New Paradigms for New Physics Searches with Machine Learning - Rutgers University - CERN Indico
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
New Paradigms for New Physics Searches
with Machine Learning
David Shih
January 13, 2021
Theory Mini-workshop
HKUST Scott
IAS Program on HEP
Thomas
Rutgers University
Where is the new physics?? Despite thousands of searches for new physics at the LHC, nothing but limits and null results so far. What if new physics is hiding in the data but we haven’t looked in the right places yet?
A Benchmark Example
LHC Olympics 2020 R&D Dataset
https://doi.org/10.5281/zenodo.2629072
q
X
Z’ q
q
Y
q
No explicit search at the LHC for this scenario!
Could be hiding in the dijet resonance search at >5sigma significance!!
The most common approach
Model specific searches
The BDT search
s search strategy is applied separately through two sets of signal regions targeting models with gluino-pair
duction with direct (BDT-GGd) or one-step (BDT-GGo) 6̃ decays. In each set, events are separated
Most NP searches at the LHC are heavily optimized with specific
0
o four categories, depending on the mass difference 3 , pmiss
T
) min > 0.4
⇢ Tmiss /< eff (#j ) > 0.2
< eff [GeV] > 1400 > 800
BDT score > 0.97 > 0.94 > 0.94 > 0.87
0
0.4BDTs) optimized
q( 91,2, (3) , pmiss
T
) min > 0.2 using simulations of signal AND
q( 98>3 , pbackground.
miss )
T min > 0.4 > 0.2
⇢ Tmiss /< eff (#j ) > 0.2
< eff [GeV] > 1400 > 800
BDT score > 0.96 > 0.87 > 0.92 > 0.84
The most common approach
Model specific searches
The BDT search
s search strategy is applied separately through two sets of signal regions targeting models with gluino-pair
duction with direct (BDT-GGd) or one-step (BDT-GGo) 6̃ decays. In each set, events are separated
Most NP searches at the LHC are heavily optimized with specific
0
o four categories, depending on the mass difference 3 , pmiss ) min
rc h > 0.4
sea
T
⇢ Tmiss /< eff (#j ) > 0.2
< eff [GeV]
of > 1400 > 800
BDT score
9 9 %> 0.97 > 0.94 > 0.94 > 0.87
0
>
0.4BDTs) optimized
q( 91,2, (3) , pmiss
T
) min > 0.2 using simulations of signal AND
q( 98>3 , pbackground.
miss )
T min > 0.4 > 0.2
⇢ Tmiss /< eff (#j ) > 0.2
< eff [GeV] > 1400 > 800
BDT score > 0.96 > 0.87 > 0.92 > 0.84
Of course, we should continue to perform these searches, because NP could always be right around the corner… But we should also recognize that perhaps new physics is on a different street, or in a different building altogether…
Existing model-independent approaches “the bump hunt” Idea: assume signal is localized in some feature (usually invariant mass) while background is smooth. Interpolate from sidebands into signal region, search for an excess.
Existing model-independent approaches
“the bump hunt”
Idea: assume signal is localized in some feature (usually invariant mass)
while background is smooth.
r i es.
Interpolate from sidebands into signal region, search
c ove
for an excess.
y di s
m an
e d i n
d, u s
e t ho
si c m
C l as
Existing model-independent searches
“the general search”
Idea: divide the phase space up into thousands of bins, compare
data
s are dominated by tt toFigures
production. SMforsimulation
additional object groupsin
caneach
be foundone
in
endix A.
0.015
0.018
0.020
0.045
0.046
0.064
0.065
0.068
0.068
0.069
0.077
0.081
0.081
0.083
0.084
0.091
0.097
0.10
0.10
0.10
Events per class
107
106
CMS
-1
Data
tt
W + jets
Higgs boson CMS ATLAS
35.9 fb (13 TeV) Drell-Yan γ + jets
105 Exclusive classes Single t Multijet
Multiboson
“Model independent
104
103
“MUSIC”
102 general search”
10
1 1807.07447
CMS-PAS-EXO-14-016
−1
10
EPJC 79:120 (2019)
2 e + 2 µ + 1b
2 e + 1b
4 µ + 1b + 1jet + pmiss
1e + 1µ + 1γ + pmiss
pmiss
1µ + 4 b + 1jet + pmiss
pmiss
pmiss
1e + 1µ + 1jet + pmiss
2 e + 1b + pmiss
3 e + 1b + 2 jets
3 µ + 1b + 3 jets
1e + 2 γ + 2 jets
1µ + 1γ + 2 b + 4 jets
1µ + 4 b + 2 jets
2 e + 1µ + 1b + 3 jets
1e + 1µ + 1γ + 1b + 1jet
1e + 1µ + 1jet
1µ + 2 γ
3 e + 1γ
T
T
T
T
T
T
T
T
1e + 1µ +
2 e + 1µ + 1jet +
1e + 1γ + 1b + 5 jets +
See also proposals by D’Agnolo, Wulzer et al (1806.02350, 1912.12155):
re 8: Data and SM predictions for the most significant exclusive event classes, where the
ficance of an event class is calculated in a single aggregated bin. Measured data are shown
train DNN on full phase space to distinguish data from background
ack markers, contributions from SM processes are represented by coloured histograms,
he shaded region represents the uncertainty in the SM background. The values above the
Existing model-independent searches
“the general search”
Idea: divide the phase space up into thousands of bins, compare
data toFigures
SMforsimulation
additional object groupsin
caneach
be foundone
st i l l
s are dominated by tt production. in
, bu t t
n
endix A.
