aaron hertzmann university of toronto siggraph 2004 tutorialhertzman/ibl2004/bayes2004.pdfcalculus,...
TRANSCRIPT
![Page 1: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/1.jpg)
Introduction to Bayesian Learning
Aaron HertzmannUniversity of Toronto
SIGGRAPH 2004 Tutorial
Evaluations: www.siggraph.org/courses_evaluation
![Page 2: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/2.jpg)
CG is maturing …
![Page 3: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/3.jpg)
… but it’s still hard to create
![Page 4: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/4.jpg)
… it’s hard to create in real-time
![Page 5: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/5.jpg)
Data-driven computer graphics
What if we can get models from the real world?
![Page 6: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/6.jpg)
Data-driven computer graphics
Three key problems:• Capture data (from video,
cameras, mocap, archives, …)• Build a higher-level model• Generate new data
Ideally, it should be automatic, flexible
![Page 7: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/7.jpg)
Example: Motion capture
mocap.cs.cmu.edu
![Page 8: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/8.jpg)
Example: character posing
![Page 9: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/9.jpg)
Example: shape modeling
[Blanz and Vetter 1999]
![Page 10: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/10.jpg)
Example: shape modeling
[Allen et al. 2003]
![Page 11: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/11.jpg)
Key problems
• How do you fit a model to data?– How do you choose weights and
thresholds?– How do you incorporate prior
knowledge?– How do you merge multiple sources
of information?– How do you model uncertainty?
Bayesian reasoning provides solutions
![Page 12: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/12.jpg)
Bayesian reasoning is …
Probability, statistics, data-fitting
![Page 13: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/13.jpg)
Bayesian reasoning is …
A theory of mind
![Page 14: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/14.jpg)
Bayesian reasoning is …
A theory of artificial intelligence
[Thrun et al.]
![Page 15: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/15.jpg)
Bayesian reasoning is …
A standard tool of computer vision
![Page 16: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/16.jpg)
and …
Applications in:• Data mining• Robotics• Signal processing• Bioinformatics• Text analysis (inc. spam filters)• and (increasingly) graphics!
![Page 17: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/17.jpg)
Outline for this course
3:45-4pm: Introduction4pm-4:45: Fundamentals
- From axioms to probability theory- Prediction and parameter estimation
4:45-5:15: Statistical shape models- Gaussian models and PCA
- Applications: facial modeling, mocap
5:15-5:30: Summary and questions
![Page 18: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/18.jpg)
More about the course
• Prerequisites– Linear algebra, multivariate
calculus, graphics, optimization• Unique features
– Start from first principles– Emphasis on graphics problems– Bayesian prediction– Take-home “principles”
![Page 19: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/19.jpg)
Bayesian vs. Frequentist
• Frequentist statistics– a.k.a. “orthodox statistics”– Probability = frequency of
occurrences in infinite # of trials– Arose from sciences with
populations– p-values, t-tests, ANOVA, etc.
• Bayesian vs. frequentist debates have been long and acrimonious
![Page 20: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/20.jpg)
Bayesian vs. Frequentist
“In academia, the Bayesian revolution is on the verge of becoming the majority viewpoint, which would have been unthinkable 10 years ago.”
- Bradley P. Carlin, professor of public health, University of MinnesotaNew York Times, Jan 20, 2004
![Page 21: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/21.jpg)
Bayesian vs. Frequentist
If necessary, please leave these assumptions behind (for today):
• “A probability is a frequency”• “Probability theory only applies
to large populations”• “Probability theory is arcane and
boring”
![Page 22: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/22.jpg)
Fundamentals
![Page 23: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/23.jpg)
What is reasoning?
• How do we infer properties of the world?
• How should computers do it?
![Page 24: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/24.jpg)
Aristotelian logic
• If A is true, then B is true• A is true• Therefore, B is true
A: My car was stolenB: My car isn’t where I left it
![Page 25: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/25.jpg)
Real-world is uncertain
Problems with pure logic:• Don’t have perfect information• Don’t really know the model• Model is non-deterministic
So let’s build a logic of uncertainty!
