Pattern recognition

→ Europe/Paris
LLR

LLR
Naomi van der Kolk (LAL), Roman Poeschl (LAL Orsay)
Description
meeting on pattern recognition
Participants
  • Balázs Kégl
  • Jean-Claude Brient
  • mehdi Cherti
  • Naomi van der Kolk
  • Roman Poeschl
  • Sviatoslav Bilokin
  • Trong Hieu TRAN
  • Vincent Boudry
  • Vladik Balagura
  • Yacine HADDAD
Cacher
Sviatoslav presented his progress up to now.
He explained his secondary track finding algorithm.
 
We  have to move now to the info directly from the MC truth, as Mehdi is needing this data.
 
We should also think about a list of observables from MC:
- they should be observable
- and they should be usefull
- e.g. outgoing charged particles (pions, muons)
- For the secondaries we should cut on the distance they have traveled, or on their energy or both
- e.g. interaction point
- e.g. the energy in the interaction region
 
For Mehdi and Balazs a proportion map of pixel energies for overlayed events would be usefull. They will make this themselves as they have all the data available.
 
I should give my visualisation macro to Mehdi, so he can see how the events look like.
 
A question from Balazs: What is the visual human error rate to classify interactions? Can we do a test?
 
Balazs introduced the concept of a data challenge: Can we define a data challenge for the machine learning community?
 
Mehdi presented his progress. He is using convolution networks. These are very well suited for images. It uses filters which sweep over the image.
The system (neural network) has to learn these filters for our problems.
It is possible to process with many filters at the same time and also with several layers of filters on top of one another.
This is a rather new technology, and it needs visualisation to help with debugging.
It does not beat AdaBoost yet, but gives 10% error rate at 2 GeV.
Which is already rather good and better than the "manual" analysis in the paper.
Next step could be to find the interaction in the convolution network.
 
Mehdi has seen some events which are inelastic but look just like a MIP, how is this possible?
Should we change the definition of the inelastic label?
Mehdi needs also data at 10 GeV. He will look at that first and then continue on the interaction point determination.
Il y a un compte-rendu associé à cet événement. Les afficher.