Virtual Machine Learning Conferences: The Good and the Bad

We are now almost two years into the COVID-19 pandemic: international travel has slowed to a trickle and all machine learning conferences have moved online. By now, most of my PhD research has taken place during the pandemic, which I’ve presented at four online conferences (ACL ’20, EMNLP ’20, NAACL ’21, and ACL ’21). I’ve also had the fortune to attend a few in-person conferences before the pandemic, so here I’ll share some of my thoughts on the advantages of each format.

Travel

One of the perks of choosing grad school for me was the chance to travel to conferences to present my work. Typically the first author of each paper gets funded by the university to present their work. The university pays for the conference fees, hotels, and airfare, which adds up to several thousand dollars per conference. With all conferences online, the school only pays for the conference fee (a trivial amount, about $25-$100). Effectively, a major part of grad student compensation has been cut without replacing it with something else.

There are some advantages though. Before the pandemic, it was mandatory to travel to present your work at the conference. This can be at an inconvenient time or location (such as a 20-hour flight to Australia), so I avoided submitting to certain conferences because of this. With virtual conferences, I can submit anywhere without location being a factor.

Another factor is climate change. The IPCC report this year announced that the earth is warming up at an alarming rate, and at an individual level, one of the biggest contributors to greenhouse gas emissions is air travel. Thousands of grad students travelling internationally several times a year adds up to a significant amount of carbon emissions. Therefore, unless there are substantial benefits to meeting in-person, the climate impact alone is probably worth keeping to virtual conferences post-pandemic.

Talks and Posters

Typically in conferences, paper presentations can be oral (12-minute talk with 3 minutes Q/​A), or poster (stand beside a poster for 2 hours). Online conferences mimiced the format: oral presentations were done in Zoom calls, while poster sessions were done in Gather Town (a game-like environment where you move around an avatar). Additionally, most conferences got the authors to record their presentations in advance, so the videos were available to watch at any time during the conference and afterwards.

For me, Gather Town was quite successful at replicating the in-person conference discussion experience. I made a list of people I wanted to talk to, and either attended their poster session if they had one, otherwise logged on to Gather Town several times a day to check if they were online, and then go talk to them. This created an environment where it was easy to enter into spontaneous conversations, without the friction of scheduling a formal Zoom call with them.

The live oral sessions on Zoom were quite pointless, in my opinion, since the videos were already available to watch asynchronously at your own pace. There was no real opportunity for interaction in the 3-minute Q/​A period, so this felt like a strictly worse format than just watching the videos (I usually watch these at 2x speed, which is impossible in a live session). Therefore I didn’t attend any of them.

The paper talk videos were by far the most useful feature of online conferences. Watching a 12-minute video is a good way of getting a high-level overview of a paper, and much faster than reading the paper: I typically watch 5 of them in one sitting. They are available after the conference, leaving a valuable resource for posterity. This is one thing we should keep even if we return to in-person conferences: the benefit is high, while the burden to authors is minimal (if they already prepared to give a talk, it is not much additional effort to record a video).

Collaborations

A commonly stated reason in favor of in-person conferences is the argument that face-to-face interaction is good for developing collaborations across universities. In my experience at in-person conferences, while I did talk to people from other universities, this never resulted in any collaborations with them. Other people’s experiences may vary though, if they’re more extroverted than me.

Virtual meetings certainly puts a strain on collaborations, but this makes more sense for collaborations within an institution. (For example, I haven’t talked to most people in my research group for the last year and a half, and have no idea what they’re working on until I see their paper published). This probably doesn’t extend to conferences though: one week is not enough time for any serious cross-institution collaboration to happen.

Final Thoughts

Like many others, I regret having missed out on so many conferences, which used to be a quintessential part of the PhD experience. There is a positive side: virtual conferences are a lot more accessible, making it feasible to attend conferences in adjacent research areas (like ICML, CVPR), since the $50 registration fee is trivial compared to the cost of travelling internationally.

Many aspects of virtual conferences have worked quite well, so we should consider keeping them virtual after the pandemic. The advantages of time, money, and carbon emissions are significant. However, organizers should embrace the virtual format and not try to mimic a physical conference. There is no good reason to have hours of back-to-back Zoom presentations, when the platform supports uploading videos to play back asynchronously. The virtual conference experience can only get better as organizers learn from past experience and the software continues to improve.

Originally posted on my blog.