Status of Online Neural Networks Bruce Denby Université

Status of Online Neural Networks Bruce Denby Université de Versailles and Laboratoire des Instruments et Systèmes, Paris, France Rapporteur’s Presentation ACAT 2000 Fermilab 16 -20 October, 2000 1

OUTLINE OF THE PRESENTATION I. III. IV. The current situation Developments foreseen Neural net hardware Conclusions 2

Acknowledgements Most of my transparencies were borrowed from the talks of: • • • Sotirios Vlachos Erez Etzion Jean-Christophe Prévotet Christian Kiesling Bertrand Granado 3

The Current Situation Neural network triggers are being used to produce physics. Examples: 1) Dirac Experiment at the CERN PS 2) H 1 Experiment at HERA 4

5

• 34 Ge. V p on target • Measure lifetime of pionium • Hodoscope input to NN 6

-The network is trained to select low Q events 7

• Net architecture 55 -2 -1 • Note that the multiply/accumulate and sigmoid evaluation are done using look-up table memories. 8

- i. e. , it works…. 9

The H 1 Neural Network Trigger Project Christian Kiesling Max-Planck-Institut für Physik München, Germany 10

The H 1 Experiment at HERA u Mission: p 920 Ge. V e 27. 6 Ge. V d u Gluon study • the structure of the nucleon • the fundamental interactions of quarks and gluons : Quantum chromodynamics (QCD) • electroweak interference search • for physics beyond the Standard Model Physics analysis : Measurements of the structure functions F 2, FL, F 3, F Jet Measurements (strong coupling constant) Charm/Beauty Production (gluon content of proton) Diffractive Vector Meson Production (gluon struct. ) Search for Instanton Effects (QCD „exotics“) D 2 Hardware (MPI): Liquid Argon Calorimeter (forward barrel section) LAr front end electronics LAr trigger (L 1) Neural Network Trigger (L 2) 11

The H 1 Trigger Scheme software hardware L 1 trigger: OR of individual subdetector triggers, such as MWPC, CJC, LAr, Spa. Cal, system. . . L 2 systems: have access to information from all subdetectors (information prepared by subtrigger processors) Neural Network at Level 2: Global Event Decision 12

Detector Information at level 2 Calorimeter (LAr) hadronic electromagnetic Spa. Cal (Pb scint. ) µ chambers Trigger towers, global energies (8 bit numbers) Trigger towers above threshold (single bits) Hits (single bits) “physics” (example of photoproduction) Central Jet chamber Nr. of tracks (8 bit numbers) MWPC’s (2 sets) Hits (single bits) z-VRTX (from MWPC) 16 bin histogram (8 bit numbers) “background” and there is much more physics in H 1. . . 13

Architecture of the H 1 Neural Network Trigger Three-Layer Feed Forward Neural Net Output (only one neuron) One hidden layer Inputs (from detector) weights discriminate „physics“ from „background“ : Central Problem: Inputs for the Neural Nets background Data Selection physics Data Transformation 14

Organization and Processing of Data from L 1 The L 2 Bus (8 subbusses, 16 bit wide) Subdetector 1 0 Subdetector 2 1 2 Subdetector 3 3 4 5 6 7 0 2 4 6 8. . . Subdetector information arrives in consecutive time slices ti („frames“, or bunch crossings BC) (tmax = 32 BC’s at present) t(BC) 1 BC = 96 ns = 10 MHz transfer rate The Data Distribution Board (preprocessing of neural input) to neural network D D B I L 2 crate backplane: L 2 Bus Selection of input data Processing (look-up, summing) data input units: Cables from subdetectors (maximum of 40) 15

The Neural Trigger System : Modular and Expandable Network processors To Final Decider Data selection and Data transformation CNAPS 11 DDB 11 CNAPS 10 DDB 10 CNAPS 9 DDB 9 CNAPS 8 DDB 8 CNAPS 7 DDB 7 CNAPS 6 DDB 6 CNAPS 5 DDB 5 CNAPS 4 DDB 4 CNAPS 3 DDB 3 CNAPS 2 DDB 2 CNAPS 1 DDB 1 CNAPS 0 DDB 0 Monitoring SBus Interface VME SUN / SBus Interface Loading and Control X 11 Terminal Set of independent networks, each one trained for a specific physics reaction Data from Detector 16

The complete System 12 independent networks Pre-processing modules (one for each neural network) Cables carrying raw input data from the detector Total of 1024 processors Integrated computing power: over 20 Giga MAC/sec 17

The Neural Network Trigger in Operation: Trigger rate Monitor (24 h) (random day in early 1999) 1: Background rejection factor > 100 ! 2: 4: 5: 6: 7: 8: 9: 10: 11: (Boxes 0 and 3 also active during 99/00) 18

Some Physics: Elastic Photoproduction of Mesons QCD xg expected small in Regge theory expected large in QCD C. Adloff et al. , Phys. Lett. B 483 (2000) 23 Due to highly selective NN trigger background is under control up to the highest HERA energies 19

Photoproduction of Mesons with Proton Dissociation Recent results on d /dt : Measurement possible due to neural trigger (publication in preparation) 20

