Machine Learning just because it is Great Fun

Transcription

1 Machine Learning just because it is Great Fun Universität Heidelberg Hamburg 2/209

2 Change of title

3 Why LHC, why jets Big jet data by ATLAS & CMS colliding protons on protons at E 3000 m p most interactions q q, gg q q, gg quarks/gluon visible as jets σ pp jj L 8 fb 80/fb events It s big data

4 Why LHC, why jets Big jet data by ATLAS & CMS colliding protons on protons at E 3000 m p most interactions q q, gg q q, gg quarks/gluon visible as jets σ pp jj L 8 fb 80/fb events It s big data Interesting physics in jets re-summed perturbative QFT prediction for QCD splittings jets as decay products 67% W jj 70% Z jj 60% H jj 67% t jjj 60% τ j... new physics in dark showers [Jennifer Thompson s talk] It s interesting

5 Why LHC, why jets Big jet data by ATLAS & CMS colliding protons on protons at E 3000 mp most interactions q q, gg q q, gg quarks/gluon visible as jets It s big data σpp jj L 8 fb 80/fb events Interesting physics in jets re-summed perturbative QFT prediction for QCD splittings jets as decay products 67% W jj 70% Z jj new physics in dark showers 60% H jj It s interesting LHC simulations QCD simulation: Pythia, Sherpa, Herwig fast detector simulation: Delphes excellent agreement with data We can simulate it 67% t jjj [Jennifer Thompson s talk] 60% τ j...

6 Inside jets Jets and machine learning from 990s to 2020s 99 NN-based quark-gluon tagger [visionary: Lönnblad, Peterson, Rögnvaldsson]

7 Inside jets Jets and machine learning from 990s to 2020s 99 NN-based quark-gluon tagger [visionary: Lo nnblad, Peterson, Ro gnvaldsson] 994 jet-algo W /top-tagger for heavy Higgs 2008 jet-algo Higgs tagger 2008 jet-algo top tagger [Seymour] [Butterworth, Davison, Rubin, Salam; Kribs, Martin, Spannowsky] [Kaplan, Rehermann, Schwartz, Tweedie] 2009 jet-algo HEPTopTagger [TP, Salam, Spannowsky; st user Gregor Kasieczka] 2009 template top tagger [Almeida, Lee, Perez, Sterman, Sung, Virzi] 20 Shower Deconstruction [Soper, Spannowsky] 205 Multi-variate HEPTopTagger [Kasieczka, TP, Schell, Strebler, Salam] 204 image recognition W -tagger 205 jet images [Cogan, Kagan, Strass, Schwartzman] [de Oliveira, Kagan), Mackey, Nachman, Schwartzman] 207 image recognition top tagger [Kasieczka, Plehn, Russell, Schell] 207 language recognition W -tagger vector-based top tagger [Louppe, Cho, Becot, Cranmer] [Butter, Kasieczka, Plehn, Russel] 208 jet autoencoder [Heinel, Kasieczka, Plehn, Thompson; Shi etal]

8 Jet-level analyses (990s) Jets as analysis objects partonic predictions from QCD jets describing partons in reality infrared safety crucial to compare with perturbative QCD rates data-to-data analyses more flexible data-to-simulation analyses similarly free? QCD recombination algorithms [FASTJET] define jet jet and jet beam distances [exclusive with resolution y cut ] k T C/A anti-k T y ij = R ij R min ( p T,i, p T,j ) y ij = R ij R y ij = R ( ij R min p T,i, p T,j () find minimum y min = min ij (y ij, y ib ) (2a) if y min = y ij merge subjets i and j, back to () (2b) if y min = y ib remove i from subjets, go to () clustering history usable? ) y ib = p T,i y ib = y ib = p T,i.

9 Fat jet taggers (2000s) R bjj 3 3 For instance: boosted tops 2 hadronic decays vs QCD splittings perfectly described by perturbative QCD labelled sample: semileptonic t t events substructure playground P T [GeV] 2

10 Fat jet taggers (2000s) For instance: boosted tops hadronic decays vs QCD splittings perfectly described by perturbative QCD labelled sample: semileptonic t t events substructure playground Simple top tagging [BDRS; TP, Salam, Spannowsky, Takeuchi] R bjj m 23 /m P T [GeV] m2 =m W m 3 =mw C/A fat jet with p T > 200 GeV 2 filtering defining 3-5 decay jets 3 top mass window m 23 = [50, 200] GeV m 23 =m W 4 A-shaped mass plane cuts probing m W not rocket science, but experimental break-through m 23 /m m arctan mm 3 =m 3 /mw 2 2 =m W m 23 =m W arctan m 3 /m 2

