Download presentation
Presentation is loading. Please wait.
Published byMaria Franklin Modified over 6 years ago
1
Inter-experimental LHC Machine Learning Working Group Activities
Sergei V. Gleyzer University of Florida
2
Big Data in Physics and Astronomy Sergei V. Gleyzer
Outline Inter-experimental LHC Machine Learning Working Group Machine Learning Applications at the LHC Community Efforts 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
3
Big Data in Physics and Astronomy Sergei V. Gleyzer
IML Inter-experimental LHC Machine Learning Working Group iml.cern.ch Sharing of expertise among LHC (and other HEP) experiments ATLAS, CMS, LHCb, ALICE, Belle-II, neutrino experiments ~450 participants Exchange between particle physics and machine learning communities Software development and maintenance Forum, Training and Education 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
4
Big Data in Physics and Astronomy Sergei V. Gleyzer
Working Format Monthly meetings around machine learning topics relevant to HEP community: Deep Learning Software and Tools Hardware Applications Unsupervised Learning Anomaly Detection Multi-class/Multi-objective Learning Bayesian ML and GANs Theory Applications 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
5
Big Data in Physics and Astronomy Sergei V. Gleyzer
IML Workshop First IML Workshop at CERN March 20-22, 2017 ~300 participants Industry session Quark/gluon tagging challenge Physics Object Tagging Workshop HEP-ML Community White Paper More workshops forthcoming 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
6
LHC Applications and Challenges
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
7
Big Data in Physics and Astronomy Sergei V. Gleyzer
LHC Applications Primarily Classification Low Level: Particle identification Photon or a jet? Pattern recognition Tracks, vertices High Level: New Physics searches Higgs/SUSY event or background? Jet sub-structure 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
8
Big Data in Physics and Astronomy Sergei V. Gleyzer
Higgs Discovery Machine Learning used in Higgs Discovery Event selection Identification of particles Identification of interactions Energy regression Improvement in analysis from all four areas 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
9
Big Data in Physics and Astronomy Sergei V. Gleyzer
Upcoming Challenges Orders of magnitude between signals and backgrounds 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
10
Big Data in Physics and Astronomy Sergei V. Gleyzer
Event Complexity 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
11
Big Data in Physics and Astronomy Sergei V. Gleyzer
Upcoming Challenges Data size: LHC 15,000,000 Тb – 2035 Resources not up as fast as data volume Unknown Physics 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
12
Big Data in Physics and Astronomy Sergei V. Gleyzer
Topics of interest 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
13
Object Identification
Interesting areas Fast Simulation Particle Tracking Object Identification Tracking: 100s particle trajectories low-level data Fast simulation: lots of computing power use, 10min/event Object id: Deep learning, use timing evolution Trigger: Throwing away 99.9%, fast, smarter Calo: posed as an image problem More: correlations Trigger Imaging Calorimetry Simulation 06/19/2017
14
Interesting areas Generative Models, Adversarial Networks Deep Kalman
RNNs FCN, Recurrent, LSTM NN Tracking: 100s particle trajectories low-level data Fast simulation: lots of computing power use, 10min/event Object id: Deep learning, use timing evolution Trigger: Throwing away 99.9%, fast, smarter Calo: posed as an image problem More: correlations Deep ML +FPGA Convolutional DNN Multiobjective Regression 06/19/2017
15
Big Data in Physics and Astronomy Sergei V. Gleyzer
Deep Learning 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
16
Big Data in Physics and Astronomy Sergei V. Gleyzer
Deep Learning Higgs Boson Example: P. Baldi, et. al. 2014 8% improvement 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
17
Big Data in Physics and Astronomy Sergei V. Gleyzer
Deep Learning Papers in HEP: Jet images and deep learning: arxiv Jet substructure and deep learning: Parton shower uncertainties and jet substructure: Deep learning for ttHhttp://inspirehep.net/record/ ?ln=en Nova Daya Bay arxiv Next: Microboone: 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
18
Big Data in Physics and Astronomy Sergei V. Gleyzer
Deep Learning Deep NN Deep NN BDT Shallow Significant improvements in performance 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
19
Big Data in Physics and Astronomy Sergei V. Gleyzer
Deep Learning Convolutional Neural Networks 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
20
Big Data in Physics and Astronomy Sergei V. Gleyzer
End-to-End Approach link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
21
Big Data in Physics and Astronomy Sergei V. Gleyzer
CVN on NOvA link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
22
