Graph Strategy for Interpretable Visual Question Answering

Abstract

n the paper, we consider the task of Visual Question An-swering – the important task for creating General Artificial Intelligencesystems. we propose an interpretable model called GS-VQA. The mainidea behind the model is that a complex compositional question could bedecomposed into sequence of simple questions about objects propertiesand their relations. We use Unified estimator to answer questions fromthat sequence and test the proposed model on CLEVR and THOR-VQAdatasets. The GS-VQA model demonstrates results comparable to thestate of the art while maintaining transparency and interpretability ofthe response generation process