Deep Neural Network Applications to Proton Decay Analysis

Food for Thoughts

Deep Neural Network Applications

to Proton Decay Analysis

Kazuhiro Terao SLAC National Accelerator Laboratory

1

Deep Neural Network Applications• Two popular types: CNN & RNN - CNN actively developed in computer vision (CV) - RNN in natural language processing (NLP) - Both are popular in physics analysis

2

K+

µ+

e+pµ+

e+

Deep Neural Network Applications• Two popular types: CNN & RNN - CNN actively developed in computer vision (CV) - RNN in natural language processing (NLP) - Both are popular in physics analysis

• What can they do? - Image classification - Object detection - Pixel segmentation - Clustering - Hierarchy/Graph

V.S.

3

Deep Neural Network Applications• Two popular types: CNN & RNN - CNN actively developed in computer vision (CV) - RNN in natural language processing (NLP) - Both are popular in physics analysis

• What can they do? - Image classification - Object detection - Pixel segmentation - Clustering - Hierarchy/Graph

4

Deep Neural Network Applications• Two popular types: CNN & RNN - CNN actively developed in computer vision (CV) - RNN in natural language processing (NLP) - Both are popular in physics analysis

• What can they do? - Image classification - Object detection - Pixel segmentation - Clustering - Hierarchy/Graph “shower” vs. “track”

separation

5

Deep Neural Network Applications• Two popular types: CNN & RNN - CNN actively developed in computer vision (CV) - RNN in natural language processing (NLP) - Both are popular in physics analysis

• What can they do? - Image classification - Object detection - Pixel segmentation - Clustering - Hierarchy/Graph

This is work in progress…

6

Deep Neural Network Applications• Two popular types: CNN & RNN - CNN actively developed in computer vision (CV) - RNN in natural language processing (NLP) - Both are popular in physics analysis

• What can they do? - Image classification - Object detection - Pixel segmentation - Clustering - Hierarchy/Graph

7

Proton Decay: Example I“ROI” Image Classification

K+

µ+

e+

pµ+

e+

Very popular “catch-all” approach. There is no reason why this cannot be combined with BDT or a different approach

(they can be complimentary discriminators)8

Proton Decay: Example IIVertex reconstruction

Q: What is the vertex reconstruction efficiency? If not great for traditional methods, this can be improved. Crucial for good clustering, thus directionality estimate.

9

Proton Decay: Example IIIMichel identification

MicroBooNE demonstrated Michel ID with high purity. Segmentation (+ point prediction) can help to improve

Michel electron search efficiency.10

Proton Decay: Example IVµ/π separation

Q: Is this useful? Study with 3mm wire pitch toy simulation (mimiced MicroBooNE) could separate µ/π at ~80% level

π- µ-

11

Proton Decay: Example VOptical signal analysis

You can surely apply DNNs on optical data analysis

Sorry, No Picture (lack of effort)

Please Imagine a Paradise

12

Venues in Techniques

All 2D techniques mentioned so far can benefit from providing multiple projection images with a proper network architecture.

Multi-plane Network

3D Pattern Recognition3D is the solution to avoid much of difficulties caused by 2D projections. All techniques mentioned so far can be applied in 3D as a simple extension.

Courtesy of Laura Domine (SLAC grad. student) Presented @ Nu2018

13

Introduction to ML@SLAC

Me Laura Domine Tracy Usher

SLAC Reconstruction Developers

ML Focused

Noah SailerSummer Student

Proton DecayGraduate Student

Generic RecoPI/Ass. Scientist

Generic RecoScientist

Generic Reco

orga

nize

r Faculty Knu subgroup lead

HiroStudent mentor

Physics analysis planningPlease feel free to contact me if you are interested in or need help for developing DL techniques. It’s part of DLP activity!

14

Next Meeting: Noah’s Summer ReportNoah has studied to apply CNNs for identifying multiple K+ decay modes as well as pµ+ topology to see how well separation can be made and investigate how to improve the performance from “simple” training approach.

Early findings include a sign of the network learning “mono-energetic µ+” and systematic comparison to show improvement by using multiple planes.

µ+ kinetic energy from mis-ID pµ+ events

Courtesy of Noah Sailer Using 2-plane network Delivered @ 5:30AM Today

15

Images for Fun (found within 50 events)

K+ decay pµ+ decay16

Typical “difficult” topology pµ+ decay mimics K+

Images for Fun (found within 50 events)

17

Radiative gamma ray from Michel

Images for Fun (found within 50 events)

18

K+ decay into π+π0

Image classification can identify this mode almost as good as µ+ decay mode (see Noah’s talk @ next meeting)

Images for Fun (found within 50 events)

19

K+ decay into e+π0π0

Pretty hard, in particular e+ radiated right away

Images for Fun (found within 50 events)

20

Images for Fun

K+ interacted with nucleus, produced 2x π0, then decayed into µ+

21