Separation of Lanthanides/ Lanthanides and Actinides
CVPR2010: Semi-supervised Learning in Vision: Part 2: Theory
1. Semi-Supervised Learning in Vision
Amir Saffari, Christian Leistner, Horst Bischof
Institute for Computer Graphics and Vision, Graz University of Technology
CVPR San Francisco, June 18, 2010
2. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Outline
1 Semi-Supervised Learning
2 Self-Training
3 Generative Models
4 Margin Assumption
5 Cluster and Manifold Assumption
6 Multi-View Learning
7 Large-Scale, Multi-Class SSL, and Online Learning
8 Related Topics
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
3. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Supervised and Semi-Supervised Learning
Supervised learning is all about finding mappings from input (feature)
space to output space:
f :X →Y
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
4. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Supervised and Semi-Supervised Learning
Supervised learning is all about finding mappings from input (feature)
space to output space:
f :X →Y
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
5. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Supervised and Semi-Supervised Learning
Supervised learning is all about finding mappings from input (feature)
space to output space:
f (x; θ ) ∈ F : X → Y
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
6. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Supervised and Semi-Supervised Learning
In Semi-supervised learning we wish to find mappings by using both
labeled and unlabeled data:
f (x; θ ) ∈ F : X → Y
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
7. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Why Should SSL Make Sense?
Learner
f (x; θ ) ∈ F : X → Y
Labeled data
Dl = {(x, y) ∈ X × Y }
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
8. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Why Should SSL Make Sense?
Learner
f (x; θ ) ∈ F : X → Y
Labeled data
Dl = {(x, y) ∈ X × Y }
Unlabeled data
Du = {x ∈ X }
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
9. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Why Should SSL Make Sense?
Learner
f (x; θ ) ∈ F : X → Y
Labeled data
Dl = {(x, y) ∈ X × Y }
Unlabeled data
Du = {x ∈ X }
Unsupervised learning: the goal is to recover the structure of the
data, eg. clusters, manifolds, low dimensional embeddings, density ...
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
10. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Why Should SSL Make Sense?
Learner
f (x; θ ) ∈ F : X → Y
Labeled data
Dl = {(x, y) ∈ X × Y }
Unlabeled data
Du = {x ∈ X }
Unsupervised learning: the goal is to recover the structure of the
data, eg. clusters, manifolds, low dimensional embeddings, density ...
Density
p(x)
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
11. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
Bayes rule:
p(y) p(x|y)
prior likelihood
p(y|x) =
p(x)
posterior
evidence
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
12. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
Bayes rule:
p(y) p(x|y)
prior likelihood
p(y|x) =
p(x)
posterior
evidence
MAP decision rule:
p(k) p(x|k)
y =arg max p(k |x) = arg max
ˆ =
k k p(x)
=arg max p(k) p(x|k)
k
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
13. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
When we expect that p(x) (structure of the data) is related to
the p(y|x), ie. they share parameters.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
14. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
When we expect that p(x) (structure of the data) is related to
the p(y|x), ie. they share parameters.
In other words, a better estimation of p(x) can improve the
estimation of p(y|x).
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
15. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
When we expect that p(x) (structure of the data) is related to
the p(y|x), ie. they share parameters.
In other words, a better estimation of p(x) can improve the
estimation of p(y|x).
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
16. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
When we expect that p(x) (structure of the data) is related to
the p(y|x), ie. they share parameters.
In other words, a better estimation of p(x) can improve the
estimation of p(y|x).
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
17. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
When unlabeled data is going to help?
When we expect that p(x) (structure of the data) is related to
the p(y|x), ie. they share parameters.
In other words, a better estimation of p(x) can improve the
estimation of p(y|x).
