quantum computing / machine learning...why machine learning? • much of particle physics and...

Quantum Computing / Machine Learning

Christian Bauer

Why machine learning?• Much of particle physics and cosmology is about dealing

with very large data sets, and ML can help• Distinguish BSM(signal) from SM(background)• Model independent analyses• Simulating the detectors• Real time analysis / triggers

• ML learning research ranges from applying state of the art techniques to developing new methods

We have several people involved in machine learning for HEP

Andreasen (Theory), Bhimji (NERSC), Calfiura (CRD), Farrell (NERSC), Gray (ATLAS), Moult (Theory), Nachman (ATLAS), Seljak (Cosmo), Wang (ATLAS)

Why quantum computing?

• HEP is continuously pushing the computing frontier• Need to obtaining precise theory predictions• Analyzing and reconstructing data• Important calculations often lack computing resources

• Scaling up resources by linear factors does not get us where we would like to be

• Hope is that quantum computing could eventually provide exponential increase in computing power

Rapidly growing field, with very strong roots in Berkeley (too many people to list).

Efforts range from building quantum hardware, building quantum sensors and developing quantum algorithms

A few examples…

Classification

RegressionGeneration

arbitrarily many

categories

map noise to structure

provide examples for training

full supervision / weak supervision

J. Collins, K. Howe, Nachman, PRD 99 (2019) 014038J. Collins, K. Howe, Nachman, PRL 121 (2018) 241803P. Komiske, E. Metodiev, Nachman, M. Schwartz, PRD 98 (2018) 011502 E. Metodiev, Nachman, J. Thaler, JHEP 10 (2017) 51 L. Dery, Nachman, F. Rubbo, A. Schwartzman, JHEP 05 (2017) 145

K. Datta, A. Larkoski, Nachman, 1902.07180 J. Lin, M. Freytsis, I. Moult, Nachman, JHEP 10 (2018) 101 L. de Oliveira, Nachman, M. Paganini, 1806.05667 Z. Jiang, Nachman, F. Rubbo, ATL-PHYS-PUB-2017-017 L. de Oliveira, M. Kagan, L. Macky, Nachman, A. Schwartzman, JHEP 07 (2016) 069 W. Bhimji, S. Farrell, T. Kurth, M. Paganini, Prabhat, E. Racah, 1711.03573

M. Paganini, L. de Oliveira, Nachman, CSBS 1 (2017) 1 M. Paganini, L. de Oliveira, Nachman, PRD 197 (2018) 014021 M. Paganini, L. de Oliveira, Nachman, PRL 120 (2018) 042003 J. Lin, W. Bhimji, Nachman, 1903.02556

A. Cukierman and Nachman, NIMA 858 (2017) 1 P. Komiske, E. Metodiev, Nachman, M. Schwartz, JHEP 12 (2017) 51 A. Cukierman, S. Haghighat, Nachman, ATL-PHYS-PUB-2018-013 S. Farrell et al. 1810.06111

BERKELEY EXPERIMENTAL PARTICLE PHYSICS

Machine learning in HEP

Classification


arbitrarily many

categories









Machine learning in HEP4

2000 2500 3000 3500 4000

mJJ / GeV

10�1

100

101

102

103

104

105

106

Eve

nts

/10

0G

eV

Signalregion

SidebandSideband

2500 3000 3500

10�12

10�10

10�8

10�6

10�4

10�2

100

p 0

3�

4�

5�

6�

7�

3�

4�

5�

6�

7�

mJJ / GeV

No signal

2500 3000 3500

10%

1%

0.2%

With signal

FIG. 1. Left: mJJ distribution of dijet events (including injected signal, indicated by the filled histogram) before and afterapplying jet substructure cuts using the NN classifier output for the mJJ ' 3 TeV mass hypothesis. The dashed red linesindicate the fit to the data points outside of the signal region, with the gray bands representing the fit uncertainties. Thetop set of markers represent the raw dijet distribution with no cut applied, while the subsequent sets of markers have cutsapplied at thresholds with e�ciency of 10�1, 10�2, 2⇥ 10�3, and 2⇥ 10�4. Right: Local p0-values for a range of signal masshypotheses in the case that no signal has been injected (left), and in the case that a 3 TeV resonance signal has been injected(right). The dashed lines correspond to the case where no substructure cut is applied, and the various solid lines correspondto cuts on the classifier output with e�ciencies of 10�1, 10�2, and 2⇥ 10�3.

to a level of discovery. There are many other possibili-ties for applying this technique directly to data, in anycase where the signal is expected to be localized in onedimension. By naturally exploiting the power of modernmachine learning, we hope that this extended bump huntwill help to expose new distance scales in nature on thequest for BSM at the LHC and beyond.

