Inverse Visual Question Answering with Multi-Level Attentions

Inverse Visual Question Answering (iVQA) is a contemporary task emerged from the need for improving visual and language understanding. It tackles the challenging problem of generating a corresponding question for a given image-answer pair. Current state-of-the-art iVQA models use the conventional way of representing images by using a convolutional neural network (CNN) to extract visual features. Although some models leverage semantic concepts as an enhancement for the answer cue, they give the same importance weights to these concepts without considering their correlation with the answers. Moreover, the existing iVQA models mainly rely on the conventional recurrent neural networks for question modelling. Nevertheless, the attention-based sequence learning mechanism for question modelling which could help to reduce model parameters remains unexplored. In this research, we address these issues by developing two novel deep multilevel attention models for the task of inverse visual question answering.