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
18. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Semi-Supervised Learning Assumptions
Semi-supervised learning often is effective, if the assumptions
regarding the relationship between the structure of the data p(x) and
the posterior p(y|x) are true for a given problem.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
19. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Computer Vision Problems: Object Recognition
Structure: similar images may contain similar objects.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
20. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Computer Vision Problems: Object Detection
Structure: similar image patches may contain similar objects. Very
close patches (over the 2D image neighborhood) may contain the
same object.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
21. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Computer Vision Problems: Object Tracking
Structure: similar image patches may contain similar objects. Very
close patches (over the 2D image neighborhood and time) may
contain the same object.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
22. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Computer Vision Problems: Object Segmentation
Structure: similar pixels may correspond to the same object. Very
close pixels (over the 2D image neighborhood and time) may belong
to the same object.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
23. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Different Semi-Supervised Learning Settings
Semi-supervised classification: f : X → Y , Y = {1, · · · , K }.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
24. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Different Semi-Supervised Learning Settings
Semi-supervised classification: f : X → Y , Y = {1, · · · , K }.
Semi-supervised regression: f : X → Y , Y = R.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
25. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Different Semi-Supervised Learning Settings
Semi-supervised classification: f : X → Y , Y = {1, · · · , K }.
Semi-supervised regression: f : X → Y , Y = R.
Semi-supervised clustering: constrainted clustering, clustering
with pair-wise must-link and cannot-link constraints.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
26. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Transductive and Inductive SSL
Transductive: find f : Dl ∪ Du → Y |Dl ∪Du | . Can not be used for
any future example which was not in the training set.
Inductive: find f : X → Y . Can be used for any future example,
beyond the training set.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
27. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Interactive Segmentation
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
28. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Other Uses of Unlabeled Data
Unsupervised preprocessing: normalization, standardization,
PCA, ICA, ...
Graz University of Technology
[Torralba ICCV 2009, Mobahi et al. ICML 2009, Ranzato NIPS 2007]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
29. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Other Uses of Unlabeled Data
Unsupervised preprocessing: normalization, standardization,
PCA, ICA, ...
Feature extraction: bag-of-words
Graz University of Technology
[Torralba ICCV 2009, Mobahi et al. ICML 2009, Ranzato NIPS 2007]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
30. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Other Uses of Unlabeled Data
Unsupervised preprocessing: normalization, standardization,
PCA, ICA, ...
Feature extraction: bag-of-words
Unsupervised feature learning: deep learning and sparse coding
Graz University of Technology
[Torralba ICCV 2009, Mobahi et al. ICML 2009, Ranzato NIPS 2007]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
31. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
The Simplest Approach to SSL
Self-training is a meta learning (wrapper) semi-supervised method.
Self-Training
Inputs: learning algorithm T, labeled set Dl , and unlabeled set
Du .
For n = 1 to N:
1 Train using the labeled set: f n = T (Dl ).
2 Use f n to classify the unlabeled set: cu = f n (Du ).
3 Create the set of m most confident examples from the unlabeled
set: C ⊂ Du .
4 Update the labeled set: Dl ← Dl ∪ {(x, f n (x ))| x ∈ C}.
5 Update the unlabeled set: Du ← Du C .
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
32. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: 1-NN Classifier
Graz University of Technology
[Zhu TPCL 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
33. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: 1-NN Classifier
Graz University of Technology
[Zhu TPCL 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
34. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with RF
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
35. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Self-Training with RF (Iteration 5)
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
36. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Self-Training with RF (Iteration 10)
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
37. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Generative Models
Joint Distribution
p(x, y|θ, π ) = p(x|y; θ ) p(y|π )
mixture component πy
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
38. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Generative Models
Joint Distribution
p(x, y|θ, π ) = p(x|y; θ ) p(y|π )
mixture component πy
Posterior
πy p(x|y; θ )
p(y|x; θ, π ) =
∑k πk p(x|k; θ )
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
39. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Expectation Maximization Approach
Log-likelihood
(θ, π; Dl , Du ) = ∑ log πy p(x|y; θ ) +λ ∑ log ∑ πk p(x|k; θ )
(x,y)∈Dl x ∈Du k
labeled data unlabeled data
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
40. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Expectation Maximization Approach
Log-likelihood
(θ, π; Dl , Du ) = ∑ log πy p(x|y; θ ) +λ ∑ log ∑ πk p(x|k; θ )
(x,y)∈Dl x ∈Du k
labeled data unlabeled data
Expectation Maximization (EM) can be used to estimate the model
parameters.
