Multimodal Residual Learning for Visual QA