Machine Learning in PandaRoot GlueX-Panda Workshop G.Washington University, May 2019 Ralf Kliemt (GSI)
Motivation • Machine Learning (ML) is about modelling data • Self-learning algorithms gain knowledge from data to make predictions Why? • Gain computation speed in online scenarios • More precision, e.g. by respecting correlations • Let algorithms do the tedious tasks of recognising patterns, structures, principal components etc. � 2
Which type of ML? � 3
Which type of ML? � 4
ML Activities at PANDA Event online PANDA FPGA Storage Selection Local Global Event sim Sim Digi Analysis Paper Reco Reco Building Raw Alignm. offline Data Calib. � 5
ML Activities at PANDA Event online PANDA FPGA Storage Selection Local Global Event sim Sim Digi Analysis Paper Reco Reco Building Raw Alignm. offline Data Calib. � 6
Key Concept � 7
Boosted Decision Tree (BDT) • Break down data by steps of decisions • Based on features in training data • Splitting by maximising information gain —> Typical application: Classification for PID � 8
Artificial Neural Networks (ANN) • Transform data meaningfully • Learn iteratively � 9
A lot to choose from… � 10
A lot to choose from… � 11
Programming Frameworks & Packages • ROOT / TMVA • NumPy • TensorFlow (Deep Learning) • Keras (on top of TensorFlow with GPUs) • MLlib Apache Spark • Sci-Kit Learn • PyTorch (Deep Learning) • DL4J (Deep Learning, Java) • R Implementations � 12
Popular Choices � 13
Machine Learning for Forward Tracking
Machine Learning For Track Finding Ma Contribution at at PANDA A FTS “Connecting the Dots 2019” Waleed Esmail, Tobias Stockmanns, and James Ritman Wa On Behalf of the PANDA Collaboration Institut für Kernphysik (IKP) Arti tific icia ial l Neural l Netw tworks: s: Application on to the FTS: Create all possible combinations of hit pairs (adjacent � layers). Train the network to predict if hit pairs are on the � same track or not. Input observables: � 1) Hit pair positions in x-z projection (vertical layers). 2) Drift radii (Isochrones). 3) Distance between hits. Output: � 1) Probability that hit pair are on the same track. � Connect hits that pass the probability cut ( threshold ). e.g. probability(h 1 -h 2 )> threshold, and probability(h 2 -h 3 )> threshold, so h 1 , h 2 , h 3 are on the same track. 03.April.2019 Page 9 � 15
ANN Tracking: Pattern Recognition with parallel layers A) first layers A B C B) middle layers C) last layers inside Magnet � 16
ANN Tracking: Residuals including skewed layers First promising results inside magnetic field � 17
RNN (LSTM) Tracking: Residuals including skewed layers Next Step: Track Fitting with RNN - stay tuned � 18
Machine Learning for Particle Identification
PID: Usual Observables DRC: vs. p DSC: vs. p EMC: E/p vs. p Θ Θ C C 0.9 0.9 [rad] [rad] E/p 10 10 1 0.8 0.8 C C 0.7 0.7 0.8 10 Θ Θ 0.6 0.6 10 10 0.6 0.5 0.5 10 0.4 0.4 0.4 1 1 0.3 0.3 0.2 0.2 0.2 1 0.1 0.1 0 2 4 6 0 2 4 6 0 2 4 6 p [GeV/c] p [GeV/c] p [GeV/c] MUO: L vs. p ( ) STT: dE/dx vs. p µ MUO: L vs. p (non- ) µ iron iron 18 80 80 dE/dx [a.u.] [cm] [cm] 10 16 70 70 10 10 iron iron 14 60 60 L L 12 50 50 10 10 1 1 40 40 8 30 30 6 1 20 20 10 10 4 10 10 2 0 2 4 6 0 2 4 6 0 2 4 6 p [GeV/c] p [GeV/c] p [GeV/c] (charged particles only) � 20
