1/ Spent the last couple weeks in quarantine obsessively coding a website for Virtual ICLR with @hen_str. We wanted to build something that was fun to browse, async first, and feels alive.
2/ We built the main interface around a chat portal with the idea the the main success in async communication has been chat apps like slack.
3/ We thought that browsing should feel quick and dynamic like walking around a poster session. To do this we utilized a recommendation system from @gneubig and John Wieting to embed and match each of the papers.
4/ We also realize that some people like old-fashioned 2D space, so @hen_str built an interactive projection to map us back down.
5/ To make things visually interesting, we used Detectron and https://github.com/ibm-aur-nlp/PubLayNet to extract images from every paper. This was useful for the conference and let me make this ridiculous thread. https://twitter.com/srush_nlp/status/1244426222349758466
6/ The conference schedule itself is made to be relatively sparse focused on invited talks and a range of targeted poster sessions. Everything is designed to be global and async.
7/ Each day includes a collection of highlighted oral presentations and invited talks. https://twitter.com/i/status/1252272318883209218
8/ Whereas each individual paper is contextualized. We have its abstract, talk, paper, code, reviews all in one place. Authors are accessible via sync zoom or by async chat.
9/ Workshops are much harder. They are really syncronous by design. We back off to a standard format of calendars and video chat here.
10/ Beta Website is now live at http://iclr.cc/virtual  so check it out and send me feedback. Thanks to @shakir_za @white_martha @kchonyc for working tirelessly on this and pushing on new ideas to try.
11/ Some more technical notes: We wanted the website to be completely static, so that we could sleep during the conference. To do this we used Flask-Freeze to write out everything as html / json / ics files. All the search is front-end with javascript. Videos / Chat are embeds.
12/ We also wanted everything to work in China, which means no Google/Twitter/Slack tools are allowed. This was surprisingly hard! It would have been really nice to use Firebase or GCP.
13/ Another note. My first pass of the website was only desktop oriented, but a surprising amount of volunteers wanted it to work on mobile. It will be interesting to see how many people use it that way.
You can follow @srush_nlp.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: