Max-Margin Latent Variable Models

Report
Max-Margin Latent Variable Models
M. Pawan Kumar
Max-Margin Latent Variable Models
M. Pawan Kumar
Kevin Miller, Rafi Witten,
Tim Tang, Danny Goodman,
Haithem Turki, Dan Preston,
Ben Packer
Daphne Koller
Dan Selsam, Andrej Karpathy
Log (Size)
Computer Vision Data
~ 2000
Segmentation
Information
Log (Size)
Computer Vision Data
~ 12000
~ 2000
Bounding Box
Information
Segmentation
Log (Size)
Computer Vision Data
> 14 M
~ 12000
~ 2000
Image-Level
Bounding Box
Segmentation
Information
“Car”
“Chair”
Computer Vision Data
Log (Size)
>6B
> 14 M
Noisy Label
~ 12000
~ 2000
Image-Level
Bounding Box
Segmentation
Information
Learn with missing information (latent variables)
Outline
• Two Types of Problems
• Latent SVM (Background)
• Self-Paced Learning
• Max-Margin Min-Entropy Models
• Discussion
Annotation Mismatch
Learn to classify an image
Image x
h
Annotation a = “Deer”
Mismatch between desired and available annotations
Exact value of latent variable is not “important”
Annotation Mismatch
Learn to classify a DNA sequence
Latent Variables h
Sequence x
Annotation a  {+1, -1}
Mismatch between desired and possible annotations
Exact value of latent variable is not “important”
Output Mismatch
Learn to segment an image
Image x
Output y
Output Mismatch
Learn to segment an image
(x, a)
Bird
(a, h)
Output Mismatch
Learn to segment an image
(x, a)
(a, h)
Cow
Mismatch between desired output and available annotations
Exact value of latent variable is important
Output Mismatch
Learn to classify actions
(x, y)
Output Mismatch
Learn to classify actions
x
ha = +1
hb
+
“jumping”
Output Mismatch
Learn to classify actions
hb
x
ha = -1
+
“jumping”
Mismatch between desired output and available annotations
Exact value of latent variable is important
Outline
• Two Types of Problems
• Latent SVM (Background)
• Self-Paced Learning
• Max-Margin Min-Entropy Models
• Discussion
Latent SVM
Andrews et al, 2001; Smola et al, 2005;
Felzenszwalb et al, 2008; Yu and Joachims, 2009
Image x
Features (x,a,h)
h
Parameters w
Annotation a = “Deer”
(a(w),h(w)) = maxa,h wT(x,a,h)
Parameter Learning
Score of
Best Completion of Ground-Truth
>
Score of
All Other Outputs
Parameter Learning
maxh wT(xi,ai,h)
>
wT(x,a,h)
Parameter Learning
min ||w||2 + CΣi ξi
maxh wT(xi,ai,h)
≥
wT(x,a,h)
Annotation Mismatch + Δ(ai,a) - ξi
Optimization
Update
hi* = argmaxh wT(xi,ai,h)
Update w by solving a convex problem
min ||w||2 + C∑i i
wT(xi,ai,hi*) - wT(xi,a,h)
≥ (ai, a) - i
Repeat until convergence
Outline
• Two Types of Problems
• Latent SVM (Background)
• Self-Paced Learning
• Max-Margin Min-Entropy Models
• Discussion
Self-Paced Learning
Kumar, Packer and Koller, NIPS 2010
1+1=2
Math is for
losers !!
1/3 + 1/6
= 1/2
eiπ+1 = 0
FAILURE … BAD LOCAL MINIMUM
Self-Paced Learning
Kumar, Packer and Koller, NIPS 2010
1+1=2
Euler was
a Genius!!