Developments Foreseen I. II. H 1 upgrade Atlas 21

H 1: New Network Preprocessing - The DDB II Why a new preprocessor? Neural Network Trigger successfully in operation since Summer 1996, promising physics results, but: So far no information from LAr trigger towers used, only global energy sums, no subdetector correlations (limitation was dictated by time schedule for the realization of the trigger) NOW: need to prepare for higher selectivity (luminosity upgrade: HERA 2000: factor 5 more physics @ constant logging rate) New Goal: separate “interesting” physics from “uninteresting”physics Need more Intelligent Preprocessing 22

Intelligent Preprocessing for Neural Networks Jean-Christophe Prévotet, MPI München Laboratoire des Instruments et Systèmes (Paris VI) 23

New Preprocessing : The DDB 2 Principle - “intelligent” preprocessing” extract physical values for the neural net (impulse, energy, particle type) - Combination of information from different subdetectors (the, phi plane) - Executed in 4 steps Clustering find regions of interest within a given detector layer Matching Ordering combination of clusters sorting of objects belonging to the same by parameter object Post Processing generates variables for the neural network 24

Description of a DDB 2 board Storage of parameters L 2 bus Data Clustering BT/TT Addresses MEM Clustering MWPC Matching Ordering Clustering CJC MEM Clustering FTT MEM Clustering Muon MEM Post Processing Clustering Spacal Matching Workable data given to the NN MEM 25

Hardware specifications Time : 8µs (Clustering, Matching, Ordering, Post Processing) Re-configurable hardware independent of data format changes Organization : 5 DDB 2 boards connected to 5 CNAPS Each board works on the same data but parameterized differently 26

Hardware resources Time : 8 µs FPGA : - Low cost (prototype board) - Speed Parallel processing Pipeline steps - Xilinx Virtex Family XCV 200, XCV 400 Data format Luts Lot of small memories Type N° gates XCV 200 236 K XCV 400 468 K Algorithm Clustering Number Type 6 to 8 XCV 200 Matching 2 XCV 400 Rams 14 20 Sel. Ram bits 75 K 153 K Ordering Post processing 1 to ? XCV 200 27

How does Physics profit from the DDB II ? Test reaction: photo-production Backgr. Physics „DDB II“ photoproduction (DDB II simulated with DDB I) Backgr. DDB I Physics Gain about a factor of 2 in efficiency with the new DDB II algorithms for this case. Expect increased selectivity also for other physics. . . 28

Momentum Reconstruction and Triggering in the ATLAS Detector Fermi. Lab, October 2000 Erez Etzion 1, Gideon Dror 2, David Horn 1, Halina Abramowicz 1 1. Tel-Aviv University, Tel Aviv, Israel. 2. The Academic College of Tel-Aviv-Yaffo, Tel Aviv, Israel. 29

ATLAS S. C Solenoid Hadron Calorimeter S. C Air core Toroids Inner Detector Muon Detectors EM Calorimeters 30

Low. Pt High Pt trigger Complicated magnetic field map => difficult problem 31

Network architecture PT Q linear output sigmoid hidden layers input parameters of straight track of muon (preprocessing LMS) 32

Testing network performance Training set 2500 events. In one octant. Test set of 1829 events. Distribution of network errors approximately gaussian. compatible with stochasticity of the data. charge is discrete!!! 95. 8% correct sign. 33

Summary & discussion • The network can successfully estimate the charge and transverse momentum of the muon. • Classification (triggering) is most efficient by specially trained network. • The data is intrinsically stochastic giving rise to approximately gaussian errors. • The simplicity of the network enables very fast hardware realization. (See presentation this workshop) 34

Neural Network Hardware • Off-the-shelf neural net hardware is scarce • Many standard products no longer exist • What should we do in HEP? 35

Analog Devices: ETANN, 1991 (Electrically Trainable Artificial Neural Network by Intel) (64 x 4 in 5 s) Neuro. Classifier, 1994 (by P. Masa, Univ. Twente, NL) Digital Devices: (70 x 6 x 1 in 20 ns) Blue color: chip no longer produced CNAPS 1993 (Adaptive Solutions, Oregon) 64 @20 MHz 8/16 MA 16 1994 (Siemens, Germany) 16 @50 MHz 16/16 TOTEM 1994 (Trento, Italy) 32 @30 MHz 16/ 8 4 @50 MHz 16/16 SAND 1 1995 (Kf. K, Germany) recent development: Maharadja, 1999 (Paris, France) details at this conference (see talk of B. Granado, AI, Sess. I) back to analog (? ) „Silicon Brain“ (Irvine Sensors Inc. ) 3 D analog FPGA array towards a complexity similar to the human brain. . . 36

- One interesting solution: use memories to evaluate NN’s 37

- Another solution: can we use a fast ‘general purpose’ NN processor implemented in FPGA’s? 38

39

40

41

42

43

44

- FPGA clock speed of 100 MHz will be available soon. - implying execution times of a few 100 ns. 45

Conclusions 1. 2. 3. 4. 5. 6. 7. Fast preprocessing is a concern – FPGA’s are one way to go H 1 NN trigger upgrade is in the works There is some NN trigger Neural net triggers exist and they work activity in LHC experiments: ATLAS muon proposal (this workshop), CMS (electron trigger, Varela et al. ) Finding NN hardware is a problem Memory or FPGA implementations may be the answer See also Neural Networks in High Energy Physics: A Ten Year Perspective, B. Denby, Comp. Phys. Comm. 119, August 1, 1999, p 219. 46