How to give your business
superpowers.
DEMYSTIFYING
MACHINE LEARNING
Expertise on demand.
Train your ideal decision-making process,
then execute it anytime, anywhere, at any scale.
WHAT WE'RE TALKING ABOUT
Fill in the gaps and squash hype around ML,
Build the case for using it now,
And provide easy ways to get started.
TODAY’S GOAL
● Why now?
● Foundation
● Use cases
● ~Technical
● Get started
● Demos
OUR JOURNEY
Who thinks machine learning is some kind of voodoo?
( That’s a good thing. )
● We’re not going to dive into the math
● My goal is to show you how easy it is to use
● It’s a tool — just another API
You don't need to understand how
an engine works to drive a car.
KEEP IT SIMPLE
● Software is eating the world and machine learning
is eating the software
● Machine learning (AI) will be the backbone of all
next generation business
“mobile first” => “AI first”
WHY IT'S IMPORTANT
Whether you want to:
● Start a new business,
● Enhance an existing business, or
● Get a new job/promotion
Machine learning will give your applications
superpowers ...for now.
(It will be the norm very soon)
WHAT IT CAN DO FOR YOU
● You don’t need a supercomputer
● You don’t need to write a ton of code
● You don’t need to invest massive amounts of time
● You don’t need a data science degree
● You don’t need to be a math whiz
● You don’t need mountains of data
MYTH BUSTING
WE’VE HEARD IT BEFORE
Is machine learning hype living up to
expectations this time around?
Everything is becoming software
● Limitless computing
● Limitless storage
● Limitless data (IoT = massive need)
● Deep learning
● Targeted machine learning SaaS (easy access)
But, more importantly...
WHY NOW?
Because Google says so :)
“Machine learning is not the future. It is now.”
~Google I/O 2016
WHY NOW?
youtube.com/watch?v=3dXQxSI3XDY
Massive strides in the past year
Just in the past few months…
● Google open sources natural language processing
platform
● Amazon open sources deep learning platform
● Google announces quantum computing works
● IBM offers access to quantum computer
● Google’s DeepMind beats Go champion
WHAT’S NEW
WILL IT STICK THIS TIME?
The Internet gave us big data (greater need)
The cloud gave us massive computing (more horsepower)
And it’s getting much, much bigger…
BIG DATAx
MASSIVE COMPUTINGx
100 million times faster...?
“I would predict that in 10 years there’s nothing but
quantum machine learning”
~Hartman Nevet
Head of Google’s Quantum AI Lab
via: technologyreview.com
via: researchgate.net
ON A PATH TO UBIQUITY
“The most profound technologies are those that
disappear. They weave themselves into the fabric of
everyday life until they are indistinguishable from it.”
~Mark Weiser
Scientific American, 1991
IN JUST 4 YEARS
Predicted for 2020...
● 13% of US households own consumer robots 1
(robotics)
● 30% of new cars will have a self-driving mode 2
(auto)
● 70% of mobile users access devices via biometrics 2
(security)
● We interact with 150+ smart devices (IoT) every day 2
(lifestyle)
All are underpinned by machine learning
1
roboticstrends.com/article/13_of_us_households_to_own_consumer_robots_by_2020
2
weforum.org/agenda/2015/02/5-predictions-for-technology-in-2020
ADDING FUEL TO THE FIRE
Think global.
tractica.com/newsroom/press-releases/artificial-intelligence-for-enterprise-applications-
to-reach-11-1-billion-in-market-value-by-2024
THE GOLDEN AGE OF AI
We’ve hit the tipping point.
Watching AI get smarter is
like watching a bullet train.
The moment you see it
coming, it’s already blown
past you.
HOW I GOT STARTED
Apache
Mahout
Decision Forest
Behavior
prediction
Suite of
mobile apps
Determine the most relevant
(highest-converting) sales offer to present to
each individual user — and the best
(highest-converting) time to present it.
Will the current user buy “Madden NFL” right now?
WHAT IS A DECISION FOREST?
is male?
is age
> 16?
is Y app
installed?
is X app
installed?
end
has used >
30 days?
was X
function
used?
was Y
function
used?
no
yes
no
yes
no
yes
no
yes
end
(better ways to do this now)
no
yes
end
do it
FOUNDATION
Exploring the basics.
“An algorithm that
can learn from
data without
relying on
rules-based
programming.”
WHAT IS MACHINE LEARNING?
analyticsvidhya.com/blog/2015/07/difference-machine-learning-statistical-modeling
Training your computer to do stuff,
just like you would train a pet.
IN OTHER WORDS...
SIMILAR TO HOW WE LEARN
Data System Output
Model
Question Answer
Life experience
Emotions
Mindset
Training data
Algorithm
Perspective
● Model — The reference data pattern (decision-making stuff)
● Algorithm — Process the computer uses to learn the model
(perspective)
● Training — Building the model from historical data (life
experience)
○ Supervised learning — Labeled training data
○ Unsupervised learning — Unlabeled training data
○ Reinforcement learning — Reward-based training
● Feature — Points of differentiation in the data
MAJOR COMPONENTS
cse.unsw.edu.au/~billw/mldict.html
ENDLESS ALGORITHMS
Different for each algorithm & platform
For Amazon Machine Learning (logistic regression)…
● Binary (Yes or no, Actionable or non-actionable)
● Pick from list (Is this tweet a question, complaint,
or praise?)
● Number (How much will this house sell for?)
Sky's the limit on how you can apply these
WHAT IS THE OUTPUT?
IN THE WILD
Recommender
(pick from list)
Classifier
(binary)
Visual
recognition
(deep learning)
“Features”
How would you teach a
child to recognize the
differences?
● Distance between eyes
● Width of nose
● Shape of cheekbones
HOW DOES IT CLASSIFY?
“Probability”
Each potential
answer gets a
numeric
probability
calculated for it.
Higher
probability
means greater
confidence.
HOW DOES IT MAKE DECISIONS?
We’re already using it.
LOOK FAMILIAR?
Understand & answer
SEARCH RESULTS
( ibm.com/smarterplanet/us/en/ibmwatson/developercloud/concept-insights.html )
PRODUCT RECOMMENDATIONS
techblog.netflix.com/2012/04/netflix-recommendations-beyond-5-stars.html
templates.prediction.io/PredictionIO/template-scala-parallel-universal-recommendation
SENTIMENT ANALYSIS
keyhole.co
ibm.com/smarterplanet/us/en/ibmwatson/developercloud/tone-analyzer.html
templates.prediction.io/pawel-n/template-scala-cml-sentiment
SPEECH RECOGNITION
cloud.google.com/speech
developer.amazon.com/public/solutions/alexa
github.com/tensorflow/models/tree/master/syntaxnet (Parsey McParseface)
Assembled by
machine learning
360° PHOTOS
bgr.com cloud.google.com/vision
TRANSLATION
AUTOMATED CAPTIONS
“A group of young
people playing a
game of frisbee.”
Great example of
deep learning —
understanding the
context of an image.
io9.gizmodo.com/computers-wrote-the-caption-for-this-photograph-and-ch-1660450610
( I believe every business will need these
2 systems moving forward. )
COMPOUNDING
FUNCTIONALITY
Speech
to Text
Sentiment
Analysis
Actionable
Analysis
Customer
Support
PREDICTIVE ENGAGEMENT
Customer
support call
recordings
Convert audio
into text
Analyze for
mood keywords
Determine if
response is required
Reach out to
customer/prospect
Blog & community comments
Social media mentions
Press & blog coverage
Customer support chat
Product reviews
Inbound emails
[ IBM Watson Speech to Text ] [ IBM Watson Tone Analyzer ] [ IBM Watson AlchemyLanguage ]
Behavior
Prediction
Interest
Tracking
PREDICTIVE PERSONALIZATION
Pages & content they’ve visited
Emails they’ve opened/clicked
Resources they’ve used/downloaded
Products they’ve viewed/wishlisted/bought
Searches they’ve made
Blog
Store
Find patterns Determine what they want to
see/do/buy next (and when)
Days/time they’re active App
Search
Devices they’ve used (& geo location)
Email
Social
• Recommended posts
• Recommended products
• Delivery day/time
• Dynamic content
• Related posts
• Sales offers
• Related products
• Cross/up sell
• Dynamic pricing
• Dynamic content
• Sales offers
• Functionality
• Query suggestions
• Results ranking
• Sales offers
• Content curation
• Delivery day/time
• Retweet/reshare
Tribe
• Recommended topics
• Topic curation
• Member introductions[ Amazon Machine Learning ]
[ Amazon Machine Learning ]
Moving into the technical details.
