Transcript PowerPoint 演示文稿

LABEL DISTRIBUTION LEARNING
AND ITS APPLICATIONS
Xin Geng （耿新）
Pattern Learning and Mining (PALM) Lab
（模式学习与挖掘实验室, http://palm.seu.edu.cn）
School of Computer Science and Engineering
Southeast University, Nanjing, China
（东南大学）
Learning with Ambiguity
Single-label
Learning
Multi-label
Learning
?
Less Ambiguity
Label Ambiguity
More Ambiguity
Label Ambiguity
• “What describes the instance?”
cloud
sky
water
building
Multi-label Learning
More Ambiguity?
• “How to describe the instance?”
some
cloud
mostly
sky
much
water
a bit of
building
How to learn?
• MLL
Thresholding
Positive labels
Not a good choice!
• Label Distribution Learning (LDL)
• Assign a real number to each label
• Importance
• Confidence
• Level
• ……
Keep more, learn more
MLL
LDL – Problem Formulation
Description Degree
A real number
WLOG
Complete label set
is assigned to the label
for the instance
Label Distribution
LDL – Problem Formulation
LDL – Algorithms
• Two Categories
• Conditional Probability Mass Function (Classification)
Model the mapping from the instance x to the label distribution d via a
conditional PMF
• Multivariate Support Vector Regression (Regression)
Model the mapping from the instance x to the label distribution d via a
multivariate support vector machine
Conditional Probability Mass Function
• Learning from Label Distribution
• Training set:
• Goal: learn a conditional mass function
label distributions similar to
K-L divergence
that can generate
given the instance
Conditional Probability Mass Function
• Directly minimizing the K-L divergence between predicted
and real LDs
• MaxEnt Model
Conditional Probability Mass Function
• IIS-LLD
[Geng, Yin, and Zhou, TPAMI’13]
[Geng, Smith-Miles, and Zhou, AAAI’10]
Conditional Probability Mass Function
• BFGS-LLD
[Geng and Ji, ICDMW’13]
Conditional Probability Mass Function
• CPNN
[Geng, Yin, and Zhou, TPAMI’13]
3
Multivariate Support Vector Regression
• Two issues
1. How to output a distribution composed by multiple components?
Multivariate Support Vector Regression (M-SVR)
[Fernandez et al., TSP’04]
2.
How to constrain each component of the distribution within the
range of a probability, i.e., [0, 1]?
Model the regression by a sigmoid function
• Solve the two problems simultaneously
• LDSVR [Geng and Hou, submitted to IJCAI’15]
Fit a sigmoid function to each component of the label distribution
simultaneously by a support vector machine
Multivariate Support Vector Regression
• Sigmoid model
• Target function of SVR
Loss Function
Multivariate Support Vector Regression
• The loss function
• Dimension by dimension
Insensitive Zone
Problem: Examples falling into the area ρ1
will be penalized once while those falling
into the area ρ2 will be penalized twice.
Multivariate Support Vector Regression
• The loss function
• Multivariate
Problem: Difficult to optimize and apply
the kernel trick
Insensitive Zone
Multivariate Support Vector Regression
• The loss function
• Measure the loss by calculating how far away from zi another point
z′i∈ Rc should move to get the same output with the ground truth
Multivariate Support Vector Regression
• The loss function
Replacing ui with u′i/4
Insensitive Zone
Age Estimation
[Geng, Yin, and Zhou, TPAMI’13]
[Geng, Smith-Miles, and Zhou, AAAI’10]
• Aging is a slow and gradual
progress
• The faces at close ages look
quite similar
• Can we use the neighboring
ages to relieve the ‘lack of
training samples’ problem?
Age Estimation
• Experiment
[Geng, Yin, and Zhou, TPAMI’13]
[Geng, Smith-Miles, and Zhou, AAAI’10]
Head Pose Estimation [Geng and Xia, CVPR’14]
Bivariate Label Distribution
Head Pose Estimation [Geng and Xia, CVPR’14]
• Experiment
Multilabel Ranking for Natural Scene Images
[Geng and Luo, CVPR’14]
• Multilabel Ranking
• A bipartition of the relevant (positive) and irrelevant (negative)
labels
• A proper ranking over relevant labels
• Multiple Rankers: Subjective Inconsistent “Ground Truth”
Multilabel Ranking for Natural Scene Images
[Geng and Luo, CVPR’14]
• Multilabel Ranking by Preference Distribution
Virtual labels as split point between relevant and irrelevant labels
Multilabel Ranking for Natural Scene Images
[Geng and Luo, CVPR’14]
• Experiment
Crowd Counting
[Wang, Zhang and Geng, Neurocomputing’15]
Crowd Counting
[Wang, Zhang and Geng, Neurocomputing’15]
Pre-release Prediction of Crowd Opinion
[Geng and Hou, submitted to IJCAI’15]
on Movies
Pre-release Metadata
Crowd Rating Distribution
Pre-release Prediction of Crowd Opinion
[Geng and Hou, submitted to IJCAI’15]
on Movies
• Experiment
Conclusion
• Label distribution learning
• More general framework than single-label and multi-label learning
• Deals with different importance of labels
• Matches certain problems better
• Needs special design
• It is useful when
• There is a natural measure of description degree
• There are multiple labeling sources for one instance
• The labels are correlated to each other
• ……
Interested?
Download the LDL Matlab package from
http://cse.seu.edu.cn/PersonalPage/xgeng/LDL
THANK YOU
http:// palm.seu.edu.cn