Automated Machine Learning,
or Machine Learning for Lazy
Dmitry Petukhov,
Machine Learning Consultant, Cloud Architect,
Microsoft Most Valuable Professional in AI && Coffee Addicted
#AI #Azure #ainights
Automated Machine Learning
 The operator decision – by hands
 Automated machine learning – the automation of automation
of automation.
 If… else – the automation
 Machine learning algorithms – the automation of automation
What is Automated Machine Learning?
The research area that targets the automation process of applying ML
 End-to-end process
 Focused on non-experts in ML
 New tools for ML experts
 Also known as Auto ML.
Auto ML: Motivation
Picture credit
Bad news 
 Too complex
 Too long.
Because life is hard!
Good news 
 Can be standardized.
Let’s automate it!
Auto ML: Frameworks
2 classes by purpose
 Optimal ML algorithm(s) and its hyperparameters search
 Optimal neural network architecture search.
Really?!
Frameworks
 Auto-WEKA (2013)
 auto-sklearn (2014)
 H2O AutoML (2014)
 AutoKeras (2017)
 Google Cloud AutoML (2017)
 Microsoft AutoML in Azure ML Services (2018)
 Microsoft ML.NET (2018)
 to be continued…
Source: https://arxiv.org/pdf/1908.05557.pdf
Auto ML: Frameworks. Part II
Framework
Open-
source
Cloud-
based
Supports
NN
architecture
search
Techniques
Training
framework
Version
auto-sklearn Y N Classification,
Regression
N Bayesian optimization + automated
ensemble construction
sklearn pre-release
AutoKeras Y N CNN, RNN,
LSTM for
classification
Y Efficient Neural Architecture Search with
Network Morphism
Keras pre-release
Google Cloud
AutoML
N Y CNN, RNN,
LSTM for
classification
Y Reinforcement learning with gradient policy
upgrade
TensorFlow beta
AutoML in Azure
ML Service
N Y Classification,
Regression
N Probabilistic Matrix Factorization + Bayesian
optimization
sklearn GA
H2O AutoML Y N Classification,
Regression
N H2O, XGBoost stable
ML.NET AutoML N N Classification,
Regression
N ML.NET preview
ML.NET AutoML Demo
ML.NET used internally by Microsoft for more than 8 years in:
 Bing Ads
 Power BI
 Azure ML Studio
 Windows Defender
 Office Excel
 Office PowerPoint
 and many internal corporate applications.
References:
1. ML.NET AutoML samples.
H2O AutoML Demo
References:
1. H2O tutorials.
Picture credit
On-prem or in the Azure, AWS, IBM’s clouds
The end is near
 Быстрая проверка гипотезы – да
 State-of-the-art – сложно
 В enterprise требует понимания на уровень абстракции ниже – без понимания это выстрел в себе ногу
 Перебор, композиция и оптимизация поиска, но не изобретение нового
 Часто сырое ПО Open source 
© 2019, Dmitry Petukhov. CC BY-SA 4.0 license. Microsoft and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.
Thank you!
Q&A
Now or later (see contacts below)
Stay connected
Facebook: @codez0mb1e
Telegram: @codez0mb1e
GitHub: @codez0mb1e
All contacts: http://0xCode.in/@codez0mb1e
Download slides from 0xcode.in/auto-ml-intro or

