*-.5cm machine learning in jet physics16 deep learning on jet-images i jet images of size 33 33. i 3...
TRANSCRIPT
![Page 1: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/1.jpg)
MACHINE LEARNING IN JET PHYSICS
Sreedevi Narayana Varma
April 11, 2018
![Page 2: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/2.jpg)
1Content
Introduction
Quark/Gluon Jets
Convolutional Neural Networks
Deep Learning on Jet-Images
Results and Future WorksDomain Adversarial Neural Netwrok (DANN)
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 3: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/3.jpg)
2Introduction
I Jets are collimated streamof particles produced byparticle collisions.
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 4: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/4.jpg)
![Page 5: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/5.jpg)
4
1. Protons collide at high energies.2. Quarks and gluons produced are hadronized.3. The decay products are clustered into jets using
algorithms.
Is the jet quark initiated or gluon initiated?
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 6: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/6.jpg)
5
I In many searches for new physics signals at the LHC, jetsare initiated by light-flavor quarks (u, d, s), while the jetsbackground processes are initiated by gluons.
I We are currently working on,
I Classification of quark and gluon initiated jets.
I Bias between the event generators.
I Bias between real and simulated data.
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 7: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/7.jpg)
6Quark/Gluon Jets
Figure: Quark jet
I Quark initiated jets arenarrower
CF =43
Figure: Gluon jet
I Gluon initiated jets aremore wide.
CA = 3
B.R.Webber, Quark and Gluon Jets in Quantum Chromodynamics, Physica Scripta, vol 25, no 1B, p 198, 1982Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 8: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/8.jpg)
Jet images are 2D representation of energy deposits in thecalorimeter.
![Page 9: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/9.jpg)
8
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 10: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/10.jpg)
9Processes
I Quark-initiated jetsI qq→ qqI qq̄ → qq̄I gg→ qq̄
I Gluon- initiated jetsI gg→ ggI qq̄ → gg
P. T. Komiske, E. M. Metodiev, M. D. Schwartz, Deep learning in color: towards automated quark/gluon jet
discrimination, arXiv:1612.01551 [hep-ph]Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 11: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/11.jpg)
10Parameters
I Radius = 0.4I Pseudorapidity |η| < 2.5I Transverse momentum in ranges 100-110 GeV, 200-220
GeV, 500-550 GeV and 1000-1100 GeV.
P. T. Komiske, E. M. Metodiev, M. D. Schwartz, Deep learning in color: towards automated quark/gluon jet
discrimination, arXiv:1612.01551 [hep-ph]Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 12: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/12.jpg)
11
Preproccessing
1. Centering: The jet is rotated and boosted so that thecentral pixel is at (0,0).
2. Crop: Crop the image with (η, φ)ε(−R,R).
3. Normalize: Total pixel intensity of the image is∑
Iij = 1.
4. Zero-center: Iij → Iij − µij , where µij is the average of thetraining set.
5. Standardize: Iij → Iij/(σij + r) where σij is the standarddeviation of the training set and r = 10−5.
P. T. Komiske, E. M. Metodiev, M. D. Schwartz, Deep learning in color: towards automated quark/gluon jet
discrimination, arXiv:1612.01551 [hep-ph]Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 13: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/13.jpg)
12
(a) Average gluon imagebefore preprocessing steps 4-5
(b) Average quark imagebefore preprocessing steps 4-5
Figure: Jet-Images
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 14: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/14.jpg)
13
(a) Average gluon image afterpreprocessing
(b) Average quark image afterpreprocessing
Figure: Jet-Images
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 15: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/15.jpg)
14Convolutional Neural Networks
I Convolutional Neural Network (CNN) are neural networksfor image recognition and image classification.
I CNN scans over the two dimensional pixel intensities of anRGB image.
Figure: Convolutional Neural Network
P. T. Komiske, E. M. Metodiev, M. D. Schwartz, Deep learning in color: towards automated quark/gluon jet
discrimination, arXiv:1612.01551 [hep-ph]Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 16: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/16.jpg)
15
Figure: Components of a CNN
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 17: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/17.jpg)
16
Deep Learning on Jet-Images
I Jet images of size 33× 33.I 3 convolutional layer and 2 fully connected layer.I ReLU (Rectified Linear Unit) activation.
Figure: ReLU and Sigmoid activations
I Filters of size 8× 8, 4× 4and 4× 4 are used.I Maxpooling layers 2× 2 is also applied to the CNN with a
stride length of 2.I The fully connected layer consists of 128 units.
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 18: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/18.jpg)
17
I The jet images produced are fed into the network asgrayscale images.
I An additional information of charge is applied to improvethe accuracy of the model.
I “Colour” images with colour channels,I RED = Transverse momentum of charge particles.I GREEN = Transverse momentum of neutral particles.I BLUE = Charge particle multiplicity.
I The “coloured” images are then trained.
P. T. Komiske, E. M. Metodiev, M. D. Schwartz, Deep learning in color: towards automated quark/gluon jet
discrimination, arXiv:1612.01551 [hep-ph]Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 19: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/19.jpg)
18
I CNN is trained on Tensorflow using NVIDIA GeForce1080Ti GPU on Cuda 9.0 platform.
I The network is trained over 100 epochs with a learningrate α of 0.001.
I 180000 jet images are used for training, 20000 images forvalidation and 40000 images are used for testing.
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 20: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/20.jpg)
19Results and Future Works
I Pythia test images are better than Herwig test images.
Comparison of machine learning algorithms trained on Herwig and Pythia.
Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 21: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/21.jpg)
20
Domain Adversarial Neural Netwrok (DANN)
I Domain adversarial neural network is a new learningapproach for data trained and tested on similar but differentdistributions.
Figure: DANN architecture
Y. Ganin et al. , Domain-Adversarial Training of Neural Networks, arXiv:1505.07818 [stat]Sreedevi Narayana Varma | MACHINE LEARNING IN JET PHYSICS
![Page 22: *-.5cm MACHINE LEARNING IN JET PHYSICS16 Deep Learning on Jet-Images I Jet images of size 33 33. I 3 convolutional layer and 2 fully connected layer. I ReLU (Rectified Linear Unit)](https://reader034.vdocuments.site/reader034/viewer/2022042312/5edace623ee6fd189f66e638/html5/thumbnails/22.jpg)
Thank you!