A BIT DEEPER
A many-layered Artificial Neural Network (~self-learning)
WHAT IS DEEP LEARNING?
“deep”
cs231n.github.io/neural-networks-1
“shallow”
(SIMPLE) NEURAL NETWORK
Each layer performs a
discrete function
≥ 1 input
neurons
≥ 1 output
neurons
≥ 1 hidden layers
Output “fires” if all
weighted inputs sum
to a set “threshold”
Each connection applies a
“weighted” influence on
the receiving neuron
Layers build on each other
(iterative)
Each input can
be a separate
“feature”
Each neuron takes in
multiple inputs
Hidden layers can’t directly
“see” or act on outside world
HOW MUCH IS A HOUSE WORTH?
Decisions based on combinations.
3 bedrooms
37 years old
1450 ft2
$191,172
Is it “old” or “historic?”
Is it “small” or “open floor plan?”
$32,108 per bedroom
$64,251 per acre
Need a lower weight for “old”
Apply initial
abstractions
Set values
● Vanilla Neural Network — nothing fancy
● Convolutional Neural Network — inspired by visual
cortex
● Deep Belief Network — undirected connections
● Recurrent Neural Network — multi-pass
MANY DIFFERENT FLAVORS
● R
● Python
● Matlab/Octave
● Java
● C / C++
kdnuggets.com/2016/06/r-python-top-analytics-data-mining-data-science-software.html
POPULAR LANGUAGES
● Amazon Machine Learning
● Google Prediction API*
● Google Cloud Machine Learning
● Microsoft Azure Machine Learning
● IBM Watson Machine Learning
● DiffBot
● Alibaba Cloud DT PAI
SaaS OPTIONS
● TensorFlow *
● Amazon DSSTNE *
● H2O *
● PredictionIO
● Apache Mahout
● Scikit Learn
● Caffe *
OPEN SOURCE OPTIONS
● Microsoft CNTK *
● Torch *
● Theano *
● MXnet *
● Chainer *
● Keras *
● Neon *
* Deep learning
How to get the ball rolling.
GET STARTED
● archive.ics.uci.edu/ml
● deeplearning.net/datasets
● mldata.org
● grouplens.org/datasets
● cs.toronto.edu/~kriz/cifar.html
● cs.cornell.edu/people/pabo/movie-review-data
● yann.lecun.com/exdb/mnist (handwriting)
● kdnuggets.com/datasets/index.html (long list)
● image-net.org (competition)
OPEN SOURCE DATASETS
● playground.tensorflow.org (neural network demo)
● cs.stanford.edu/people/karpathy/convnetjs
● github.com/awslabs/machine-learning-samples
● ibm.com/smarterplanet/us/en/ibmwatson/devel
opercloud/starter-kits.html
● templates.prediction.io
EASY STARTING POINTS
Start now
● It’s here, today
● It’s evolving exponentially
● Build “AI-First”
RECAP
Let’s see some action.
DEMOS
● AlchemyLanguage
● Dialog
● Natural Language Classifier
● Personality Insights
● Relationship Extraction
● Tradeoff Analytics
ibm.com/smarterplanet/us/en/ibmwatson/developercloud/services-catalog.html
IBM WATSON
UNLEASH YOUR BUSINESS
EMBRACE EXPONENTIAL
10xnation.com

Demystifying Machine Learning - How to give your business superpowers.