<latexit sha1_base64="3ZvoEMUZT2fNdKs/18Cr9fQEUpU=">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</latexit> <latexit sha1_base64="woYPUKldMCmRspq6e3C9QClfum8=">AC3icbVC7SgNBFJ31GeNr1dJmSBiE3ZjoY0QTKGFRQTzgGxYZmdnk2FnH8zMBsOaXgQLf8TGQhFbP0A7f0acPIqYeODC4Zx7ufceJ2ZUSMP41hYWl5ZXVjNr2fWNza1tfWe3LqKEY1LDEYt40GCMBqSmqSkWbMCQocRhqOXxn6jR7hgkbhtezHpB2gTkg9ipFUkq3nLgtWj+D0ZnDbPTy1Yh65th/b/pRq63mjaIwA54k5Ifny+X3lx3r8rNr6l+VGOAlIKDFDQrRMI5btFHFJMSODrJUIEiPsow5pKRqigIh2OvplA+U4kIv4qpCUfq9ESKAiH6gaM6AyS7YtYbiv95rUR6J+2UhnEiSYjHi7yEQRnBYTDQpZxgyfqKIMypuhXiLuISxVfVoVgzr48T+qlonlULF2pNM7AGBmwD3KgAExwDMrgAlRBDWBwB57AC3jVHrRn7U17H7cuaJOZPfAH2scvmcmfKw=</latexit> PID Approaches Bayes: Y Combination of measurements: L ( ~ x | h ) = p k ( ~ x | h ) k = MVD dE/dx, DRC theta C … k Probability that a given track with parameters x corresponds to particle type h L ( ~ x | h ) × P ( h ) P ( ~ x | h ) = P L ( ~ x | h ) × P ( h ) h = e,µ, π ,K,p Machine Learning: A. Boosted Decision Tree (BDT) B. “Deep Learning” Artificial Neural Network (ANN) —> gain performance by considering correlations � 21
Input Features Input: • 𝑞𝑞→𝑌𝑌𝑍𝑍 , • where X and Y = e ∓ , 𝜌 ∓ , 𝜈 ∓ ,k ∓ ,p ∓ • Beam momentum: 15GeV/c � 22
Performance Plots Boosted Decision Tree Artificial Neural Network � 23
Confusion Matrices Boosted Decision Tree Artificial Neural Network � 24
Pions Boosted Decision Tree Artificial Neural Network � 25
Kaons Boosted Decision Tree Artificial Neural Network � 26
Machine Learning for the Software Trigger
Expected Data Rates • PANDA will run with a continuous beam • Event rates will be high, some events will overlap • Storage constraints in size & bandwidth • Data rate has to be reduced by 1/1000 • No specific trigger on hardware possible • Event topology of signals similar to background —> no “Jets” or similarly obvious features Solution: An online physics filter (“software trigger”) � 28
Software Trigger EvtGen DPM Event Generation Physics Channel 1 • Sign al Physics Channel 2 Background • Background ... Physics Channel m Simulation & Reconstruction Toy MC Full MC Event Filtering • Com binatorics ... Trigger 1 Trigger 2 Trigger 3 Trigger n • Mass Window Selection • Trigger Specific Selection → E vent Tagging reduce bg. Trigger Decision (Logical OR) Global Trigger Tag by 1/1000 � 29
Software Trigger - Cuts DPM - + - + - - - + + K K + K K η → π J/ e e J/ φ → ψ → ψ → µ µ p p e e ( = 21.9%) → ∈ S t c 500 2200 1800 180 Entries 15610 Entries 15610 Entries 25069 Entries 25069 Entries 15610 Entries 15610 Entries Entries 6232 6232 Entries 210195 Entries 210195 4000 2000 1600 160 3500 400 1800 = 0.0% = 0.8% = 3.4% = 0.0% = 0.0% ∈ ∈ ∈ ∈ ∈ 1400 140 1600 3000 1200 120 1400 300 2500 1000 100 1200 2000 1000 800 80 200 800 1500 600 60 600 1000 400 40 100 400 500 200 20 200 0 0 0 0 0 0 1 2 3 4 5 6 7 0.5 1 1.5 2 2.5 3 0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6 + 2 - - + - 2 + 2 + - 2 2 m(K K ) [GeV/c ] - π + m(e e ) [GeV/c ] m(K K ) [GeV/c ] m(e e ) [GeV/c ] m( ) [GeV/c ] µ µ S - - - - + + - + + 0 + p pK + + D K K Λ → π Λ → π + → π D K D K → π π → π c s 4500 1600 8000 10000 Entries 74779 Entries 74779 Entries 90669 Entries 90669 Entries 295962 Entries 295962 Entries 224819 Entries 224819 Entries 667078 Entries 667078 1400 4000 1400 7000 1200 3500 8000 = 7.2% = 8.6% = 2.8% = 7.6% = 5.2% ∈ ∈ ∈ ∈ ∈ 1200 6000 3000 1000 1000 5000 6000 2500 800 800 4000 2000 600 4000 600 3000 1500 400 400 2000 1000 2000 200 200 1000 500 0 0 0 0 0 0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6 0.5 1 1.5 2 2.5 3 0 1 2 3 4 5 6 - - - - + 2 2 2 - 2 2 + + + + + m(K ) [GeV/c ] m(K ) [GeV/c ] m(K K ) [GeV/c ] m(p ) [GeV/c ] m(pK ) [GeV/c ] π π π π π π • Cut based approach, optimised on signal & bg. MC • Many observables taken into account • Correlations are not respected 2014 � 30
Software Trigger - Cuts • Trade-o ff between signal e ffi ciency & background suppression • Many channels = much feedthrough & cross-tagging Full MC - Efficiencies - all cuts (high suppression) Full MC - Background fraction 1 e+e- J/ (2e) J/ (2 ) φ η ψ ψ µ 0.7 c 0 ± D D D Λ Λ s c 0.6 -1 10 0.5 mass cut only 0.4 all cuts (high supression) 0.3 -2 10 0.2 Goal: BG suppression 1/1000 0.1 -3 10 0 2 2.5 3 3.5 4 4.5 5 5.5 6 2 2.5 3 3.5 4 4.5 5 5.5 6 s [GeV] s [GeV] 2014 � 31
Recommend
More recommend