The datasets and code used for the case study can befound at Refs. [48, 49].

We appreciate helpful discussions with and useful feed-back on the manuscript from Timothy Cohen, AvivCukierman, Patrick Fox, Jack Kearney, Zhen Liu, EricMetodiev, Brian Nord, Bryan Ostdiek, Francesco Rubbo,and Jesse Thaler. We would also like to thank PeizhiDu for providing the UFO file for the benchmark sig-nal model. The work of JHC is supported by NSFunder Grant No. PHY-1620074 and by the MarylandCenter for Fundamental Physics (MCFP). The work ofB.N. is supported by the DOE under contract DE-AC02-05CH11231. This manuscript has been authored byFermi Research Alliance, LLC under Contract No. DE-AC02-07CH11359 with the U.S. Department of Energy,O�ce of Science, O�ce of High Energy Physics. TheUnited States Government retains and the publisher, byaccepting the article for publication, acknowledges thatthe United States Government retains a non-exclusive,paid-up, irrevocable, world-wide license to publish or re-produce the published form of this manuscript, or allowothers to do so, for United States Government purposes.

⇤ [email protected]† [email protected]‡ [email protected]

[1] Georges Aad et al. (ATLAS), “Observation of a new par-ticle in the search for the Standard Model Higgs bo-son with the ATLAS detector at the LHC,” Phys. Lett.B716, 1–29 (2012), arXiv:1207.7214 [hep-ex].

[2] Serguei Chatrchyan et al. (CMS), “Observation of a newboson at a mass of 125 GeV with the CMS experi-ment at the LHC,” Phys. Lett. B716, 30–61 (2012),arXiv:1207.7235 [hep-ex].

[3] ATLAS Collaboration, “Supersymmetry searches,”(2018), https://twiki.cern.ch/twiki/bin/view/

AtlasPublic/SupersymmetryPublicResults.[4] ATLAS Collaboration, “Exotic physics searches,”

(2018), https://twiki.cern.ch/twiki/bin/view/

AtlasPublic/ExoticsPublicResults.[5] CMS Collaboration, “Cms exotica public physics re-

sults,” (2018), https://twiki.cern.ch/twiki/bin/

view/CMSPublic/PhysicsResultsEXO.[6] CMS Collaboration, “Cms supersymmetry physics re-


view/CMSPublic/PhysicsResultsSUS.[7] CMS Collaboration, “Cms beyond-two-generations (b2g)

public physics results,” (2018), https://twiki.cern.ch/twiki/bin/view/CMSPublic/PhysicsResultsB2G.

[8] see e.g. Ellis, John R., “Beyond the standard model withthe LHC,” Nature 448, 297–301 (2007).

[9] A model independent general search for new phenomenawith the ATLAS detector at

ps = 13 TeV, Tech. Rep.

ATLAS-CONF-2017-001 (CERN, Geneva, 2017).[10] Model Unspecific Search for New Physics in pp Collisions

4

2000 2500 3000 3500 4000

mJJ / GeV

10�1

100

101

102

103

104

105

106

Eve

nts

/10

0G

eV

Signalregion

SidebandSideband

2500 3000 3500

10�12

10�10

10�8

10�6

10�4

10�2

100

p 0

3�

4�

5�

6�

7�

3�

4�

5�

6�

7�

mJJ / GeV

No signal

2500 3000 3500

10%

1%

0.2%

With signal





















4

2000 2500 3000 3500 4000

mJJ / GeV

10�1

100

101

102

103

104

105

106

Eve

nts

/10

0G

eV

Signalregion

SidebandSideband

2500 3000 3500

10�12

10�10

10�8

10�6

10�4

10�2

100

p 0

3�

4�

5�

6�

7�

3�

4�

5�

6�

7�

mJJ / GeV

No signal

2500 3000 3500

10%

1%

0.2%

With signal





















4

2000 2500 3000 3500 4000

mJJ / GeV

10�1

100

101

102

103

104

105

106

Eve

nts

/10

0G

eV

Signalregion

SidebandSideband

2500 3000 3500

10�12

10�10

10�8

10�6

10�4

10�2

100

p 0

3�

4�

5�

6�

7�

3�

4�

5�

6�

7�

mJJ / GeV

No signal

2500 3000 3500

10%

1%

0.2%

With signal





















Given that the LHC has found nothing, it is critical that we

broaden our search program.

Weak supervision are a set of techniques to train on unlabeled data.

Classification


arbitrarily many

categories









Machine learning in HEP9

Depth-weighted total energy ld

FIG. 12: Comparison of shower shape variables, introduced in Table IV, and other variables of interest, such as thesparsity level per layer, for the Geant4 and CaloGAN datasets for e

+, � and ⇡+.

