This thesis is focused on the development of novel approaches to improve the explainability of Deep Neural Network models in High Energy Physics to reduce the size of a certain model by dropping irrelevant input information. We show that it is possible to reduce the size of a signal-background classification problem by automatically ranking the relative importance of available particle jet input features. Variables are importance-sorted with a decision tree algorithm. The selected features can be used as input quantities for the classification problem at hand. A k-fold cross-validation is applied to raise the confidence in the extracted ranking. On the same line, a new Neural Network layer, called CancelOut, is presented as a tool to reduce the input parameter size by keeping the performance the highest during the training of the model. Both strategies are tested with the case of highly boosted di-jet resonances decaying to two b-quarks, to be selected against an overwhelming QCD background with a Deep Neural network. The data are produced via a pseudo experiment simulation.
Deep Learning Models Resizing for High Energy Physics experiments / Di Luca, Andrea. - (2022 Apr 14), pp. 1-141. [10.15168/11572_338814]
Deep Learning Models Resizing for High Energy Physics experiments
Di Luca, Andrea
2022-04-14
Abstract
This thesis is focused on the development of novel approaches to improve the explainability of Deep Neural Network models in High Energy Physics to reduce the size of a certain model by dropping irrelevant input information. We show that it is possible to reduce the size of a signal-background classification problem by automatically ranking the relative importance of available particle jet input features. Variables are importance-sorted with a decision tree algorithm. The selected features can be used as input quantities for the classification problem at hand. A k-fold cross-validation is applied to raise the confidence in the extracted ranking. On the same line, a new Neural Network layer, called CancelOut, is presented as a tool to reduce the input parameter size by keeping the performance the highest during the training of the model. Both strategies are tested with the case of highly boosted di-jet resonances decaying to two b-quarks, to be selected against an overwhelming QCD background with a Deep Neural network. The data are produced via a pseudo experiment simulation.File | Dimensione | Formato | |
---|---|---|---|
thesis_diluca.pdf
accesso aperto
Tipologia:
Tesi di dottorato (Doctoral Thesis)
Licenza:
Tutti i diritti riservati (All rights reserved)
Dimensione
25.33 MB
Formato
Adobe PDF
|
25.33 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione