third generation machine intelligence christopher m. bishop microsoft research, cambridge microsoft...

30
Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Post on 21-Dec-2015

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Third Generation Machine

IntelligenceChristopher M. BishopMicrosoft Research, Cambridge

Microsoft Research Summer School 2009

Page 2: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

First Generation

“Artificial Intelligence” (GOFAI)

Within a generation ... the problem of creating ‘artificial intelligence’ will largely be solved

Marvin Minsky (1967)

Expert Systems– rules devised by humans

Combinatorial explosion

General theme: hand-crafted rules

Page 3: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Second Generation

Neural networks, support vector machines

Difficult to incorporate complex domain knowledge

General theme: black-box statistical models

Page 4: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Third Generation

General theme: deep integration of domainknowledge and statistical learning

Probabilistic graphical models– Bayesian framework– fast inference using local message-passing

Origins: Bayesian networks, decision theory, HMMs, Kalman filters, MRFs, mean field theory, ...

Page 5: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Bayesian Learning

Consistent use of probability to quantify uncertainty

Predictions involve marginalisation, e.g.

Page 6: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Why is prior knowledge important?

y

x

?

Page 7: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Probabilistic Graphical Models

1. New insights into existing models

2. Framework for designing new models

3. Graph-based algorithms for calculation and computation (c.f. Feynman diagrams in physics)

4. Efficient software implementation

Directed graphs to specify the model

Factor graphs for inference and learning

Probability theory + graphs

Page 8: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Directed Graphs

Page 9: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Example: Time Series Modelling

Page 10: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009
Page 11: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Manchester Asthma and Allergies Study

Chris BishopIain BuchanMarkus SvensénVincent TanJohn Winn

Page 12: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009
Page 13: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Factor Graphs

Page 14: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

From Directed Graph to Factor Graph

Page 15: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Local message-passing

Efficient inference by exploiting factorization:

Page 16: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Factor Trees: Separation

v w x

f1(v,w) f2(w,x)

y

f3(x,y)

z

f4(x,z)

Page 17: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Messages: From Factors To Variables

w x

f2(w,x)

y

f3(x,y)

z

f4(x,z)

Page 18: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Messages: From Variables To Factors

x

f2(w,x)

y

f3(x,y)

z

f4(x,z)

Page 19: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

What if marginalisations are not tractable?

True distribution Monte Carlo Variational Bayes

Loopy belief propagation

Expectation propagation

Page 20: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Illustration: Bayesian Ranking

Ralf HerbrichTom MinkaThore Graepel

Page 21: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Two Player Match Outcome Model

y1

2

1 2

s1 s2

Page 22: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Two Team Match Outcome Model

y1

2

t1

t2

s2

s3

s1

s4

Page 23: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Multiple Team Match Outcome Model

s1

s2

s3

s4

t1

y1

2

t2

t3

y2

3

Page 24: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Efficient Approximate Inference

s1

s2

s3

s4

t1

y1

2

t2

t3

y2

3

Gaussian Prior Factors

Ranking Likelihood Factors

Page 25: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Convergence

0

5

10

15

20

25

30

35

40L

ev

el

0 100 200 300 400

Number of Games

char (Elo)

SQLWildman (Elo)

char (TrueSkill™)

SQLWildman (TrueSkill™)

Page 26: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

TrueSkillTM

Page 27: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

John WinnChris Bishop

Page 28: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

research.microsoft.com/infernet

Tom MinkaJohn WinnJohn GuiverAnitha Kannan

Page 29: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

Summary

New paradigm for machine intelligence built on:– a Bayesian formulation– probabilistic graphical models– fast inference using local message-passing

Deep integration of domain knowledge and statistical learning

Large-scale application: TrueSkillTM

Toolkit: Infer.NET

Page 30: Third Generation Machine Intelligence Christopher M. Bishop Microsoft Research, Cambridge Microsoft Research Summer School 2009

http://research.microsoft.com/~cmbishop