事例を通じて機械学習とは何かを説明する

Machine Learning explained with
Examples
Heli Helskyaho

Copyright © Miracle Finland Oy
* Graduated from University of Helsinki (Master of Science, computer science), currently a doctoral
student, researcher and lecturer (databases, Big Data, Multi-model Databases, methods and tools for
utilizing semi-structured data for decision making) at University of Helsinki
* Worked with Oracle products since 1993, worked for IT since 1990
* Data and Database!
* CEO for Miracle Finland Oy
* Oracle ACE Director, Oracle Groundbreaker Ambassador
* Ambassador for EOUC (EMEA Oracle Users Group Community)
* Listed as one of the TOP 100 influencers on IT sector in Finland (2015, 2016, 2017, 2018)
* Public speaker and an author
* Author of the book Oracle SQL Developer Data Modeler for Database Design Mastery (Oracle Press,
2015), co-author for Real World SQL and PL/SQL: Advice from the Experts (Oracle Press, 2016)
Introduction, Heli

3 Membership Tiers
• Oracle ACE Director
• Oracle ACE
• Oracle ACE Associate
bit.ly/OracleACEProgram
500+ Technical Experts
Helping Peers Globally
Connect:
Nominate yourself or someone you know: acenomination.oracle.com
@oracleace
Facebook.com/oracleaces
oracle-ace_ww@oracle.com

What is Machine Learning?
* An important part of Artificial Intelligence (AI)
* Machine learning (ML) teaches computers to learn from experience
(algorithms)
* “field of study that gives computers the ability to learn without being
explicitly programmed“ -- Arthur Samuel, 1959
* A systematic study of algorithms and systems that improve their
knowledge or performance with experience

* Spam filters, Log filters/alarms
* Data analytics
* Image recognition, Speech recognition
* Medical diagnosis
* Robotics
* Fraud protection/detection (credit card)
* Product / music / movie recommendation
* …
Real life use cases for ML

* ”How to serve customers better in simple tasks like
* - finding the bus schedule
* - order pizza
* - help students with their courses, studies etc
* - ”therapy” with Woebot
* …”
A simple example, Chatbot

* Online shopping (Amazon, Search, recommendations)
* Voice-to-Text, Smart Personal Assistants (mobile services: ”recipe for
bread”, ”find the nearest grocery”)
* Siri, Google Assistant, Alexa, Echo, Cortana,…
* Facebook
* …
Real life use cases for ML

* The face recognition (demo) for our house alarm system
* (Python)
* ”If there are human faces, take a photo, identify those and let me know
how many faces there are.”
My real life use case

* Oracle Database Advanced Analytics (OAA) =
* Oracle DB + Oracle Data Mining (ODM) (+Data Miner GUI in Oracle SQL
Developer) +
* Oracle R Enterprise (ORE)
* Predictive Analytics with Oracle Data Mining (ODM)
* Predictive Queries with Oracle Analytic Functions
* “Oracle Machine Learning” is a Zeppelin based SQL notebook that is
available with ADWC
In-database Machine Learning

* Why to move the data?
* It usually is a lot of data…
* Preparing the data is the hard part (80/20)
* Database is the natural environment for handling data
* Maybe you already know SQL and PL/SQL?
* You can use APEX for visualizing the data and the process
* The deployment (table, view, PL/SQL Package, Function, Procedure,…) is easily
used with many technologies
* …
Why in-database?

* Oracle SQL Developer is a free tool from Oracle
* Has an add-on called Data Miner
* Advanced analytics (Data Miner uses that) is a licensed product (in the EE
database separately licensed, in the Cloud: Database Service either High
Performace Package or Extreme Performance Package)
* Oracle Data Miner GUI Installation Instructions
http://www.oracle.com/technetwork/database/options/advanced-
analytics/odm/odmrinstallation-2080768.html
* Tutorial
http://www.oracle.com/webfolder/technetwork/tutorials/obe/db/12c/BigDataDM/
ODM12c-BDL4.html
Oracle SQL Developer, Data Miner

* ”Which customers would most likely buy the insurance?”
* (SQL Developer)
The Task