EM
Inputs: labeled set Dl , and unlabeled set Du , initial model
parameters θ0 , π0 .
For n = 1 to N:
1 E step: Estimate the posterior p (y | x; θn−1 , πn−1 ) for Du .
2 M step: Update the model parameters, given the posteriors for
unlabeled data: θn , πn = arg max (θ, π; Dl , Du ).
θ,π Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
41. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Two Gaussians
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
42. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Supervised GMM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
43. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Semi-Supervised EM with GMM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
44. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Assumptions
[Jaakkola et al.]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
45. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Assumptions
[Jaakkola et al.]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
46. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Linear Classifier
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
47. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Maximum Margin Linear Classifier
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
48. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
SVM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
49. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
S3VM and TSVM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
50. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
S3VM
SVM with soft margins: binary case y ∈ {−1, 1} without the bias
term
λ 1
min w 2 +
w 2 2 ∑ max(0, 1 − y w, x )
|Dl | (x,y)∈D
l
margin
hinge loss
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
51. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
S3VM
SVM with soft margins: binary case y ∈ {−1, 1} without the bias
term
λ 1
min w 2 +
w 2 2 ∑ max(0, 1 − y w, x )
|Dl | (x,y)∈D
l
margin
hinge loss
Prediction rule:
ˆ
y = sign( w, x )
Margin of an unlabeled data:
mu (w, x) = | w, x |
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
52. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
S3VM: Loss Functions
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
53. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
S3VM: Formulation
S3VM
λ 1
min
w 2
w 2
2 + ∑ max(0, 1 − y w, x ) +
|Dl | (x,y)∈D
l
margin
hinge loss
γ
+ ∑ max(0, 1 − | w, x |)
|Du | x∈Du
sym. hinge
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
54. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
S3VM: Formulation
S3VM
λ 1
min
w 2
w 2
2 + ∑ max(0, 1 − y w, x ) +
|Dl | (x,y)∈D
l
margin
hinge loss
γ
+ ∑ max(0, 1 − | w, x |)
|Du | x∈Du
sym. hinge
Balancing constraint:
1 1
∑ y = |Du | ∑ w, x |
|Dl | (x,y)∈D
lx∈Du
[Vapnik 1998, Joachims ICML 1999] Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
55. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Entropy Regularization
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
56. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Linear SVM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
57. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Linear TSVM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
58. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Linear TSVM
Graz University of Technology
[Zhu TPCL 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
59. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Assumption and Margin Maximization
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
60. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Kernels
Linear SVM:
f (x) = w, x
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
61. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Kernels
Linear SVM:
f (x) = w, x
Nonlinear SVM:
f (x ) = w, φ(x ) = ∑ yα x φ(x), φ(x )
(x,y)∈Dl
nonlinear mapping K (x,x )
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
62. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Kernels
Linear SVM:
f (x) = w, x
Nonlinear SVM:
f (x ) = w, φ(x ) = ∑ yα x φ(x), φ(x )
(x,y)∈Dl
nonlinear mapping K (x,x )
Cluster Kernel:
∑ p I(c(x) == c(x ))
Kc (x, x ) = K (x, x )
n
[Weston et al. NIPS 2003]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
63. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting
Boosting
M
f (x) = ∑ wm g(x; θm )
m =1
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
64. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting
Boosting
M
f (x) = ∑ wm g(x; θm )
m =1
Base Learner
g(x; θm ) ∈ G : X → Y
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
65. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting
