Jesse Thaler IAIFI Director
Total Page:16
File Type:pdf, Size:1020Kb
The NSF AI Institute for Artificial Intelligence and Fundamental Interactions Jesse Thaler IAIFI Director High Energy Physics Advisory Panel — December 4, 2020 Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 1 NSF: National AI Research Institutes 5 Inaugural NSF AI Institute for Research on Trustworthy AI in Institutes: Weather, Climate, and Coastal Oceanography +2 USDA-NIFA Institutes NSF AI Institute for Foundations of Machine Learning NSF AI Institute for Student-AI Teaming NSF AI Institute for Molecular Discovery, Synthetic Strategy, and Manufacturing NSF AI Institute for Artificial Intelligence and Fundamental Interactions Human-AI Interaction and Collaboration AI Institute in Dynamic Systems 8 Themes for AI Institute for Advances in Optimization AI-Augmented Learning AI and Advanced Cyberinfrastructure AI to Advance Biology Next Round: Advances in AI and Computer and Network Systems AI-Driven Innovation in Agriculture and the Food System [NSF Announcement, August 26, 2020; Call for New Proposals] Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 2 The NSF AI Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) “eye-phi” [http://iaifi.org/, MIT News Announcement] Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 3 The NSF AI Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) “eye-phi” Advance physics knowledge — from the smallest building blocks of nature to the largest structures in the universe — and galvanize AI research innovation Physics Physics Theory IAIFI Fellows Experiment Training, education & outreach at Physics/AI intersection IAIFI Fellows Cultivate early-career talent (e.g. IAIFI Fellows) Foster connections to physics facilities and industry IAIFI Fellows Build strong multidisciplinary collaborations Advocacy for shared solutions across subfields AI Foundations E.g. Analyzing Collision Debris ⇔ Geometric Data Processing [Harris, Schwartz, JDT, Williams] [Wang, Sun, Liu, Sarma, Bronstein, Solomon, TOG 2019] Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 4 The NSF AI Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) “eye-phi” Senior Investigators: 20 Physicists + 7 AI Experts + IAIFI Affiliates Junior Investigators: ≈20 PhD Students, ≈7 IAIFI Fellows in steady state Pulkit Agrawal Phiala Shanahan Demba Ba James Halverson Lisa Barsotti Tracy Slatyer Edo Berger Brent Nelson Isaac Chuang Marin Soljacic Cora Dvorkin William Detmold Justin Solomon Daniel Eisenstein Bill Freeman Washington Taylor Doug Finkbeiner Philip Harris Max Tegmark Matthew Schwartz Kerstin Perez Jesse Thaler Yaron Singer Alexander Rakhlin Mike Williams Todd Zickler Taritree Wongjirad Boston Area: Critical Mass for Transformative Research in “Ab Initio AI” Heavy HEP involvement across experiment, phenomenology & theory Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 5 Can we teach a machine to “think” like a physicist? (Have you ever tried to reason with a toddler?) Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 6 The IAIFI Vision Deep Learning meets Deep Thinking Sophisticated networks, increased Continued progress requires exploiting computational power & large data sets the structure of physics problems & have led to extraordinary advances time-tested strategies of physical reasoning “Deeper understanding, not just deeper networks” Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 7 (backup slide from my Extensive Use of ML in HEP May 2019 HEPAP talk) CMS: Z → e+e– calibration MicroBooNE: Object Identification 60 With machine-learning correction With clustering 50 Raw counts ) 3 40 30 133.1 cm Number of events (10 20 10 293.3 cm 0 50 60 70 80 90 100 110 120 130 Mass (GeV) ATLAS: H → μ+μ– search NOvA: Object Classification a 104 Data Model, P = 1.4 Model, P = 1 Z → W+W– 103 Other decays Fake tau particles Uncertainty 102 Number of events 101 b 1.5 Qx neutral current 1.0 QW charged current 0.5 and models charged current Normalized data Qe –1 –0.5 010.5 Q charged current BDT output P Fig. 4 | Exploring NOvA’s event-selection neural network using The subplots show example event topologies from points in the two- Machine learning is transforming many aspects of society, including fundamental physics research [figures from Radovic, Williams, Rousseau, Kagan, Bonacorsi, Himmel, Aurisano, Terao, Wongjirad, Nature 2018] Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 8 Off-the-Shelf ML for HEP? 2D Images? Appropriate for fixed-grid calorimeters, but less ideal