d en t nd e
e pe n d e p e
d
0.015
0.018
0.020
0.045
0.046
0.064
0.065
0.068
0.068
0.069
0.077
0.081
0.081
0.083
0.084
0.091
0.097
n n
0.10
0.10
0.10
e l i l at i o
Events per class
d
107
CMS
u ATLAS
o
Data W + jets
m
CMS
m si
106 tt Higgs boson
5
35.9 fb-1 (13 TeV)
gn al nd )
Drell-Yan γ + jets
i
10 Exclusive classes Single t Multijet
s u
Multiboson
ly ro “Model independent
104
Tru (backg “MUSIC”
103
102 general search”
h ly
hig
10
1 1807.07447
CMS-PAS-EXO-14-016
10−1
EPJC 79:120 (2019)
2 e + 2 µ + 1b
2 e + 1b
4 µ + 1b + 1jet + pmiss
1e + 1µ + 1γ + pmiss
pmiss
1µ + 4 b + 1jet + pmiss
pmiss
pmiss
1e + 1µ + 1jet + pmiss
2 e + 1b + pmiss
3 e + 1b + 2 jets
3 µ + 1b + 3 jets
1e + 2 γ + 2 jets
1µ + 1γ + 2 b + 4 jets
1µ + 4 b + 2 jets
2 e + 1µ + 1b + 3 jets
1e + 1µ + 1γ + 1b + 1jet
1e + 1µ + 1jet
1µ + 2 γ
3 e + 1γ
T
T
T
T
T
T
T
T
1e + 1µ +
2 e + 1µ + 1jet +
1e + 1γ + 1b + 5 jets +
See also proposals by D’Agnolo, Wulzer et al (1806.02350, 1912.12155):
re 8: Data and SM predictions for the most significant exclusive event classes, where the
ficance of an event class is calculated in a single aggregated bin. Measured data are shown
train DNN on full phase space to distinguish data from background
ack markers, contributions from SM processes are represented by coloured histograms,
he shaded region represents the uncertainty in the SM background. The values above the2 An Overview of Model (In)dependent Searches
A viable search for new physics generally must have two essential com
New paradigms for model-agnostic searches
sensitive to new phenomena and it must also be able to estimate the b
null hypothesis (Standard Model only). The categorization of a searc
(in)dependence requires consideration of both of these components. Figu
Can advances in machine learning open up new avenues for
to characterize model independence for both BSM sensitivity and SM b
model-independent
We will nowsearches?
consider each in turn.
from Nachman & DS 2001.04990
background (SM) model independence
background (SM) model independence
autoencoders Direct De
Some searches LDA estimation, S
(train signal ANODE
versus data) ABCD
CWoLa
SALAD
Control
region
Most searches MUSiC (CMS), method
(train with General Search
simulations) (ATLAS) Pure MC
prediction
signal model independence signal model ind2 An Overview of Model (In)dependent Searches
A viable search for new physics generally must have two essential com
New paradigms for model-agnostic searches
sensitive to new phenomena and it must also be able to estimate the b
null hypothesis (Standard Model only). The categorization of a searc
(in)dependence requires consideration of both of these components. Figu
Can advances in machine learning open up new avenues for
to characterize model independence for both BSM sensitivity and SM b
model-independent
We will nowsearches?
consider each in turn.
from Nachman & DS 2001.04990
background (SM) model independence
background (SM) model independence
autoencoders Direct De
Some searches estimation, S
(train signal
LDA
Many new
ANODE
versus data) CWoLa ideas recently!
ABCD
SALAD
Control
region
Most searches MUSiC (CMS), method
(train with General Search
simulations) (ATLAS) Pure MC
prediction
signal model independence signal model indMany new approaches inspired by (or at least demonstrated