Boosting
M
f (x) = ∑ wm g(x; θm )
m =1
Base Learner
g(x; θm ) ∈ G : X → Y
Boosting is a linear classifier over the space of base learners G :
f (x) = G (x; θ )w
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
66. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting: Learning with Functional Gradient Descent
M
1
f (x; β∗ ) = arg min ∑ (x, y; f ) ⇒ f (x; β∗ ) = ∑ wm g(x; θm )
|Xl | (x,y)∈X
∗ ∗
β l m =1
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
67. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting: Learning with Functional Gradient Descent
M
1
f (x; β∗ ) = arg min ∑ (x, y; f ) ⇒ f (x; β∗ ) = ∑ wm g(x; θm )
|Xl | (x,y)∈X
∗ ∗
β l m =1
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
68. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting with Prior
Priors
∀x ∈ Du , k ∈ Y : q(k |x)
1
min
w,θ |Dl |
∑ (x, y; w, θ ) +
(x,y)∈Dl
Supervised
γ
|Du | x∑ u
+ p (x, q; w, θ )
∈D
Prior
[Saffari et al. ECCV 2008, CVPR 2009, ECCV 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
69. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Prior
Graz University of Technology
[Saffari et al. CVPR 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
70. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Prior
Graz University of Technology
[Saffari et al. CVPR 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
71. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Cluster Prior
Graz University of Technology
[Saffari et al. CVPR 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
72. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: SSL Boosting with Cluster Prior
0.60
VOC2006
0.55
0.50
Class. Acc.
0.45
0.40
0.35
RMSB
SER
0.30 AML
SVM
TSVM
0.25
0.0 0.1 0.2 0.3 0.4 0.5
Labeled Samp. Ratio, r
Graz University of Technology
[Saffari et al. CVPR 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
73. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: SSL Boosting with Cluster Prior
10000
Computation Time
r =0.1
r =0.5
8000
Time (sec)
6000
4000
2000
0 RMSB SER TSVM
Graz University of Technology
[Saffari et al. CVPR 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
74. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Boosting
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
75. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Cluster Prior
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
76. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Manifold Assumption
[Hein and von Luxburg MLSS 2007]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
77. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Manifold Assumption
[Hein and von Luxburg MLSS 2007] Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
78. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Manifold Assumption
[Hein and von Luxburg MLSS 2007]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
79. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Label Propagation: Supervised
[Zhu TPCL 2009]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
80. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Label Propagation: Semi-Supervised
[Zhu TPCL 2009]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
81. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Label Propagation: Semi-Supervised
[Kveton et al. CVPR OLCV 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
82. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Label Propagation: Semi-Supervised
[Kveton et al. CVPR OLCV 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
83. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Mincut
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
84. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Mincut
Mincut energy:
min ∑
{y x } x∈D (x ,y )∈D
∑ s(x, x )(yx − y )2 + ∑ s(x, x )(yx − yx )2
u l x ∈Du
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
85. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Random Walks on Graph
Graz University of Technology
[Zhu TPCL 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
86. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Manifold Regularization: LapSVM
LapSVM
λ 1
min
w 2
w 2
2 + ∑ max(0, 1 − y w, x )+
|Dl | (x,y)∈D
l
γ
+
|Du |2 ∑ ∑ s(x, x )( w, x − w, x )2
x∈Du x ∈Dl ∪Du
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
87. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Manifold Regularization: LapSVM
LapSVM
λ 1
min
w 2
w 2
2 + ∑ max(0, 1 − y w, x )+
|Dl | (x,y)∈D
l
γ
+
|Du |2 ∑ ∑ s(x, x )( w, x − w, x )2
x∈Du x ∈Dl ∪Du
Graph Laplacian
λ 1
min
w 2
w 2
2 + ∑ max(0, 1 − y w, x )+
|Dl | (x,y)∈D
l
γ
+ f, Lf
|Du |2
where f is the response vector of the classifier to both labeled and
unlabeled data.