![Page 26: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/26.jpg)
Beliefs
Let B(A) = “belief A is true”B(¬A) = “belief A is false”
e.g., A = “my car was stolen”B(A) = “belief my car was stolen”
![Page 27: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/27.jpg)
Reasoning with beliefs
Cox Axioms [Cox 1946]1. Ordering exists
– e.g., B(A) > B(B) > B(C)
2. Negation function exists– B(¬A) = f(B(A))
3. Product function exists– B(A ∧ Y) = g(B(A|Y),B(Y))
This is all we need!
![Page 28: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/28.jpg)
The Cox Axioms uniquely define a complete system of reasoning:
This is probability theory!
![Page 29: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/29.jpg)
“Probability theory is nothing more than common sense reduced to calculation.”
- Pierre-Simon Laplace, 1814
Principle #1:
![Page 30: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/30.jpg)
Definitions
P(A) = “probability A is true”= B(A) = “belief A is true”
P(A) 2 [0…1]P(A) = 1 iff “A is true”P(A) = 0 iff “A is false”
P(A|B) = “prob. of A if we knew B”P(A, B) = “prob. A and B”
![Page 31: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/31.jpg)
Examples
A: “my car was stolen”B: “I can’t find my car”
P(A) = .1P(A) = .5
P(B | A) = .99P(A | B) = .3
![Page 32: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/32.jpg)
Sum rule:
P(A) + P(¬A) = 1
Basic rules
Example:A: “it will rain today”
p(A) = .9 p(¬A) = .1
![Page 33: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/33.jpg)
Sum rule:
∑i P(Ai) = 1
Basic rules
when exactly one of Ai must be true
![Page 34: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/34.jpg)
Product rule:
P(A,B) = P(A|B) P(B)= P(B|A) P(A)
Basic rules
![Page 35: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/35.jpg)
Basic rules
Conditioning
∑i P(Ai) = 1 ∑i P(Ai|B) = 1
P(A,B) = P(A|B) P(B)
P(A,B|C) = P(A|B,C) P(B|C)
Sum Rule
Product Rule
![Page 36: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/36.jpg)
Summary
Product ruleSum rule
All derivable from Cox axioms; must obey rules of common sense
Now we can derive new rules
P(A,B) = P(A|B) P(B)
∑i P(Ai) = 1
![Page 37: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/37.jpg)
Example
A = you eat a good meal tonightB = you go to a highly-recommended
restaurant¬B = you go to an unknown restaurant
Model: P(B) = .7, P(A|B) = .8, P(A|¬B) = .5
What is P(A)?
![Page 38: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/38.jpg)
Example, continued
Model: P(B) = .7, P(A|B) = .8, P(A|¬B) = .5
1 = P(B) + P(¬B)1 = P(B|A) + P(¬B|A)P(A) = P(B|A)P(A) + P(¬B|A)P(A)
= P(A,B) + P(A,¬B)= P(A|B)P(B) + P(A|¬B)P(¬B)= .8 .7 + .5 (1-.7) = .71
Sum rule
Conditioning
Product rule
Product rule
![Page 39: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/39.jpg)
Basic rules
Marginalizing
P(A) = ∑i P(A, Bi)for mutually-exclusive Bi
e.g., p(A) = p(A,B) + p(A, ¬B)
![Page 40: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/40.jpg)
Given a complete model, we can derive any other probability
Principle #2:
![Page 41: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/41.jpg)
Model: P(B) = .7, P(A|B) = .8, P(A|¬B) = .5
If we know A, what is P(B|A)?(“Inference”)
Inference
P(A,B) = P(A|B) P(B) = P(B|A) P(A)
P(B|A) =P(A|B) P(B)
P(A) = .8 .7 / .71 ˜ .79
Bayes’ Rule
![Page 42: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/42.jpg)
Inference
Bayes Rule
P(M|D) =P(D|M) P(M)
P(D)
Posterior
LikelihoodPrior
![Page 43: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/43.jpg)
Describe your model of the world, and then compute the
probabilities of the unknowns given the observations
Principle #3:
![Page 44: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/44.jpg)
Use Bayes’ Rule to infer unknown model variables from observed data
Principle #3a:
P(M|D) =P(D|M) P(M)
P(D)
LikelihoodPrior
Posterior
![Page 45: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/45.jpg)
Discrete variables
Probabilities over discrete variables
C 2 { Heads, Tails }
P(C=Heads) = .5
P(C=Heads) + P(C=Tails) = 1
![Page 46: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/46.jpg)
Continuous variables
Let x 2 RN
How do we describe beliefs over x?e.g., x is a face, joint angles, …
![Page 47: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/47.jpg)
Continuous variables
Probability Distribution Function (PDF)a.k.a. “marginal probability”
p(x) P(a · x · b) = sab p(x) dx
x
Notation: P(x) is probp(x) is PDF
![Page 48: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/48.jpg)
Continuous variables
Probability Distribution Function (PDF)Let x 2 Rp(x) can be any function s.t.
