A. Sokolov - Moscow

abstract: Applications of Neural Networks (NN) in off-line data
analysis in High Energy Physics (HEP) use mostly the ability of NN to
solve intricate pattern recognition problems. The power of NN
technique was demonstrated in diverse fields of particle physics.  For
the next generation of accelerators an use of NN have to become much
broader. Higher energies mean higher multiplicities and thus the next
generation of HEP experiments will have to deal with a wealth of
information both on-line and off-line. Vast amounts of data require an
increase in the speed of data processing and make massive parallelism
unavoidable.  A neural network is a natural tool here.
The applications of NN for particle identification, event
classification, energy reconstruction as well as current trends of
investigations in this field are reviewed.

        J. M. Lopez, F. Matorras, A. Ruiz - Univ. de Cantabria

abstract: A neural network approach has been used to identify
different exclusive $\tau$ decays among the $\tt$ events collected by
the DELPHI experiment at LEP. With the use of two feed forward NN up
to ten one and three prong decays were analysed.  The method operates
with better efficiency-purity than the classical methods for the
previously studied channels and allows the possibility of analysing
additional ones.
It has been applied to the whole data sample taken around the $Z^0$
resonance to obtain the branching fractions of these channels. A
detailed study of the systematic errors has been performed.

       G. Dror, E. Etzion, D. Horn - Tel Aviv

abstarct: We devise a neural network solution for a complicated
inverse problem in particle physics experiments.  In spite of
intrinsic stochasticity of the data and the abundance of missing and
redundant values, simple networks are efficient in performing robust
estimation and classification.  We use second order learning methods
to train the network over a set comprising a few thousand simulated
events over a wide range of parameters.  The trained network produces
an unbiased estimate of the transverse momentum and charge of the
particles.  On a test set the error in transverse momentum estimate
stems primarily from intrinsic noise in the data.  Similar methods are
applied for classification of the data into 'background' and
'interesting', totally providing 96\% accuracy.

        G. Dror, H. Abramowicz, D. Horn - Tel Aviv

abstract: In High Energy Physics experiments one has to sort through a
high flux of events, at a rate of tens of MHz, and select the few that
are of interest.  One of the factors in making this decision is the
location of the vertex of the event.  Here we present an
unconventional solution to the problem of finding the the vertex,
based on two feed-forward neural networks with fixed architectures,
whose parameters are chosen so as to obtain a high accuracy.  The
system is tested on simulated data sets, and is shown to perform
better than conventional algorithms.

         Harpreet Singh - University of California, Riverside

abstract:Neural Networks  provide a powerful  tool with  the  potential to give
excellent   signal   to    background   discrimination   in    complex
environments. In this paper,  we will give  an overview of  the neural
network  techniques used for  studies  of $t\bar t$  decays  in the D0
experiment  at Fermilab. The  neural  network studies in the  dilepton
channel ($t\bar t\rightarrow e\nu\mu\nu b\bar b$) will be discussed in
detail.  Earlier work on  the measurement of production cross  section
in   the all-jets ($t\bar t\rightarrow  qqqqb\bar  b$) channel and the
measurement of top quark mass  in the lepton+jets ($t\bar t\rightarrow
l\nu qqb\bar b$) channel will be reviewed.

        S. Tentindo-Repond - D0

abstract: D0 has searched for Leptoquarks associated with the three generations,
 using both conventional and neural network techniques. Use of advanced
 search techniques (NN) has allowed to maxinize discrimination (signal
 versus background) and to set best limits on cross sections and on LQ
 Results of these searches will be presented. The Tevatron Combined
 (D0 + CDF) mass limit for First Generation Leptoquark (beta=1) is
 242 Gev/c2.

         R. M. de Lucas

abstract: An Artificial Neural Networks is used to search for the Higgs
 Boson of the Standard Model and Minimal SUSY extension
  of the SM, MSSM, in the DELPHI detector at LEP2 e+e- collider at
  CERN, at high energies.
  The use of this thechnique is compared  with other
  multidimensional methods and results are presented for the
  selection of the Higgs boson signal of different masses.

        A. Klier, M. Kupper

abstract: A search for Standard Model Higgs boson through the process
e^+e^- -> Z^0H^0 is conducted using data from the OPAL detector at
LEP.  The search in the Missing-Energy channel is aimed at identifying
final states in which the Z^0 decays into a neutrino pair and the
Higgs goes to a b-quark pair, resulting in two b-flavoured jets and
large missing energy.  A new ANN-based analysis for the Missing-Energy
channel is presented and compared with the currently used
likelyhood-based analysis.

