OpenTalks.AI - Андрей Устюжанин, Предиктивная аналитика - обзор текущего состояния и что произошло важного за 2019 год

Ad8ae7af280edaecb09bd73a551b5e5f?s=47 OpenTalks.AI
February 21, 2020

OpenTalks.AI - Андрей Устюжанин, Предиктивная аналитика - обзор текущего состояния и что произошло важного за 2019 год

Ad8ae7af280edaecb09bd73a551b5e5f?s=128

OpenTalks.AI

February 21, 2020
Tweet

Transcript

  1. 2.

    Quick self-intro Andrey Ustyuzhanin 2 ▌ Head of LAboratory of

    Methods for Big Data Analysis, LAMBDA at HSE ▌ Head of Yandex School of Data Analysis team at LHCb and SHiP at CERN › Applications of Machine Learning to natural science challenges › Playground for advanced methods and technologies ▌ Co-organizer of several data science competitions (Flavours of Physics, TrackML, IDAO) ▌ Education (MLHEP, ICL, ClermonFerrand, URL Barcelona, Coursera) ▌ Core expertise: › Data analytics, simulation, generative models, complex optimization ▌ Industry predictive analytics projects with “YADRO”, “MMK”, “Yandex”
  2. 3.

    Predictive analytics – how to turn data into future insights.

    Andrey Ustyuzhanin 3 https://wisdomschema.com/analytics-capability-maturity/
  3. 6.

    Predictive Analytics Key Drivers ▌ Transition from analog to digital

    › IOT – data abundance › Dataism – mindset, developed by significance of Big Data (flows) ▌ Global AI race › AI technologies transit from ‘nice to have’ to ‘must have’ for companies and governments › Changing nature of power ▌ Sustainable solutions, service personalization › From offline to real-time › “What-if … “ analytics, process-oriented analytics Andrey Ustyuzhanin 6
  4. 7.

    Big Data Problem ▌ Complexity of the system › Data

    pieces are always missing › Noise to signal ratio may get high › No single expert knowledge ▌ Process-agnostic › Data is only part of the truth › Relying on part of the past we cannot predict all future scenarios › Future could be a very special version of the past Andrey Ustyuzhanin 7
  5. 8.

    Data is great, but is it the final answer? Andrey

    Ustyuzhanin 8 http://bit.ly/2T4YNJv
  6. 10.

    Simulation Andrey Ustyuzhanin 10 • add any additional rules from

    experts • build a trustworthy model of the system, including those rough edges that your data might miss • build and verify models with historical data • replay it with slightly different conditions and random variations • take into account unexpected interactions (think butterfly effect) http://bit.ly/2T4YNJv
  7. 11.

    Simulation toolkits ▌ Anylogic ▌ NetLogo ▌ Flexsim ▌ Simio,

    ▌ Simul8 ▌ Arena ▌ Salabim ▌ Hash.ai Andrey Ustyuzhanin 11
  8. 12.

    Simulation for industry – Digital Twins Reduced CO emission by

    factor of 3.5 by CompMechLab, http://bit.ly/37DPl4Y Andrey Ustyuzhanin 12 DOI 10.1109/ACCESS.2018.2890566 Optimisation of: ▌ Design ▌ Logistic ▌ Supply chain ▌ New materials Testing of ▌ Maintenance ops ▌ Anomaly ▌ “What-if” scenarios http://bit.ly/2SJfTNO
  9. 14.

    Simulation in Science (Particle Physics) ▌ Toolkits: › Pythia ›

    GEANT4 ▌ Applications: › Rare events simulation › Background process › Tuning of software › Design of hardware Andrey Ustyuzhanin 14
  10. 15.

    Interesting Questions to Explore ▌ Simulation speed-up by either solving

    ODEs or approximating routine simulator calls by Neural Nets, arXiv:1812.01319v2 ▌ Transfer Learning for Machine Fault Diagnosis, http://bit.ly/37PYCal , http://bit.ly/2T4qT7l ▌ Optimisation of computationally expensive hardware design https://arxiv.org/abs/2002.04632 ▌ Tuning of heavy simulators to match historical data Simulation of realistic anomalies - https://arxiv.org/abs/1912.00520 ▌ Fast simulation of physics process by neural networks https://arxiv.org/abs/1903.11788 Andrey Ustyuzhanin 15
  11. 16.

    Key technologies towards Prescriptive Analytics › Simulation, Simulation tuning, Simulation

    speed-up › Transfer Learning (from another domain, from prior knowledge) › NLP (Process Mining, Network Analysis) › Extrapolating models, causality › Interpretability, uncertainty modelling › Few-shot learning (see talk of Sergey Bartunov) Andrey Ustyuzhanin 16
  12. 22.

    Conclusion ▌ Predictive Analytics (PA) is a multifaceted and ubiquitous

    technology ▌ Data-driven PA is not enough ▌ PA is going to expand/adapt a variety of new aspiring technologies ▌ … while pushing AI quite far: › Man-in-the-loop learning › Advanced simulation › AI “scientist” › Few-shot learning ▌ Scientific collaborations (e.g. with CERN, SKA) serve as a great testbed for future industry cases Andrey Ustyuzhanin 22 http://cs.hse.ru/lambda/ anaderiRu@twitter austyuzhanin@hse.ru