s-11 p(x) dx = 1
p(x) ¸ 0
Define P(a · x · b) = sab p(x) dx
![Page 49: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/49.jpg)
Uniform distribution
x » U(x0, x1)p(x) = 1/(x0 – x1) if x0 · x · x1
= 0 otherwise
x0 x1
p(x)
![Page 50: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/50.jpg)
Gaussian distributions
x » N(µ, σ2)p(x|µ,σ2) = exp(-(x-µ)2/2σ2) / p 2πσ2
µ
σ
![Page 51: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/51.jpg)
Why use Gaussians?
• Convenient analytic properties• Central Limit Theorem• Works well• Not for everything, but a good
building block• For more reasons, see
[Bishop 1995, Jaynes 2003]
![Page 52: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/52.jpg)
![Page 53: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/53.jpg)
Rules for continuous PDFs
Same intuitions and rules apply
“Sum rule”: s-11 p(x) dx = 1
Product rule: p(x,y) = p(x|y)p(x)Marginalizing: p(x) = s p(x,y)dy
… Bayes’ Rule, conditioning, etc.
![Page 54: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/54.jpg)
Multivariate distributions
Uniform: x » U(dom) Gaussian: x » N(µ, Σ)
![Page 55: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/55.jpg)
Inference
How do we reason about the world from observations?
Three important sets of variables:• observations• unknowns• auxiliary (“nuisance”) variables
Given the observations, what are the probabilities of the unknowns?
![Page 56: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/56.jpg)
Inference
Example: coin-flippingP(C = heads|θ) = θp(θ) = U(0,1)
Suppose we flip the coin 1000 times and get 750 heads. What is θ?
Intuitive answer: 750/1000 = 75%
![Page 57: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/57.jpg)
What is θ?
p(θ) = Uniform(0,1)P(Ci = h|θ) = θ, P(Ci = t|θ) = 1-θP(C1:N | θ) = ∏i P(Ci = h | θ)p(θ | C1:N) = P(C1:N| θ) p(θ)
P(C1:N)Bayes’ Rule
= ∏i P(Ci | θ) P(θ) / P(C1:N)/ θH (1-θ)T
H = 750, T = 250
![Page 58: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/58.jpg)
What is θ?
θ
p(θ | C1, … CN) / θ750 (1-θ)250
“Posterior distribution:” new beliefs about θ
![Page 59: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/59.jpg)
Bayesian prediction
What is the probability of another head?
P(C=h|C1:N) = s P(C=h,θ|C1:N) dθ= s P(C=h|θ, C1:N) P(θ | C1:N) dθ= (H+1)/(N+2)= 751 / 1002 = 74.95 %
Note: we never computed θ
![Page 60: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/60.jpg)
Parameter estimation
• What if we want an estimate of θ?• Maximum A Posteriori (MAP):
θ* = arg maxθ p(θ | C1, …, CN)= H / N = 750 / 1000 = 75%
![Page 61: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/61.jpg)
A problem
Suppose we flip the coin onceWhat is P(C2 = h | C1 = h)?