  • 1.
    How to giveyour business superpowers. DEMYSTIFYING MACHINE LEARNING
  • 2.
    Expertise on demand. Trainyour ideal decision-making process, then execute it anytime, anywhere, at any scale. WHAT WE'RE TALKING ABOUT
  • 3.
    Fill in thegaps and squash hype around ML, Build the case for using it now, And provide easy ways to get started. TODAY’S GOAL
  • 4.
    ● Why now? ●Foundation ● Use cases ● ~Technical ● Get started ● Demos OUR JOURNEY
  • 5.
    Who thinks machinelearning is some kind of voodoo? ( That’s a good thing. ) ● We’re not going to dive into the math ● My goal is to show you how easy it is to use ● It’s a tool — just another API You don't need to understand how an engine works to drive a car. KEEP IT SIMPLE
  • 6.
    ● Software iseating the world and machine learning is eating the software ● Machine learning (AI) will be the backbone of all next generation business “mobile first” => “AI first” WHY IT'S IMPORTANT
  • 7.
    Whether you wantto: ● Start a new business, ● Enhance an existing business, or ● Get a new job/promotion Machine learning will give your applications superpowers ...for now. (It will be the norm very soon) WHAT IT CAN DO FOR YOU
  • 8.
    ● You don’tneed a supercomputer ● You don’t need to write a ton of code ● You don’t need to invest massive amounts of time ● You don’t need a data science degree ● You don’t need to be a math whiz ● You don’t need mountains of data MYTH BUSTING
  • 9.
    WE’VE HEARD ITBEFORE Is machine learning hype living up to expectations this time around?
  • 10.
    Everything is becomingsoftware ● Limitless computing ● Limitless storage ● Limitless data (IoT = massive need) ● Deep learning ● Targeted machine learning SaaS (easy access) But, more importantly... WHY NOW?
  • 11.
    Because Google saysso :) “Machine learning is not the future. It is now.” ~Google I/O 2016 WHY NOW? youtube.com/watch?v=3dXQxSI3XDY
  • 12.
    Massive strides inthe past year Just in the past few months… ● Google open sources natural language processing platform ● Amazon open sources deep learning platform ● Google announces quantum computing works ● IBM offers access to quantum computer ● Google’s DeepMind beats Go champion WHAT’S NEW
  • 13.
    WILL IT STICKTHIS TIME? The Internet gave us big data (greater need) The cloud gave us massive computing (more horsepower) And it’s getting much, much bigger…
  • 14.
  • 15.
    MASSIVE COMPUTINGx 100 milliontimes faster...? “I would predict that in 10 years there’s nothing but quantum machine learning” ~Hartman Nevet Head of Google’s Quantum AI Lab via: technologyreview.com via: researchgate.net
  • 16.
    ON A PATHTO UBIQUITY “The most profound technologies are those that disappear. They weave themselves into the fabric of everyday life until they are indistinguishable from it.” ~Mark Weiser Scientific American, 1991
  • 17.
    IN JUST 4YEARS Predicted for 2020... ● 13% of US households own consumer robots 1 (robotics) ● 30% of new cars will have a self-driving mode 2 (auto) ● 70% of mobile users access devices via biometrics 2 (security) ● We interact with 150+ smart devices (IoT) every day 2 (lifestyle) All are underpinned by machine learning 1 roboticstrends.com/article/13_of_us_households_to_own_consumer_robots_by_2020 2 weforum.org/agenda/2015/02/5-predictions-for-technology-in-2020
  • 18.
    ADDING FUEL TOTHE FIRE Think global. tractica.com/newsroom/press-releases/artificial-intelligence-for-enterprise-applications- to-reach-11-1-billion-in-market-value-by-2024
  • 19.
