Towards Transparent AI Systems: Interpreting Visual Question Answering Models