NeurIPS is the most popular machine learning conference every year, but unless you’ve been steadily going for a while, you wouldn’t realize how much bigger it was this year (this is my first time, but folks I met noted this sudden spike in interest).

There are now so many papers that it seems impractical to walk around and catch everything in the poster sessions.

I’m glad I stayed for the workshops because it was there that I had the opportunity to speak to paper authors directly, working on ideas actually thematically relevant to my interest area.

I found the cognitive science talk discussing how babies are more sample efficient than models pretty enlightening—it raised many questions, and I think I’ll revisit that talk for the gems. Apparently, it’s very important in babies’ cognitive development to see sharp lines and edges, and babies are more sample efficient because they get to move their heads and adjust their gaze—in effect, choosing the data they’re training on.

I also made many friends: this is the real value of these kinds of conferences. The talks are an excuse to make new friends and catch up with old ones.

When we go next year, we’ll have published something.