RNN Tau Identi f ication IN THE ATLAS High-Level Trigger M ARIEL P - - PowerPoint PPT Presentation

β–Ά
rnn tau identi f ication
SMART_READER_LITE
LIVE PREVIEW

RNN Tau Identi f ication IN THE ATLAS High-Level Trigger M ARIEL P - - PowerPoint PPT Presentation

Mar 04, 2024 460 likes β€’646 views

RNN Tau Identi f ication IN THE ATLAS High-Level Trigger M ARIEL P ETTEE October 26 th , 2018 US LUA Annual Meeting 2 BDT Identi f ication 3 BDT Tau Identi f ication BDT = Boosted Decision Tree Inputs: 12 high-level

slide-1
SLIDE 1

RNN Tau Identification

IN THE

ATLAS High-Level Trigger

MARIEL PETTEE October 26th, 2018 US LUA Annual Meeting

slide-2
SLIDE 2
  • 2
slide-3
SLIDE 3
  • 3

BDT 𝜐 Identification

slide-4
SLIDE 4
  • 4

BDT Tau Identification

  • BDT = Boosted Decision Tree
  • Inputs: 12 β€œhigh-level” tau ID variables
  • Separate models for 1- and 3-prong taus
  • ET / pT of leading

charged track

  • Significance of

leading track transverse impact parameter

  • Transverse

flight path significance

  • Maximum Ξ”R
  • Mass of

charged tracks Central energy fraction Mean track radius Momentum fraction

  • f isolation tracks

Momentum ratio of track + EM system Energy flow pT ratio Energy flow mass

1-prong 3-prong

Table: ATLAS-CONF-2017-061

slide-5
SLIDE 5
  • 5

BDT Tau Identification

arXiv:1704.07657

  • After training, classify each tau candidate with a single score ∈ (0,1):

1 = signal-like

https://xkcd.com/210/

0 = background-like

slide-6
SLIDE 6
  • 6

BDT Tau Identification

Benefits of the BDT architecture:

  • Highly-optimized classification
  • Physical intuition & insight

But... are we throwing away information by only considering these select high-level variables? Could adding low-level input variables into the mix improve our signal/background discrimination? BDT Tau ID Performance in the HLT (2017)

Plot: ATLAS-CONF-2017-061

slide-7
SLIDE 7
  • 7

RNN 𝜐 Identification

slide-8
SLIDE 8
  • 8
  • RNN = Recurrent Neural Network
  • Approaches 𝜐 identification as a sequence classification problem
  • Inputs:
  • Track-level variables (for at most 10 tracks; full list in backup slides)
  • Cluster-level variables (for at most 6 clusters; full list in backup slides)
  • BDT input variables

Image: C. Deutsch

RNN Tau Identification

Track 1 Track 2 pT βˆ†Ξ· βˆ†Ο• d0

. . .

Track N pT βˆ†Ξ· βˆ†Ο• d0

. . .

Reconstructed hadronic tau decay pT βˆ†Ξ· βˆ†Ο• d0

. . .

(

Ordered by pT

. . . . . .

Dense layer with shared weights RNN Output neuron S B

slide-9
SLIDE 9
  • 9

Images: C. Deutsch

RNN Tau Identification

Track 1 Track 2 pT βˆ†Ξ· βˆ†Ο• d0 . . . Track N pT βˆ†Ξ· βˆ†Ο• d0 . . . Reconstructed hadronic tau decay pT βˆ†Ξ· βˆ†Ο• d0 . . .

(

. . . . . . Dense layer with shared weights RNN

Tracks

Track 1 Track 2 pT βˆ†Ξ· βˆ†Ο• d0 . . . Track N pT βˆ†Ξ· βˆ†Ο• d0 . . . Reconstructed hadronic tau decay pT βˆ†Ξ· βˆ†Ο• d0 . . .

(

. . . . . . Dense layer with shared weights RNN

Clusters

+ +

BDT Input Variables

=

slide-10
SLIDE 10
  • 10
  • Shared dense & LSTM (Long Short-Term Memory) layers preserve

contextual information from multiple tracks/clusters to improve decision-making

Image: C. Deutsch

RNN Tau Identification

Cell state Input #1

Image: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Input #2 Input #3

slide-11
SLIDE 11
  • 11
  • BDT 𝜐 identification: 1-prong and 3-prong taus
  • RNN 𝜐 identification: 0-prong, 1-prong and β‰₯2-prong taus

Image: C. Deutsch

RNN Tau Identification

Goal: recover true 1p taus for which the charged track has been poorly reconstructed, especially at low-pT and high-ΞΌ. Goal: recover true 3p taus for which at least one charged track has been mis-reconstructed.

slide-12
SLIDE 12
  • 12
  • Training samples:
  • Ξ³*β†’πœπœ and dijets
  • Fewer statistics than for offline, but same # of trainable parameters
  • 0-prong: Signal ~100k events, Background ~50k events
  • 1-prong: Signal ~2M events, Background ~175k events
  • β‰₯2-prong: Signal ~700k events, Background ~3M events
  • Reweight signal pT spectrum to match background pT spectrum for determining same-rejection

working points as the BDT (since trigger rates are dominated by low-pT jets).

RNN Tau Identification: Online

slide-13
SLIDE 13
  • 13

RNN Tau Identification: Outlook

  • RNN tau identification successfully deployed in the

ATLAS high-level trigger in mid-2018! (Public results forthcoming)

  • Will likely be the default ATLAS tau identification algorithm

in the Run 3 trigger

  • Until then:
  • Continue optimizing models
  • Investigate RNN input variable modeling
  • Consider training on data rather than MC
slide-14
SLIDE 14
  • 14

Backup Slides

slide-15
SLIDE 15
  • 15
  • Track-level variables
  • pt_log
  • pt_jetseed_log
  • d0_abs_log
  • z0sinThetaTJVA_abs_log
  • dEta
  • dPhi
  • nIBLHitsAndExp
  • nPixelHitsPlusDeadSensors
  • nSCTHitsPlusDeadSensors
  • Cluster-level variables
  • et_log
  • pt_jetseed_log
  • dEta
  • dPhi
  • SECOND_R
  • SECOND_LAMBDA
  • CENTER_LAMBDA

Tau ID RNN Input Variables

# of tracker (pixel, SCT) hits β†’ # of tracker hits + # of dead sensors

  • Designed to protect against varying

detector conditions # of IBL hits β†’ If a hit is expected, use actual #

  • f IBL hits. If not, set # IBL hits = 1.
  • If # of IBL hits were set to 0 in the latter

case, might be wrongly classified as a bad track

slide-16
SLIDE 16
  • 16

Tau ID RNN Input Variables

slide-17
SLIDE 17
  • 17

CALORIMETER CLUSTERING INITIAL TRACKING PRECISION TRACKING TAU IDENTIFICATION

  • Inclusive MVA TES:

MVA Tau Energy Scale calibration applied to all tau candidates

  • Minimum pT cut
  • β€œFast Track

Finder” (FTF): Look for a track in a narrow cone (Ξ”R < 0.1) around the center of the cluster & along the full beam line

  • RNN

0-/1-/multi-prong RNN with tau ID variables + track + cluster variables as inputs

  • Track refit:

Run precision tracking on FTF seeds

  • nTracks cut

nTracks Cut: only pass taus with… βœ… 1-3 tracks in core region βœ… ≀1 tracks in isolation region

Tau HLT in 2018