NeurIPS warm up

The first “real” day on NeurIPS has been like a warm up. There has been three tutorial sessions followed by two invited talks. I have not been bussy at all.

I just sat down at the first tutorial talk that I wanted to hear, when I saw Lior Kamma and Casper Freksen walking passed me two meteres ahead of me. I joined them and we followed each other most of the day, as none of the tutorials was of particular interest for any of us.

From the morning, we heard a talk about networks that could be made robust to adversarial attacks (Guess it relates to this article [1]). Even bounds for the generalization error on new data (at most $\epsilon$ away from the images in the training set for some measure $L_p$) had a proven worst case, which was pretty cool. Afterwards, we heard an other talk about fairness of models, which was in a room with very bad acoustics. We didn’t get much from that talk. Then we heard a talk by Susan Athey from Stanford that I personally had very high expectations to, as it was on the topic ‘Counterfactual Inference’. However, she never really got into the flesh of the topic and so I didn’t really learn anything new from that talk. It was mostly about RCTs, Diff-in-diff, etc. from econometrics, which I already knew.

Being a bit tiered of talks, I went to collect some swag with Lior and Casper. I got t-shits, a scarf, a cap, an expensive pen, a thermo bottle and some stickers:

Finally, we heard the opening talk of the conference and a talk by the CEO of Atipica, Laura Gomez, before we went out to get something to eat.

[1] Kolter, J.Z. and Wong, E., 2017. Provable defenses against adversarial examples via the convex outer adversarial polytope. arXiv preprint arXiv:1711.00851, 1(2), p.3.