{"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/best-of-both-worlds-transferring-knowledge","title":"Best of Both Worlds: Transferring Knowledge from Discriminative Learning to a Generative Visual Dialog Model","arxiv_id":"1706.01554","date":"2017-06-05","proceeding":"NeurIPS 2017 12","authors":["Jiasen Lu","Anitha Kannan","Jianwei Yang","Devi Parikh","Dhruv Batra"],"abstract":"We present a novel training framework for neural sequence models,\nparticularly for grounded dialog generation. The standard training paradigm for\nthese models is maximum likelihood estimation (MLE), or minimizing the\ncross-entropy of the human responses. Across a variety of domains, a recurring\nproblem with MLE trained generative neural dialog models (G) is that they tend\nto produce 'safe' and generic responses (\"I don't know\", \"I can't tell\"). In\ncontrast, discriminative dialog models (D) that are trained to rank a list of\ncandidate human responses outperform their generative counterparts; in terms of\nautomatic metrics, diversity, and informativeness of the responses. However, D\nis not useful in practice since it cannot be deployed to have real\nconversations with users.\n  Our work aims to achieve the best of both worlds -- the practical usefulness\nof G and the strong performance of D -- via knowledge transfer from D to G. Our\nprimary contribution is an end-to-end trainable generative visual dialog model,\nwhere G receives gradients from D as a perceptual (not adversarial) loss of the\nsequence sampled from G. We leverage the recently proposed Gumbel-Softmax (GS)\napproximation to the discrete distribution -- specifically, an RNN augmented\nwith a sequence of GS samplers, coupled with the straight-through gradient\nestimator to enable end-to-end differentiability. We also introduce a stronger\nencoder for visual dialog, and employ a self-attention mechanism for answer\nencoding along with a metric learning loss to aid D in better capturing\nsemantic similarities in answer responses. Overall, our proposed model\noutperforms state-of-the-art on the VisDial dataset by a significant margin\n(2.67% on recall@10). The source code can be downloaded from\nhttps://github.com/jiasenlu/visDial.pytorch.","url_abs":"http://arxiv.org/abs/1706.01554v2","url_pdf":"http://arxiv.org/pdf/1706.01554v2.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":[{"paper_slug":"best-of-both-worlds-transferring-knowledge","repo_url":"https://github.com/jiasenlu/visDial.pytorch","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"informativeness","task_name":"Informativeness"},{"task_slug":"metric-learning","task_name":"Metric Learning"},{"task_slug":"transfer-learning","task_name":"Transfer Learning"},{"task_slug":"visual-dialogue","task_name":"Visual Dialog"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/visual-dialog-on-visdial-v09-val","task":"Visual Dialog","dataset":"VisDial v0.9 val","model":"HCIAE-NP-ATT","rank_in_archive_order":8,"of":18,"metrics":{"MRR":"62.22","Mean Rank":"4.81","R@1":"48.48","R@10":"87.59","R@5":"78.75"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/1706.01554","atlas_url":"https://app.syntology.ai/?focus=1706.01554","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}