11 Multi-variate subjet physics (20s) OptimalR and N-Subjettiness [Kasieczka, TP, Salam, Schell, Strebler] multivariate analysis old idea [Lonnblad, Peterson, Rognvaldsson] HEPTopTaggerv2 to keep up with shower deconstruction [Soper, Spannowsky] optimal fat jet size R opt [large to decay jets, small to avoid combinatorics, compute from kinematics] m 23 m (Rmax) 23 < 0.2 m (Rmax) 23 R opt add N-subjettiness [Thaler, van Tilburg] {m 23, f W, R opt R (calc) opt, τ j, τ (filt) } j

12 Multi-variate subjet physics (20s) OptimalR and N-Subjettiness [Kasieczka, TP, Salam, Schell, Strebler] multivariate analysis old idea [Lonnblad, Peterson, Rognvaldsson] HEPTopTaggerv2 to keep up with shower deconstruction [Soper, Spannowsky] optimal fat jet size R opt [large to decay jets, small to avoid combinatorics, compute from kinematics] m 23 m (Rmax) 23 < 0.2 m (Rmax) 23 R opt add N-subjettiness [Thaler, van Tilburg] {m 23, f W, R opt R (calc) opt, τ j, τ (filt) } j Fat jet and top kinematics FSR major problem for Z search tag and reconstruction in each other s way {..., m tt, p T,t, m (filt), p (filt) jj T,j } expected performance increase / ε B ED[PRD89] HTT[JHEP] filtered fat jets (2.3) variable masses (2.4) optimalr (3.2) N-subjettiness (3.4) Qjets (3.7, 0.x0. cells) s = 4 TeV ε S

13 Jet images (2020s) Deep learning = modern architectures on low-level observables wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] impact of shower? [Barnard etal] combining calorimeter and tracking? [Komiske etal] understanding additional information? [Datta & Larkosky] link to infrared safety? [Choi, Lee, Perelstein; Friday speakers] φ Calorimeter E [GeV] η

14 Jet images (2020s) Deep learning = modern architectures on low-level observables 0 5 wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] impact of shower? [Barnard etal] combining calorimeter and tracking? [Komiske etal] understanding additional information? [Datta & Larkosky] 0 link to infrared safety? [Choi, Lee, Perelstein; Friday speakers] 5 5 Convolutional network [Kasieczka, TP, Russell, Schell; Macaluso, Shih] run on 2-D jet images [p T = 350,..., 450 GeV] 35 colored image as input binning through calorimeter resolution [ η = 0. vs φ = 5 ]

15 Inside Benchmarking image-based top tagger 2+2 convolutional layers [Kasieczka, TP, Russell, Schell; Macaluso & Shih]

16 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers [Kasieczka, TP, Russell, Schell; Macaluso & Shih]

17 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images [Kasieczka, TP, Russell, Schell; Macaluso & Shih] η pixels φ pixels Pearson correlation coefficient r

18 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images comparison to MotherOfTaggers BDT understandable performance gain [Kasieczka, TP, Russell, Schell; Macaluso & Shih] Background rejection /ɛb SoftDrop+N-subjettiness MotherOfTaggers full minimal Signal efficiency ɛs

19 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) [Kasieczka, TP, Russell, Schell; Macaluso & Shih] images minimal comparison to MotherOfTaggers BDT understandable performance gain jets 5 Training Architecture Preprocessing 4 Sample size /ϵb ϵs

20 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images comparison to MotherOfTaggers BDT understandable performance gain [Kasieczka, TP, Russell, Schell; Macaluso & Shih] Typical reaction: Fuck you, you fucking machine full control for supervised learning easy checks for correctly identified signal/background events MC truth vs MotherOfTaggers vs fat jet mass Signal MotherOfTaggers Background truth mfat [GeV] mfat [GeV]

21 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images comparison to MotherOfTaggers BDT understandable performance gain [Kasieczka, TP, Russell, Schell; Macaluso & Shih] Typical reaction: Fuck you, you fucking machine full control for supervised learning easy checks for correctly identified signal/background events MC truth vs MotherOfTaggers vs fat jet mass N-subjettiness Signal MotherOfTaggers truth τ3/τ Background τ3/τ2

22 Inside Benchmarking image-based top tagger 2+2 convolutional layers 3 fully connected layers Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images comparison to MotherOfTaggers BDT understandable performance gain [Kasieczka, TP, Russell, Schell; Macaluso & Shih] Typical reaction: Fuck you, you fucking machine full control for supervised learning easy checks for correctly identified signal/background events MC truth vs MotherOfTaggers vs fat jet mass N-subjettiness transverse momenta it works and we know why Signal MotherOfTaggers truth Background pt,j [GeV] pt,j [GeV]