Introduction to Auto ML

  • 1.
    Automated Machine Learning, orMachine Learning for Lazy Dmitry Petukhov, Machine Learning Consultant, Cloud Architect, Microsoft Most Valuable Professional in AI && Coffee Addicted #AI #Azure #ainights
  • 2.
    Automated Machine Learning The operator decision – by hands  Automated machine learning – the automation of automation of automation.  If… else – the automation  Machine learning algorithms – the automation of automation
  • 3.
    What is AutomatedMachine Learning? The research area that targets the automation process of applying ML  End-to-end process  Focused on non-experts in ML  New tools for ML experts  Also known as Auto ML.
  • 4.
    Auto ML: Motivation Picturecredit Bad news   Too complex  Too long. Because life is hard! Good news   Can be standardized. Let’s automate it!
  • 5.
    Auto ML: Frameworks 2classes by purpose  Optimal ML algorithm(s) and its hyperparameters search  Optimal neural network architecture search. Really?! Frameworks  Auto-WEKA (2013)  auto-sklearn (2014)  H2O AutoML (2014)  AutoKeras (2017)  Google Cloud AutoML (2017)  Microsoft AutoML in Azure ML Services (2018)  Microsoft ML.NET (2018)  to be continued…
  • 6.
  • 7.
    Auto ML: Frameworks.Part II Framework Open- source Cloud- based Supports NN architecture search Techniques Training framework Version auto-sklearn Y N Classification, Regression N Bayesian optimization + automated ensemble construction sklearn pre-release AutoKeras Y N CNN, RNN, LSTM for classification Y Efficient Neural Architecture Search with Network Morphism Keras pre-release Google Cloud AutoML N Y CNN, RNN, LSTM for classification Y Reinforcement learning with gradient policy upgrade TensorFlow beta AutoML in Azure ML Service N Y Classification, Regression N Probabilistic Matrix Factorization + Bayesian optimization sklearn GA H2O AutoML Y N Classification, Regression N H2O, XGBoost stable ML.NET AutoML N N Classification, Regression N ML.NET preview
  • 8.
    ML.NET AutoML Demo ML.NETused internally by Microsoft for more than 8 years in:  Bing Ads  Power BI  Azure ML Studio  Windows Defender  Office Excel  Office PowerPoint  and many internal corporate applications. References: 1. ML.NET AutoML samples.
  • 9.
    H2O AutoML Demo References: 1.H2O tutorials. Picture credit On-prem or in the Azure, AWS, IBM’s clouds
  • 10.
    The end isnear  Быстрая проверка гипотезы – да  State-of-the-art – сложно  В enterprise требует понимания на уровень абстракции ниже – без понимания это выстрел в себе ногу  Перебор, композиция и оптимизация поиска, но не изобретение нового  Часто сырое ПО Open source 
  • 11.
    © 2019, DmitryPetukhov. CC BY-SA 4.0 license. Microsoft and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. Thank you!
  • 12.
    Q&A Now or later(see contacts below) Stay connected Facebook: @codez0mb1e Telegram: @codez0mb1e GitHub: @codez0mb1e All contacts: http://0xCode.in/@codez0mb1e Download slides from 0xcode.in/auto-ml-intro or

Editor's Notes

  • #3 Кейс с оттоком Оператор > программист > data scientist
  • #4 Что это такое: Область исследований, чьей целью Желаемыми характеристиками являются Следствие – все при деле Далее auto ml
  • #5 Современный workflow гипотеза, чтение, предобработка, feature engineering…. Плохие новости: как говорил наш колоритный современник, это долго дорого… Хорошие новости: если что-то делается похожим образом, то это надо попытаться автоматизировать
  • #6 2 класс по конечной цели: Feature engineering - сам описываешь признаки, описывающие процесс Architecture engineering – описываешь архитектура, которая будет заниматься поиском признаков Развитие идет довольно бурно
  • #7 Поезд уехал? Нет! Пруф: Не тупиковая ветвь/сумасшедшая идея Интерес со стороны вендоров и community Идет развитие из preview What we can?
  • #8 Самая большая фишка – как идет поиск моделей Байесовская оптимизацию Meta-learning на основе датасетов, которые уже видели Auto ensemble generation Feature selection Скромный опыт в CS (10 лет), т.е. написании либо собственного велосипеда, либо чужих новых инновационных либ, подсказывает… На вкус и цвет фломастеры разные, но они действительно разные
  • #9 ML.NET: already used by MS evolution Ideally for .NET developer +Model builder = AI democratization
  • #10 H2O: is ecosystem is one of leader everywhere: on-prem and in clouds
  • #11 Выводы: под капотом стало еще только сложнее Выстрел в ногу, если не знаешь на уровень абстракции ниже TODO: перевести на английский