Pattern recognition

mardi 1 avr. 2014, 15:00 → 16:30 Europe/Paris

100 bat. 208 (LAL)

Naomi van der Kolk (LAL), Roman Poeschl (LAL Orsay)

meeting on pattern recognition

Mehdi presented his project and current results and the next steps. The results which Frank has gotten before are not reproducible. The best results so far is an error rate of 8.5% on classifying inelastic/elastic events. The performance of the algorithms depends on the satting of internal parameters (hyper parameters) and these should be optimized. Plans are to try some new (other) algorithms for deep learning.

 

Mehdi will need at some point to receive data with for each event an array of properties which the algorithms can then predict. These properties (labels) can be: interaction layer, number of produced secondaries, interaction process, incoming energy, number of incoming particles, etc...

 

Naomi discussed her progress with the generative model. A first working version in RooFit is there, which uses a multi-variate Gaussian. This is very slow. As once the mean for each pixel has been determined, they are not dependent, we should use independent Gaussians which should be much quicker. The sigma of the energy deposition in x,y is governed by multiple scattering. Once we have a good model for one Mip, we should add a second Mip, or a "blob" to the Mip.

 

Sviatoslav explained his work on finding the number of outgoing tracks (secondaries) by clustering. This can be a feature for Multiboost. He will also get the number of secondaries directly from GEANT4. These will be input also for Mehdi.

1 Progress report Mehdi

Slides MeetingILC_-_Apr01_2014.pdf

2 Progress report Naomi

3 Progress report Sviatoslav