1/3 + 1/6
= 1/2
eiπ+1 = 0
SUCCESS … GOOD LOCAL MINIMUM
Optimization
Update
hi* = argmaxh wT(xi,ai,h)
Update w by solving a convex problem
vi  {0,1}
min ||w||2 + C∑i i vi - λ∑i vi
wT(xi,ai,hi*) - wT(xi,a,h)
≥ (ai, a) - i
λ  λμ
Repeat until convergence
Image Classification
Mammals Dataset
271 images, 6 classes
90/10 train/test split
5 folds
Image Classification
Kumar, Packer and Koller, NIPS 2010
4.75
4.7
4.65
4.6
4.55
4.5
4.45
4.4
17.5
17
16.5
16
CCCP
15.5
SPL
15
CCCP
SPL
14.5
Objective
Test Error
HOG-Based Model. Dalal and Triggs, 2005
Image Classification
PASCAL VOC 2007 Dataset
~ 5000 images
Car vs. Not-Car
50/50 train/test split
5 folds
Image Classification
Witten, Miller, Kumar, Packer and Koller, In Preparation
Objective
HOG + Dense SIFT + Dense Color SIFT
SPL+ – Different features choose different “easy” samples
Image Classification
Witten, Miller, Kumar, Packer and Koller, In Preparation
Mean Average Precision
HOG + Dense SIFT + Dense Color SIFT
SPL+ – Different features choose different “easy” samples
Motif Finding
UniProbe Dataset
~ 40,000 sequences
Binding vs. Not-Binding
50/50 train/test split
5 folds
Motif Finding
Kumar, Packer and Koller, NIPS 2010
140
120
100
80
60
40
20
0
CCCP
SPL
Objective
36
35
34
33
32
31
30
29
28
CCCP
SPL
Test Error
Motif + Markov Background Model. Yu and Joachims, 2009
Semantic Segmentation
VOC Segmentation 2009
Stanford Background
+
Train - 1274 images
Validation - 225 images
Test - 750 images
Train - 572 images
Validation - 53 images
Test - 90 images
Semantic Segmentation
VOC Detection 2009
ImageNet
+
Bounding Box Data
Train - 1564 images
Image-Level Data
Train - 1000 images
Semantic Segmentation
Kumar, Turki, Preston and Koller, ICCV 2011
30
29
28
27
26
25
24
23
22
SPL
SUP
CCCP
VOC Overlap
55.5
55
54.5
54
53.5
53
52.5
52
SPL
SUP
CCCP
SBD Overlap
SUP – Supervised Learning (Segmentation Data Only)
Region-based Model. Gould, Fulton and Koller, 2009
Action Classification
PASCAL VOC 2011
+
Bounding Box Data
Train – 3000 instances
Test – 3000 instances
Noisy Data
Train - 10000 images
Action Classification
Packer, Kumar, Tang and Koller, In Preparation
62.8
62.6
62.4
62.2
62
61.8
61.6
61.4
61.2
61
60.8
SPL
CCCP
SUP
Mean Average Precision
Poselet-based Model. Maji, Bourdev and Malik, 2011
Self-Paced Multiple Kernel Learning
Kumar, Packer and Koller, In Preparation
1+1=2
Integers
1/3 + 1/6
= 1/2
Rational
Numbers
eiπ+1
Imaginary
Numbers
=0
USE A FIXED MODEL
Self-Paced Multiple Kernel Learning
Kumar, Packer and Koller, In Preparation
1+1=2
Integers
1/3 + 1/6
= 1/2
Rational
Numbers
eiπ+1
=0
Imaginary
Numbers
ADAPT THE MODEL COMPLEXITY
Optimization
hi* = argmaxh wT(xi,ai,h)
Update
and c
Update w by solving a convex problem
^
vi  {0,1}
min ||w||2 + C∑i i vi - λ∑i vi
wT(xi,ai,hi*) - wT(xi,a,h)
≥ (ai, a) - i
Kij = (xi,ai,hi)T (xj,aj,hj)
λ  λμ
Repeat until convergence
K = Σk ck Kk
Image Classification
Mammals Dataset
271 images, 6 classes
90/10 train/test split
5 folds
Image Classification
Kumar, Packer and Koller, In Preparation
1
20
0.8
15
0.6
0.4
10
FIXED
FIXED
5
0.2
SPMKL
SPMKL
0
0
Objective
Test Error
HOG-Based Model. Dalal and Triggs, 2005
Motif Finding
UniProbe Dataset
~ 40,000 sequences
Binding vs. Not-Binding
50/50 train/test split
5 folds
Motif Finding
Kumar, Packer and Koller, NIPS 2010
78
77
76
75
74
73
72
71
70
69
11.5
11
10.5
10
FIXED
9.5
SPMKL
9
FIXED
SPMKL
8.5
Objective
Test Error
Motif + Markov Background Model. Yu and Joachims, 2009
Outline
• Two Types of Problems
• Latent SVM (Background)
• Self-Paced Learning
• Max-Margin Min-Entropy Models
• Discussion
MAP Inference
Pr(a,h|x) = exp( wT(x,a,h))
Z(x)
Pr(a1,h|x)
0.00
0.00
0.00
0.00
0.25
0.00
0.25
0.00
0.25
MAP Inference
Pr(a,h|x) = exp( wT(x,a,h))
Z(x)
mina,h – log (Pr(a,h|x))
Value of latent variable?