Physics simulation is slow, neural network evaluation is fast.

Generative models use neural network to enhance / extend

physics-based models.

Classification


arbitrarily many

categories









Machine learning in HEP

Constructing the graph

• Select hits in neighborhood of true track

• Label the seeds

• Target is the true binary labels for every hit

21

Hit classification Segment classification

• Connect hits on adjacent layers using crude geometric constraints

• delta(phi) < pi/4

• delta(z) < 300mm

QCD data with pileup µ=10, pt>1GeV, barrel only, and duplicate hits removed

Investigating advanced deep learning for pattern

recognition

Classification


arbitrarily many

categories









Machine learning in HEPThese are just a few examples!

We have a broad and deep program, with a close connection to

our colleagues at NERSC.

Always looking for students in this endeavor

MACHINE LEARNING AND PHYSICS WORKSHOP UROŠ SELJAK

Cosmology initial condition reconstruction

27

We use optimization that finds the best solution in terms of final data (optimal filter). This 3-d example optimizes in 2 million dimensions. Galaxy are sparse tracers, so we loose small scale info

Track Pattern Recognition• The High-Luminosity LHC (HL-LHC) is expected to pose a challenge for

computing

• Increased luminosity

• Increased read-out rates (trigger+detector upgrades)

• Increased pile up

• Currently project to need more CPU time than will be available

• Dominated by track reconstruction: algorithms to reconstruct the trajectories of charged particles passing through the detector

�11

Vital to explore new algorithms and technologies

HEP.QPR: can quantum computers play a role

Quantum Pattern Recognition (HEP.QPR)• Project exploring algorithm for pattern recognition on currently available

quantum computers

• Recent publications on quantum annealing (D-Wave) and quantum associative memory

• Berkeley/LBL Team: Heather Gray (PI), Paolo Calafiura, Wim Lavrijsen, Wahid Bhimji, Alex Smith and collaborators in Switzerland, Canada and Japan

• More details: https://sites.google.com/lbl.gov/hep-qpr/

• Openings for graduate students (including this summer!)

�12

Copyright © D-Wave Systems Inc. 13

D-Wave Container –Faraday Cage - No RF Interference

IBM 20Q Tokyo

D Wave

https://arxiv.org/abs/1902.08324



https://sites.google.com/lbl.gov/hep-qpr/

Quantum computing for simulations

Simulations for scattering can not be performed with high accuracy for high multiplicity final states

Main limitation is that quantum interference effects can not be included for high multiplicity

Goal to develop quantum algorithms that can do high multiplicity simulations including quantum interference effects

Can quantum algorithms allow more precise calculations?Bauer, Nachman, Provasoli, deJong

Quantum computing for simulationsConsider simple toy model which exhibits quantum interference

L = f̄1i(/@ +m1)f1 + f̄2(i/@ +m2)f2 + (@µ�)2 + g1f̄1f1�+ g2f̄2f2�+ g12

⇥f̄1f2 + f̄2f1

⇤�

<latexit sha1_base64="jkUgo4e0na1AMh/B2ykAUPhKrR8=">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</latexit><latexit sha1_base64="jkUgo4e0na1AMh/B2ykAUPhKrR8=">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</latexit><latexit sha1_base64="jkUgo4e0na1AMh/B2ykAUPhKrR8=">AAAEO3icnVPLbhMxFHVneJTwSmHJxhAhJSqK7CEvFkgVbFiwaCXSVsoMI4/rmVj1PGR7kKLR/BcbfoIdGzYsQIgte+xkBCWhqKotWcf3ce6513JUCK40Qp+2HPfK1WvXt2+0bt66fedue+feocpLSdmU5iKXxxFRTPCMTTXXgh0XkpE0EuwoOn1p/UfvmFQ8z97oRcGClCQZjzkl2pjCHeeg8ikR8HUNn0MDbcJMJlFQoT4aDQd48gT1hwhP8NAAb4gR8mo/IhLGYYVryGHXV4Kouf/Q7soviNSciBruwjTEvVWYva2RP5uMvMHIcCI0xh62wBsPng5+k3s17HL4H3Kv14TtGg2NJ/TTEvrFnPfeGgfcLHtuT8myHXimtdVpyWrYurj+JPT+0Hh1c65ojB5bx94Fi/XsEhOvLyxkWfhyk98ocq6wRpMveTLXge0Shu2O5bYLbgLcgA5o1n7Y/uif5LRMWaapeW01w6jQQWVflApWt/xSsYLQU5KwmYEZSZkKqqW+Gj42lhMY50Z5nmm4tJ7NqEiq1CKNTGRK9Fyt+6zxX75ZqeNJUPGsKDXL6KpQXAqoc2g/EjzhklEtFgYQKrnRCumcSEK1+W4tMwS83vImOPT6GPXxwaCz96IZxzZ4AB6BLsBgDPbAK7APpoA6753Pzlfnm/vB/eJ+d3+sQp2tJuc++Gu5P38BlhpAzg==</latexit><latexit sha1_base64="jkUgo4e0na1AMh/B2ykAUPhKrR8=">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</latexit>