MAP estimate: θ* = H/N = 1This is absurd!Bayesian prediction:
P(C2 = h | C1 = h) = (H+1)/(N+2) = 2/3
![Page 62: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/62.jpg)
What went wrong?
p(θ | C1)p(θ | C1:N)
![Page 63: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/63.jpg)
Over-fitting
• A model that fits the data well but does not generalize
• Occurs when an estimate is obtained from a “spread-out” posterior
• Important to ask the right question: estimate CN+1, not θ
![Page 64: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/64.jpg)
Parameter estimation is not Bayesian. It leads to errors,
such as over-fitting.
Principle #4:
![Page 65: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/65.jpg)
p(x|D) = s p(x, θ | D) dθ
Advantages of estimation
Bayesian prediction is usually difficult and/or expensive
![Page 66: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/66.jpg)
Q: When is estimation safe?
A: When the posterior is “peaked”• The posterior “looks like” a spike• Generally, this means a lot more data
than parameters• But this is not a guarantee (e.g., fit a
line to 100 identical data points)• Practical answer: use error bars
(posterior variance)
![Page 67: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/67.jpg)
Parameter estimation is easier than prediction. It works well
when the posterior is “peaked.”
Principle #4a:
![Page 68: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/68.jpg)
Learning a Gaussian
?
µ, σ2{x2}
![Page 69: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/69.jpg)
Want: max p(x1:K|µ, σ2)= min –ln p(x1:K|µ, σ2)= ∑i (x-µ)2/2σ2 + K/2 ln 2π σ2
Closed-form solution:µ = ∑i xi / Nσ2 = ∑i (x - µ)2/N
Learning a Gaussian
p(x|µ,σ2) = exp(-(x-µ)2/2σ2) / p 2πσ2
p(x1:K|µ, σ2) = ∏ p(xi | µ, σ2)
![Page 70: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/70.jpg)
Stereology
[Jagnow et al. 2004 (this morning)]
Model:
PDF over solids
Problem: What is the PDF over solids?Can’t estimate individual solid shapes:
arg max p(θ, S | I) is underconstrained)
p(θ) p(S | θ) p(I|S)
![Page 71: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/71.jpg)
Stereology
PDF over solids
Marginalize out S:p(θ | I) = s p(θ, S | I) dS
can be maximized
p(θ) p(S | θ) p(I|S)
![Page 72: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/72.jpg)
When estimating variables, marginalize out as many
unknowns as possible.
Principle #4b:
Algorithms for this:•Expectation-Maximization (EM)•Variational learning
![Page 73: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/73.jpg)
Regression
![Page 74: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/74.jpg)
Regression
Curve fitting
?
![Page 75: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/75.jpg)
Linear regression
x
y y = a x + b + εε » N(0, σ2I)
Model:
Or:
p(y|x,a,b,σ2) =N(ax + b, σ2I)x
![Page 76: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/76.jpg)
Linear regression
p(y|x, a, b,σ2) = N(ax + b, σ2I) p(y1:K | x1:K, a, b, σ2) = ∏i p(yi | xi, a, b, σ2)Maximum likelihood:a*,b*,σ2* = arg max ∏i p(yi|xi,a,b,σ2)
= arg min –ln ∏i p(yi|x, a, b, σ2)Minimize:
∑i (yi-(axi+b))2/(2σ2) + K/2 ln 2 π σ2
Sum-of-squared differences: “Least-squares”
![Page 77: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/77.jpg)
Linear regression
Same idea in higher dimensionsy = Ax + µ + ε
x y
x
y
![Page 78: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/78.jpg)
Nonlinear regression
y = f(x;w) + εε » N(0, σ2I)
Model:
Or:
p(y|x,w,σ2) =N(f(x;w), σ2I)x
y
Curve parameters
![Page 79: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/79.jpg)
Typical curve models
Linef(x;w) = w0 x + w1
B-spline, Radial Basis Functionsf(x;w) = ∑i wi Bi(x)
Artificial neural networkf(x;w) = ∑i wi tanh(∑j wj x + w0)+w1
![Page 80: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/80.jpg)
Nonlinear regression
p(y|x, w, σ2) = N(f(x;w), σ2I) p(y1:K | x1:K, w, σ2) = ∏i p(yi | xi, a, b, σ2)Maximum likelihood:w*,σ2* = arg max ∏i p(yi|xi,a,b,σ2)
= arg min –ln ∏i p(yi|x, a, b, σ2)Minimize:
∑i (yi-f(xi;w))2/(2σ2) + K/2 ln 2 π σ2
Sum-of-squared differences: “Least-squares”
![Page 81: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/81.jpg)
Least-squares estimation is a special case of maximum
likelihood.
