Practical Machine LearningDavid Jones

“Field of study that gives computers the ability to learn without being explicitly programmed”

Arthur Samuel, 1959

“Write a program to make this helicopter hover”

Pitch

Yaw

Roll

helicopter.rb

while helicopter.flying

if helicopter.pitch < 0 helicopter.pitchBy(0.1) else helicopter.pitchBy(-0.1) end

end

helicopter.rb

while helicopter.flying

if helicopter.pitch < 0 helicopter.pitchBy(0.1) else helicopter.pitchBy(-0.1) end

if helicopter.yaw < 0 helicopter.yawBy(0.1) else helicopter.yawBy(-0.1) end

end

helicopter.rbwhile helicopter.flying

if helicopter.pitch < 0 helicopter.pitchBy(0.1) else helicopter.pitchBy(-0.1) end

if helicopter.yaw < 0 helicopter.yawBy(0.1) else helicopter.yawBy(-0.1) end

if helicopter.roll < 0 helicopter.rollBy(0.1) else helicopter.rollBy(-0.1) end

end

OK, but what if…• it’s about to hit a tree?

• one of the main rotor blades is broken?

• power is running low?

• there is wind?

What if the helicopter was upside down?

helicopter.rbwhile helicopter.flying

if helicopter.pitch < 0 helicopter.pitchBy(0.1) else helicopter.pitchBy(-0.1) end

if helicopter.yaw < 0 helicopter.yawBy(0.1) else helicopter.yawBy(-0.1) end

if helicopter.roll < 0 helicopter.rollBy(0.1) else helicopter.rollBy(-0.1) end

end

Fail

Observe new exception case

Write code to handle exception

Helicopter Flying CodebaseHelicopter Flying Codebase

You will soon realise you can’t explicitly handle every

exception.

“Field of study that gives computers the ability to learn without being explicitly programmed”

Arthur Samuel, 1959

Autonomous RC HelicopterFlown using machine learning algorithms

That was 8 years ago…How good is machine learning today?

Germany wins

All 15 match outcomes predicted correctlyNo “luck” here.

Google SearchNetflix

Face DetectionSpam Detection

Medical Diagnosis AdvertisingFraud Detection

Product Recommendations

Siri

OCR

Priority Inbox

Dictation

Autonomous Cars

Video Games

Finance

Sentiment Analysis

So, how does it work?

Collect Data

Train Model

Make Predictions

Two distinct algorithm types• Supervised algorithms

• Unsupervised algorithms

Supervised

Supervised Learning

Trai

ning

Dat

a

estimate_sales_price.rbdef estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood) price = 0 # In my area, the average house costs $200 per sqft price_per_sqft = 200

if neighborhood == "hipsterton": # but some areas cost a bit more price_per_sqft = 400 elsif neighborhood == "skid row": # and some areas cost less price_per_sqft = 100 end

# start with a base price estimate based on how big the place is price = price_per_sqft * sqft

# now adjust our estimate based on the number of bedrooms if num_of_bedrooms == 0 # Studio apartments are cheap price = price - 20000 else # places with more bedrooms are usually # more valuable price = price + (num_of_bedrooms * 1000) end

priceend

estimate_sales_price_ml.rbdef estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood) do_some_maths(num_of_bedrooms, sqft, neighborhood)end

estimate_sales_price_ml.rbdef estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood) price = 0 # a little pinch of this price += num_of_bedrooms * .841231951398213 # and a big pinch of that price += sqft * 1231.1231231 # maybe a handful of this price += neighborhood * 2.3242341421 # and finally, just a little extra salt for good measure price += 201.23432095end

estimate_sales_price_ml.rbdef estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood) price = 0 # a little pinch of this price += num_of_bedrooms * 1.0 # and a big pinch of that price += sqft * 1.0 # maybe a handful of this price += neighborhood * 1.0 # and finally, just a little extra salt for good measure price += 1.0end

…500

Square Feet

Number of Bedrooms

estimate_sales_price_ml.rbdef estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood) price = 0 # a little pinch of this price += num_of_bedrooms * .841231951398213 # and a big pinch of that price += sqft * 1231.1231231 # maybe a handful of this price += neighborhood * 2.3242341421 # and finally, just a little extra salt for good measure price += 201.23432095end

$300,000

Unsupervised

“Computer, tell me what’s interesting about this data”

Trai

ning

Dat

a

Machine Learning>

Explicit Programming

x = sqr feety = price

Selecting Features

Force applied, weight, colour, wind, material, who threw it, day of week

Force applied, weight, colour, wind, material, who threw it, day of week

Practical Machine LearningHow do I use this as a developer?

Algorithm SelectionHow do I know what algorithm to use?

Algorithm ImplementationHow do I implement an algorithm? Don’t.

Algorithm PerformanceLarge amounts of training data changing in

realtime

HostingHow am I going to run special software

required to successfully use ML?

No Data?Start logging today.

ML for DevelopersSo you don’t need to get a PHD in maths

Prediction.IO• Open Source

• Deploy on your own servers or instantly on Amazon’s Cloud

• Cheap to run

• Developer friendly API

• Easy to use admin UI

Prediction.IO• Ignore the maths

• Helps you find the best algorithm for your problem

• Easily hosted and performant

• Uses scalable services such as MapReduce and Hadoop.

• You don’t need to know how to work this stuff though.

Prediction.IO• Specialises in two use cases

• recommendations

• similarity

• more being added…

Product ratingProduct viewsPurchases

Selecting Features

Selecting Features

Selecting Features

Selecting Features

Ruby SDK

A/B Test Results• 45% longer average session

• 22% increase in conversion rate

• 37% increase in average order value

• 71% increase in revenue

Machine Learning• Extremely powerful at solving complex

problems

• Increasingly important for developers to know about it

• Don’t need to know the maths to get the benefit

More InformationStanford Machine Learning https://www.coursera.org/course/ml

Bootstrapping Machine Learning http://www.louisdorard.com/machine-learning-book/

Machine Learning is Fun https://medium.com/@ageitgey/machine-learning-is-fun-80ea3ec3c471

Building The Smart Shophttp://info.resolvedigital.com/building-the-smart-spree-shop

David Jones@d_jones

Questions?