Status of NNW's in HEP Experiments

CDF Neural Network Triggers

The CDF experiment at the Tevatron proton-antiproton collider at Fermilab in the USA has had several calorimeter neural network triggers running for several of years now. See Badget 1992 for some early results. Three ETANN fastbus cards , receive analog signals from 50 trigger towers. Two of the ETANN's do not use trained networks. One of these identifies isolated photon showers in the central calorimeter and the other looks for isolated electron showers in the plug calorimeter. Differences in the weighted sums of central and surrounding towers are calculated by the chips and then thresholds cuts on the outputs provide the triggers. The third network implements a trained network to identify b-jets in the central calorimeter.

A tau finding network with a trainied network has been developed recently and implemented in the 1994-1995 run by a group from Rutgers. The network has performed satisfactorily at triggering on tau 1-prong hadronic decay events but not well on 3-prong. Although the chip emulation worked well with both types of events, the chip-in-the-loop training of the ETANN diverged badly (all weights went to max values) from the initially loaded emulation network. So they have used only the initial downloaded network, which, because of the inexact nature of the analog chip, does not perform as well as a network tuned with chip-in-the-loop training. There are now plans to examine the possibilities of a digital hardware network for future runs.

CP-LEAR Neural Networks

Neural networks for tracking and triggering have been developed for possible use in the CP-LEAR experiment. The hardware networks can

• count and locate tracks in the tracking detector within 75ns using NNW's implemented in 16 cards with commercial ECL chips (Athanasiu, Leimgruber);
• do on-line event selection, using information from the tracking network, within 40microsecs with a NNW implemented on a commercial DSP board (Pavalopoulos).

In both cases, the NNW's were found to perform better than the traditional methods. For further details see the above references, a brief hardware description , and the CP-LEAR/Basel NNW Activities Report.

H1 1st Level Neural Network Trigger

The H1 experiment at the HERA ep Collider at the DESY Laboratory in Hamburg, Ge., has developed a 1st level trigger that runs in 50ns in addition to the 2nd level trigger discussed below. It is based on the Neuroclassifier chip that was specifically designed for this experiment. This chip has 70 analog inputs, 6 hidden neurons with 5-bit digitally loaded weights and sigmoidal transfer functions, and 1 summed output (transfer function is off-chip to give the option of combining chips to increase the number of hidden units). A fast real-time tracking system provides a 16 bin histogram of track vertices along the beamline. A true electron-proton interaction will create a sharp peak in this histogram at the location of the collision. Whereas, a background event of a proton-beamgas or proton-beampipe collision will usually result in a flat distribution of vertices. These 16 bins are fed into the chip, which then gives yes/no answer to whether there is a good peak in the proper range. This trigger will go online in the fall of 1995. For futher info, see Tong et al and Schiek and Schmidt .

H1 2nd Level Neural Network Trigger

The H1 experiment at the HERA ep Collider at the DESY Laboratory in Hamburg, Ge., has implemented a level 2 trigger based on hardware neural networks. Here is the abstract from The Realization of a Second Level Neural Network Trigger for the H1 Experiment at HERA, Fent et al presented by J.H.Koehne at AIHENP'96, Lausanne, Switzerland, Sept.2-8, 1996:

The 1996 period of H1 data taking at the electron proton collider HERA at DESY (Hamburg) will be the first year with a fully equipped second level trigger (L2). Two independent systems of pattern recognition machines are operating as event rejectors within 20 microseconds deadtime after the keep signal of the pipelined (2.1 microseconds) deadtime-free first level (L1) trigger producing an input rate to L2 of up to 1 kHz.
One of the two systems is the neural network trigger. It is based upon an array of VME-boards with a CNAPS 1064 chip (20 MHz) by Adaptive Solutions. The single instruction multiple data (SIMD) architecture of the 64-processor-nodes-chip allows to compute a 64-64-1-feed-forward neural network. With 8-bit input values and 16-bit weights stored in 4kb look up tables on each individual node the output value is computed in 8 microseconds. Both systems use a 8x16 bit wide L2 20-MHz-bus to transmit the L1 information from the subdetectors to the L2-machines. The total transmission time is 3.1 microseconds. The data for the neural network are prepared for the CNAPS input by a 'Data Distribution Board' providing several algorithms realized by transfer functions (look up tables) and simple arithmetic funtions implemented in a field programmable gate array XILINX 4008. An internet server with two child processes for loading and monitoring of the system runs on a Themis VME-Sun with two HyperSpark processors. The client application allows full control and loading of the machine from a X11-graphical user interface.

They have used 10 of the Adaptive Solutions CNAPS VME digital neural network cards. See the L2 Neuro Group home page and their hypertext paper for more details.

NEMO Experiment: Tracking with Elastic Networks and Cellular Automata

From the NEMO Software Group home page:

The goal of NEMO collaboration is to study double-beta decay without the emission of neutrinos of 100Mo and other nuclei to probe the effective Majorananeutrino mass down to 0.1 eV. The peculiarity of the NEMO experiment is presence of very low energy electrons, so we must take into account quite large multiple scattering effects in the heliumgas as well as hard interactions with different wires. We use a cellular automaton for searching for and an elastic net for fitting of charged particles trajectories with multiple scattering and breaks on wires. The advantages of the methods are simplicity of the algorithms, fast and stable convergency to real tracks, and reconstruction efficiency close to 100%. Both these methods were used for event reconstruction in the NEMO-2 experiment. Results of the application are successful. The next stage of our investigation will be taking into account a magnetic field to develop the algorithms for the NEMO-3 experiment.

WA-92 Neural Network Demonstration Experiment

Here is the abstract from Baldanza et al. 1994 preprint:

Results from a non-leptonic neural-network trigger hosted by experiment WA92, looking for beauty particle production from 350 GeV negative pions on a fixed Cu target are presented. The neural trigger has been used to send events selected by means of a non-leptonic signature based on microvertext detector information to a special stream, meant for early analysis. The non-leptonic signature, defined in a neural-network fashion, was devised so as to enrich the selected sample in the number of events containing C3 secondary vertices (i.e. vertices having three tracks with sum of electric charges equal to +1 or -1), which are sought for further analysis to identify charm and beauty non-leptonic decays. The neural trigger module consists of a VME crate hosting two MA16 digital neural chips from Siemens and two ETANN analog neural chips from Intel. During the experimental run, only the ETANN chips were operational. The neural trigger operated for two continuous weeks during the WA92 1993 run. For an acceptance of 15% for C3 events, the neural trigger yields a C3 enrichment factor of 6.6-7.1 (depending on the event sample considered), which multiplied by that already provided by the standard trigger leads to a global C3 enrichment factor of ~150. In the event sample selected by the neural trigger, one every ~7 events contains a C3 vertex. The response time of the neural trigger module is 5.8microsecs.