on) the LHC Olympics 2020 Data Challenge
1 The LHC Olympics 2020
2 A Community Challenge for Anomaly
3 Detection in High Energy Physics
4
5 Gregor Kasieczka (ed),1 Benjamin Nachman (ed),2,3 David Shih (ed),4 Oz Amram,5
6 Anders Andreassen,6 Kees Benkendorfer,2,7 Blaz Bortolato,8 Gustaaf Brooijmans,9
7 Florencia Canelli,10 Jack H. Collins,11 Biwei Dai,12 Felipe F. De Freitas,13 Barry M.
8 Dillon,8,14 Ioan-Mihail Dinu,5 Zhongtian Dong,15 Julien Donini,16 Javier Duarte,17 D.
9 A. Faroughy10 Julia Gonski,9 Philip Harris,18 Alan Kahn,9 Jernej F. Kamenik,8,19
10 Charanjit K. Khosa,20,30 Patrick Komiske,21 Luc Le Pottier,2,22 Pablo
11 Martı́n-Ramiro,2,23 Andrej Matevc,8,19 Eric Metodiev,21 Vinicius Mikuni,10 Inês
12 Ochoa,24 Sang Eon Park,18 Maurizio Pierini,25 Dylan Rankin,18 Veronica Sanz,20,26
13 Nilai Sarda,27 Uros̆ Seljak,2,3,12 Aleks Smolkovic,8 George Stein,2,12 Cristina Mantilla
14 Suarez,5 Manuel Szewc,28 Jesse Thaler,21 Steven Tsan,17 Silviu-Marian Udrescu,18
15 Louis Vaslin,16 Jean-Roch Vlimant,29 Daniel Williams,9 Mikaeel Yunus18
1
16 Institut für Experimentalphysik, Universität Hamburg, Germany
2
Physics Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA
to appear on the arXiv next week…
17
3
18 Berkeley Institute for Data Science, University of California, Berkeley, CA 94720, USA
4
19 NHETC, Department of Physics & Astronomy, Rutgers University, Piscataway, NJ 08854, USA
5
20 Department of Physics & Astronomy, The Johns Hopkins University, Baltimore, MD 21211,
21 USA66 Contents
67 1 Introduction 2
Many new approaches inspired by (or at least demonstrated
68 2 Dataset and Challenge 5
69 2.1 R&D Dataset 5
on) the LHC Olympics 2020 Data Challenge
70
71
2.2 Black Box 1
2.3 Black Box 2
6
7
72 2.4 Black Box 3 7
73 Individual Approaches 9
74 3 Unsupervised 11
75 3.1 Anomalous Jet Identification via Variational Recurrent Neural Network 11
76 3.2 Anomaly Detection with Density Estimation 16
77 3.3 BuHuLaSpa: Bump Hunting in Latent Space 19
78 3.4 GAN-AE and BumpHunter 24
79 3.5 Gaussianizing Iterative Slicing (GIS): Unsupervised In-distribution Anomaly
80 Detection through Conditional Density Estimation 29
81 3.6 Latent Dirichlet Allocation 34
82 3.7 Particle Graph Autoencoders 38
83 3.8 Regularized Likelihoods 42
84 3.9 UCluster: Unsupervised Clustering 46
85 4 Weakly Supervised 51
86 4.1 CWoLa Hunting 51
87 4.2 CWoLa and Autoencoders: Comparing Weak- and Unsupervised methods
88 for Resonant Anomaly Detection 55
89 4.3 Tag N’ Train 60
90 4.4 Simulation Assisted Likelihood-free Anomaly Detection 63
91 4.5 Simulation-Assisted Decorrelation for Resonant Anomaly Detection 68
92 5 (Semi)-Supervised 71
93 5.1 Deep Ensemble Anomaly Detection 71
94 5.2 Factorized Topic Modeling 77
95 5.3 QUAK: Quasi-Anomalous Knowledge for Anomaly Detection 81
96 5.4 Simple Supervised learning with LSTM layers 85
97 6 Discussion 88LHC Olympics 2020: Black Boxes
https://doi.org/10.5281/zenodo.3547721
In 2019, Gregor Kasieczka, Ben Nachman and I initiated the LHC
Olympics 2020 Challenge.
It consisted of three “black boxes” of simulated data (bg dominated!):
• 1 million events each
• 4-vectors of every reconstructed particle (all hadronic) in the event
• Particle ID, charge, etc not included
• Single R=1 jet trigger pT>1.2 TeV
The goal of the challenge was for participants to analyze each box and
1. Decide whether or not it contains new physics
2. Characterize the new physics, if it’s thereLHC Olympics 2020: Black Boxes
https://doi.org/10.5281/zenodo.3547721
In 2019, Gregor Kasieczka, Ben Nachman and I initiated the LHC
Olympics 2020 Challenge.
xes
It consisted of three “black boxes” of simulated data (bg dominated!):
e b o
s i n t h
• 1 million events each at w a
e e w h n g e
k t o s h al l e
• 4-vectors of every reconstructed
e n ’s ta l particle (all
f t e c
hadronic)
h in the event
fo r B om e o
• taParticle
t u n e
ID, u t c
dcharge, etcenotoincluded
S y a n d t h
• Single R=1 jet trigger pT>1.2 TeV
The goal of the challenge was for participants to analyze each box and
1. Decide whether or not it contains new physics
2. Characterize the new physics, if it’s thereLHC Olympics 2020: R&D Dataset
https://doi.org/10.5281/zenodo.2629072
Prior to the challenge, we also released a labeled R&D dataset consisting
of 1M QCD dijet events and 100k signal events
q
No explicit search at the
mX=500 GeV LHC for this scenario!
X
Z’ q
mZ’=3.5 TeV q
Y
mY=100 GeV
Unofficial theme of LHCO2020: q
“enhancing the bump hunt”LHC Olympics 2020: R&D Dataset
Resonant feature
m = mZ 0 = mJJ
(null)
Benchmark S=500, B=500,000, BSR=61,000
re 2. Histograms for the invariant mass of the leading two jets for the Standard Model background
signal strength: S/BSR~6x10-3, S/√BSR~1.5
ell as the injected signal. There are 1 million background events and 1000 signal events.LHC Olympics 2020: R&D Dataset
Additional features:
Figure 3. The four features used for classification: mJ1x = m(m
(top left),
left), and · J2 (bottom right). These histograms are inclusive in m
≠ mJ1 ,m
(top right),
J1
J·2 ,(bottom
J2
⌧21 , ⌧21 )
(null)
J1 J2
J1
21
. There are 1 million backgroundEnhancing the bump hunt
⇥(null)
~x 2 Rd
(null)
primary resonant feature (mJJ) additional features
If the signal is localized in additional features, can we find it in a model-
independent way?sideband SR sideband
Enhancing the bump hunt
The optimal model-agnostic discriminant would be Figure 2. Histograms for the invariant mass of the leading two jets for the Standard Model ba
as well as the injected signal. There are 1 million background events and 1000 signal events
epochs results in a stable result. Averaging over more epochs does not further imp
P(x|data) stability. All results with ANODE present the SB density estimator with this averagin
R(x) = for the last 10 epochs.
