COVID-19 officially became a global pandemic on Wednesday. As public health officials and governments respond; businesses brace for losses; and events like trade shows, SXSW, and Google’s I/O shutter around the world, the disease is also impacting scientific conferences. Ironically, a coronavirus conference got canceled this week, and on Tuesday the International Conference on Learning Representations (ICLR), one of the fastest-growing machine learning conferences in the world, shared that it will now be a virtual event held entirely online. Papers will be presented in prerecorded five-minute videos with a slide deck, while researchers invited to make longer presentations can submit 15-minute videos.
In a post about the change to an all-digital conference, organizers called the cancellation of an in-person event an “… opportunity to innovate on how to host an effective remote conference.”
ICLR general chair and Cornell University researcher Alexander Rush told VentureBeat board members are interested in the evolution of AI research conference experiences that do not require travel.
“We figured we should use these events to document and test out different ways to make that happen,” Rush said.
Initially scheduled to take place in Addis Ababa, Ethiopia, ICLR was set to be the first international AI research conference held in Africa. Members of the machine learning community on the conference board had made a deliberate effort to host the conference in Africa to support the continent’s burgeoning machine learning community. Members of the group Black in AI told VentureBeat they plan to request that ICLR host a conference in Addis Ababa in 2021 or 2022.
“We have been really excited to host a conference in Addis [Ababa] this whole year — it felt like a special chance and something outside the typical ML conference. Over the last couple weeks, it was really hard to face having to let go of that possibility,” Rush said.
Transitioning from bummed out that ICLR Addis is not happening => kind of getting hyped for ICLR Virtual. https://t.co/7HiIiu2nLE
— Sasha Rush (@srush_nlp) March 10, 2020
A silver lining for attendees will be reduced prices. ICLR’s online event registration fees now stand at $50 for students and $100 for nonstudents. By comparison, a 2019 ICLR in-person ticket cost $450 for students and $550 for nonstudents.
Learnings from ICLR 2020 could be shared with other scientific or AI research conferences interested in encouraging remote participation or with other big conferences forced to adapt to the pandemic. Another major machine learning event, the Computer Vision and Pattern Recognition (CVPR) conference, is set to take place in Seattle in June, but that could still change. As the state was one of the areas hardest hit by coronavirus cases in the U.S., Washington Governor Jay Inslee this week banned all gatherings of more than 250 people in Seattle. Inslee said if changes aren’t made, the state of Washington could have 64,000 cases by May.
An online-only version of ICLR may be the biggest such experiment yet, and pulling it off will be no small feat. According to the 2019 AI Index report, ICLR attracted about 3,000 attendees last year and attendance grew 15 times between 2014 and 2019. After considering a record number of papers, ICLR accepted 680 papers for publication in 2020. More than a dozen workshops attached to the conference may also consider remote options.
The largest machine learning research conferences, like CVPR and leader NeurIPS, can attract more than 10,000 attendees. As part of a series of changes to NeurIPS submissions, researchers must now include a three-minute video with their paper to make content accessible for those unable to attend the conference.
Inclusion and sustainability
Before COVID-19 forced conference organizers to reconsider international travel, machine learning researchers had started looking at remote access as a way to combat climate change.
Last month, AI researcher Carl Johann Simon-Gabriel of the Max Planck Institute for Intelligent Systems, drafted a petition urging organizers of all machine learning conferences to allow remote paper and poster presentations in order to reduce their carbon footprint. The petition was created with support from luminaries like Jeff Dean and Geoffrey Hinton and suggests a hybrid model in which conference attendees can use telepresence or scan QR-codes to log into a video session with a paper’s author.
An online element may also enable access for researchers who can’t afford to travel or who experience challenges obtaining visas from immigration officials in the host country, something that has been an ongoing issue in North America and Europe.
ICLR board members are considering an approach in which researchers publish videos about their paper to a conference website and authors respond to questions via remote sessions, but members of the board headed by Facebook chief AI scientist Yann LeCun plan to share more details in the coming days.
Deep learning pioneer Yoshua Bengio is also on the ICLR board and, like LeCun, one of the most cited researchers on Earth. Bengio recently dedicated the first posts on his new blog to the assertion that it’s “time for rethinking the overall publication process in the field of machine learning.”
Prior to the impact of COVID-19, Bengio had asserted the need to rethink conferences and workshops and to take collective responsibility for their environmental impact.
“It’s great and it is important to have these meetings, which bring minds from all over the world to exchange [ideas], brainstorm, build on top of each other’s ideas, [and] educate each other. But the carbon footprint associated with all the … air travel is by far the greatest source of greenhouse gases deriving from us as individuals because of our job as scientists and scholars,” he wrote.
Conferences used to help facilitate faster turnaround of AI research, Bengio said, but the developments of tools like preprint repository arXiv fulfill that role now, and he believes conferences should focus on highlighting work worthy of oral presentation.
He also criticized conference culture as beholden to a steady stream of deadlines, which leads to incremental work that fails to focus on long-term trends. Speaking on a climate change panel last December, Bengio labeled research culture today as unhealthy and oppressive.
“The field has almost completely switched to a conference publication model (in fact a large part of computer science has done so too), and each conference paper does not get the chance to be cleaned up as well as a typical journal paper, rarely benefitting from the repeated iterations to improve the paper [that] is typical of journals,” he wrote. “So we are more productive, on the surface, but this stress, productivity, and fast pace have a price on the depth and quality of the papers we produce.”
Instead, he suggests organizers try a model in which research papers are submitted to a journal for quick turnaround and then program organizers choose the papers they like most. Bengio shared the same thoughts with NeurIPS organizers last year as a member of an advisory committee.
He argued experimentation should start now with entirely digital or hybrid in-person and virtual events to make progress toward reducing the carbon footprint of the growing machine learning community, though he believes entirely remote events can lose the social value delivered by in-person conferences.
“Let’s put our heads together and explore how we can at the same time improve the quality of our science and improve our lives as human beings,” he wrote.