Principle #5:
![Page 82: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/82.jpg)
Because it is maximum likelihood, least-squares suffers
from overfitting.
Principle #5a:
![Page 83: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/83.jpg)
Overfitting
![Page 84: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/84.jpg)
Smoothness priors
Assumption: true curve is smooth
Bending energy:p(w|λ) ~ exp( -s kr fk2 / 2 λ2)
Weight decay:p(w|λ) ~ exp( -kwk2 / 2 λ2)
![Page 85: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/85.jpg)
MAP estimation:arg max p(w|y) = p(y | w) p(w)/p(y)=arg min –ln p(y|w) p(w) =∑i (yi – f(xi; w))2/(2σ2) + kwk2/2λ2 + K ln σ
Smoothness priors
Sum-of-squares differences Smoothness
![Page 86: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/86.jpg)
Underfitting
![Page 87: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/87.jpg)
Underfitting
![Page 88: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/88.jpg)
MAP estimation with smoothness priors leads to under-fitting.
Principle #5b:
![Page 89: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/89.jpg)
Applications in graphics
Two examples:
[Rose III et al. 2001]
[Grzeszczuk et al. 1998]
Shape interpolation Approximate physics
![Page 90: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/90.jpg)
Choices in fitting
• Smoothness, noise parameters• Choice of basis functions• Number of basis functions
Bayesian methods can make these choices automatically and effectively
![Page 91: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/91.jpg)
Learning smoothness
Given “good” data, solveλ* , σ2* = arg max p(λ, σ2 | w, x1:K, y1:K)
Closed-form solutionShape reconstruction
in vision [Szeliski 1989]
![Page 92: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/92.jpg)
Learning without shape
Q: Can we learn smoothness/noise without knowing the curve?
A: Yes.
![Page 93: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/93.jpg)
Learning without shape
λ* , σ2* = arg max p(λ, σ2 | x1:K, y1:K)(2 unknowns, K measurements)
p(λ, σ2 | x1:K, y1:K) = s p(λ, σ2, w | x1:K,y1:K) dw/ s p(x1:K,y1:K|w,σ2,λ)p(w|λ,σ2)dw
![Page 94: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/94.jpg)
Bayesian regression
don’t fit a single curve, but keep the uncertainty in the curve:
p(x | x1:N, y1:N)
x
![Page 95: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/95.jpg)
Bayesian regression
MAP/Least-squares(hand-tuned λ, σ2,basis functions)
Gaussian Process regression(learned parameters λ, σ2)
![Page 96: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/96.jpg)
Bayesian regression
![Page 97: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/97.jpg)
Bayes’ rule provide principle for learning (or marginalizing out)
all parameters.
Principle #6:
![Page 98: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/98.jpg)
Prediction variances
More info: D. MacKay’s Introduction to Gaussian Processes
![Page 99: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/99.jpg)
NIPS 2003 Feature Selection Challenge
• Competition between classification algorithm, including SVMs, nearest neighbors, GPs, etc.
• Winners: R. Neal and J. Zhang• Most powerful model they could
compute with (1000’s of parameters) and Bayesian prediction
• Very expensive computations
![Page 100: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/100.jpg)
Summary of “Principles”
1.Probability theory is common sense reduced to calculation.
2.Given a model, we can derive any probability
3.Describe a model of the world, and then compute the probabilities of the unknowns with Bayes’ Rule
![Page 101: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/101.jpg)
Summary of “Principles”
4. Parameter estimation leads to over-fitting when the posterior isn’t “peaked.” However, it is easier than Bayesian prediction.
