CNN with residual learning extensions in neutrino high energy physics

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21340%2F21%3A00353091" target="_blank" >RIV/68407700:21340/21:00353091 - isvavai.cz</a>

Výsledek na webu

<a href="https://doi.org/10.1088/1742-6596/1730/1/012133" target="_blank" >https://doi.org/10.1088/1742-6596/1730/1/012133</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1088/1742-6596/1730/1/012133" target="_blank" >10.1088/1742-6596/1730/1/012133</a>

Jazyk výsledku

angličtina

Název v původním jazyce

CNN with residual learning extensions in neutrino high energy physics

Popis výsledku v původním jazyce

As many reconstruction steps in neutrino high energy physics (HEP) are similar to image pattern recognition tasks, we explore the potential of Convolutional Neural Networks (CNN) combined with residual machine learning algorithm. Characteristic features from neutrino track image pixelmaps are extracted at different scales and these features are used for classification of the type of neutrino interaction. In this contribution, we sumarize observed performance of the residual neural networks (ResNet) for neutrino charged current (CC) interaction detections using image-like Monte Carlo simulated data for muon and electron neutrinos. The two topologies depicted at the neutrino detectors differ, muon neutrino CC interaction is dominated by a slowly ionizing muon, while electron neutrino CC interaction is usually recorded as a wide shower. For the ResNet performance evaluation, we use area under ROC curve (AUC) as the evaluation metric. We observe an improvement while using residual learning compared to general CNN architecture, which is caused by a more stable training with lesser vulnerability to the vanishing gradient of the ResNets. Moreover, stacking other hidden layers within our ResNet model greatly increased the AUC value on the test neutrino dataset without the signs of unstable training or overfitting.

Název v anglickém jazyce

CNN with residual learning extensions in neutrino high energy physics

Popis výsledku anglicky

As many reconstruction steps in neutrino high energy physics (HEP) are similar to image pattern recognition tasks, we explore the potential of Convolutional Neural Networks (CNN) combined with residual machine learning algorithm. Characteristic features from neutrino track image pixelmaps are extracted at different scales and these features are used for classification of the type of neutrino interaction. In this contribution, we sumarize observed performance of the residual neural networks (ResNet) for neutrino charged current (CC) interaction detections using image-like Monte Carlo simulated data for muon and electron neutrinos. The two topologies depicted at the neutrino detectors differ, muon neutrino CC interaction is dominated by a slowly ionizing muon, while electron neutrino CC interaction is usually recorded as a wide shower. For the ResNet performance evaluation, we use area under ROC curve (AUC) as the evaluation metric. We observe an improvement while using residual learning compared to general CNN architecture, which is caused by a more stable training with lesser vulnerability to the vanishing gradient of the ResNets. Moreover, stacking other hidden layers within our ResNet model greatly increased the AUC value on the test neutrino dataset without the signs of unstable training or overfitting.

Druh

J_ost - Ostatní články v recenzovaných periodicích

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

Výsledek vznikl pri realizaci vícero projektů. Více informací v záložce Projekty.

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)<br>S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2021

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název periodika

Journal of Physics Conference Series

ISSN

1742-6588

e-ISSN

—

Svazek periodika

1730

Číslo periodika v rámci svazku

1

Stát vydavatele periodika

GB - Spojené království Velké Británie a Severního Irska

Počet stran výsledku

6

Strana od-do

—

Kód UT WoS článku

—

EID výsledku v databázi Scopus

2-s2.0-85101573067