Probabilistic framework for solving visual dialog

In this paper, we propose a probabilistic framework for solving the task of ‘Visual Dialog’. Solving requires reasoning and understanding visual modality, language common sense knowledge to answer. Various architectures have been proposed solve by variants multi-modal deep learning techniques that combine representations. However, believe it is crucial understand analyze sources uncertainty task. Our approach allows estimating also aids diverse generation answers. The obtained through representation module provides us with representations image, question conversation history, ensures latent candidate answers are given an chooses appropriate answer minimizes uncertainty. We thoroughly evaluate model detailed ablation analysis, comparison state art visualization in method. Using framework, thus obtain improved dialog system more explainable.