The state of machine learning research moves incredibly fast. There are dozens of new papers published on arxiv every day, and it’s overwhelming trying to keep up to date. Conferences are a great way to get a lot of signals in a condensed amount of time. Conference papers and sessions are pretty vetted, which saves you from judging each new publication that comes out. As someone more recent to the field of AI research, I think it’s super beneficial to try to go to conferences, even as just a participant. I was lucky to go to NeurIPs last year with my previous startup.
One thing I like to do before conferences, though this might get harder and harder to do as conference size increases and the number of accepted papers increases, is skim all the abstracts of papers I find interesting. I’ll do an initial run through based on titles, keywords, and author affiliations to determine if I want to read the abstracts. I’m pretty generous at this stage. In the next step, I’ll read through the abstracts of all the papers I’ve selected. Then from there, I’ll (skim) read the papers of the abstracts I’ve found interesting. This process then gives me a shortlist of papers and authors I’ll try to seek out during poster sessions. Poster sessions are a great way to meet researchers and ask questions. Last year people were very excited to share their work and answer any small or big questions you may have.
For me, the most fun and educational bits of conferences are when I get to engage in 1:1 or small discussions with researchers. I’ll try to look for events happening during the meeting that are interest-oriented or find events where researchers I admire are speaking at. For example, last year, an AI Safety Unconference was an unofficial event and had fascinating discussions. Conference workshops are also a great way to find experts in a specific subject.
Conferences, whether or not they are virtual, are intimidatingly packed with events. NeurIPs this year has events 24/7 through the week. I like to spend some time figuring out which events I want to prioritize and optimize for discussion-oriented events first. Most talks are recorded, so you can always catch those later (at 2x speed) on your own time.
I’ll update this post with tips for how to navigate online conferences after next week’s virtual NeurIPs.