Figure 4 shows a scatter plot of R(x|m) versus log pbackground (x|m) for the test s
P(x|bg)
(null)
SR. As desired, the background is mostly concentrated around R(x|m) = 1, while there
tail for signal events at higher values of R(x|m) and between ≠2 < log pbackground (x
This is exactly what is expected for this signal: it is an over-density (R > 1) in a
phase space that is relatively rare for the background (pbackground (x|m) π 1).
The background density in Fig. 4 also shows that the R(x|m) is narrower aroun
Key idea: use sidebands in mJJ to model P(x|bg) pbackground (x|m) is large and more spread out when pbackground (x|m) π 1. This is
that the density estimation is more accurate when the densities are high and wo
the densities are low. This is also to be expected: if there are many data points clo
another, it should be easier to estimate their density than if the data points are ver
Another view of the results is presented in Fig. 5, with one-dimensional info
about R(x|m) in the SR. The left plot of Fig. 5 shows that the background is cent
approximately symmetric around R = 1 with a standard deviation of approximat
This width is due to various sources, including the accuracy of the SR density, the ac
the SB density, and the quality of the interpolation from SB to SR. Each of these so
contributions from the finite size of the datasets used for training, the neural network fl
and the training procedure. The right plot of Fig. 5 presents the number of backgro
signal events as a function of a threshold R > Rc . The starting point are the original
– 12 –sideband SR sideband
Enhancing the bump hunt
The optimal model-agnostic discriminant would be Figure 2. Histograms for the invariant mass of the leading two jets for the Standard Model ba
as well as the injected signal. There are 1 million background events and 1000 signal events
epochs results in a stable result. Averaging over more epochs does not further imp
P(x|data) stability. All results with ANODE present the SB density estimator with this averagin
R(x) = for the last 10 epochs.
Figure 4 shows a scatter plot of R(x|m) versus log pbackground (x|m) for the test s
P(x|bg)
(null)
SR. As desired, the background is mostly concentrated around R(x|m) = 1, while there
tail for signal events at higher values of R(x|m) and between ≠2 < log pbackground (x
This is exactly what is expected for this signal: it is an over-density (R > 1) in a
phase space that is relatively rare for the background (pbackground (x|m) π 1).
The background density in Fig. 4 also shows that the R(x|m) is narrower aroun
Key idea: use sidebands in mJJ to model P(x|bg) pbackground (x|m) is large and more spread out when pbackground (x|m) π 1. This is
that the density estimation is more accurate when the densities are high and wo
the densities are low. This is also to be expected: if there are many data points clo
another, it should be easier to estimate their density than if the data points are ver
Another view of the results is presented in Fig. 5, with one-dimensional info
about R(x|m) in the SR. The left plot of Fig. 5 shows that the background is cent
approximately symmetric around R = 1 with a standard deviation of approximat
This width is due to various sources, including the accuracy of the SR density, the ac
the SB density, and the quality of the interpolation from SB to SR. Each of these so
contributions from the finite size of the datasets used for training, the neural network fl
and the training procedure. The right plot of Fig. 5 presents the number of backgro
signal events as a function of a threshold R > Rc . The starting point are the original
– 12 –ANODE: Anomaly Detection with Density Estimation
Nachman & DS 2001.04990
Example of a new approach inspired by LHCO2020.
(See Ben’s talk for additional new approaches!)
Use neural density estimation to directly learn the conditional probability
densities from the data
P(x|data; mJJ 2 SR)
(null) (null)
P(x|data; mJJ 2
/ SR) = P(x|bg; mJJ 2
/ SR)
interpolate in (x,mJJ)
(null)
P(x|bg; mJJ 2 SR)
P(x|data; mJJ 2 SR)
Construct the likelihood ratio: R(x) =
P(x|bg; mJJ 2 SR) (null)ANODE: Results on LHCO R&D
Figure 4. Scatter Dataset
plot of R(x|m) versus log p background (x|
Nachman & DS 2001.04990 events are shown (as a two-dimensional histogram) in gray
in red.
Figure 5.
Left: Histogram of R(x|m) evaluated on the
igure 4. Scatter plot of R(x|m) versus log pbackground (x|m) across the test set in the SR. Background
ents are shown (as a two-dimensional histogram) in grayscale and individual signal events are shown
red.
The method works! ANODE is sensitive events that
survive atothreshold on R(x|m). The two distr
the signal!
