[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-140371-en":3,"detail-sidebar-cat-0-en-105":31,"doc-seo-140371-105":82},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":28,"seo_description":14,"update_tm":29,"read_time":30},140371,8796095462418,"Noah","https://ap-avatar.wpscdn.com/avatar/80000253c1241d02b47?x-image-process=image/resize,m_fixed,w_180,h_180&k=1778826106357471780",8,"Research & Report","Learning to Draw - Emergent Communication through Sketching","Evidence links early visual communication to the emergence of written language, with pictorial traces preceding formal text. Refer To emergent communication research, agents learn cooperative communication to solve tasks, typically via learned discrete-token channels. This paper studies a visual communication channel where agents draw using simple strokes. Using a differentiable drawing process and a referential communication game, agents learn to communicate by sketching, and with suitable inductive biases produce sketches humans can interpret, encouraging further work on directly interpretable visual protocols.","Learning to Draw: Emergent Communication through Sketching  \nDaniela Mihai  \nElectronics and Computer Science The University of Southampton Southampton, UK [adm1g15@soton.ac.uk](adm1g15@soton.ac.uk)  \nJonathon Hare  \nElectronics and Computer Science The University of Southampton Southampton, UK [jsh2@soton.ac.uk](jsh2@soton.ac.uk)  \nAbstract  \nEvidence that visual communication preceded written language and provided a basis for it goes back to prehistory, in forms such as cave and rock paintings depicting traces of our distant ancestors. Emergent communication research has sought to explore how agents can learn to communicate in order to collaboratively solve tasks. Existing research has focused on language, with a learned communication channel transmitting sequences of discrete tokens between the agents. In this work, we explore a visual communication channel between agents that are allowed to draw with simple strokes. Our agents are parameterised by deep neural networks, and the drawing procedure is differentiable, allowing for end-to-end training. In the framework of a referential communication game, we demonstrate that agents can not only successfully learn to communicate by drawing, but with appropriate inductive biases, can do so in a fashion that humans can interpret. We hope to encourage future research to consider visual communication as a more flexible and directly interpretable alternative of training collaborative agents.  \n1 Introduction  \nImagine you and a friend are playing a game where you have to get your friend to guess an object in the room by you sketching the object. No other communication is allowed beyond the sketched image. This is an example of a referential communication game. To play this game you need to have learned how to draw in a way that your friend can understand. This paper explores how artificial agents parameterised by neural networks can learn to play similar drawing games. More specifically, we reformulate the traditional referential game such that one agent draws a sketch of a given photo and the second agent guesses, based on the drawing, the corresponding photo from a set of images.  \nSpurred by innovations in artificial neural networks, deep and reinforcement learning techniques, recent work in multi-agent emergent communication [4, 17, 19, 29, 37] pursues interactions in the form of gameplay between agents to induce human-like communication. Artificial communicating agents can collaborate to solve various tasks: image referential games with realistic visual input [19, 28, 29], negotiation [2], navigation of virtual environments [7, 22], reconstruction of missing input [4, 17] and, more recently, drawing games [11] . The key to achieving the shared goal in many of these games is collaboration, and implicitly, communication. To date, studies on communication emergence in multi-agent games have focused on exploring a language-based communication channel, with messages represented by discrete tokens or token sequences [6, 17, 19, 25, 28, 29, 35] . However, these communication protocols can be difficult for a human to interpret [3, 26, 33] . In this work we propose a direct and potentially self-explainable means of transmitting knowledge: sketching. Evidence suggests pre-and early-humans were able to communicate by drawing long before developing the various stages of written language [20, 38] . Drawings such as petrograms and petroglyphs  \n35th Conference on Neural Information Processing Systems (NeurIPS 2021) .  \nexist from the oldest palaeolithic times and may have been used to record past experiences, events, beliefs or simply the relation with other beings [13, 21] . These pictorial characters which are merely impressions of real objects or beings stand at the basis of all writing [16] . This leads us to question if drawing is a more natural way of starting to study emergent communication and if it could lead to better written communication later on.  \nThis idea has recently gained interest","cbCaiqX1oeUz5w3m","https://ap.wps.com/l/cbCaiqX1oeUz5w3m","pdf",2517837,3,1,14,"English","en",105,"# Introduction\n## Communication between agents","[{\"question\":\"What communication channel do the authors study instead of discrete tokens?\",\"answer\":\"They study a visual communication channel where agents communicate by drawing simple strokes in a referential communication game.\"},{\"question\":\"How are the agents trained to draw and communicate?\",\"answer\":\"Agents use deep neural networks, and the drawing procedure is differentiable, enabling end-to-end training within a referential communication game framework.\"},{\"question\":\"What improves whether humans can interpret the sketches?\",\"answer\":\"Adding an appropriate inductive bias—especially a perceptual loss—improves human interpretability without substantially reducing gameplay success.\"}]","Learning to Draw - Emergent Communication through Sketching | PDF",1787573811,35,{"code":4,"msg":5,"data":32},[33,38,43,48,53,58,63,66,71,74,78],{"id":21,"doc_module":4,"doc_module_name":34,"category_name":35,"show_sort_weight":36,"slug":37},"Document","Story & Novel",90,"story-novel",{"id":39,"doc_module":4,"doc_module_name":34,"category_name":40,"show_sort_weight":41,"slug":42},2,"Literature",80,"literature",{"id":44,"doc_module":4,"doc_module_name":34,"category_name":45,"show_sort_weight":46,"slug":47},4,"Exam",70,"exam",{"id":49,"doc_module":4,"doc_module_name":34,"category_name":50,"show_sort_weight":51,"slug":52},5,"Comic",60,"comic",{"id":54,"doc_module":4,"doc_module_name":34,"category_name":55,"show_sort_weight":56,"slug":57},6,"Technology",50,"technology",{"id":59,"doc_module":4,"doc_module_name":34,"category_name":60,"show_sort_weight":61,"slug":62},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":34,"category_name":12,"show_sort_weight":64,"slug":65},30,"research-report",{"id":67,"doc_module":4,"doc_module_name":34,"category_name":68,"show_sort_weight":69,"slug":70},9,"Religion & Spirituality",20,"religion-spirituality",{"id":69,"doc_module":4,"doc_module_name":34,"category_name":72,"show_sort_weight":69,"slug":73},"World Cup","world-cup",{"id":75,"doc_module":4,"doc_module_name":34,"category_name":76,"show_sort_weight":75,"slug":77},10,"Lifestyle","lifestyle",{"id":79,"doc_module":4,"doc_module_name":34,"category_name":80,"show_sort_weight":49,"slug":81},19,"General","general",{"code":4,"msg":83,"data":84},"ok",{"site_id":25,"language":24,"slug":85,"title":13,"keywords":86,"description":14,"schema_data":87,"social_meta":140,"head_meta":142,"extra_data":144,"updated_unix":29},"learning-to-draw-emergent-communication-through-sketching","",{"@graph":88,"@context":139},[89,102,122],{"@type":90,"itemListElement":91},"BreadcrumbList",[92,96,98,100],{"item":93,"name":94,"@type":95,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":97,"name":34,"@type":95,"position":39},"https://docshare.wps.com/document/",{"item":99,"name":12,"@type":95,"position":20},"https://docshare.wps.com/document/research-report/",{"item":101,"name":13,"@type":95,"position":44},"https://docshare.wps.com/document/learning-to-draw-emergent-communication-through-sketching/140371/",{"url":101,"name":13,"@type":103,"image":104,"author":109,"headline":13,"publisher":111,"fileFormat":114,"inLanguage":24,"description":14,"dateModified":115,"datePublished":116,"encodingFormat":114,"isAccessibleForFree":117,"interactionStatistic":118},"DigitalDocument",{"url":105,"@type":106,"width":107,"height":108},"https://docshare.wps.com/thumbnails/learning-to-draw-emergent-communication-through-sketching/140371.png","ImageObject",300,407,{"name":9,"@type":110},"Person",{"url":93,"name":112,"@type":113},"DocShare","Organization","application/pdf","2026-09-13","2026-08-24",true,{"@type":119,"interactionType":120,"userInteractionCount":20},"InteractionCounter",{"@type":121},"ViewAction",{"@type":123,"mainEntity":124},"FAQPage",[125,131,135],{"name":126,"@type":127,"acceptedAnswer":128},"What communication channel do the authors study instead of discrete tokens?","Question",{"text":129,"@type":130},"They study a visual communication channel where agents communicate by drawing simple strokes in a referential communication game.","Answer",{"name":132,"@type":127,"acceptedAnswer":133},"How are the agents trained to draw and communicate?",{"text":134,"@type":130},"Agents use deep neural networks, and the drawing procedure is differentiable, enabling end-to-end training within a referential communication game framework.",{"name":136,"@type":127,"acceptedAnswer":137},"What improves whether humans can interpret the sketches?",{"text":138,"@type":130},"Adding an appropriate inductive bias—especially a perceptual loss—improves human interpretability without substantially reducing gameplay success.","https://schema.org",{"og:url":101,"og:type":141,"og:title":13,"og:site_name":112,"og:description":14},"article",{"robots":143,"canonical":101},"index,follow",{"doc_id":7,"site_id":25}]