        Y. Rozen, Sh. Tarem and A.  Harel

abstract: Rare B decays offer a good opportunity to discover new physics
beyond the standard model. One such possibility is the R parity violating B
decay, $b\rightarrow ss\overline{d}$.
A search for the R parity violating decay channel B$^\pm\rightarrow$
K$^\pm$K$^\pm\pi^\mp$ was conducted using data collected by the OPAL
detector at LEP. Since no such signal has been reported in the past, a
powerful signal extraction technique is needed. We have used artificial
neural network to improve the significance of a possible signal detection.
The analysis presented here optimise the signal significance as well as
keeping the systematic uncertainties associated with this technique at
a low level. Preliminary results are given.

     Hardware applications

         G. Wrochna - Warsaw

abstract: The trigger system of the CMS detector has to reduce the
initial rate of 40 MHz proton-proton interactions down to 100 Hz
acceptable for mass storage.  The reduction is done in several steps.
The first level trigger is based on hardware processors (ASIC, FPGA,
look-up tables). It recognizes muons, electrons/photons, jets, and
missing energy.  Applied momentum (energy) cuts reduce the rate below
100 kHz.  Higher level triggers are performed by a farm of commercial
processors. This gives a lot of freedom for choosing algorithms. Neural
nets are one of the possibilities

          S. Gonzalez - CERN


         Joao Varela - LIP/IST and CERN

abstract: We discuss the application of Modular neural Networks (MNN)
for high-performance, high rate event triggering in HEP experiments,
both in terms of the algorithms involves and their hardware
implementation. Two different problems were treated successfully with
the MNN framework, namely the trigger of electrons and photons in the
CMS experiment and the classification of Cherenkov rings in a RICH

         F. Spinella - Pisa

abstract: The Silicon Vertex Tracker (SVT) is an online tracker for
the CDF upgrade which will reconstruct 2D tracks using information
from the Silicon VerteX detector (SVXII) and the Central Outer Tracker
(COT).  The SVT will work in the level 2 of the CDF trigger chain, to
refine the level 1 tracking information from the eXtra Fast Tracker
(XFT), which uses data from the central drift chamber COT. It will
combine XFT tracks with hit coordinates from SVXII.The level 2 latency
time is about 20 microseconds, therefore the design of the SVT has
concentrated on parallelizing the various tasks, from the
reconstruction of the hit coordinates from the single strip pulse
heights to the pattern recognition and the final precision track
fitting. The result is a data driven architecture in which many
functions overlap in the internal processor pipeline and which
comprises several different modules, built on 9U Eurocard boards with
VMEbus implementation for diagnostic and control. The SVT
specifications require 30 MHz operation for each module.  The
precision measurement of the track impact parameter will be used to
select and record large samples of B hadrons.  We discuss the overall
architecture, algorithms, and hardware implementation of the system.

         C. Kiesling - MPI Munich

abstract: At the HERA ep collider the H1 experiment is
successfully operating a set of hardwired digital neural networks at the
second trigger level. The latency of the network triggers, which are of
the feed-forward type, is about 20 microseconds. The inputs to the
networks are suitably preprocessed 8 bit quantities, based on the trigger
information available from the various subdetectors of H1 at level 1. We
present the principles of the neural trigger, its hardware and its use
within the overall H1 trigger strategy.  We present some recent physics
results from ep reactions triggered specifically with the neural trigger
and indicate some future developments on the preprocessing of the neural
input, designed for the future high luminosity running at HERA beyond the
year 2000.

          Bruce Denby - Versailles

abstract: The talk will survey past, present, and future hardware platforms
useful for implementing neural network architectures, with an
emphasis on those amenable to use in high energy physics experiments.
Included will be ETANN, CNAPS, TOTEM, NeuroClassifier, and an
FPGA solution called Maharadja, among others.

           Bruce Denby - Versailles

abstract: This paper presents the architecture of Maharadja, a low-power
consumption system for performing real time simulation of RBF networks
using three possible distances: Manhattan, Euclidian, Mahalanobis. For the
Manhattan distance the system performance is similar to that of existing RBF
dedicated hardware such as Zisc or Ni1000, but, unlike those chips,
Maharadja can also simulate Euclidian and Mahalanobis distances in real

        Jose M. de Seixas - Rio de Janeiro

abstract:  Neural networks have been extensively used in pattern
recognition problems, including triggering systems for experimental physics.
There is a clear tendency to consider neural networks as powerful
processing blocks that interface with other non-neural blocks within
an overall hybrid processing system.  By making the neural network
to work in conjunction with well understood techniques (preprocessing
methods), improvements on the overall performance can be achieved.
Considering this approach, recent neural network designs for
online triggering systems are discussed, with emphasis on calorimeter
triggers.  It is also presented a number of successful applications
on other fields of research of techniques that have originally been
developed for the experimental physics environment.


         D. Horn - Tel Aviv