    THE GOLDEN AGEOF AI We’ve hit the tipping point. Watching AI get smarter is like watching a bullet train. The moment you see it coming, it’s already blown past you.
  • 20.
    HOW I GOTSTARTED Apache Mahout Decision Forest Behavior prediction Suite of mobile apps Determine the most relevant (highest-converting) sales offer to present to each individual user — and the best (highest-converting) time to present it.
  • 21.
    Will the currentuser buy “Madden NFL” right now? WHAT IS A DECISION FOREST? is male? is age > 16? is Y app installed? is X app installed? end has used > 30 days? was X function used? was Y function used? no yes no yes no yes no yes end (better ways to do this now) no yes end do it
  • 22.
  • 23.
    “An algorithm that canlearn from data without relying on rules-based programming.” WHAT IS MACHINE LEARNING? analyticsvidhya.com/blog/2015/07/difference-machine-learning-statistical-modeling
  • 24.
    Training your computerto do stuff, just like you would train a pet. IN OTHER WORDS...
  • 25.
    SIMILAR TO HOWWE LEARN Data System Output Model Question Answer Life experience Emotions Mindset Training data Algorithm Perspective
  • 26.
    ● Model —The reference data pattern (decision-making stuff) ● Algorithm — Process the computer uses to learn the model (perspective) ● Training — Building the model from historical data (life experience) ○ Supervised learning — Labeled training data ○ Unsupervised learning — Unlabeled training data ○ Reinforcement learning — Reward-based training ● Feature — Points of differentiation in the data MAJOR COMPONENTS cse.unsw.edu.au/~billw/mldict.html
  • 27.
  • 28.
    Different for eachalgorithm & platform For Amazon Machine Learning (logistic regression)… ● Binary (Yes or no, Actionable or non-actionable) ● Pick from list (Is this tweet a question, complaint, or praise?) ● Number (How much will this house sell for?) Sky's the limit on how you can apply these WHAT IS THE OUTPUT?
  • 29.
    IN THE WILD Recommender (pickfrom list) Classifier (binary) Visual recognition (deep learning)
  • 30.
    “Features” How would youteach a child to recognize the differences? ● Distance between eyes ● Width of nose ● Shape of cheekbones HOW DOES IT CLASSIFY?
  • 31.
    “Probability” Each potential answer getsa numeric probability calculated for it. Higher probability means greater confidence. HOW DOES IT MAKE DECISIONS?
  • 32.
    We’re already usingit. LOOK FAMILIAR?
  • 33.
    Understand & answer SEARCHRESULTS ( ibm.com/smarterplanet/us/en/ibmwatson/developercloud/concept-insights.html )
  • 34.
  • 35.
  • 36.
  • 37.
    Assembled by machine learning 360°PHOTOS bgr.com cloud.google.com/vision
  • 38.
  • 39.
    AUTOMATED CAPTIONS “A groupof young people playing a game of frisbee.” Great example of deep learning — understanding the context of an image. io9.gizmodo.com/computers-wrote-the-caption-for-this-photograph-and-ch-1660450610
  • 40.
    ( I believeevery business will need these 2 systems moving forward. ) COMPOUNDING FUNCTIONALITY
  • 41.
    Speech to Text Sentiment Analysis Actionable Analysis Customer Support PREDICTIVE ENGAGEMENT Customer supportcall recordings Convert audio into text Analyze for mood keywords Determine if response is required Reach out to customer/prospect Blog & community comments Social media mentions Press & blog coverage Customer support chat Product reviews Inbound emails [ IBM Watson Speech to Text ] [ IBM Watson Tone Analyzer ] [ IBM Watson AlchemyLanguage ]
  • 42.
