Beyond Bilinear: Generalized Multimodal Factorized High-Order Pooling for Visual Question Answering