The ICLR 2020 virtual conference wrapped up this weekend, with generally favourable reviews from participants and a number of areas for future improvements identified by organizers.
A surprise came from ICLR 2020 General Chair Alexander (Sasha) Rush of Cornell Tech, who revealed without elaboration in an April 30 conversation on the conference general group chat that “PCs [program chairs] decided against having best paper this year.”
Like other AI conferences impacted by Covid-19, this year’s International Conference on Learning Representations (ICLR 2020) was moved completely online, where it ran relatively smoothly from April 26 to 30. The ICLR yesterday made the entire virtual conference available in open-access, enabling anyone to access the content and explore the virtual conference portal.
The conference was originally scheduled for Addis Ababa, Ethiopia, but this aspect was cancelled on March 10. Organizers made a number of adjustments to enable and encourage networking for conference participants, such as opening various virtual rooms and scheduling socials. “Organising the 8th international conference on learning representations (ICLR 2020) was highly challenging, but ultimately, highly rewarding for our organising committees,” the organizers wrote in an official blog post.
One of the world’s major machine learning conferences, ICLR 2020 accepted 687 out of 2,594 papers and drew over 5,600 participants from nearly 90 countries — more than double from 2,700 physical attendees of ICLR 2019. Each of the papers was presented by its authors through pre-recorded videos, and every paper was presented twice (in two separate sessions) considering global time zone differences.
ICLR 2020 received more than a million page views and over 100,000 video watches over its five-day run. All posters were widely viewed as part of a live conference, with an average of 200 unique views per paper page.
The virtual conference created a new environment in which organizers were able to try out — albeit in a more rushed manner than they might have liked — many of the machine learning approaches that the conference itself showcases. “We used a latent variable model for review score calibration, a vision model to extract thumbnails from each paper to be used on the web, natural language tools to visualise related papers, and recommender systems to create sets of balanced recommendations for papers and participants,” explained the organizers.
“We wanted to build something that was fun to browse, async first, and feels alive,” tweeted Rush, who’s also responsible for coding the virtual ICLR website. The main interface for example was inspired by the idea that “the main success in async communication has been chat apps like Slack.”
Considering that the entire virtual conference was put together within about six weeks, Graham Neubig thinks the organizers did a “ridiculously amazing job.” An assistant professor with the Language Technologies Institute of Carnegie Mellon University, Neubig tweeted: “I’m in awe that they were able to pull this together in such a short time.”
Neubig also opined that the interactions aspect of such virtual conference remains a challenge, tweeting “I was able to efficiently browse papers, but less chat with friends about new ideas… I had nice conversations with junior researchers presenting papers, but don’t think I met any other professors in video chat at all.”
The conference hosted a total of 29 official socials throughout the week, although the organizers say these were not as globally diverse as they would have liked. The coverage of topics was also lacking, and organizers believe that given more time they could have made better calls for particular socials topics to facilitate improved interaction between attendees.
For future virtual conferences, the organizers also recommend pre-arranged video meet-and-greet sessions, mentorship programs, and even standing virtual coffee sessions, and hope the community can explore additional ways for improving interactivity.
There was plenty of positive feedback from participants, with Landing AI Founder and CEO Andrew Ng tweeting “I’m really enjoying the new #ICLR2020 website, and am having fun browsing the workshop & poster talks. Congrats to the organizers and speakers on what’s happening so far in this all-virtual conference! Great job!” There were even some welcome lighthearted moments, with NLP company Hugging Face for example rebranding as “(not Hugging for the moment) Face,” and also giving the virtual conference top marks.
The organizers say ICLR 2021 will be held the first week of May, but it remains undecided whether this will be a physical or virtual conference or some combination of the two. Synced will provide updates as further information becomes available.
Journalist: Yuan Yuan | Editor: Michael Sarazen