Report of the neural network activities of the last three years
G. Athanasiu, P. Pavlopoulos and S. Vlachos
February 1995
Since 1992 we have been involved in developing several neural network
applications. Most of them were facing problems of the CP-LEAR experiment
which however, as usual with neural network algorithms,
could be easily modified to accommodate various needs.
In our group we are dealing with all the parts of the development
of such applications. We start from the detailed software simulation
using our own computer tools (including a small neural network
development system), and we go on up to the point of constructing
the actual hardware that executes the algorithms on-line.. Particular
attention is payed every time to the real-time performance of any
system as our studies are focused on trigger applications.
In what follows there is a description of our neural network projects
since 1992.
- A neural network model has been developed to recognize tracks in
the CP-LEAR detector [1]. The detector was represented as
a set of 11 individual superimposed sub-detectors, each one having 64
sectors. The final neural network solution obtained could not only
count the number of tracks in an events but also locate them in the
detector volume. As the system was aimed to be used as an on-line event
selection process the simplest working architecture had to be
established This lead to a network with only two nodes in a single
hidden layer. The function of each of those two hidden nodes
resembles that on neurons in the first layer of the mammalian retina.
- Following the work described above, specialized hardware
was built to implement that network as an actual trigger stage of
the CP-LEAR experiment [2]. In parallel a detailed software
simulation of the hardware was developed. It was used to verify the
function the hardware as well as to compare this trigger decision
with other similar ones that are already implemented.
The hardware design was based on commercially available ECL chips.
16 cards were built in a first phase in order to cover a quarter
of the whole CP-LEAR detector. Each card needs only 60 ns for
a track decision, and as all of the cards work in parallel,
the whole system evaluates an event in less than 75 ns. It should
be noted here that conventional track-following techniques,
already in use in the same experiment, need about 600 ns for
the evaluation of a whole event.
The complete system assembled in a crate with special
interconnection cards has been already tested with CP-LEAR
events and its performance matches that from its detailed
software simulation.
Furthermore the system was tested against the track-following
assembly and against elaborated off-line track counting methods.
Clearly superior in terms of execution time, the neural network
system was only marginally worse in event selection than the
off-line one.
It should also be underlined that the hardware was developed
in such a way as to be able to accommodate any neural network
model based on the same neural network architecture.
- Again as a part of the CP-LEAR trigger a neural network
was developed to select on-line events of physics interest [3].
It based its decisions on the momenta of all the detected tracks
of an event. Events were selected if they contained a decaying
neutral kaon. The network established had not only
to select the proper events but to ensure that no biases were
introduced in the selected event sample. Again after comparing
with traditional methods the neural network turned out to perform better.
Once the proper model was found, it was implemented in hardware to be
actually used on-line. As the time limits at the corresponding
trigger level were not critical, the algorithm was implemented
on a commercial Digital Signal Processor (DSP) board.
However special modules were designed and built, for interfacing
the DSP board to the outside trigger world.
The trigger decision for each event is provided 38 mus after
the input data are available. The system is already in use
in the CP-LEAR trigger. It provides a factor of 2 improvement
of the signal to noise ratio of the sample of selected events.
Detailed studies have shown that the network's selection
is quite robust against calibration variations and missing
information.
- Independently, a trigger stage was designed for a new
experiment (DIRAC, a fixed-target experiment studying
pi^+pi^- atomic states). It was based on neural network
algorithms and will use, when installed, the hardware cards
developed for the track recognition.
References
- Real time track identification with artificial
neural networks, G. Athanasiu et. al.,
Nucl. Instr. and Meth. A324 (1993) 320-329.
- Hardware realization of a fast neural network trigger
step for real-time tracking in HEP experiments,
F. Leimgruber et. al.,
submitted to Nucl. Instr. and Meth. A.
- Identification of neutral kaon events using
artificial neural networks, G. Polivka et. al.,
to be published in Nucl. Instr. and Meth. A.
NNW/HEP Home Page
Authors:
Clark S. Lindsey ,
Bruce Denby &
Thomas Lindblad
Curator:
Clark S. Lindsey (lindsey@particle.kth.se)
Latest revision: 15 Feb 1994