System Test with SGV input

1
System Test with SGV input

• Understanding results of previous system test
  with SGV input status end of February
• different code for comparing SGV and Marlin gave
  different results
• this had been shown to arise from the flavour
  tagging part of the chain
• suspected there might be a problem with the
  neural network code
• Since then
• a bug in the newest NN code was found and
  identified as source of the problems since
  starting
• to look at MARLIN results (i.e. since before
  Valencia)
• all previously seen odd effects explained,
  including the temporary vanishing of the
  problems
• at the time of the RAL collaboration meeting
• error easy to fix this has been done and
  results agree perfectly with former version of NN
  code
• after correction, excellent agreement between
  MARLIN and SGV,
• with MARLIN doing somewhat better for some of
  the tags

2
The source of the problem

• NN problem calculation of the scalar product
  from inputs and weights used for finding
• activation of each neuron uses the dimension
  of the first input vector to determine
• how many components to use
• vector storing the weights also stores the
  weight for the bias dim (weights) dim (inputs)
  1,
• this was ignored in the call, so in addition to
  what would have been needed one had the term
• weights dim(inputs) 1 arbitrary number
• in the calculation this contribution can be
  zero, by chance (looking as if problem vanished)
• The corrected version reads (file Neuron.cpp)
• double Neuronactivation(const
  stdvectorltdoublegt inputs) const
• double result stdinner_product(inputs.begin(),
  inputs.end(),_weights.begin(),0.0)
• result _bias_weights_numberOfInputs
• return result
• Ben checked carefully that the NN code is free
  of further problems of this type

3
Revisit NN output distributions

• previously saw deviation of outputs from new NN
  version as compared to old version,
• for identical inputs after correction, these
  agree 100,
• plots below show 2 histograms each
• red MARLIN flavour tag inputs fed into
  MARLIN flavour tag with corrected new NN code
• black MARLIN flavour tags fed into
  standalone flavour tag with old NN code

• checked printing out values for a few hundred
  jets that identical results are found,
• agreement also seen in 2D distributions of the
  output variables shown above (new vs old)

4
Resulting Purity vs Efficiency at the Z-peak

• excellent agreement of Marlin with SGV using
  identical input events,
• Marlin performing slightly better

5
Purity vs Efficiency at sqrt(s) 500 GeV

• excellent performance also at sqrt(s) 500 GeV,
  
• MARLIN result very close to SGV, better in some
  places