv.org
I can already train something like OVER(CUP(X), TABLE(X)) as an ordinary classifier. But OVER() wouldn't generalize to predicates other than CUP or TABLE, and it couldn't take independent image fragments: OVER(CUP(X), TABLE(Y)), where X < Z and Y < Z, where < means (rectangular?) subset.
I'd like to train a GAN or similar to generate an image that satisfies arbitrary formulas in predicate logic, with a dataset of images that are tagged with formulas. Anyone have any ideas for that?