    Behavior Prediction Interest Tracking PREDICTIVE PERSONALIZATION Pages &content they’ve visited Emails they’ve opened/clicked Resources they’ve used/downloaded Products they’ve viewed/wishlisted/bought Searches they’ve made Blog Store Find patterns Determine what they want to see/do/buy next (and when) Days/time they’re active App Search Devices they’ve used (& geo location) Email Social • Recommended posts • Recommended products • Delivery day/time • Dynamic content • Related posts • Sales offers • Related products • Cross/up sell • Dynamic pricing • Dynamic content • Sales offers • Functionality • Query suggestions • Results ranking • Sales offers • Content curation • Delivery day/time • Retweet/reshare Tribe • Recommended topics • Topic curation • Member introductions[ Amazon Machine Learning ] [ Amazon Machine Learning ]
  • 43.
    Moving into thetechnical details. A BIT DEEPER
  • 44.
    A many-layered ArtificialNeural Network (~self-learning) WHAT IS DEEP LEARNING? “deep” cs231n.github.io/neural-networks-1 “shallow”
  • 45.
    (SIMPLE) NEURAL NETWORK Eachlayer performs a discrete function ≥ 1 input neurons ≥ 1 output neurons ≥ 1 hidden layers Output “fires” if all weighted inputs sum to a set “threshold” Each connection applies a “weighted” influence on the receiving neuron Layers build on each other (iterative) Each input can be a separate “feature” Each neuron takes in multiple inputs Hidden layers can’t directly “see” or act on outside world
  • 46.
    HOW MUCH ISA HOUSE WORTH? Decisions based on combinations. 3 bedrooms 37 years old 1450 ft2 $191,172 Is it “old” or “historic?” Is it “small” or “open floor plan?” $32,108 per bedroom $64,251 per acre Need a lower weight for “old” Apply initial abstractions Set values
  • 47.
    ● Vanilla NeuralNetwork — nothing fancy ● Convolutional Neural Network — inspired by visual cortex ● Deep Belief Network — undirected connections ● Recurrent Neural Network — multi-pass MANY DIFFERENT FLAVORS
  • 48.
    ● R ● Python ●Matlab/Octave ● Java ● C / C++ kdnuggets.com/2016/06/r-python-top-analytics-data-mining-data-science-software.html POPULAR LANGUAGES
  • 49.
    ● Amazon MachineLearning ● Google Prediction API* ● Google Cloud Machine Learning ● Microsoft Azure Machine Learning ● IBM Watson Machine Learning ● DiffBot ● Alibaba Cloud DT PAI SaaS OPTIONS
  • 50.
    ● TensorFlow * ●Amazon DSSTNE * ● H2O * ● PredictionIO ● Apache Mahout ● Scikit Learn ● Caffe * OPEN SOURCE OPTIONS ● Microsoft CNTK * ● Torch * ● Theano * ● MXnet * ● Chainer * ● Keras * ● Neon * * Deep learning
  • 51.
    How to getthe ball rolling. GET STARTED
  • 52.
    ● archive.ics.uci.edu/ml ● deeplearning.net/datasets ●mldata.org ● grouplens.org/datasets ● cs.toronto.edu/~kriz/cifar.html ● cs.cornell.edu/people/pabo/movie-review-data ● yann.lecun.com/exdb/mnist (handwriting) ● kdnuggets.com/datasets/index.html (long list) ● image-net.org (competition) OPEN SOURCE DATASETS
  • 53.
    ● playground.tensorflow.org (neuralnetwork demo) ● cs.stanford.edu/people/karpathy/convnetjs ● github.com/awslabs/machine-learning-samples ● ibm.com/smarterplanet/us/en/ibmwatson/devel opercloud/starter-kits.html ● templates.prediction.io EASY STARTING POINTS
  • 54.
    Start now ● It’shere, today ● It’s evolving exponentially ● Build “AI-First” RECAP
  • 55.
    Let’s see someaction. DEMOS
  • 56.
    ● AlchemyLanguage ● Dialog ●Natural Language Classifier ● Personality Insights ● Relationship Extraction ● Tradeoff Analytics ibm.com/smarterplanet/us/en/ibmwatson/developercloud/services-catalog.html IBM WATSON
  • 57.
    UNLEASH YOUR BUSINESS EMBRACEEXPONENTIAL 10xnation.com