Denis Uzhva at NOvA

From AstroNuWiki
Revision as of 20:25, 31 August 2018 by OlegSamoylov (Talk | contribs)

(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Working time :

  • 6 weeks July-August 2018 within SSP-2018

Advisors :

Tasks description

Title: Investigation of Deep Learning methods for the classification of events in the NOvA experiment

Abstract: With the rise of questions concerning the problems of the theory behind the phenomenon of neutrino oscillations, more and more neutrino experiments are under development and underway. One such experiment, NOvA, is supposed to tell the scientific comunity more about ν_µ → ν_e (as well as anti-ν_µ → anti-ν_e) oscillations, ν_µ and anti-ν_µ disappearance channels, to determine the order of neutrino masses, CP-violation phase in the lepton sector, and to measure precisely mixing angles, as well as ∆m^2 (squared mass differences). A brand new approach of using convolutional neural networks has been applied in order to improve the quality of NOvA's data analysis, bypassing the standard event reconstruction procedures. The idea of combining the reconstruction with Convolutional Neural Networks (CNN) is interesting, as it may help with increasing the degree of precision and perfomance speed. Since the data is generally composed of particle tracks, emitted from an interaction point, we have attempted to train NOvA CNN on images that have had a polar transformation applied to each event, with the transformation origin located at the event's interaction vertex. In this way tracks emitted from the vertex will be represented by a horisontal line, perhaps presenting an easier data set for the network to learn from. We have shown that such a dataset decreases the overtraining of the network, though, slightly reducing the validation accuracy. Therefore, the future development of this method may be fruitful. Moreover, after removing parallel inception layers, a simplified version of the original NOvA network has been shown to be much faster with little cost in accuracy. Therefore, simpler networks should also be investigated further.

Details of his work will be appeared soon here.