Dear All,
I forward to you a recent message from our ILD guest members working on strange tagging, which should be interesting
for many of you for various reasons.
Since the whole complex of charged hadron ID (dE/dx, ToF) - strange tagging - flavour tagging with ML is creating quite some activity,
I'm wondering whether it would make sense to form a dedicated team / task force / .. in this area - probably that would however presumably
happen under the guidance of the new physics coordination team.
Concerning their question about the fixed dE/dx likelihoods, we have to tell them to wait for the upcoming iLCSoft release, I guess.
For the track impact parameters, I see in principle two possibilities:
a) tell them to switch to the DST
b) add the track impact parameters to the miniDST, eg in the reference point of the charged PFO or as additional parameters,
and then train them to re-do their own miniDSTs
I have a slight preference for a), because it is probably only a matter of time until they realize that they want more track information,
or need to improve V0 finding, or, or, or. I guess this is an agenda item for the next software convener meeting.
Cheers,
Jenny.
-------- Forwarded Message --------
| Subject: |
Status of hadronic Higgs samples with corrected dE/dx? |
| Date: |
Thu, 18 Feb 2021 15:47:07 -0800 |
| From: |
Matthew <matthew.joseph.basso@cern.ch> |
| To: |
Jenny List <jenny.list@desy.de> |
| CC: |
Valentina Cairo <valentina.maria.cairo@cern.ch>, Jan Fridolf Strube <jstrube@uoregon.edu>, Tomohiko Tanabe <tomohiko.tanabe@kek.jp> |
Hi Jenny,
I hope you've been well!
I just wanted to follow up on the status of the corrected samples from the ILD analysis meeting on January 13th: have there been any updates? As far as we understand, Jan has already contacted you about this. Ideally, the samples would include the dE/dx likelihoods, the b/c/other tagger info per jet, and the impact parameters for the associated tracks of the particles in each jet (only available in DSTs, afaik). It might be a lot to ask for the IPs, but it could be potentially useful - maybe a smaller production with fewer statistics for s, u, and d jets would be sufficient at first just to check if there's any discriminating power?
We've got further along with our own studies. We're in the process of understanding and training the EdgeConv network. We've also fully trained a smaller network relying on RNNs (in particular, GRUs) for the 50 leading particles in a jet combined with an MLP for the jet-level variables. The network uses the jet's kinematics and tagger scores as well as the kinematics, charge, and number of tracks/clusters per particle in the jet. I've attached some performance plots.
We have discrimination for b and c jets (expected, the NN should do as good as the tagger scores) as well as some discrimination for g and s+u/d jets. But we unfortunately have no separation power for s versus u/d at the moment. But I suppose this is what the likelihood info should help with ;)
Thanks for all of your help, it is appreciated!
Best,
Matt
