Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Mlse20190208

950 views

Published on

AE: A domain-agnostic platform for adaptive experimentation
MLSE NeurIPS 2018 Systems for ML 論文読む会

Published in: Technology
  • Be the first to comment

Mlse20190208

  1. 1. AE: A domain-agnostic platform for adaptive experimentation 2019-02-08@ MLSE NuerIPS2018 Sys for ML
  2. 2. Yuta Kashino ( ) BakFoo, Inc. CEO Astro Physics /Observational Cosmology Zope / Python Realtime Data Platform for Enterprise PoC/Prototyping
  3. 3. AE: A domain-agnostic platform for adaptive experimentation Eytan Bakshy, Lili Dworkin, Brian Karrer, Konstantin Kashin, Benjamin Letham, Ashwin Murthy, Shaun Singh Facebook
  4. 4. AE: • Google Vizier / Ray.Tune ML + Python • Vizier: http://bit.ly/2DZoh4I • Ray.Tune: http://bit.ly/2DZ4D8U • FB (OSS • FB Learner Flow http://bit.ly/2DXHGD5 … • sequential experiment: BO,
  5. 5. AE: 1. 1. DNN Web UI A/B / HHVM JIT 2. HHVM JIT: PHP C++/LLVM http://bit.ly/2DYOqQO 2. human-in-the-loop 3. A/B 4.
  6. 6. 1. : BO • Bayesian Optimization • • Spearmint http://bit.ly/2E0pM2p • BayesOpt http://bit.ly/2DZ1mGB • MOE http://bit.ly/2DZOCiS • Hyperopt http://bit.ly/2DYIGXx • Optuna https://optuna.org/ • Ray.Tune http://bit.ly/2DZ4D8U • SigOpt https://sigopt.com/ • Vizier http://bit.ly/2DZoh4I • Kaggle Hyperopt
  7. 7. 1.seqential expriment • sequential expriment: • box 1992 http://bit.ly/2E0qPiR • Bayesian Optimization Bayesian Optimization
  8. 8. 1. :BO • • Gaussian Process(GP) • A/B • x: parameter config. • x*: param. • f: • Expected Improvement(EI) • EI http://bit.ly/2GA7Opj EI(x) = ∫ ∞ −∞ max(0,y − y*)p(y|x)dy y = f(x), y* = f(x*)
  9. 9. 1.Hyperopt • BO : Tree Parzen Estimator (TPE) • PFN http://bit.ly/2GtAm3s • • • fmin
  10. 10. 1.AE: Core API
  11. 11. 2.AE: human-in-the-loop
  12. 12. 3.AE: • Bayesian Optimization • Constrained Bayesian Optimization with Noisy Experiments http://bit.ly/2E1pJTZ • Expected Improvement • botorch • PyTorch/GPyTorch •
  13. 13. 4.AE: • Runner Batch • Online/Offline • MySQL • AnalysisPlan Batch • AnalysisPlan • AnalysisPlan
  14. 14. AE: FB • HHVM JIT • PlanOut UI A/B • DNN • • FB Video • FB http://bit.ly/2GwMAbI
  15. 15. • AE Facebook
  16. 16. • AE Runner: Python shell • Ray + Tune • DNN Kubernetes • Kubeflow + Seldon • Polyaxon • 2018 ML http://bit.ly/2GwocH8 http://bit.ly/2GuCxEa
  17. 17. BakFoo, Inc. • Software 2.0 http://bit.ly/2DZ4z96 • Sofware 2.0 = ML + Software 1.0
  18. 18. BakFoo, Inc. • Software • Kubernetes • FaaS + • FaaS: AWS Lambda, Google Functions … • kubeflow polyaxon • DB/ • Distributed Functions ( ) •
  19. 19. Questions kashino@bakfoo.com @yutakashino
  20. 20. BakFoo, Inc. NHK NMAPS: +
  21. 21. BakFoo, Inc. PyConJP 2015 Python
  22. 22. BakFoo, Inc.
  23. 23. BakFoo, Inc. : SNS +

×