5. Least-squares estimation is a special case of MAP, and can suffer from over- and under-fitting
6. You can learn (or marginalize out) all parameters.
![Page 102: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/102.jpg)
Statistical shape and appearance models with PCA
![Page 103: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/103.jpg)
Key vision problems
• Is there a face in this image?
• Who is it?• What is the 3D
shape and texture?
Turk and Pentland 1991
![Page 104: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/104.jpg)
Key vision problems
• Is there a person in this picture?• Who?• What is their 3D pose?
![Page 105: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/105.jpg)
Key graphics problems
• How can we easily create new bodies, shapes, and appearances?
• How can we edit images and videos?
![Page 106: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/106.jpg)
The difficulty
• Ill-posed problems– Need prior assumptions– Lots of work for an artist
![Page 107: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/107.jpg)
Outline
• Face modeling problem– Linear shape spaces– PCA– Probabilistic PCA
• Applications– face and body modeling
![Page 108: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/108.jpg)
Background: 2D models
• Eigenfaces– Sirovich and Kirby 1987, Turk and
Pentland 1991
• Active Appearance Models/Morphable models– Beier and Neely 1990– Cootes and Taylor 1998
![Page 109: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/109.jpg)
Face representation
• 70,000 vertices with (x, y, z, r, g, b)• Correspondence precomputed
[Blanz and Vetter 1999]
![Page 110: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/110.jpg)
Data representation
yi = [x1, y1, z1, …, x70,000, y70,000, z70,000]T
Linear blends:
ynew = (y1 + y2) / 2
a.k.a. blendshapes, morphing
.5 . + .5. =y1
y2 y3
![Page 111: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/111.jpg)
Linear subspace model
y = ∑i wi yi (s.t., ∑i wi = 1)= ∑i xi ai + µ= A x + µ
Problem: can we learn this linear space?
x y
![Page 112: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/112.jpg)
Principal Components Analysis (PCA)
Same model as linear regressionUnknown x
x y
x
y
![Page 113: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/113.jpg)
Conventional PCA(Bayesian formulation)
x, A, µ » Uniform, AT A = I ε » N(0, σ2 I)y = A x + µ + εGiven training y1:K, what are A, x, µ, σ2?
Maximum likelihood reduces to:∑i k yi – (A xi + µ) k2 / 2σ2 + K/2 ln 2 π σ2
Closed-form solution exists
![Page 114: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/114.jpg)
PCA with missing data
x y
Linear constraint
ML point
Problems:•Estimated point far from data if data is noisy•High-dimensional y is a uniform distribution•Low-dimensional x is overconstrainedWhy? Because x » U
![Page 115: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/115.jpg)
Probabilistic PCA
x y
x » N(0,I)y = Ax + b + ε
[Roweis 1998, Tipping and Bishop 1998]
![Page 116: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/116.jpg)
Fitting a Gaussian
y » N(µ, Σ)easy to learn, and nice properties… but Σ is a 70,0002 matrix
y
![Page 117: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/117.jpg)
PPCA vs. Gaussians
However…PPCA: p(y) = s p(x,y) dx
= N(b, A AT + σ2 I)This is a special case of a Gaussian!PCA is a degenerate case (σ2=0)
![Page 118: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/118.jpg)
Face estimation in an image
p(y) = N(µ, Σ)p(Image | y) = N(Is(y), σ2 I)
-ln p(S,T | Image) = kImage – Is(y)k2 /2σ2 + (y-µ)TΣ-1(y-µ)/2
Image fitting term Face likelihood
Use PCA coordinates for efficiencyEfficient editing in PCA space
y
Image
[Blanz and Vetter 1999]
![Page 119: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/119.jpg)
![Page 120: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/120.jpg)
Comparison
PCA: unconstrained latent space –not good for missing data
Gaussians: general model, but impractical for large data
PPCA: constrained Gaussian – best of both worlds
![Page 121: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/121.jpg)
Estimating a face from video
[Blanz et al. 2003]
![Page 122: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/122.jpg)
The space of all body shapes
[Allen et al. 2003]
![Page 123: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/123.jpg)
The space of all body shapes
[Allen et al. 2004]
![Page 124: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/124.jpg)
Non-rigid 3D modeling from video
What if we don’t have training data?
