Search result?Word?Who’s the best? ClicksGesturesGame results.

Infer.NETBuilding software with intelligenceJohn Winn and John GuiverMicrosoft Research, Cambridge, UK

VTL03

Intelligent Software

Search result?Word? Who’s the best?

ClicksGestures Game results

> It should be easier to write software that can adapt, learn and reason…

Reasoning Backwards

Word?

Gestures

> Need to reason backwards from things we can measure to things we can’t

Search result? Who’s the best?

Clicks Game results

Probability

• “Hello” 70%• “Halo” 20%• “Hall” 5%…

Gestures

> Can be multiple possible interpretations of some measurements

Code for Reasoning Backwards

100s-1000s of lines of code

Ordinary program

20-30 lines of ‘simulation’ code (in any .NET language)

Probabilistic program

with Infer.NET

Hours, not months!

Some Infer.NET Probabilistic Programs

Electronic health records

New understandin

g of the causes of asthma

...in Healthcare

...in Social Networks

Clinical simulation of

asthma

Network and

music tastes

Recommended songs

Simulation of sharing of musical tastes between

friends

Some Other Infer.NET Applications

> Program verification> Personalisation and recommendation> Data form entry checking> Gene expression analysis> Judgement calibration> Population modelling> Extracting plots of story books> … and many more

A Simple Probabilistic Program

> I toss two fair coins

> What is the probability both are heads?

C# 'Probabilistic' Program

T F T

F

T

T FT

F T FF

…

…

…

bool firstCoin = random.NextDouble()>0.5;

bool secondCoin = random.NextDouble()>0.5;

bool bothHeads = firstCoin & secondCoin;

After a Very Large Number of Runs…

T

F

~50%

~50%

T

F

~50%

~50%

T

F

~25%

~75%




Reasoning Backwards

> What is the probability the first coin was heads?

> Suppose I did not get two heads

Probabilistic Program

T F T

F

T

T FT

F T FF

…

…

…

We observe that bothHeads is F




Two Coins in C#

example

After a Very Large Number of Runs…

T

F

~33%

~67%

T

F

~33%

~67%

T

F

~0%

~100%




Multiple Runs Are Very Inefficient

> Is there a practical approach?

Infer.NET

> Illustrates how a prob. program works> But we want to reason about

complex situations with 1000s of variables e.g. observing 20 binary variables needs

~220 million runs

Random Variables in Infer.NET

50%T

50%T

25%F

var firstCoin = Variable.Bernoulli(0.5);

var secondCoin = Variable.Bernoulli(0.5);

var bothHeads = firstCoin & secondCoin;

Getting the Distribution of ‘bothHeads’

var engine = new InferenceEngine();

Bernoulli result = engine.Infer<Bernoulli>(bothHeads);

double probTrue = result.GetProbTrue();// ‘probTrue’ is now exactly 0.25

Adding an Observation

We observe that bothHeads is F

bothHeads.ObservedValue = false;

Bernoulli firstDist = engine.Infer<Bernoulli>(firstCoin);

double newProb = firstDist.GetProbTrue();// ‘newProb’ is now exactly 0.333…

Two Coins in Infer.NET

example

How Infer.NET Works

firstCoin

Bernoulli(0.5)

secondCoin

Bernoulli(0.5)

bothHeads

&

Normal executionBackwards messages

Observe F

Almost Done with the Coins!

> For ‘tossing a coin’ think:> Clicking on a link> Choosing a menu option> Buying a product

…

> Want to learn the probability of these events> Like having a biased coin

Biased Coins

10% 50% 90%Probability of heads (p)

FTFFT

TFTTTT

FFFFFT T

10% 50% 90%

Reasoning Backwards

FTFFT

TFTTTT

FFFFFT T

Beta distributions

Reasoning Backwards

…

T

F

…

// a flat Beta distributionvar p = Variable.Beta(1,1);

var toss1 = Variable.Bernoulli(p);toss1.ObservedValue = false;

var toss2 = Variable.Bernoulli(p);toss2.ObservedValue = true;

Beta result = engine.Infer<Beta>(p);// gives a Beta curve like the ones// on the last slide

Example : Search Log Analysis

The Click Log

1

2

3

4

T

F

F

T

Let’s look at the next result …… and see if it’s worth clicking

on

Let’s look at the page …… and see if it’s useful

Aaargh!It’s relevant!

