Jets in the 21st Century

Transcription

1 LoLa Jets in the 2st Century Universität Heidelberg Mainz /28

2 LoLa Rise of the Machines Brief history of jets 994 jet-algo W -tagger for heavy Higgs [Seymour] 994 jet-algo top tagger for fun [Seymour] 28 jet-algo Higgs tagger [Butterworth, Davison, Rubin, Salam; Kribs, Martin, Spannowsky] 28 jet-algo top tagger [Kaplan, Rehermann, Schwartz, Tweedie] 29 jet-algo HEPTopTagger [TP, Salam, Spannowsky] 29 template top tagger [Almeida, Lee, Perez, Sterman, Sung, Virzi] 2 Shower Deconstruction [Soper, Spannowsky] 25 Multi-variate HEPTopTagger [Kasieczka, TP, Schell, Strebler, Salam] 24 image recognition W -tagger [Cogan, Kagan, Strass, Schwartzman] 25 jet images [de Oliveira, Kagan), Mackey, Nachman, Schwartzman] 27 image recognition top tagger [Kasieczka, Plehn, Russell, Schell] 27 language recognition W -tagger [Louppe, Cho, Becot, Cranmer] 27 4-vector-based top tagger [Butter, Kasieczka, Plehn, Russel] 28 adversarial autoencoder [Heinel, Kasieczka, Plehn, Thompson; Shi etal]

3 LoLa Jet-level analyses (99s) Jets as analysis objects jet parton duality what are partons in detector? crucial for any LHC analysis [really???] infrared safety the main issue any of this true for LHC? QCD recombination algorithms [FASTJET: Cacciari, Salam, Soyez] 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 () fat jets/substructure: use clustering history ) [k T, C/A only] y ib = p T,i y ib = y ib = p T,i.

4 Heavy resonance taggers (2s) R bjj 3 LoLa Tagging boosted tops hadronic decays vs QCD splittings SM sample: semileptonic t t events substructure playground P T [GeV]

5 LoLa Heavy resonance taggers (2s) Tagging boosted tops hadronic decays vs QCD splittings SM sample: semileptonic t t events substructure playground HEPTopTagger [BDRS; TP, Salam, Spannowsky, Takeuchi] R bjj m 23 /m P T [GeV] m m 3 =m 2 =m W W.8 C/A fat jet, R =.5 and p T > 2 GeV [FastJet limitation] 2 mass drop, cutoff m sub > 3 GeV 3 filtering leading to hard substructure triple 4 top mass window m 23 = [5, 2] GeV 5 A-shaped mass plane cuts as function of m W /m t 6 consistency condition p (tag) T > 2 GeV m 23 /m m 23 =m W.5.5 m arctan mm 3 =m 3 /mw 2 2 =m W m 23 =m W arctan m 3 /m 2

6 LoLa Heavy resonance taggers (2s) Tagging boosted tops hadronic decays vs QCD splittings SM sample: semileptonic t t events substructure playground HEPTopTagger [BDRS; TP, Salam, Spannowsky, Takeuchi] C/A fat jet, R =.5 and p T > 2 GeV [FastJet limitation] 2 mass drop, cutoff m sub > 3 GeV 3 filtering leading to hard substructure triple 4 top mass window m 23 = [5, 2] GeV 5 A-shaped mass plane cuts as function of m W /m t 6 consistency condition p (tag) T LHC break-through > 2 GeV R bjj m 23 /m 23 m 23 /m P T [GeV] m m 3 =m 2 =m W W m 23 =m W.5.5 m arctan mm 3 =m 3 /mw 2 2 =m W m 23 =m W arctan m 3 /m 2

7 LoLa N-Subjettiness N-Jettiness to count subjets inside fat jet [Thaler, van Tilburg] how many subjets do the calo entries correspond to? choice of reference axes τ N = α jet p T,αR β p T,α min ( R k,α) β k=,...,n α jet - from subjet algorithm 2- from minimization of τ N τ N means many calo entries away from N axes τ N means perfect matching systematics cancelled in ratios τ N+ /τ N for N + subjets easily added to any other tagger Relative occurence GeV < p T < 6 GeV, 6 GeV < m < 24 GeV top jets (min axes) top jets (k T axes) QCD jets (min axes) QCD jets (k T axes) τ 3 /τ 2 (β = ).8.2

8 LoLa Multi-variate top taggers (2s) 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 <.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

9 LoLa Multi-variate top taggers (2s) 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 <.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 } Z W b W b

10 LoLa Multi-variate top taggers (2s) 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 <.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 } outperforming deterministic taggers / 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,.x. cells) s = 4 TeV ε S

11 LoLa Jet images (22s) Deep learning = modern architectures on low-level observables why high-level observables as input? wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] top-tagging attempt [Almeida, Backovic, Cliche, Lee, Perelstein] QCD and shower study [Barnard etal] quark-gluon discrimination including tracks [Komiske etal] new hammer for jet nails

12 LoLa Jet images (22s) Deep learning = modern architectures on low-level observables why high-level observables as input? wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] top-tagging attempt [Almeida, Backovic, Cliche, Lee, Perelstein] QCD and shower study [Barnard etal] quark-gluon discrimination including tracks [Komiske etal] new hammer for jet nails Experimental questions does it work? what is the training sample? [Metodiev, Nachman, Thaler] how do we get it past the jets experts? Theoretical questions how much of all this is QCD? how can be improve the setup? [the future has not been written] what does the network learn?