500,000 total background events and 500 total signal even
to have the same number of events as each other and iJ1 ANODE: Results on LHCO R&D Dataset
(left) and mJ2 ≠ mJ1 (right) in the signal region after applying a
Nachman & DS 2001.04990
Can enhance the significance of the bump hunt by a factor of up to 7!
ng Characteristic (ROC) curve (left) and Significance Improvement
ht). 1.5σ (dijet bump hunt) => 10σ (ANODE+dijet bump hunt)determine the background efficiency in the SR after a requirement on R > Rc . Figure 8 presents
a comparison between integration methods (direct integration and importance sampling)
described in Sec. 3.2 and the true background yields. Qualitatively, both methods are able to
ANODE: Results on LHCO R&D Dataset
characterize the yield across several orders of magnitude in background efficiency. However,
both methods
Nachman & DSdiverge from the truth in the extreme tails of the R distribution. The right plot
2001.04990
of Fig. 8 offers a quantitative comparison between methods. For efficiencies down to about
10≠3 , both methods are accurate within about 25%. The direct integration method has a
smaller
Novelbias of about
aspect 10%. This
of ANODE: canis consistent with Fig. 5, for
estimate backgrounds which the
directly withstandard deviation is
P(x|bg; m∈SR)
between 10-20%.
Figure 8. Left: The number of events after a threshold requirement R > Rc using the two integration
methods described in Sec. 3.2, as well as the true background yield. Right: The ratio of the predicted
and true background yields from the left plot, as a function of the actual number of events that surviveVia Machinae: ANODE IN SPACE
DS, Matt Buckley, Lina Necib & John Tamanas, in preparation
ANODE is a completely general method for identifying multivariate
overdensities in data using sidebands. It could have many diverse applications.
We are currently using it to find stellar streams in Gaia data.
Gaia satellite:
• Launched in 2013; extended to 2025
• Mission: map out the full 6d phase space of
the stars in our galaxy
• Angular positions, velocities, color and
magnitude of over 1 billion stars in our
galaxy
• radial positions and velocities for a smaller
subset of nearby stars (not used in this
work)Stellar streams
Cold stellar streams are tidally-
stripped remnants of globular
clusters and dwarf galaxies, falling
into and orbiting our galaxy.
They are very interesting objects
of study for astrophysicists and
particle physicists.
In particular, they could
be unique probes into
dark matter substructure.Stream finding
https://github.com/cmateu/galstreams
Existing methods (eg STREAMFINDER, Malhan & Ibata 2018) have found many
new streams in the Gaia data, but they make a number of model-dependent
assumptions (form of the galactic potential, orbits, isochrones, …).
We were interested in whether unsupervised ML could be used to find
streams more model-independently.Ibata & Martin
Stream finding
from Malhan et al 2018
position velocity photometry
ties
ties of
of aa sample
sample of
of previously-discovered
Figure 4. Properties of a sample
previously-discovered streams,
streams, ofas recovered
recovered by by the
previously-discovered
as the STREAMFINDER. The
The first,
streams, as recovered
STREAMFINDER. second,
first, by third
third and
the STREAMFINDER.
second, and fourth The first, second
fourth
perties of
operties of the GD-1,
therows
GD-1, Jhelum,
show
Jhelum, Indus
Indus and
the properties and ofOrphan
the GD-1,
Orphan streams,
Jhelum,
streams, respectively.
Indus andThe
respectively. columns
Orphan
The reproduce,
streams,
columns from
from left
respectively.
reproduce, The
left to right,
right, the
tocolumns thereproduce, from
ates of
nates of the
the structures,
equatorialthe
structures, distance
distance solutions
coordinates
the found
found by
of the structures,
solutions the
bythe algorithm
thedistance
algorithm (for
(for representative
solutions found by the
representative metallicity
algorithmvalues),
metallicity the
the proper
(for representative
values), proper metallicity v
on
on (with
(with observations
observations in
in red,
red, model
motion distribution model solutions
solutions
(with in
in blue,
observations blue,inand
red,the
and the full
modelfull DR2
DR2 sample
sample
solutions in
in grey),
in blue, and and
grey), the
and full
the colour-magnitude
the DR2
colour-magnitude
sample in grey), and the
e stars
stars (with
(with observations
observations
distribution of in red
in the and
and template
red stars template model
model in
(with observations in blue)
blue)
in red selected
selected by
by STREAMFINDER.
and template model in blue)
STREAMFINDER. The
The distance
distanceby
selected solutions
solutions found
found The distanc
STREAMFINDER.
match
match closely the
closely by distance
thethe
distance
algorithmvalues
values that have
have been
thatclosely
match the previously
been previously
distance derived
derived
values that for these
forhave
these streams:
streams:
been DD ⇠
previously 88kpc
kpc for
⇠derived forGD-1
for these(Grillmair
GD-1 (Grillmair
streams: D ⇠ 8 kpc fo
D ⇠
,, D ⇠ 13.2
13.2 kpc and ⇠
and
kpcDionatos
& ⇠ 16.6
16.6 kpc
kpcD
2006), for
for Jhelum
⇠Jhelum and
13.2 kpcandandIndus,
Indus, respectively
respectively
⇠ 16.6 kpc for Jhelum(Shipp
(Shipp andet
et al.
al. 2018) and
and D D =
2018)respectively
Indus, = [33
[33 38]
(Shipp etkpc
38] al. for
kpc for
2018) and D =
g etet al.
al. 2010). The
The CMD
2010).Orphan CMD template
template
(Newberg models,
2010).shown
models,
et al. shown
The CMDin
in blue
blue in
in the
templatethe last
last column,
models,column,
shown have
have in been
blue plotted
been plotted
in at
the lastat the
the appropriate
appropriate
column, have been plotted a
spective streams.
espective streams. The
Thefor
distance colour-magnitude
colour-magnitude diagram
diagram The
the respective streams. of
of the
the Orphan
Orphan stream
stream might
colour-magnitude might seem
diagram seem peculiar,
peculiar,
of the Orphan but
but here
here we
stream we only see
onlyseem
might see the
the
peculiar, but he
due
due to to the
the trimming
trimming
red-giantof the
the data
of branch data sample
sample
due to the below
below GG=
trimming = 19.5.