Done!That looks promising …

… let’s click

Imagine One User and One Query

Click?

Examine

View

Next?

Relevant?

Next?

Next?

Y

N

N

N

N

N

Y

Y

Y

Y

var click = Variable.Bernoulli (appeal[d]);var next = Variable.Bernoulli (0.2); var doNext = Variable.Bernoulli (0.9); var isRel = Variable.Bernoulli (relevance[d]);

appeal

relevance

A Snippet of Infer.NET code

// Is user examining this item?examine[d] = examine[d - 1] & (((!click[d - 1]) & nextIfNotClick) | (click[d - 1] & nextIfClick));

// Flip the biased coins!click[d] = examine[d] & Variable.Bernoulli(appeal[d]);

isRelevant[d] = click[d] & Variable.Bernoulli(relevance[d]);

Reasoning Backwards

T

T

F

F

F

F

F

T

T

F

F

T

T

T

F

F

for (int d = 0; d < nRanks; d++) click[d].ObservedValue = user.clicks[d];

Click Analysis in Infer.NET

example

How Good Are You at Halo?Xbox Live

> 12 million players> 2 million matches per day> 2 billion hours of gameplay

The Challenge> Tracking how good each player is

to match players of similar skill.

TrueSkill™> Months of work, 100s of lines of

codeGamertag Score

Sully 25

SniperEye 22

DrSlowPlay 17

New Estimates of

Players’ Skills

Old Estimates of

Players’ Skills

Inferring Skills

0 10 20 30 40 50

Skill Level

Belie

f in

Skill

Level

1st Place

2nd Place

3rd Place

Game Outcome

DrSlowPlay

SniperEye

Sully

Probabilistic Program

// Gaussian random variables for skills var skill1 = Variable.Gaussian(oldMean1, oldStdDev1);var skill2 = Variable.Gaussian(oldMean2, oldStdDev2);var skill3 = Variable.Gaussian(oldMean3, oldStdDev3);// Players’ performances are centred around their skillsvar perf1 = Variable.Gaussian(skill1, beta);var perf2 = Variable.Gaussian(skill2, beta);var perf3 = Variable.Gaussian(skill3, beta);// OutcomesVariable.ConstrainPositive(perf1 – perf2);Variable.ConstrainPositive(perf2 - perf3);// Now we update the players’ skillsvar newSkill1 = engine.Infer<Gaussian>(skill1);var newSkill2 = engine.Infer<Gaussian>(skill2);var newSkill3 = engine.Infer<Gaussian>(skill3);

‘Language’ Elements of Infer.NET

Variable<bool>

Variable.If

Variable.Case Variable.IfNot

Variable.SwitchVariable.ForEach

var coin = Variable.Bernoulli(bias);

var bias = Variable.Beta(1,1);

var h = Variable.GaussianFromMeanAndPrecision(m, p);

var z = x + y;

var a = b > c;

Gaussian Dirichlet

Beta

BernoulliGamma

Poisson

Wishart

Discrete

>>FUTURE

Sometime in the Future?var firstCoin = Variable.Bernoulli(0.5);

var secondCoin = Variable.Bernoulli(0.5);

var bothHeads = c1 & c2;

bothHeads.ObservedValue = false;

var ie = new InferenceEngine();

Bernoulli result = ie.Infer<Bernoulli>(firstCoin);

Infer.NET API

Probabilistic

language?

http://research.microsoft.com/infernet


Thank you




YOUR FEEDBACK IS IMPORTANT TO US!

Please fill out session evaluation

forms online atMicrosoftPDC.com

Learn More On Channel 9

> Expand your PDC experience through Channel 9

> Explore videos, hands-on labs, sample code and demos through the new Channel 9 training courses

channel9.msdn.com/learnBuilt by Developers for Developers….

http://channel9.msdn.com/learn

© 2009 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

Search result?Word?Who’s the best? ClicksGesturesGame results.

Documents

bernoullirelevanced

clicksd slide

ttffttffttttffttffttffff

tf slide

variable variable

bernoulli relevanced

t fft slide

clickd variable