23 LoLa Our version of graph network sparsely filled picture: graph CNN [Butter, Kasieczka, TP, Russell; many more by now] physics objects from calorimeter and tracker distance measure known from e&m Events (normalized) 2 3 Calo p T Calo p T > 300 GeV > 350 GeV 4 5 PF p T PF p T > 300 GeV > 350 GeV N const

24 LoLa Our version of graph network [Butter, Kasieczka, TP, Russell; many more by now] sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer input 4-vectors k 0, k 0,2 k 0,N k (k µ,i ) =, k,2 k,n k 2, k 2,2 k 2,N k 3, k 3,2 k 3,N

25 LoLa Our version of graph network [Butter, Kasieczka, TP, Russell; many more by now] sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer input 4-vectors on-shell conditions for top tag k 2 µ, = (k µ, + k µ,2 + k µ,3 ) 2 k 2 µ,2 = (k µ, + k µ,2 ) 2 =! m 2 W! = m 2 t

26 LoLa Our version of graph network [Butter, Kasieczka, TP, Russell; many more by now] sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer input 4-vectors on-shell conditions for top tag combined 4-vectors 0 0 C,N+2 C,M CoLa k µ,i k 0. C 2,N+2 C 2,M µ,j = k µ,i C ij C = C N,N+2 C N,M after combination of input 4-vectors original momenta k i M N trainable linear combinations physics step, easy to interpret [M-N=5]

27 LoLa Our version of graph network sparsely filled picture: graph CNN [Butter, Kasieczka, TP, Russell; many more by now] physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer combined 4-vectors k µ,i CoLa k µ,j = k µ,i C ij Inspired by Jackson Lorentz layer DNN on Lorentz scalars m 2 ( k j ) LoLa k j ˆk p T ( k j ) j = w (E) jm E( k m) w (d) jm d 2 jm learn Minkowski metric g =diag(0.99±0.02,.0±0.0,.0±0.02, 0.99±0.02)

28 Simple questions ML4Jets 207: what architecture? ML4Jets 208: top tagging study

29 Simple questions ML4Jets 207: what architecture? ML4Jets 208: top tagging study lots of architectures work 5 ParticleNet ResNeXt PFN NSub(8) LBN RecNN CNN P-CNN LoLa EFN EFP TopoDNN nsub+m Background rejection B Signal efficiency S 0.8.0

30 Simple questions ML4Jets 207: what architecture? ML4Jets 208: top tagging study lots of architectures work More questions what about uncertainties? how stable are taggers in experimental reality? can we go beyond fully supervised learning? how do we go beyond jets? is classification all we can use ML for? are there analyses only ML will allow us to do? what is the particle nature of dark matter? etc

31 When reality hits ML-Life is not always nice to us [Kasieczka, Kiefer, TP, Thompson] Quark-gluon tagging a problem since 99 quark jets typical for resonance searches gluon jets typical as dark matter recoil BDT/NN on high-level variables established deep-learning advantage gone after detector simulation, REALLY??? / q 3 2 LoLa, Delphes LoLa, particle BDT, Delphes BDT, particle BDT, reduced BDT, Delphes, reduced g

32 Getting seriously inspired Anomaly search only trained on background established ML concept: autoencoder reconstruct typical QCD jet image from many QCD jets reduce weights in central layer compress information on typical search for outliers hard to describe benchmark on top jets, search for Higgs or dark showers [Heimel, Kasieczka, TP, Thompson; Farina, Macari, Shih] Background rejection / B Signal efficiency S Bottleneck size

33 Getting seriously inspired Anomaly search only trained on background established ML concept: autoencoder reconstruct typical QCD jet image from many QCD jets reduce weights in central layer compress information on typical search for outliers hard to describe benchmark on top jets, search for Higgs or dark showers [Heimel, Kasieczka, TP, Thompson; Farina, Macari, Shih] De-correlate background shaping established concept: adversary

34 Getting seriously inspired Anomaly search only trained on background established ML concept: autoencoder reconstruct typical QCD jet image from many QCD jets reduce weights in central layer compress information on typical search for outliers hard to describe benchmark on top jets, search for Higgs or dark showers [Heimel, Kasieczka, TP, Thompson; Farina, Macari, Shih] De-correlate background shaping established concept: adversary atypical QCD jets typially with large jet mass remove jet mass from network training normalized distribution % 0% 70% least QCD-like % 5% normalized distribution % 70% 40% % 5% least QCD-like jet mass [GeV] jet mass [GeV]

35 The future Times are moving fast......jets are containers for subjet physics [was 990s]...deterministic taggers are established/old/boring [was 2000s]...multi-variate taggers are an intermediate step [dying with the 20s]...imagine recognition is a starting point [will be 2020s]...deep learning is not just classification Join the fun! Background rejection B ParticleNet ResNeXt PFN NSub(8) LBN RecNN CNN P-CNN LoLa EFN EFP TopoDNN nsub+m Signal efficiency S