for tracking detectors Natural Language? Clustering can yield “semantic” structure, but identical particles have no intrinsic ordering 3D Objects? Much closer to particle physics, though doesn’t capture all symmetries Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 9 AI2: Ab Initio Artificial Intelligence Machine learning that incorporates first principles, best practices, and domain knowledge from fundamental physics Symmetries, conservation laws, scaling relations, limiting behaviors, locality, causality, unitarity, gauge invariance, entropy, least action, factorization, unit tests, exactness, systematic uncertainties, reproducibility, verifiability, … Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 10 AI2: Ab Initio Artificial Intelligence Convolutional Neural Networks ⇔ Translational Equivariance ⇒ Momentum Conservation Identical Particles (QM) Energy Flow Networks ⇔ Infrared/Collinear Safety (QFT) Powerful strategy to AI analyze LHC collisions Efficient neural network × AI for point clouds 2 Shared solution = AI across disciplines 4060810 [Komiske, Metodiev, JDT, JHEP 2019] Progress driven by early-career talent with cross-disciplinary expertise: Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 11 IAIFI Postdoctoral Fellowships Recruit and train a talented and diverse group of early-career researchers Spark interdisciplinary, multi-investigator, multi-subfield collaborations Physics Physics Theory IAIFI Fellows Experiment IAIFI Fellows IAIFI Fellows AI Foundations 133 applicants for 2021–2024 IAIFI Fellows [https://iaifi.org/fellows.html; https://academicjobsonline.org/ajo/jobs/16695] Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 12 IAIFI Research Plan AI2 for Theoretical Physics Standard Model of Nuclear & Particle Physics String Theory & Physical Mathematics Astroparticle Physics Automated Discovery of Physics Models Physics Physics Theory IAIFI Fellows Experiment 2 IAIFI Fellows AI for Experimental Physics Particle Physics Experiments IAIFI Fellows Gravitational Wave Interferometry (Multi-Messenger) Astrophysics AI Foundations AI2 for Foundational AI Symmetries & Invariance Speeding up Control & Inference Physics-Informed Architectures Neural Networks Theory Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 13 AI2 for Theoretical Physics E.g. Lattice Field Theory for Nuclear/Particle Physics Equations governing the strong nuclear force are known, but precision computations are extremely demanding (>10% of open supercomputing in US) Industry collaboration to develop custom AI tools Custom generative models based on normalizing flows achieve 1000-fold acceleration while preserving symmetries & guaranteeing exactness Tools designed for physics find interdisciplinary applications [Kanwar, Albergo, Boyda, Cranmer, Hackett, Racanière, Rezende, Shanahan, PRL 2020] Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 14 AI2 for Experimental Physics E.g. Gravitational Wave Interferometry at LIGO Potential to enhance the physics potential of flagship experiments via improved calibrations, better quantification of uncertainties, enhanced interpretability, and sub-microsecond inference BH-BH Merger LHC Fast-ML to speed Theory Inputs: AI2 to Speed Up up multi-messenger Numerical GR Calculations (Tegmark) astrophysics (Harris) Autoencoders improve LIGO’s sensitivity by 20% [2005.06534] LIGO Data Detector(s) 2017 Nobel Prize Knowledge Propose using RL for Noise Reduction (Barsotti, Rakhlin) Jesse Thaler — The NSF AI Institute for Artificial Intelligence and Fundamental Interactions 15 AI2 for Experimental Physics E.g. Gravitational Wave Interferometry at LIGO Potential to enhance the physics potential of flagship experiments via improved calibrations, better quantification of uncertainties, enhanced interpretability, and sub-microsecond inference BH-BH Merger Outgrowth of Sub-Microsecond LHC Fast-ML to speed Theory Inputs: AI2 to Speed Up up multi-messenger Inference for LHCNumerical Triggering GR Calculations (Tegmark) CMS TriggerCMS Trigger astrophysics (Harris) 1 1 1 KHz Autoencoders CMS100100 KHzkHz T1001 MB kHz 1 MB improve LIGO’s 100 kHz High-Leve High-Leve sensitivity by 20% Trigg 1 MB/evtTrigg [2005.06534] High-Level Offline 40 MHz MHz trigger reconstruction L1 triggerL1 Trigger LIGO ! " 40 MHz Data # Detector(s) hls 4 ml 4 Computing time 1 ns 1 μs ! 100 ms 1 s 2017 Nobel Prize ! !! [talk by Ngadiuba, Fast ML Workshop, Nov 30, 2020; Knowledge Duarte, Han, Harris, Jindariani, Kreinar,Propose Kreis,Ngadiuba, using RL Pierini,