Oracle SQL Developer demo

* ”Students that will enroll?”
* (Oracle Autonomous DW)
The Task

* Apache Zeppelin based Machine Learning SQL Notebook
SQL, Oracle Autonomous DW

SQL, Oracle Autonomous DW

* ”Predict an overall rating for a beer.”
* (SQL, PL/SQL, APEX)
The Task

SQL, APEX

”What is the probability for beer 43548 to get overall 5?”
SELECT PREDICTION_PROBABILITY(Beer_DT, 5 USING *) beer_overall_prob
FROM beer_test
WHERE idindex = 43658;
1.5087463556851313E-001
SELECT PREDICTION_PROBABILITY(Beer_DT, 5 USING STYLE) beer_overall_prob
FROM beer_test
WHERE idindex = 43658;
2.4140018157974377E-001
USING: Predictors and/or Expressions
Real-time scoring, single record scoring

”What are the brewers that I should contact (making most of the overall 5
beers)?"
SELECT brewerid, count(*)
FROM beer_test
WHERE PREDICTION(Beer_DT USING *) = 5
group by brewerid
order by 2 desc;
Real-time scoring, batch scoring

* “Predict the Telco customers about to leave”
* (Python)
Task

* ”Predict the house prices”
* (R)
Task

An example of something more complicated

Kim Hazelwood, Sarah Bird, David Brooks, Soumith Chintala, Utku Diril, Dmytro Dzhulgakov,
Mohamed Fawzy, Bill Jia, Yangqing Jia, Aditya Kalro, James Law, Kevin Lee, Jason Lu, Pieter
Noordhuis, Misha Smelyanskiy, Liang Xiong, Xiaodong Wang, “Applied Machine Learning at
Facebook: A Datacenter Infrastructure Perspective”, Facebook, Inc.
X. He, J. Pan, O. Jun, T. Xu, B. Liu, T. Xu, Y. Shi, A. Atallah, R. Herbrich, S. Bowers, and J.
Quinonero Candela, “Practical lessons from predicting clicks on ads at facebook,” in
Proceedings of the Eighth International Workshop on Data Mining for Online Advertising, ser.
ADKDD’14. New York, NY, USA: ACM, 2014, pp. 5:1–5:9.
J. Dunn, “Introducing FBLearner flow: Facebook’s AI backbone,” May 2016, https://fb.me/dunn
2016.
https://code.facebook.com/posts/1072626246134461/introducing-fblearner-flow-facebook-s-ai-
backbone/
Example Facebook, References

* “Give people the power to build community and bring the world closer together.”
* Facebook connects more than two billion people as of December 2017
* Could not be done without ML
* The massive amount of data required by machine learning services presents challenges to
Facebook’s datacenters.
* Several techniques are used to efficiently feed data to the models including decoupling of
data feed and training, data/compute co-location, and networking optimizations.
* Disaster recovery planning is essential
* actively evaluating and prototyping new hardware solutions while remaining cognizant of
game changing algorithmic innovations
Facebook’s mission

* News Feed ranking
* Ads
* Search
* Sigma
* Lumos
* Facer
* Language Translation
* Speech Recognition
Facebook, some use cases for ML, the Products

* ML is used for
* ranking and personalizing News Feed stories
* filtering out offensive content
* highlighting trending topics
* ranking search results, and much more.
* General models are trained to determine various user and environmental factors
that should ultimately determine the rank order of content.
* The model is used to generate a personalized set of the best posts, images, and
other content to display from thousands of candidates, and the best ordering of
this chosen content.
News Feed

* Online advertising allows advertisers to only bid and pay for measurable user
responses, such as clicks on ads.
* As a consequence, click prediction systems are central to most online advertising
systems.
* General Ads models are trained to learn how user traits, user context, previous
interactions, and advertisement attributes can be most predictive of the
likelihood of clicking on an ad, visiting a website, and/or purchasing a product.
* Inputs are run through a trained model to immediately determine which ads to
display to a particular Facebook user.
Ads

* The click prediction system needs to be robust and adaptive, and capable
of learning from massive volumes of data.
* At Facebook they use a model which combines decision trees with logistic
regression
* Based on their experience: the most important thing is to have the right
features (those capturing historical information about the user or ad
dominate other types of features) and the right model
* Measures: the accuracy of prediction
Predicting the Clicks

X. He, J. Pan, O. Jun, T. Xu, B. Liu, T. Xu, Y. Shi, A. Atallah, R. Herbrich, S. Bowers, and J. Quinonero Candela, “Practical lessons from predicting clicks on ads at facebook,” in
Proceedings of the Eighth International Workshop on Data Mining for Online Advertising, ser. ADKDD’14. New York, NY, USA: ACM, 2014, pp. 5:1–5:9.