13 LoLa Jet images (22s) Deep learning = modern architectures on low-level observables why high-level observables as input? wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] top-tagging attempt [Almeida, Backovic, Cliche, Lee, Perelstein] QCD and shower study [Barnard etal] quark-gluon discrimination including tracks [Komiske etal] new hammer for jet nails Neural network [Kasieczka, TP, Russell, Schell; Macaluso, Shih] colored image as input binning through calorimeter resolution [ η =. vs φ = 5 ] analyze geometric patterns [Facebook: convolutional network] Background rejection /ɛb nc-kernel = nc-kernel = 6 nc-size = 2 nc-size = 6 nc-size = 8 nc-block = nc-block = 3 nc-block = 4 nc-layer = 3 nc-layer = 4 nc-layer = 5 nd-layer = 2 nd-layer = 4 nd-node = 28 nd-node = 32 default Signal efficiency ɛs

14 Inside Benchmarking image-based top tagger [Kasieczka, TP, Russell, Schell] 2+2 convolutional layers probing 2D structure with kernel matrix LoLa

15 Inside Benchmarking image-based top tagger [Kasieczka, TP, Russell, Schell] 2+2 convolutional layers probing 2D structure with kernel matrix 3 fully connected layers weight function linking input and output LoLa

16 LoLa Inside Benchmarking image-based top tagger [Kasieczka, TP, Russell, Schell] 2+2 convolutional layers probing 2D structure with kernel matrix 3 fully connected layers weight function linking input and output signal-ness vs background-ness output [probability?] Arbitrary units Background 2 Signal Signal probability

17 LoLa Inside Benchmarking image-based top tagger [Kasieczka, TP, Russell, Schell] 2+2 convolutional layers probing 2D structure with kernel matrix 3 fully connected layers weight function linking input and output signal-ness vs background-ness output [probability?] Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images η pixels φ pixels Pearson correlation coefficient r

18 LoLa Inside Benchmarking image-based top tagger [Kasieczka, TP, Russell, Schell] 2+2 convolutional layers probing 2D structure with kernel matrix 3 fully connected layers weight function linking input and output signal-ness vs background-ness output [probability?] Pearson input-output correlation [pixel x vs label y] r ij ( ) xij x ij (y ȳ) images comparison to MotherOfTaggers {m sd, m fat, m rec, f rec, R opt, τ 2, τ 3, τ sd 2, τ sd 3 } understandable performance gain Background rejection /ɛb SoftDrop+N-subjettiness MotherOfTaggers full minimal Signal efficiency ɛs

19 LoLa Just a new tool Typical reaction: Fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs - fat jet mass and N-subjettiness Signal MotherOfTaggers Background.8.4 truth mfat [GeV] Signal 3.5 MotherOfTaggers truth τ3/τ mfat [GeV] Background τ3/τ2

20 LoLa Just a new tool Typical reaction: Fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs - fat jet mass and N-subjettiness 2- soft drop mass Signal MotherOfTaggers truth Background msd [GeV] msd [GeV]

21 LoLa Just a new tool Typical reaction: Fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs - fat jet mass and N-subjettiness 2- soft drop mass 3- transverse momenta it works and we know why Signal MotherOfTaggers truth Background msd [GeV] msd [GeV] Signal MotherOfTaggers truth Background pt,j [GeV] pt,j [GeV]

22 LoLa LoLa Our version of graph network sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] Events (normalized) 2 3 Calo p T Calo p T > 3 GeV > 35 GeV 4 5 PF p T PF p T > 3 GeV > 35 GeV N const

23 LoLa LoLa Our version of graph network sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] input 4-vectors k, k,2 k,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

24 LoLa LoLa Our version of graph network sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] 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

25 LoLa LoLa Our version of graph network sparsely filled picture: graph CNN physics objects from calorimeter and tracker distance measure known from e&m Inspired by jet algorithm combination layer [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] input 4-vectors on-shell conditions for top tag combined 4-vectors C,N+2 C,M CoLa k µ,i k. 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]

26 LoLa LoLa Our version of graph network sparsely filled picture: graph CNN 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 [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] 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 fun: measure Minkowski metric g =diag(.99±.2,.±.,.±.2,.99±.2) extend to particle flow network with little but Lorentz invariance

27 LoLa Getting inspired Anomaly search trained on background established concept: autoencoder [Heimel, Kasieczka, TP, Thompson; Farina, Macari, Shih] 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 Background rejection / B Signal efficiency S Bottleneck size

28 LoLa Getting inspired Anomaly search trained on background established concept: autoencoder [Heimel, Kasieczka, TP, Thompson; Farina, Macari, Shih] 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 De-correlate background shaping established concept: adversary [Google Go]

29 LoLa Getting inspired Anomaly search trained on background established concept: autoencoder [Heimel, Kasieczka, TP, Thompson; Farina, Macari, Shih] 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 De-correlate background shaping established concept: adversary [Google Go] atypical QCD jets typially with large jet mass remove jet mass from network training normalized distribution % % 7% least QCD-like % 5% normalized distribution % 7% 4% % 5% least QCD-like jet mass [GeV] jet mass [GeV]

30 LoLa The future Times are moving fast......jets are containers for subjet physics [was 99s]...deterministic taggers established/old/boring [was 2s]...multi-variate taggers an intermediate step [dying with the 2s]...imagine recognition a starting point [will be 22s]... is not a black box...lola took Jackson...s really change analyses Join the fun!