19.5.
of the data sample below G = 19.5.Ibata & Martin
Stream finding
from Malhan et al 2018
position velocity photometry
ties
ties of
of aa sample
sample of
of previously-discovered
Figure 4. Properties of a sample
previously-discovered streams,
streams, ofas recovered
recovered by by the
previously-discovered
as the STREAMFINDER. The
The first,
streams, as recovered
STREAMFINDER. second,
first, by third
third and
the STREAMFINDER.
second, and fourth The first, second
fourth
perties of
operties of the GD-1,
therows
GD-1, Jhelum,
show
Jhelum, Indus
Indus and
the properties and ofOrphan
the GD-1,
Orphan streams,
Jhelum,
streams, respectively.
Indus andThe
respectively. columns
Orphan
The reproduce,
streams,
columns from
from left
respectively.
reproduce, The
left to right,
right, the
tocolumns thereproduce, from
ates of
nates of the Idea: streams are local overdensities in position, velocity and photometric
the structures,
equatorialthe
structures, distance
distance solutions
coordinates
the found
found by
of the structures,
solutions the
bythe algorithm
thedistance
algorithm (for
(for representative
solutions found by the
representative metallicity
algorithmvalues),
metallicity values), the
the proper
(for representative
proper metallicity v
on
on (with
(with observations in
in red,
red, model solutions in
in blue, inand
red,the full
full DR2
DR2 sample in
in grey),
and and the colour-magnitude
e stars
stars (with
space.
observations
motion distribution
(with observations
observations
distribution of in
model
red
in the and
red stars
solutions
(with observations
and template
template model
blue,
model in
(with observations
and
in blue)
blue)
the
in red
model
selected
selected by
sample
solutions in blue,
by STREAMFINDER.
and template
grey),
model in blue)
STREAMFINDER. The
and full
the the DR2
The distance
distanceby
selected
colour-magnitude
sample in grey), and the
solutions
solutions found
found The distanc
STREAMFINDER.
match
match closely the
closely by distance
thethe
distance
algorithmvalues
values that have
have been
thatclosely
match the previously
been previously
distance derived
derived
values that for these
forhave
these streams:
streams:
been DD ⇠
previously 88kpc
kpc for
⇠derived forGD-1
for these(Grillmair
GD-1 (Grillmair
streams: D ⇠ 8 kpc fo
D ⇠
,, D ⇠ 13.2
13.2 kpc and ⇠
and
kpcDionatos
& ⇠ 16.6
16.6 kpc
kpcD
2006), for
for Jhelum
⇠Jhelum and
13.2 kpcandandIndus,
Indus, respectively
respectively
⇠ 16.6 kpc for Jhelum(Shipp
(Shipp andet
et al.
al. 2018) and
and D D =
2018)respectively
Indus, = [33
[33 38]
(Shipp etkpc
38] al. for
kpc for
2018) and D =
g etet al.
al. 2010). The
The CMD
2010).Orphan CMD template
template
(Newberg models,
2010).shown
models,
et al. shown
The CMDin
in blue
blue in
in the
templatethe last
last column,
models,column,
shown have
have in been
blue plotted
been plotted
in at
the lastat the
the appropriate
appropriate
column, have been plotted a
spective streams.
espective streams. The
Thefor
distance colour-magnitude
colour-magnitude diagram
diagram The
the respective streams. of
of the
the Orphan
Orphan stream
stream might
colour-magnitude might seem
diagram seem peculiar,
peculiar,
of the Orphan but
but here
here we
stream we only see
onlyseem
might see the
the
peculiar, but he
due
due to to the
the trimming
trimming
red-giantof the
the data
of branch data sample
sample
due to the below
below GG=
trimming = 19.5.
19.5.
of the data sample below G = 19.5.Ibata & Martin
Stream finding
from Malhan et al 2018
position velocity photometry
ties
ties of
of aa sample
sample of
of previously-discovered
Figure 4. Properties of a sample
previously-discovered streams,
streams, ofas recovered
recovered by by the
previously-discovered
as the STREAMFINDER. The
The first,
streams, as recovered
STREAMFINDER. second,
first, by third
third and
the STREAMFINDER.
second, and fourth The first, second
fourth
perties of
operties of the GD-1,
therows
GD-1, Jhelum,
show
Jhelum, Indus
Indus and
the properties and ofOrphan
the GD-1,
Orphan streams,
Jhelum,
streams, respectively.
