Machine Learning applications in Machine Learning applications in HEPHEP

Dr. Carlos Javier Solano SalinasJefe Laboratorio de Altas Energías y Simulación Computacional

Investigador RENACyT de CONCYTECLider UNI en experimentos MINERvA y DUNE en FERMILAB

Escuela Profesional Ciencia de la Computación

rumbo Acreditación

OUTLINE

1. Machine Learning in HEP

2. DCNN and DANN in MINERvA

3. Pandora

4. Feynman Computer Center - FERMILAB

Machine Learning Machine Learning

in High Energy in High Energy

Physics (HEP)Physics (HEP)(Nature Vol 560. 2 ag 2018)

Large Hadron Collider – LHCFrontier between Switzerland and France

Machine learning for calorimetry at CMS. The mass distribution of Z bosons decay (Z → e+e−)

Separating signal from background in ATLAS experiment. BDT-score distribut for search for Higgs decay (H → τ+τ−)

FERMILAB - USA

Neutrino selection and isolation in MicroBooNE

Exploring NOvA’s event-selection neural network using t-distributed stochastic neighbour embedding (t-SNE)

MINERMINERA (A (MMain ain ININjector jector EExpexpeRRiment iment -A-A))

(using DCNN and DANN)(using DCNN and DANN)

Beam-line

MINERvA MINOSDIS2010

NuMI Beamline Graphic courtesy B. Zwaska

Detector MINERA

VetoWall

LHe¼ ton

Blancos Nucleares con He, C, Fe, Pb, H2O,CHEn mismo experimento reduce errores sistematicos entre nucleos

Blanco centellador finamente segmentado ycompletamente activo. 8.3 tons, 3 tons fiducial

120 “modulos” planos. Masa total: 200 tons. Total canales: ~32K

MINOS Near Detector(Muon Spectrometer)

CHALLENGE FOR ANALYSIS IN

PARTICLE PHYSICS THIS CENTURY

So much data, both in channels and in number of events poses

unique challenges

Inspiration from vision and images: use Deep Convolutional

Neural Networks to extract geometric features

Requires a huge number of parallel processes and so was

enabled by the advent of GPUs.

Inspiration from vision and images: domain-adversarial training

Machine Learning algorithms are complicated

MINERvA AT FERMILAB

120 modules for tracking and

calorimetry (32k readout

channels)

MINOS near detector serves as

a muon spectrometer.

Made up of planes of strips in 3

orientations: X, U, and V.

Includes He target, water target

and 5 nuclear targets made up

of C, Fe, Pb

MINERVA VERTEX FINDINGPROCEDURAL ALGORITHM WALKS BACK MAIN TRACK AND USESSECONDARY TRACKS. IN DIS EVENTS LARGE AND COMPLICATEDHADRONIC SHOWERS MAY MASK THE PRIMARY VERTEX

MINERvA VERTEX FINDINGTreat localization as a classification

problem: DNN gives prediction which

segment out of the 11 segments(or which

plane out of 67 plane number) an

interaction is from.

Only 4 planes between most targets (2

planes between targets 4 and 5). That

means only a single U or V view `pixel’.

We want to keep our resolution in z to be

able to find the segment or plane that the

interaction took place in so we use non-

square kernels so that pooling is only

along U, V or X.

ML: DEEP CONVOLUTIONAL NEURAL NETWORKS

Feature extraction is realized within the ML Algorithm possibly in the same layers as the non-linear combination of features or possibly separately. In a NN the output is compared to a loss function (when the neural network is used as a classifier).

It is possible to use the Convolutional NN purely for feature extraction and feed the extracted features into a different sort of MLA for classification or regression.

DEEP NEURAL NETWORKIn a deep NN, the early layers of the

network `learn’ local features while the

later layers `learn’ global features.

This is the `hierarchal model’ where the

representations in early layers are

combined in the later layers.

These deep layers allow for more

complicated combinations of the

features fed in from the inputs.

For deep networks, non-linear layers

are required in order to have an

advantage over a shallow linear

network.

CONVOLUTIONAL NEURAL NETWORK

• These types of networks are well suited for feature extraction for

things like images with geometric structures.

Particle physics events have geometric structures which are

procedural algorithms (or scanners) identify.

• Convolutional networks have fewer parameters that are fit due to

having only a single parameter across the space (for a given kernel).

Parameters describe how the kernel is applied.

• In MINERvA we have time and energy information (obvious use of

`depth’)

● Final convolutional layer is a `semantic’ representation rather than

a spatial representation.

DEEP CONVOLUTIONAL NN FOR VERTEX FINDING

• We have three separate convolutional towers that look at each of the X, U, and Vimages.• These towers feature image maps of different sizes at different layers of depth to reflect the different information density in the different views.• The output of each convolutional tower is fed to fully connected layer, thenconcatenated and fed into another fully connected layer before being fed into the lossfunction.

PandoraPandora

Machine Learning is more than just Neural Networks!!!

Pandora team

FERMILABFERMILAB

Grid Computing - LHC-CERN

Grid Computing - LHC-CERN

FERMILAB - USA

FEYNMAN COMPUTING CENTER

Since 2006 until 2018 we sent 7 master students to FERMILAB to work in their

thesis projects in one year stays (with finantial help of FERMILAB),

They were students from Physics and Engineering Physics.

Now it is time for Computer Science Master students!

Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Slide 22Slide 23Slide 24Slide 25Pandora teamSlide 27Slide 28Slide 29Slide 30Slide 31Slide 32Slide 33Slide 34