* Launches a series of distinct and specialized sub-searches to the various
verticals, e.g., videos, photos, people, events, etc.
* A classifier layer is run atop the various search verticals to predict which
of the many verticals to search (searching all possible verticals would be
inefficient)
* The classifier and these search verticals consist of
* an offline stage to train the models
* and an online stage to run the models and perform the classification
and search
Search

* General classification and anomaly detection framework that is used for a
variety of internal applications (site integrity, spam detection, payments,
registration, unauthorized employee access, and event
recommendations)
* Sigma includes hundreds of distinct models running in production everyday
* each model is trained to detect anomalies (e.g. classify content)
Sigma

* Extract high-level attributes and embeddings from an image and its
content
* That data can be used as input to other products and services
* for example as it were text.
Lumos

* Facebook’s face detection and recognition framework
* Given an image
* finds all of the faces in that image
* runs a user-specific facial-recognition algorithm to determine the likelihood of
that face belonging to one of your top-N friends who have enabled face
recognition
* This allows Facebook to suggest which of your friends you might want to
tag within the photos you upload.
Facer

* Service that manages internationalization of Facebook content
* Supports translations for more than 45 languages (as the source or
target language)
* supports more than 2000 translation directions
* serves 4.5B translated post impressions every day
* Each language pair direction has its own model
* multi-language models are being considered
Language Translation

* Converts audio streams into text
* Provides automated captioning for video
* Most streams are English language
* other languages will be available in future
* Additionally, non-language audio events are also detected with a similar
system (simpler model).
Speech Recognition

Algorithms Facebook uses for these services

How do they do all this at Facebook?

FBLearner Platform

* The starting point for a ML modeling task is to gather and generate
features.
* The Feature Store is a catalog of several feature generators
* can be used both for training and real-time prediction
* serves as a marketplace that multiple teams can use to share and discover
features
FBLearner Feature Store

* Facebook’s machine learning platform for model training
* Workflows: A workflow is a single pipeline defined within FBLearner Flow and is the entry point for all
machine learning tasks.
* Each workflow performs a specific job, such as training and evaluation of a specific model.
* Workflows are defined in terms of operators and can be parallelized.
* Operators: Operators are the building blocks of workflows
* In FBLearner Flow, operators are the smallest unit of execution and run on a single machine.
* Channels: Channels represent inputs and outputs, which flow between operators within a workflow.
* All channels are typed using a custom type system.
* Flow has tooling for experiment management.
* The user interface keeps track of all of the artifacts and metrics generated by each workflow execution
or experiment.
* The user interface makes it simple to compare and manage these experiments.
FBLearner Flow

* The platform consists of three core components:
* an authorship and execution environment for custom distributed workflows
* an experimentation management UI for launching experiments and viewing
results
* numerous predefined pipelines for training the most commonly used machine
learning algorithms at Facebook.
FBLearner Flow

* Facebook’s internal inference engine that uses the models trained in FBLearner
Flow to provide predictions in real time.
* Can be used
* as a multitenancy service
* or as a library that can be integrated in product specific backend services
* Is used by multiple product teams at Facebook, many of which require low latency
solutions.
* The direct integration between Flow and Predictor also helps with
* running online experiments
* managing multiple versions of models in productions
FBLearner Predictor

* Two distinct but synergistic frameworks for deep learning at Facebook:
* PyTorch, which is optimized for research
* Caffe2, which is optimized for production
Frameworks for deep learning

* PyTorch is the framework for AI research at Facebook which enables rapid
experimentation
* Flexibility
* Debugging
* Dynamic neural networks
* Not optimized for production and mobile deployments (Python)
* When research projects produce valuable results, the models need to be
transferred to production.
* Traditionally, rewriting the training pipeline in a product environment with other
frameworks.
PyTorch

* Facebook’s in-house production framework
* For training and deploying large-scale machine learning models
* Focuses on several key features required by products:
* Performance
* cross-platform support
* coverage for fundamental machine learning algorithms (convolutional neural
networks (CNNs), recurrent networks (RNNs), and multi-layer perceptrons
(MLPs)) and up to tens of billions of parameters
Caffe2