Graz University of Technology
[Belkin et al. JMLR 2006]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
88. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting with Priors and Manifolds
1
min
w,θ
∑ (x, y; w, θ ) +
|Dl | (x,y)∈D
l
Supervised
γ s(x, x )
+ ∑ λ p (x, q; w, θ ) +(1 − λ) ∑ z(x) m (x, x ; w, θ )
|Du | x∈Du x ∈Du
Prior x =x
Manifold
Graz University of Technology
Unsupervised
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
91. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Cluster Prior
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
92. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Cluster Prior and Manifolds
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
93. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Learning from Different Views
Data is represented by multiple views: x = [x1 | · · · |xV ] T
T T
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
94. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Learning from Different Views
Data is represented by multiple views: x = [x1 | · · · |xV ] T
T T
There is a classifier per view: F = { f v }V=1
v
Multi-View Learning
F ∗ = arg min ∑ (x, y; F ) + γ ∑ φ(x; F )
F (x,y)∈Dl x∈Du
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
95. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training
Co-Training
Inputs: learning algorithm T, labeled set Dl , and unlabeled set
Du .
For n = 1 to N:
1 1 2
1 Train using the labeled set: f n = T (Dl ) and f n = T (Dl ).2
2 Use f n1 and f 2 to classify the unlabeled set: c1 = f 1 (D 1 ) and
n u n u
c2 = f n (Du ).
u
2 2
3 Create the set of m most confident examples from each view of
the unlabeled set: C 1 ⊂ Du andC 2 ⊂ Du .
1 2
4 Update the labeled set:
Dl ← Dl ∪ {(x, f n (x))| x ∈ C 1 } ∪ {(x, f n (x))| x ∈ C 2 }.
1 2
5 Update the unlabeled set: Du ← Du (C 1 ∪ C 2 ).
[Blum and Mitchell COLT 1998]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
96. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training and Assumptions
We can represent the data in two views: x = [x1 , x2 ].
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
97. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training and Assumptions
We can represent the data in two views: x = [x1 , x2 ].
We can train a good classifier only using either x1 or x2 .
[Blum and Mitchell COLT 1998, Sindhwani et al. ICML 2005]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
98. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training and Assumptions
We can represent the data in two views: x = [x1 , x2 ].
We can train a good classifier only using either x1 or x2 .
x1 and x2 are conditionally independent given the class.
[Blum and Mitchell COLT 1998, Sindhwani et al. ICML 2005]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
99. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training and Assumptions
We can represent the data in two views: x = [x1 , x2 ].
We can train a good classifier only using either x1 or x2 .
x1 and x2 are conditionally independent given the class.
[Blum and Mitchell COLT 1998, Sindhwani et al. ICML 2005]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
100. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training: Visual and EEG Data
Graz University of Technology
[Kapoor et al. CVPR 2008]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
101. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training: Visual and EEG Data
[Kapoor et al. CVPR 2008]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
102. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Co-Training: Visual and EEG Data
Graz University of Technology
[Kapoor et al. CVPR 2008]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
103. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-View Boosting with Priors
Multi-View Priors
1
V − 1 s∑
qv (k |xv ) = ps (k |xs ), ∀v ∈ {1, · · · , V }, k ∈ Y
=v
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
104. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-View Boosting with Priors
Multi-View Priors
1
V − 1 s∑
qv (k |xv ) = ps (k |xs ), ∀v ∈ {1, · · · , V }, k ∈ Y
=v
Boosting with Priors:
1
arg min ∑ (x, y; w, θ ) +
|Dl | (x,y)∈D
w,θ l
Supervised
γ
+ ∑ p (x, q; w, θ )
|Du | x∈Du
Prior
[Saffari et al. ECCV 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
105. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Robust Loss Functions
q + =0.5
5 5
0-1 KL
Hinge SKL
4
Exponential 4
JS
Logit
Savage
3 3
(x,y;f)
(x,q;f)
2 2
1 1
0 4 2 0 2 4 0 4 2 0 2 4
m(x,y;f) f + (x)
q + =0.75
5
KL
SKL
4
JS
3
(x,q;f)
2
1
0 4 2 0 2 4
f + (x)
Graz University of Technology
[Saffari et al. ECCV 2010]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
106. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with RF
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
107. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Linear SVM
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
108. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Multi-View
Classifiers
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
109. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Example: Interactive Segmentation with Multi-View
Classifiers
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
110. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Large-Scale Applications