Beyond Classification
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
23
Single-Objective Regression
Train learning model to estimate a single function target or “objective” Ex. photon energy/muon momentum With a machine learning algorithm Decision tree, random forest, neural network etc. 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
24
Big Data in Physics and Astronomy Sergei V. Gleyzer
Photon Energy Single Target Example: Inputs: shower information, photon coordinates, median event energy Target Output: EMEASURED/ETRUE ~10-30% improvement in resolution 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
25
Deep Learning Regression
Deeper Higher is better Shallow 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
26
Multi-Objective Regression
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
27
Multi-Objective Regression
Simultaneous estimate of multiple functions or “targets” Possibly additionally correlated N single-target models not as optimal lingo: “multi-task” learning and more cumbersome Train a single model to simultaneously predict all targets 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
28
Big Data in Physics and Astronomy Sergei V. Gleyzer
Applicable Models Methods: Regression decision trees Decision rules Decision rule ensembles Random forest Neural networks… Trade-offs: accuracy, model size, interpretability 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
29
Multi-objective Example
X input variables {a, b, c, d…} K of them strongly correlated Y target outputs to estimate {A, B, C, D…} N of them strongly correlated Challenge: build a predictive model to describe simultaneously all the outputs {A,B,C,D…}, provided a corresponding set of inputs. 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
30
Big Data in Physics and Astronomy Sergei V. Gleyzer
Illustrative Example 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
31
Big Data in Physics and Astronomy Sergei V. Gleyzer
Target Correlations Target Correlations Prediction-Target Difference Very close to Zero 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
32
Physics Object Tagging
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
33
Big Data in Physics and Astronomy Sergei V. Gleyzer
Object Tagging link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
34
Object Tagging link 06/19/2017
35
Big Data in Physics and Astronomy Sergei V. Gleyzer
Tracking 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
36
HEP.TrackX link 06/19/2017
37
HEP.TrackX link 06/19/2017
38
Big Data in Physics and Astronomy Sergei V. Gleyzer
Software and Tools 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
39
Big Data in Physics and Astronomy Sergei V. Gleyzer
DNN Apache Spark link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
40
Big Data in Physics and Astronomy Sergei V. Gleyzer
Deep Learning Throughput Comparison 2.7 * Theano Single precision Excellent throughput compared to Theano on same GPU 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
41
Hardware Applications
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
42
Big Data in Physics and Astronomy Sergei V. Gleyzer
link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
43
Big Data in Physics and Astronomy Sergei V. Gleyzer
Neuromorphic link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
44
Theory and Phenomenology
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
45
Big Data in Physics and Astronomy Sergei V. Gleyzer
link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
46
Big Data in Physics and Astronomy Sergei V. Gleyzer
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
47
Big Data in Physics and Astronomy Sergei V. Gleyzer
link 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
48
Big Data in Physics and Astronomy Sergei V. Gleyzer
Other areas Unsupervised Learning and Anomaly Detection Generative adversarial models for fast detector simulation Multi-class applications Understanding uncertainties associated with decision-making in machine learning applications 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
49
HEP Community White Paper in Machine Learning
06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
50
Big Data in Physics and Astronomy Sergei V. Gleyzer
Community White Paper HEP Software Foundation HSF link Community White Paper link to CWP Machine Learning Identification of challenges Roadmap to address them Important to think of these issues now Impact on how we dedicate resources and design our software 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
51
Big Data in Physics and Astronomy Sergei V. Gleyzer
Summary LHC physics and computing challenges will require significant progress: Higher backgrounds and pileup, data volume, unknown new physics Machine learning offers a promising direction An opportunity to examine new areas of ML applications to HEP IML an inter-experimental effort to foster collaboration and progress in HEP-ML 06/19/2017 Big Data in Physics and Astronomy Sergei V. Gleyzer
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.