Pr(a1,h|x)
0.00
0.00
0.00
0.00
0.25
0.00
Pr(a2,h|x)
0.25
0.00
0.25
0.00
0.00
0.00
0.00
0.24
0.00
0.01
0.00
0.00
Min-Entropy Inference
mina – log (Pr(a|x))
+ Hα (Pr(h|a,x))
Renyi entropy of generalized distribution
Q(a; x, w) = Set of all {Pr(a,h|x)}
mina Hα(Q(a; x, w))
Max-Margin Min-Entropy Models
Miller, Kumar, Packer, Goodman and Koller, AISTATS 2012
min ||w||2 + C∑i i
Hα(Q(a; x, w))- Hα(Q(ai; x, w)) ≥ (ai, a) - i
i ≥ 0
Like latent SVM, minimizes (ai, ai(w))
In fact, when α = ∞...
Max-Margin Min-Entropy Models
Miller, Kumar, Packer, Goodman and Koller, AISTATS 2012
min ||w||2 + C∑i i
maxhwT(x,ai,h)-maxhwT(x,a,h) ≥ (ai, a) - i
i ≥ 0
Like latent SVM, minimizes (ai, ai(w))
In fact, when α = ∞... Latent SVM
Image Classification
Mammals Dataset
271 images, 6 classes
90/10 train/test split
5 folds
Image Classification
Miller, Kumar, Packer, Goodman and Koller, AISTATS 2012
HOG-Based Model. Dalal and Triggs, 2005
Image Classification
Miller, Kumar, Packer, Goodman and Koller, AISTATS 2012
HOG-Based Model. Dalal and Triggs, 2005
Image Classification
Miller, Kumar, Packer, Goodman and Koller, AISTATS 2012
HOG-Based Model. Dalal and Triggs, 2005
Motif Finding
UniProbe Dataset
~ 40,000 sequences
Binding vs. Not-Binding
50/50 train/test split
5 folds
Motif Finding
Miller, Kumar, Packer, Goodman and Koller, AISTATS 2012
Motif + Markov Background Model. Yu and Joachims, 2009
Outline
• Two Types of Problems
• Latent SVM (Background)
• Self-Paced Learning
• Max-Margin Min-Entropy Models
• Discussion
Very Large Datasets
• Initialize parameters using supervised data
• Impute latent variables (inference)
• Select easy samples (very efficient)
• Update parameters using incremental SVM
• Refine efficiently with proximal regularization
Output Mismatch
Σh Prθ(h|a,x) Δ(a,h,a(w),h(w)) + A(θ)
C. R. Rao’s Relative Quadratic Entropy
Minimize over w and θ
Output Mismatch
Σh Prθ(h|a,x) Δ(a,h,a(w),h(w)) + A(θ)
Prθ(h,a|x)
C. R. Rao’s Relative Quadratic Entropy
(a1,h)
Minimize over w
(a2,h)
Output Mismatch
Σh Prθ(h|a,x) Δ(a,h,a(w),h(w)) + A(θ)
Prθ(h,a|x)
C. R. Rao’s Relative Quadratic Entropy
(a1,h)
Minimize over w
(a2,h)
Output Mismatch
Σh Prθ(h|a,x) Δ(a,h,a(w),h(w)) + A(θ)
Prθ(h,a|x)
C. R. Rao’s Relative Quadratic Entropy
(a1,h)
Minimize over θ
(a2,h)
Output Mismatch
Σh Prθ(h|a,x) Δ(a,h,a(w),h(w)) + A(θ)
Prθ(h,a|x)
C. R. Rao’s Relative Quadratic Entropy
(a1,h)
Minimize over θ
(a2,h)
Output Mismatch
Σh Prθ(h|a,x) Δ(a,h,a(w),h(w)) + A(θ)
Prθ(h,a|x)
C. R. Rao’s Relative Quadratic Entropy
(a1,h)
Minimize over θ
(a2,h)
Questions?

similar documents