Hsu:

Congratulations for the paper submission! It's great to see nice presentations to this analysis using several novel ML at jet level, and analysis level.

If possible, a big take away of this talk would be to quantify how much improvement of search sensitivity or systematic deduction, or MC/data agreement thanks to ML techniques, in comparison to the previous version of analysis.

Here are a few slide by slide comments:
p3 Cosmic ray background was mentioned once here, but seems disappear in the rest of talk. How relevant is it? Nice to mention it.

p6 It's good to briefly describe jet clustering algorithm and radius.

     Definition of "Control Region" per jet. Normally, CR is defined as event level. It's interesting that you have a per-jet definition  in a CR

     Is the MC in the CR = multi-jet + BIB + Cosmic ray?

p9/p10  Sum dR_min is a critical variable in this analysis. It will be nice to define it in a cartoon earlier

p11 Here is a good place to clarify how much improvement coming from ML in addition to luminosity gain (more data)
 

p14 Backup

One nature questions that audiences would like to know is what variables go into BDT, and how good MC/data agreement of those input variable.

The other question is whether you have sufficient statistics of events in MC and Data for the  per-jet NN adversarial training.