3.5 Jet Substructure Tagging
91
)
sig
∈
Signal efficiency (
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
)
bkg
∈
Background rejection (1 /
1
10
2
10
3
10
4
10
ATLAS Simulation
s
V
e
T
3
1
=
= 1.0 jets
R
t
k
Trimmed anti| < 2.0
true
η
|
= [1500, 2000] GeV
true
T
p
Top tagging
DNN top
BDT top
Shower
Deconstruction
2-var optimised
tagger
HEPTopTagger v1
> 60 GeV
comb
m
,
32
τ
TopoDNN
0
0.2
0.4
0.6
0.8
1
Signal efficiency
4
−
10
3
−
10
2
−
10
1
−
10
1
10
Background efficiency
(13 TeV)
CMS
Simulation
DeepAK8
DeepAK8-MD
ImageTop
ImageTop-MD
32
τ
+
SD
m
+ b
32
τ
+
SD
m
BEST
HOTVR
-BDT (CA15)
3
N
Top quark vs. QCD multijet
| < 2.4
gen
η
< 1500 GeV, |
gen
T
1000 < p
< 210 GeV
AK8
SD
105 < m
< 210 GeV
CA15
SD
110 < m
< 220 GeV
HOTVR
140 < m
Fig. 3.13 Performance comparison of top taggers in ATLAS (left) and CMS (right) at high p T .
Taken from [525] (left) and [526] (right)
perform better than the HTT, a two-variable optimised tagger or the shower deconstruction algorithm, but do not achieve the optimal performance. A DNN based on
four-vectors of a fixed-number of topoclusters (TopoDNN) shows the best performance. The algorithm consists of a fully connected neural network with four hidden
layers, following closely the approach outlined in [548]. The gain in background
rejection at low signal efficiency with respect to the DNN top tagger indicates that a
DNN based on low-level inputs can learn correlations hidden to algorithms based on
high-level observables. CMS has studied multi-class taggers and the use of jet images
to further improve the performance of ML taggers. Figure 3.13 (right) shows the ROC
curves of nine different top taggers, out of which six are based on ML techniques.
The best performance is achieved with the DeepAK8 algorithm, which exploits the
particle-level information to obtain a multiclass classifier for the identification of the
five main categories, W /Z /H /t/other. In addition to particle four-vectors, information
such as the energy deposit, the charge, the displacement and quality of the tracks, etc.
are included for each particle. A secondary vertex list helps the network to extract
features related to the heavy-flavour content of the jet. The particle and secondary
vertex lists are processed with two CNNs. The outputs of these are combined using
a fully connected network to perform the jet classification [526]. In the case of H
and t tagging, the secondary vertex information helps to improve the performance
significantly due to the heavy-flavour content of the jet. The other ML algorithms
studied by CMS are a top tagger based on jet images (ImageTop), a multiclass classifier based on event shapes in the rest frames of W , Z , H and t [549, 550], called
the boosted event shape tagger (BEST) [229], and a BDT based on eleven ECF
ratios (N 3 -BDT). Their performance lies between the cut-based approaches and the
DeepAK8 tagger. It should be noted that the N 3 -BDT has been designed for low p T ,
such that it shows the best performance for 300 < p T < 500 GeV. Interestingly, in
this region the HOTVR algorithm has a comparable performance to the ML taggers,
91
)
sig
∈
Signal efficiency (
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
)
bkg
∈
Background rejection (1 /
1
10
2
10
3
10
4
10
ATLAS Simulation
s
V
e
T
3
1
=
= 1.0 jets
R
t
k
Trimmed anti| < 2.0
true
η
|
= [1500, 2000] GeV
true
T
p
Top tagging
DNN top
BDT top
Shower
Deconstruction
2-var optimised
tagger
HEPTopTagger v1
> 60 GeV
comb
m
,
32
τ
TopoDNN
0
0.2
0.4
0.6
0.8
1
Signal efficiency
4
−
10
3
−
10
2
−
10
1
−
10
1
10
Background efficiency
(13 TeV)
CMS
Simulation
DeepAK8
DeepAK8-MD
ImageTop
ImageTop-MD
32
τ
+
SD
m
+ b
32
τ
+
SD
m
BEST
HOTVR
-BDT (CA15)
3
N
Top quark vs. QCD multijet
| < 2.4
gen
η
< 1500 GeV, |
gen
T
1000 < p
< 210 GeV
AK8
SD
105 < m
< 210 GeV
CA15
SD
110 < m
< 220 GeV
HOTVR
140 < m
Fig. 3.13 Performance comparison of top taggers in ATLAS (left) and CMS (right) at high p T .
Taken from [525] (left) and [526] (right)
perform better than the HTT, a two-variable optimised tagger or the shower deconstruction algorithm, but do not achieve the optimal performance. A DNN based on
four-vectors of a fixed-number of topoclusters (TopoDNN) shows the best performance. The algorithm consists of a fully connected neural network with four hidden
layers, following closely the approach outlined in [548]. The gain in background
rejection at low signal efficiency with respect to the DNN top tagger indicates that a
DNN based on low-level inputs can learn correlations hidden to algorithms based on
high-level observables. CMS has studied multi-class taggers and the use of jet images
to further improve the performance of ML taggers. Figure 3.13 (right) shows the ROC
curves of nine different top taggers, out of which six are based on ML techniques.
The best performance is achieved with the DeepAK8 algorithm, which exploits the
particle-level information to obtain a multiclass classifier for the identification of the
five main categories, W /Z /H /t/other. In addition to particle four-vectors, information
such as the energy deposit, the charge, the displacement and quality of the tracks, etc.
are included for each particle. A secondary vertex list helps the network to extract
features related to the heavy-flavour content of the jet. The particle and secondary
vertex lists are processed with two CNNs. The outputs of these are combined using
a fully connected network to perform the jet classification [526]. In the case of H
and t tagging, the secondary vertex information helps to improve the performance
significantly due to the heavy-flavour content of the jet. The other ML algorithms
studied by CMS are a top tagger based on jet images (ImageTop), a multiclass classifier based on event shapes in the rest frames of W , Z , H and t [549, 550], called
the boosted event shape tagger (BEST) [229], and a BDT based on eleven ECF
ratios (N 3 -BDT). Their performance lies between the cut-based approaches and the
DeepAK8 tagger. It should be noted that the N 3 -BDT has been designed for low p T ,
such that it shows the best performance for 300 < p T < 500 GeV. Interestingly, in
this region the HOTVR algorithm has a comparable performance to the ML taggers,