* Different tools are better for different subset of problems and have varying tradeoffs on
flexibility, performance, and supported platforms . As a result, there should be a way to exchange
trained models between different frameworks or platforms.
* ONNX is a format to represent deep learning models in a standard way to enable interoperability
across different frameworks and vendor-optimized libraries.
* ONNX is designed as an open specification
* Within Facebook, ONNX is used for transferring research models from the PyTorch environment
to high-performance production environment in Caffe2.
* ONNX provides the ability to automatically capture and translate static parts of the models.
* An additional toolchain facilitates transfer of dynamic graph parts from Python by either mapping
them to control-flow primitives in Caffe2 or reimplementing them in C++ as custom operators.
Open Neural Network Exchange, ONNX

Caffe2 and PyTorch join forces to create a Research + Production platform
PyTorch 1.0:
https://caffe2.ai/blog/2018/05/02/Caffe2_PyTorch_1_0.html
“We realized that in order to deliver the best user experience, it makes
sense to combine the beneficial traits of Caffe2 and PyTorch into a single
package and enable a smooth transition from fast prototyping to fast
execution. It’d also improve our developer efficiency by more easily
utilizing a shared set of tools.”
Caffe2 and PyTorch projects are merging

* success is predicated on the availability of extensive, high-quality data
* complex preprocessing logic is applied to ensure that data is cleaned and
normalized to allow efficient transfer and easy learning
* The ability to rapidly process and feed these data to the training machines
is important for ensuring that we have fast and efficient offline training.
* These impose very high resource requirement especially on storage,
network, and CPU.
* actively evaluating and prototyping new hardware solutions while
remaining cognizant of game changing algorithmic innovations
The success factors 1/2

* https://research.fb.com/category/machine-learning/
Facebook, research

* Knowing what to measure to know what to improve
The success factors 2/2

* Number of positives, number of negatives, number of true positives, number of
false positives, number of true negatives, number of false negatives
* Portion of positives, portion of negatives
* Class ratio
* Accuracy, Error rate
* ROC curve, coverage curve,
* …
* It all depends on how we define the performance for the answer to our
question (experiment): experimental objective
What to measure?

* “we noticed that the largest improvements in accuracy often came from
quick experiments, feature engineering, and model tuning rather than
applying fundamentally different algorithms”
* An engineer may need to attempt hundreds of experiments before
finding a successful new feature or set of hyperparameters.
Facebook

* AI and machine learning is here and it’s the future
* So many interesting areas to learn
* Pick your area and START LEARNING!
The future and now!

* Python
* R
* C/C++
* Java
* JavaScript
* Julia, Scala, Ruby, Octave, MATLAB, SAS
* …
* https://medium.com/towards-data-science/what-is-the-best-programming-language-for-
machine-learning-a745c156d6b7
So many more languages to learn…

* Reinforcement learning
* the machine or software agent learns based on feedback from the environment
* Preference learning
* inducing predictive preference models from empirical data
* Multi-task learning
* multiple learning tasks are solved at the same time, while exploiting commonalities
and differences across tasks
* Online machine learning
* data becomes available in a sequential order and is used to update our best
predictor for future data at each step
And there’s so much more in Machine Learning
too!

* Active learning
* A learning algorithm is able to interactively query the user (or some other
information source) to obtain the desired outputs at new data points
* Deep learning
* Images and anything that is in ”several layers”
* Adaptive Intelligence
* People and machines

* ML can be used ”everywhere”:
* Spam filters
* Log filters (and alarms)
* Data analytics
* Image recognition
* Speech recognition
* Medical diagnosis
* Robotics
* Chatbots
* …
Conclusion

* Facebook uses ML ”everywhere”
* News Feed ranking
* Ads
* Search
* Sigma
* Lumos
* Facer
* Language Translation
* Speech Recognition
Conclusion

* You can use ML ”everywhere”
* Start small and when you learn more do more
* Define a Task and let ML solve it
* Machines are not taking our jobs but helping us to do more interesting things
* With ML we can understand our data better and make better decisions
Conclusion

THANK YOU!
QUESTIONS?
Email: heli@miracleoy.fi
Twitter: @HeliFromFinland
Blog: Helifromfinland.com

事例を通じて機械学習とは何かを説明する

More Related Content

Similar to 事例を通じて機械学習とは何かを説明する (20)

More from Insight Technology, Inc. (20)

Recently uploaded (20)

事例を通じて機械学習とは何かを説明する