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
111. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Large-Scale Applications
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
112. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Large-Scale Applications
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
113. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Learning from Gigantic Image Collections
Original Objective
J (f) = (f − y)T A(f − y) + f T Lf
Objective with eigenvectors using f = Uα
J (α) = (Uα − y)T A(Uα − y) + α T Σα T
Graz University of Technology
[Fergus et al. NIPS 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
114. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Learning from Gigantic Image Collections
80 million tiny images
[Torralba PAMI 2008, Fergus et al. NIPS 2009]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
115. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Learning from Gigantic Image Collections
Graz University of Technology
[Fergus et al. NIPS 2009]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
116. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Deep Learning
Embedding loss
∑ ( f ( x ), y ) + γ ∑ ∑ s(x, x ) ( f (x), f (x ))
(x,y)∈Dl x∈Du x ∈D
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
117. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Deep Learning
Semantic Role Labeling:
“The cat eats the fish in the pond” →
“TheARG0 catARG0 eatsREL theARG1 fishARG1 inARGM-LOC
theARGM-LOC pondARGM-LOC ”
Trained with stochastic gradient descent on 1 million labeled data
and 631 million unlabeled data from Wikipedia.
Graz University of Technology
[Weston et al. ICML 2008]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
118. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Online Learning and SSL
More about online semi-supervised learning in the second part of the
tutorial.
[Goldberg et al. ECML 2009, Zeisl et al. CVPR 2010, Leistner et al. PR 2010, Saffari et al. ECCV 2010, Kveton CVPR
OLCV 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
119. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-Class Problems
Many interesting problems are multi-class.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
120. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-Class Problems
What is wrong with 1-vs-all?
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
121. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-Class Problems
What is wrong with 1-vs-all?
Computational problems.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
122. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-Class Problems
What is wrong with 1-vs-all?
Computational problems.
Calibration problems [B. Schoelkopf, A. Smola, 2002].
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
123. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multi-Class Problems
What is wrong with 1-vs-all?
Computational problems.
Calibration problems [B. Schoelkopf, A. Smola, 2002].
Artificial unbalanced binary problems.
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
124. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Boosting with Priors and Manifolds
Boosting with Priors and Manifolds is an inherently multi-class
algorithm.
Graz University of Technology
[Saffari et al. CVPR 2009, CVPR 2010, ECCV 2010]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
125. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Transfer Learning
[Pan and Yang TKDE 2009]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
126. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Transfer Learning
[Pan and Yang TKDE 2009]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
127. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Transfer Learning with Attributes
[Farhadi CVPR 2009, Lampert CVPR 2009] Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
128. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Transfer Learning with Attributes
Graz University of Technology
[Farhadi et al. CVPR 2010]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
129. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Transfer Learning with Attributes
[Rohbach et al. CVPR 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
130. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Learning from Latent (Hidden) Variables
Latent SVM
f w (x) = max w, φ(x, z)
z
λ 1
min
w 2
w 2
2 + ∑ max(0, 1 − y f w (x))
|Dl | (x,y)∈D
l
Graz University of Technology
[Felzenszwalb et al. CVPR 2008]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
131. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Indoor Scene Labeling
[Wang et al. ECCV 2010]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
132. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Latent Structural SVM
[Yu and Joachims 2009]
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
133. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Latent Structural SVM for Indoor Scene Labeling
Graz University of Technology
[Wang et al. ECCV 2010]
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision
134. SSL Self-Training Generative Models Margin Manifold Multi-View Large-Scale Related
Multiple Instance Learning
-
-
- -
+ -
-
+ -
- -
-
+
-
Graz University of Technology
Amir Saffari, Christian Leistner, Horst Bischof Semi-Supervised Learning in Vision