Indus andThe
respectively. columns
Orphan
The reproduce,
streams,
columns from
from left
respectively.
reproduce, The
left to right,
right, the
tocolumns thereproduce, from
ates of
nates of the Idea: streams are local overdensities in position, velocity and photometric
the structures,
equatorialthe
structures, distance
distance solutions
coordinates
the found
found by
of the structures,
solutions the
bythe algorithm
thedistance
algorithm (for
(for representative
solutions found by the
representative metallicity
algorithmvalues),
metallicity values), the
the proper
(for representative
proper metallicity v
on
on (with
(with observations in
in red,
red, model solutions in
in blue, inand
red,the full
full DR2
DR2 sample in
in grey),
and and the colour-magnitude
e stars
stars (with
space.
observations
motion distribution
(with observations
observations
distribution of in
model
red
in the and
red stars
solutions
(with observations
and template
template model
blue,
model in
(with observations
and
in blue)
blue)
the
in red
model
selected
selected by
sample
solutions in blue,
by STREAMFINDER.
and template
grey),
model in blue)
STREAMFINDER. The
and full
the the DR2
The distance
distanceby
selected
colour-magnitude
sample in grey), and the
solutions
solutions found
found The distanc
STREAMFINDER.
match
match closely the
closely by distance
thethe
distance
algorithmvalues
values that have
have been
thatclosely
match the previously
been previously
distance derived
derived
values that for these
forhave
these streams:
streams:
been DD ⇠
previously 88kpc
kpc for
⇠derived forGD-1
for these(Grillmair
GD-1 (Grillmair
streams: D ⇠ 8 kpc fo
D ⇠
,, D ⇠ 13.2 Since they are cold, the stars in the stream are clustered in velocity.
13.2 kpc and ⇠
and
kpcDionatos
& ⇠ 16.6
16.6 kpc
kpcD
2006), for
for Jhelum
⇠Jhelum and
13.2 kpcandandIndus,
Indus, respectively
respectively
⇠ 16.6 kpc for Jhelum(Shipp
(Shipp andet
et al.
al. 2018) and
and D D =
2018)respectively
Indus, = [33
[33 38]
(Shipp etkpc
38] al. for
kpc for
2018) and D =
g etet al.
al. 2010). The
The CMD
2010).Orphan CMD template
template
(Newberg models,
2010).shown
models,
et al. shown
The CMDin
in blue
blue in
in the
templatethe last
last column,
models,column,
shown have
have in been
blue plotted
been plotted
in at
the lastat the
the appropriate
appropriate
column, have been plotted a
spective streams.
espective streams. The
Thefor
distance colour-magnitude
colour-magnitude diagram
diagram The
the respective streams. of
of the
the Orphan
Orphan stream
stream might
colour-magnitude might seem
diagram seem peculiar,
peculiar,
of the Orphan but
but here
here we
stream we only see
onlyseem
might see the
the
peculiar, but he
due
due to to the
the trimming
trimming
red-giantof the
the data
of branch data sample
sample
due to the below
below GG=
trimming = 19.5.
19.5.
of the data sample below G = 19.5.Ibata & Martin
Stream finding
from Malhan et al 2018
position velocity photometry
ties
ties of
of aa sample
sample of
of previously-discovered
Figure 4. Properties of a sample
previously-discovered streams,
streams, ofas recovered
recovered by by the
previously-discovered
as the STREAMFINDER. The
The first,
streams, as recovered
STREAMFINDER. second,
first, by third
third and
the STREAMFINDER.
second, and fourth The first, second
fourth
perties of
operties of the GD-1,
therows
GD-1, Jhelum,
show
Jhelum, Indus
Indus and
the properties and ofOrphan
the GD-1,
Orphan streams,
Jhelum,
streams, respectively.
Indus andThe
respectively. columns
Orphan
The reproduce,
streams,
columns from
from left
respectively.
reproduce, The
left to right,
right, the
tocolumns thereproduce, from
ates of
nates of the Idea: streams are local overdensities in position, velocity and photometric
the structures,
equatorialthe
structures, distance
distance solutions
coordinates
the found
found by
of the structures,
solutions the
bythe algorithm
thedistance
algorithm (for
(for representative
solutions found by the
representative metallicity
algorithmvalues),
metallicity values), the
the proper
(for representative
proper metallicity v
on
on (with
(with observations in
in red,
red, model solutions in
in blue, inand
red,the full
full DR2
DR2 sample in
in grey),
and and the colour-magnitude
e stars
stars (with
space.
observations
motion distribution
(with observations
observations
distribution of in
model
red
in the and
red stars
solutions
(with observations
and template
template model
blue,
model in
(with observations
and
in blue)
blue)
the
in red
model
selected
selected by
sample
solutions in blue,
by STREAMFINDER.
and template
grey),
model in blue)
STREAMFINDER. The
and full
the the DR2
The distance
distanceby
selected
colour-magnitude
sample in grey), and the
solutions
solutions found
found The distanc
STREAMFINDER.
match
match closely the
closely by distance
thethe
distance
algorithmvalues
values that have
have been
thatclosely
match the previously
been previously
distance derived
derived
values that for these
forhave
these streams:
streams:
been DD ⇠
previously 88kpc
kpc for
⇠derived forGD-1
for these(Grillmair
GD-1 (Grillmair
streams: D ⇠ 8 kpc fo
D ⇠
,, D ⇠ 13.2 Since they are cold, the stars in the stream are clustered in velocity.
