{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/making-history-matter-gold-critic-sequence","title":"Making History Matter: History-Advantage Sequence Training for Visual Dialog","arxiv_id":"1902.09326","date":"2019-02-25","proceeding":"ICCV 2019 10","authors":["Tianhao Yang","Zheng-Jun Zha","Hanwang Zhang"],"abstract":"We study the multi-round response generation in visual dialog, where a\nresponse is generated according to a visually grounded conversational history.\nGiven a triplet: an image, Q&A history, and current question, all the\nprevailing methods follow a codec (i.e., encoder-decoder) fashion in a\nsupervised learning paradigm: a multimodal encoder encodes the triplet into a\nfeature vector, which is then fed into the decoder for the current answer\ngeneration, supervised by the ground-truth. However, this conventional\nsupervised learning does NOT take into account the impact of imperfect history,\nviolating the conversational nature of visual dialog and thus making the codec\nmore inclined to learn history bias but not contextual reasoning. To this end,\ninspired by the actor-critic policy gradient in reinforcement learning, we\npropose a novel training paradigm called History Advantage Sequence Training\n(HAST). Specifically, we intentionally impose wrong answers in the history,\nobtaining an adverse critic, and see how the historic error impacts the codec's\nfuture behavior by History Advantage-a quantity obtained by subtracting the\nadverse critic from the gold reward of ground-truth history. Moreover, to make\nthe codec more sensitive to the history, we propose a novel attention network\ncalled History-Aware Co-Attention Network (HACAN) which can be effectively\ntrained by using HAST. Experimental results on three benchmarks: VisDial\nv0.9&v1.0 and GuessWhat?!, show that the proposed HAST strategy consistently\noutperforms the state-of-the-art supervised counterparts.","url_abs":"http://arxiv.org/abs/1902.09326v3","url_pdf":"http://arxiv.org/pdf/1902.09326v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[],"tasks":[{"task_slug":"answer-generation","task_name":"Answer Generation"},{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"},{"task_slug":"response-generation","task_name":"Response Generation"},{"task_slug":null,"task_name":"Triplet"},{"task_slug":"visual-dialogue","task_name":"Visual Dialog"},{"task_slug":"visual-reasoning","task_name":"Visual Reasoning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/visual-dialog-on-visdial-v09-val","task":"Visual Dialog","dataset":"VisDial v0.9 val","model":"HACAN","rank_in_archive_order":10,"of":18,"metrics":{"MRR":"0.6792","Mean Rank":"3.97","R@1":"54.76","R@10":"90.68","R@5":"83.03"},"uses_additional_data":false},{"leaderboard":"/sota/visual-dialog-on-visual-dialog-v1-0-test-std","task":"Visual Dialog","dataset":"Visual Dialog v1.0 test-std","model":"HACAN","rank_in_archive_order":59,"of":80,"metrics":{"MRR (x 100)":"64.22","Mean":"4.20","NDCG (x 100)":"57.17","R@1":"50.88","R@10":"89.45","R@5":"80.63"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1902.09326","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}