Addressing Big Science challenges with Machine Learning

Natural Science offers us quite a bit of serious challenges for understanding our universe that can be addressed from Machine Learning point of view. Those slides show just a few examples of fruitful ML methods applications that show just a tip of an iceberg. Real challenge for AI (or generic AI) is about finding something no-one has ever thought could exist.

Transcript

1. Addressing Big Science challenges with Machine Learning (ML) 1 7 Feb 2018, openatalks.ai Andrey Ustyuzhanin NRU HSE YSDA ICL
2. Andrey Ustyuzhanin 2 From D.Whiteson, J Cham book “We have no idea”
3. Andrey Ustyuzhanin 3
4. Dark Matter Andrey Ustyuzhanin 4 Illustrations from J. Cham; D. Whiteson. “We Have No Idea” https://en.wikipedia.org/wiki/Dark_matter
5. Gravitational Lenses Andrey Ustyuzhanin 5 http://www.spacetelescope.org/videos/heic0701j/
6. Lens modeling Andrey Ustyuzhanin 6 ▌ How Does The Background Source Truly Look Like? What Is The Undistorted Image? ▌ How Is Matter Distributed In The Lensing Structure?
7. Andrey Ustyuzhanin 7
8. Andrey Ustyuzhanin 8
9. Low-hanging fruits for ML Andrey Ustyuzhanin 9 ▌ Classify: lenses vs non-lenses ▌ Non-gradient optimization for parameters search or ▌ Train CNNs to predict (regression) parameters of the background object and mass distribution: Fast automated analysis of strong gravitational lenses with CNNs Hezaveh, Perreault Levasseur, Marshall, Nature Aug. 2017 : “10 million times faster than regular lens modeling: 0.01 seconds on GPU”
10. Quick self-intro Andrey Ustyuzhanin 10 ▌ Head of Laboratory of Methods for Big Data Analysis › Applying ML to natural sciences probmels › http://cs.hse.ru/lambda ▌ Yandex School of Data Analysis › Course on solving High Energy Physics problems with ML approaches ▌ Collaboration with experiments at CERN: › Particle physics: LHCb , SHiP (CERN) › Dark matter search: NEWSdm (Gran Sasso) › Cosmic ray (CRAYFIS)
11. New Physics (Dark Matter) search at CERN Andrey Ustyuzhanin 11 ▌ Standard Model is cool but is not the final answer: › Neutrino has mas › Why antimatter is so rare? › What is dark matter? ▌ Laboratory search for deviations: › SHiP
12. Search for Hidden Particles Andrey Ustyuzhanin 12 ▌ Physics cases: › Variety of Hidden Sector portals explored › Tau neutrino physics › Light Dark Matter (LDM) Search ▌ ML cases: › Experiment design (shield, emulsion optimization) › Signal/background separation in emulsion › Fast simulation
13. Muon Shield Design Bayesian optimization: “designed” magnets, that are 25% lighter Andrey Ustyuzhanin 13 http://iopscience.iop.org/article/10.1088/1742- 6596/934/1/012050/meta
14. Nuclear Emulsion Andrey Ustyuzhanin 14 ▌ After the passage of charged particles through the emulsion, a latent image is produced ▌ The emulsion chemical development makes Ag grains visible with an optical microscope Compton electron
15. Emulsion, electromagnetic showers Development of new software tools based on ML techniques for electron identification and energy measurement in photo emulsions Andrey Ustyuzhanin 15
16. Fast physics simulation Andrey Ustyuzhanin 16 https://github.com/hep-lbdl Arxiv: 1705.02355, 1701.05927
17. Cosmic rays (OMG particles) Andrey Ustyuzhanin 17
18. Cosmic Ray + Smart Phone = CRAYFIS Andrey Ustyuzhanin 18
19. Generating realistic CMOS-camera images Andrey Ustyuzhanin 19
20. ML challenges Andrey Ustyuzhanin 20 1. Realistic signal simulation 2. Fast data processing on smart phones 3. Tuning of simulator parameters 4. Fraud detection
21. Galaxy Zoo: science + human IQ Andrey Ustyuzhanin 21 https://arxiv.org/abs/0907.4155 http://bit.ly/2o4d2P8 ▌ Regular challenge: › Identify galaxies ▌ ’Occasional’ discoveries: › Green peas and voorwerps
22. Conclusion Andrey Ustyuzhanin 22 ▌ There is quite a bit of cool stuff going on in fundamental science › Dark matter, gravitational lenses, “OMG “cosmic rays, dwarf galaxies, … ▌ There are serious challenges for ML in Natural Sciences Domain › Experiment design, simulation tuning, anomaly/fraud detection, … ▌ AI challenge: › Find something no human has considered possible (green pea, voorwerps, etc) ▌ Summer school on Machine Learning for High Energy Physics, Oxford: › http://bit.ly/mlhep2018 http://cs.hse.ru/lambda anaderiRu@twitter
23. Backup Andrey Ustyuzhanin 23
24. Andrey Ustyuzhanin 24
25. References - Machine Learning for High Energy Physics, https://bit.ly/mlhep2018 - M. Nielson, Reinventing the discovery - C. Cardamone et al, Galaxy Zoo Green Peas: Discovery of A Class of Compact Extremely Star-Forming Galaxies https://arxiv.org/abs/0907.4155 - Google exa planet search: https://www.theverge.com/2017/12/14/16777394/google-nasa-ai- machine-learning-planets-astronomy
26. Andrey Ustyuzhanin 26
27. Andrey Ustyuzhanin 27 https://www.youtube.com/watch?v=Y9yQOb94yl0 https://www.youtube.com/watch?v=2qeT4DkEX-w