Unifying Vision-and-Language Tasks via Text Generation

                  Jaemin Cho 1 Jie Lei Hao Tan Mohit Bansal
                         UNC Chapel Hill
                {jmincho,jielei,haotan,mbansal}@cs.unc.edu

             Abstract
    Existing methods for vision-and-language learn-
    ing typically require designing task-specific ar-
    chitectures and objectives for each task. For ex-
    ample, a multi-label answer classifier for visual
    question answering, a region scorer for ref ...