Online Electron/Jet Neural High-level Trigger Over Independent Calorimetry Information

1
Online Electron/Jet Neural High-level Trigger
Over Independent Calorimetry Information

• R.C. Torres1,2, J.M. Seixas1,
• A.R. dos Anjos3, D.V. Cunha1

1 Federal University of Rio de Janeiro 2 CERN 3
University of Wisconsin Madison
2
Summary

• The ATLAS Experiment
• The High Level Trigger
• The proposed discrimination system
• Feature extraction
• The ringing process
• Independent component analysis
• Hypothesis making
• Achieved results
• Conclusions
• Future work

3
The ATLAS Experiment

• International collaboration (32 countries).
• Built to analyze the collisions promoted by the
  Large Hadron Collider (LHC) at CERN, which can
  collide protons with 14 Tev in its center of
  mass.
• One of its main goal is to experimentally prove
  the existence of the Higgs Boson.
• Composed by multiple, high resolution
  subdetectors
• Inner detectors (tracking).
• Electromagnetic calorimeter.
• Hadronic calorimeter.
• Muon chamber.
• Due to detector segmentation and the high LHC
  event rate (40 MHz), a data stream of 60 TB/s is
  expected.
• Events of interest occurs at a frequency smaller
  than 1 Hz.
• To cope with these stringent conditions, an
  online trigger system is required.

4
The ATLAS Trigger Overview
5
ATLAS Calorimetry for LVL2
6
Level 2 Trigger Facts

• The ATLAS Trigger is predominantly inclusive
  (searches for high-pT representative objects and
  accepts events based on those).
• Important objects to be identified in this system
  are high-pT electrons.
• In average, for every 25,000 high-pT electrons
  accepted by LVL1, it is expected that only 1 is a
  true electron.
• Todays algorithms for electron identification
  using calorimetry information _at_ LVL2 (T2Calo
  EGammaHypo) use basic clustering strategies and
  apply linear cuts to identify electrons and
  reject jets.
• If approved by the cuts applied after T2Calo,
  Inner Detector (tracker) algorithms are applied
  to the object. These algorithms take longer
  processing times.
• The average processing time per event (not per
  RoI) should be in the order of 10 ms.

7
Things We Have Tried to Address

• Provide an electron/jet separation algorithm.
• Operation split into feature extraction and
  hypothesis making sections.
• Provide a feature extraction module that keeps
  the physics interpretation of the events, while
  isolates the relevant information for the
  discrimination.
• Use neural networks to perform non-linear
  correlation of a high dimension input space in
  order to improve detection efficiency.
• Try to improve early jet rejection so the HLT
  system looses less time on uninteresting objects.
• Try to keep ourselves within the time budget
  defined by LVL2.

8
The Ringer

• Ring formatting (1,000 cells for 0.4 x 0.4 RoI).
• The ring sum strategy is adapted to layer
  granularity and leads to 100 values, using full
  RoI data.
• This approach, like T2Calo, preserves the physics
  interpretation of the events.
• Energy-based normalization can use the energy per
  layer, section (E.M. and Hadronic) or of the
  whole object.

9
Independent Component Analysis

• Analysis which relies on high order statistics,
  in order to separate independent sources.
• Therefore, it can be used to isolate incoming
  signal sources from noise and pile-up, making the
  target pattern more evident to the hypothesis
  algorithm.
• The independent components are applied over the
  ring sums.

Independent Component Transformation
M
Ind. Sources (S)
Acquired Signal (P)
Calorimeter
Pre-Processing
10
ICA Applied on Ring Formatted Cells
Cells
Electrons
Jets
ICA
11
Hypothesis Section

• Receives independent sources extracted from ring
  sums.
• Uses a neural network to perform the
  discriminator.
• Provides non-linear cuts, being able to better
  separate electrons and jets.
• Robust against noise and loss of information.
• Easy to implement and fast to execute.
• May help identifying unknown (new) patterns.

12
Achieved Results with Ring Sums and ICA
T2Calo EGammaHypo

• A network with a single linear node achieves
  almost the same result as a network with 12
  non-linear hidden nodes fed with only the ring
  sums.
• ICA efficiently applies non-linear uncorrelation
  of the input information.
• Best results achieved using only 4 non-linear
  hidden nodes, resulting in 96.4 of electron
  detection, for a false alarm of 3.9.
• For the same detection efficiency as T2Calo
  EgammaHypo, the proposed algorithm rejects almost
  7 times more jets.
• ICA, therefore, allows a simpler neural network
  design.
• 1 false alarm reduction means 250 less fake
  electrons propagated to the event filter
  (improved early rejection).

13
? and ? Analysis
14
Timing Analysis

• Executed on a Intel Core 2 Duo _at_ 2.33 GHz with 2
  Gbytes of RAM.
• Considering the RoI information already retrieved
  from the ROSs.
• Results show that the algorithm is able to cope
  with the time budget (10 ms) for the LVL2.

Step Mean (?s) Std (?s)
Peak Find 8.32 1.30
Build Rings 188.40 49.61
Normalize 12.95 1.90
ICA Proj. 24.93 1.45
Discriminate 95.33 5.02
Total 329.93 50.84
15
Conclusions

• ICA ring sums helped to unmask the relevant
  information from the discrimination point of
  view.
• This resulted in a simpler neural network design
  for the hypothesis making section.
• The achieved results allowed, for the same
  detection efficiency, a false alarm 7 times
  smaller than T2Calo EgammaHypo.
• The time analysis showed that the proposed
  algorithm is a feasible LVL2 option.

16
Future Work

• Evaluate segmented ICA discrimination approach,
  in order to fully exploit detection segmentation.
• Apply relevance analysis in order to select the
  most relevant independent sources, for further
  signal compaction and trigger tuning.
• Start beam misalignment analysis.
• Start pile-up analysis.