[Torresani and Hertzmann 2004]
![Page 125: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/125.jpg)
Non-rigid 3D modeling from video
• Approach: learn all parameters– shape and motion– shape PDF– noise and outliers
• Lots of missing data (depths)– PPCA is essential
• Same basic framework, more unknowns
![Page 126: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/126.jpg)
Results
Lucas-Kanade tracking
Tracking result
3D reconstruction
Reference frame
![Page 127: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/127.jpg)
Results
Robust algorithm
3D reconstruction
[Almodovar 2002]
![Page 128: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/128.jpg)
Inverse kinematics
Constraints
DOFs (y)
[Grochow et al. 2004 (tomorrow)]
![Page 129: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/129.jpg)
Problems with Gaussians/PCA
Space of poses may is nonlinear, non-Gaussian
x y
![Page 130: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/130.jpg)
Non-linear dimension reduction
y = f(x;w) + εLike non-linear regression w/o x
x y
f(x;w)
NLDR for BRDFs: [Matusik et al. 2003]
![Page 131: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/131.jpg)
Problem with Gaussians/PPCA
![Page 132: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/132.jpg)
Style-based IK
f(x;w)
Walk cycle:
Details: [Grochow 2004 (tomorrow)]
y
![Page 133: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/133.jpg)
Discussion and frontiers
![Page 134: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/134.jpg)
Designing learning algorithms for graphics
Write a generative modelp(data | model)
Use Bayes’ rule to learn the model from data
Generate new data from the model and constraints
(numerical methods may be required)
![Page 135: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/135.jpg)
What model do we use?
• Intuition, experience, experimentation, rules-of-thumb
• Put as much domain knowledge in as possible– model 3D shapes rather than pixels– joint angles instead of 3D positions
• Gaussians for simple cases; nonlinear models for complex cases (active research area)
![Page 136: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/136.jpg)
Q: Are there any limits to the power of Bayes' Rule?
A: According to legend, one who fully grasped Bayes' Rule would gain the ability to create and physically enter an alternate universe using only off-the-shelf equipment. One who fully grasps Bayes' Rule, yet remains in our universe to aid others, is known as a Bayesattva.
http://yudkowsky.net/bayes/bayes.html:
![Page 137: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/137.jpg)
Problems with Bayesian methods
1. The best solution is usually intractable
• often requires expensive numerical computation
• it’s still better to understand the real problem, and the approximations
• need to choose approximations carefully
![Page 138: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/138.jpg)
Problems with Bayesian methods
2. Some complicated math to do• Models are simple, algorithms
complicated• May still be worth it• Bayesian toolboxes on the way
(e.g., VIBES, Intel OpenPNL)
![Page 139: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/139.jpg)
Problems with Bayesian methods
3. Complex models sometimes impede creativity
• Sometimes it’s easier to tune• Hack first, be principled later• Probabilistic models give insight
that helps with hacking
![Page 140: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/140.jpg)
Benefits of the Bayesian approach
1. Principled modeling of noise and uncertainty
2. Unified model for learning and synthesis
3. Learn all parameters4. Good results from simple models5. Lots of good research and
algorithms
![Page 141: Aaron Hertzmann University of Toronto SIGGRAPH 2004 Tutorialhertzman/ibl2004/bayes2004.pdfcalculus, graphics, optimization ... dx x Notation: P(x) is prob p(x) is PDF. Continuous variables](https://reader033.vdocuments.mx/reader033/viewer/2022043013/5f5156d8e5f918157102c072/html5/thumbnails/141.jpg)
Course notes, slides, links:http://www.dgp.toronto.edu/~hertzman/ibl2004
Course evaluationhttp://www.siggraph.org/courses_evaluation
Thank you!