Exploring Artificial Neural Networks to Discover Higgs at LHC

1
Exploring Artificial Neural Networks to Discover
Higgs at LHC

• Using Neural Networks for B-tagging
• By Rohan Adur
• www.hep.ucl.ac.uk/radur

2
Exploring Artificial Neural Networks to Discover
Higgs at LHC

• Outline
• What are Neural Networks and how do they work?
• How can Neural Networks be used in b-jet tagging
  to discover the Higgs boson?
• What results have I obtained using Neural
  Networks to find b-jets?

3
Neural Networks - Introduction

• Neural Networks simulate neurons in biological
  systems
• They are made up of neurons connected by synapses
  
• They are able to solve non-linear problems by
  learning from experience, rather than being
  explicitly programmed for a particular problem

4
The Simple Perceptron
Output layer

• The Simple Perceptron is the simplest form of a
  Neural Network
• It consists of one layer of input units and one
  layer of output units, connected by weighted
  synapses

Synapses connected by weights
Input layer
5
The Simple Perceptron contd.

• Requires a training set, for which the required
  output is known
• Synapse weights start at random values. A
  learning algorithm then changes the weights until
  they give the correct output and the weights are
  frozen
• The trained network can then be used on data it
  has never seen before

Output layer
Synapses connected by weights
Input layer
6
The Multilayer Perceptron
Output layer

• The main drawback of the simple perceptron is
  that it is only able to solve linearly-separable
  problems
• Introduce a hidden layer to produce the
  Multilayer Perceptron
• The Multilayer Perceptron is able to solve
  non-linear problems

Synapses
Hidden Layer
Synapses
Input layer
7
Finding Higgs

• The Higgs boson is expected to decay to b-quarks,
  which will produce b-jets
• b-jet detection at LHC is important in detecting
  Higgs
• 40 million events happening per second
• b-taggers must reject light quark jets

8
b-tagging

• B mesons are able to travel a short distance
  before decaying, so b-jets will originate away
  from the primary vertex
• Several b-taggers exist
• IP3D tagger uses the Impact Parameter of the
  b-jets

1mm
Primary Vertex
B
B-jets
Secondary Vertex
IP

• SecVtx tagger reconstructs the secondary vertex
  and rejects jets which have a low probability of
  coming from this vertex

9
IP3D Tagger

• Good amount of separation between b-jets and
  light jets

10
b-tagger performance
11
Neural Network for b-tagging

• The current best tagger is a combination of IP3D
  and SV1 tag weights
• Using Neural Networks, can this tagger be
  combined with others to provide better separation?

12
The Multilayer Perceptron and b-tagging

• The TMultiLayerPerceptron class is an
  implementation of a Neural Network built into the
  ROOT framework
• It contains several learning methods. The best
  was found to be the default BFGS method
• Train with output 1 for signal and output 0
  for background
• The b-tagging weights were obtained using the
  ATHENA 10.0.1 release
• The data was obtained from Rome ttbar AOD files
• Once extracted, the weights were used to train
  the Neural Network

13
Results

• 5 Inputs used Transverse momentum, IP3D tag, SV1
  tag, SecVtx Tag and Mass
• 12 Hidden units and 1 Output unit

14
Results Contd.
15
Results Contd.
Rejection rates
Efficiency IP3DSV1 Neural Network
60 88 136
50 175 387
Mistagging efficiency
Efficiency IP3DSV1 Neural Network
60 1.14 0.73
50 0.57 0.26
At fixed rejection
Rejection IP3DSV1 Neural Network
100 57 62
16
Discussion of Results

• Using a Neural Network, b-taggers can be combined
  to provide up to double the purity at fixed
  efficiency
• At fixed rejection rate, the Neural Network
  provides 5 more signal than the IP3DSV1 tagger
  alone
• Neural Network performance is not always
  reproducible. Each time training is undertaken a
  different network is produced

17
Conclusions

• Neural Networks are a powerful tool for b-jet
  classification
• Neural Networks can be used to significantly
  increase b-tagging efficiency/rejection ratios
  and could be useful in the search for Higgs
• Training a Neural Network on real data will be
  the next hurdle