I attended the NeurIPS (recently renamed from NIPS) conference this year and here are some observations:
- I find it remarkable how for many people machine learning is equivalent to deep learning and deep learning is equivalent to computer vision.
- Susan Athey talked about causal inference. She’s been doing a lot of work combining econometrics with machine learning, but it was very interesting to see this presented from the other side of the fence. When she asked who knew about instrumental variables, a reasonable number of hands (maybe 5% of a very large crowd) went up.
- Edward Felten (Princeton) talked about how experts should talk to decision makers. Basically, don’t come with the “truth” and don’t provide “just the facts”. Structure the decision problem by inquiring about the decision makers preferences and provide information and an option menu of recommendations based on that.
- Many of the sessions touched on economics: There were papers on market design, value function iteration or - more close to home - on the social effects of automation.
Differences to economics conferences:
- Many more consumers than producers of research
- Different demographics: Younger, more nerdy, fewer women, more Asian
- Fewer parallel sessions, instead “tutorials” (really lectures) by famous people that last 1.5 hours
- Videos/GIFs on slides. Works suprisingly well and engages people. Examples include simulations of how fruit flies move, 3d shapes circling around, bubbling formations of networks
- The anthropomorphizing is a bit weird: They say “The algorithm should be able to reason about its own uncertainty”, instead of “We need accurate standard errors”
- The money involved is different. We don’t have huge recruiting parties by Uber or have Yo-Yo Ma lead a session on the intersection of music and deep learning.
Montreal is the way I wish France was. People are laid back and cosmopolitan. Everything is in French and people switch between languages without effort.