13.2 kpc and ⇠
and
kpcDionatos
& ⇠ 16.6
16.6 kpc
kpcD
2006), for
for Jhelum
⇠Jhelum and
13.2 kpcandandIndus,
Indus, respectively
respectively
⇠ 16.6 kpc for Jhelum(Shipp
(Shipp andet
et al.
al. 2018) and
and D D =
2018)respectively
Indus, = [33
[33 38]
(Shipp etkpc
38] al. for
kpc for
2018) and D =
g etet al.
al. 2010). The
The CMD
2010).Orphan CMD template
template
(Newberg models,
2010).shown
models,
et al. shown
The CMDin
in blue
blue in
in the
templatethe last
last column,
models,column,
shown have
have in been
blue plotted
been plotted
in at
the lastat the
the appropriate
appropriate
column, have been plotted a
spective streams.
espective streams. The
Thefor
distance colour-magnitude
colour-magnitude diagram
diagram The
the respective streams. of
of the
the Orphan
Orphan stream
stream might
colour-magnitude might seem
diagram seem peculiar,
peculiar,
of the Orphan but
but here
here we
stream we only see
onlyseem
might see the
the
peculiar, but he
due
due to to the Can sideband in one of the velocities and use ANODE to look for local
the trimming
trimming
red-giantof the
the data
of branch data sample
sample
due to the below
below GG=
trimming = 19.5.
19.5.
of the data sample below G = 19.5.
overdensities!Via Machinae: preliminary results
The method successfully finds
known streams!Via Machinae: preliminary results
The method also finds ~600 other
stream candidates
We are currently investigating
these further to see if they
could be real. Stay tuned!Summary and Outlook
• Advances in machine learning are opening up new and exciting
avenues for model independent new physics searches at the LHC.
• Ideas and methods inspired by LHC problems are being applied
successfully to other fields such as astronomy and astrophysics.
• The LHC Olympics 2020 provided a very useful testing ground for the
development and common benchmarking of new approaches.
• Much work remains to be done in order to port these ideas over to
ATLAS and CMS and implement them as actual analyses on real data.
• We need more ideas for model-independent searches at the LHC.
This is just the beginning!Current Organization of Physics Analysis Groups at the LHC
B physics B2G / Exotics/
HDBS Exotica
Search
Statistics
Groups
forum ML
forum
SM
Supporting
organizations
SUSY
Measurement Top
Higgs
Groups
Q: Why is there no model independent search group???A vision for the future…
Future Organization of Physics Analysis Groups at the LHC??
Unsupervised
B2G /
B physics Weakly Supervised
Model HDBS
Agnostic?
(Semi) Supervised
Statistics
Search
forum ML Groups
forum Exotics/
SM
Supporting Exotica
organizations
Measurement Top SUSY
Higgs
Groups
from G. Kasieczka, B. Nachman, DS (eds), et al 2101.nextweekThanks for your attention!
9. A comparison of the four features x between the SR and two nearby sidebands defined by
.1, 3.3] TeV (lower sideband) and mjj œ [3.7, 3.9] TeV (upper sideband).
ANODE: Results on LHCO R&D Dataset
mate the background. To study this sensitivity in a controlled fashion, correlations
Ben Nachman & DS 2001.04990
oduced artificially. In practice, adding more features to x will inevitably result in
pendence with mjj ; the artificial example here illustrates the challenges already in low
Can also consider performance on a feature set which is not
ons. New jet independent
mass observables are created, which are linearly shifted:
of m. We introduced artificial correlations just as proof
of concept: mJ1,2 æ mJ1,2 + c mJJ , (5.1)
= 0.1 for this study. The resulting shifted lighter jet mass is presented in Fig. 10.
w ANODE and CWoLa models are trained using the shifted dataset and their perfor-
s quantified in Fig. 11. As expected, the fully supervised classifier is nearly the same as
ANODE is still able to significantly enhance the signal, with a maximum significance
ment near 4. While in principle ANODE could achieve the same classification accuracy
hifted and nominal datasets, the performance on the shifted examples is not as strong
g. 7. In practice the interpolation of pbackground into the SR is more challenging now
he linear correlations. This could possibly be overcome with improved training, better
of hyperparameters, or more sophisticated density estimation techniques.
construction, there are now bigger differences between the SR and SB than between
background and the SRANODEsignal. Therefore, the CWoLa
is robust while hunting classifier
CWoLa completely fails! is not able to
Figure 11. ROC (left) and SIC (right) curves in the signal region using the shifted dataset specifiedSearching for NP with deep autoencoders
Heimel et al 1808.08979; Farina, Nakai & DS 1808.08992
Latent layer
Figure 1: The schematic diagram of an autoencoder. The input is mapped into a low(er) dimensional
An autoencoder
representation, maps
in this case anand
6-dim, input into a reduced “latent representation”
then decoded.
and then attempts to reconstruct the original input from it.
threshold.
CanForuse reconstruction
concreteness, error
we will focus as work
in this an anomaly threshold!
on distinguishing “fat” QCD jets from
other types of heavier, boosted resonances decaying to jets. Building on previous work
on top tagging [12], we will concentrate on machine learning algorithms that take jet
images as inputs. For signal, we will consider all-hadronic top jets, as well as 400 GeV
See also:
Hajergluinos
et al “Novelty Detection
decaying to 3Meets
jets Collider
via RPV.Physics” 1807.10261
Obviously, this is not meant to be an exhaustive
Cerri study
et al “Variational Autoencoders
of all possible for New Physics
backgrounds Mining atand
and signals the Large Hadron
methods Collider”
but is just1811.10276
meant to be asignal (tops and 400
background (QCD)
GeV RPV gluinos)
Performance
It works as an anomaly detector!
Robust against
Figure contamination
2: Distribution with
of reconstruction errorsignal —with
computed cana CNN
use autoencoder
in fully unsupervised
on test samples of mode
QCD background (gray) and two signals: tops (blue) and 400 GeV gluinos (orange).You can also read