Simulation on classical computer takes 2N computations

Can be simulated on quantum computer with N4 scaling

Exciting cutting edge research with opening for graduate students

3

� ! ff̄ , which still follows the Markov Chain of ampli-tudes in Eq. (5). The core idea of the quantum algo-rithm is to first rotate to a particle basis where there isno mixing between fermion states (Appendix B). In thissuperposition basis, emissions between states are uncor-related. Sudakov factors can then be used to govern theno emission probability o↵ of the uncorrelated fermions.The bulk of the quantum circuitry will then be dedicatedto book-keeping, to encode the emission history and de-cide which fermions/bosons radiate/split at a given step✓ in the shower.

Figure 1 is the quantum circuit implementing thequantum final state radiation algorithm. The circuit callsfor six registers, which are are detailed in Appendix Aand summarized in Table I. The initial state is a singleparticle, in the f1/2 basis. As a first step, one rotates thisinitial particle from the f1/2 basis to the fa/b basis, usinga simple unitary R operation discussed in Appendix A.After this rotation, the quantum shower proceeds in Nsteps, with the same block repeated for each step. Thesub-circuit describing the one-step operations is shown inFig. 2. At each step, there are four operations, which aresummarized in Table II.

|pi / R⌦N

U(1)step U

(2)step

. . .

U(N)step

R†⌦N

|hi / . . .

|ei . . .

|n�i / . . .

|nAi / . . .

|nBi / . . .

FIG. 1: Full circuit schematic in terms of the circuit blocksdefined in Figure 2.

Register Purpose # of qubits

|pi Particle state 3(N + 1)

|hi Emission history N log2(N + 1)

|ei Did emission happen? 1

|n�i Number of bosons log2(N + 1)

|nai Number of fa log2(N + 1)

|nbi Number of fb log2(N + 1)

TABLE I: All of the registers in the quantum circuit.

At the end of the N -step evolution we must rotateback into the f1/f2 basis. We do so by applying theR† gate to all of the three-qubit particle registers in |pi.This creates interferences between equivalent final stateswhich had di↵erent intermediate fermions. Finally, wemeasure all the qubits thereby generating one event. Byrepeating the process over and over we can generate alarge number of events which we can then use to com-

|pi / p p U(m)p

U(m)step

|hi / Uh h

|ei U(m)e e ⌘

|n�i /

Ucount

n�

Uh|nai / nA

|nbi / nB

FIG. 2: Circuit for the mth step

Operation Complexity Operator Appendix

count particles N log2(N) Ucount C

decide emission N3

Ue D

create history N4

Uh E

adjust particles N Up F

TABLE II: Complexity of the various quantum circuitoperations.

pute physical observables for our theory. The number ofstandard quantum gates (single qubit gates and CNOTgates) required at each step are summarized in Table IV.

Operation Number of Standard Gates

Count Particles 5⇥ [Plog2(k)+3

n=3 34n� 27 ]

Decide Emission (k + 1)3 ⇥ [99 log2(k + 1)� 27]

Create History 5⇥ [Plog2(k)+3

n=3 32n� 27] +P

b[6(k + 1)3

⇥(2b� 1)[33(3 log2(k + 1) + b+ 3)� 27]]

Adjust Particles k ⇥M ⇥ [5 + 34 +Plog2(k)

n=3 34n� 27]

TABLE III: Number of standard quantum gates (singlequbit gates and CNOT gates) necessary for each of the four

main operations.

The practical challenge with above circuit is that itrequires more connected qubits and operations than arecurrently available in state-of-the-art hardware. There-fore, we consider a special case that is amenable tomeasurement on existing technology, which ignores the� ! ff̄ splitting (naturally suppressed in gauge theo-ries, but not in the scalar-only theory). This results ina much simpler circuit since there is only one fermion,but an arbitrary number of scalars (Appendix G). A de-composition of the resulting circuit into single qubit andCNOT gates requires ngates = 12N + 2 (Appendix ??).This model is however still su�ciently complex that theclassical MCMC described earlier1 fails to capture im-

1 While the standard parton shower-inspired MCMC algorithm

Bauer, Nachman, Provasoli, deJong

If you have any more questions, you can talk to a few of us in a little bit

during the “lab tour”

quantum computing / machine learning...why machine learning? • much of particle physics and...

Documents