BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory...

17
BrainStorming 樊樊樊 2015-7-20

Transcript of BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory...

Page 1: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

BrainStorming

樊艳波 2015-7-20

Page 2: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Outline

• Several papers on icml15 & cvpr15

• PALM

• Information Theory Learning.

Page 3: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Multi-view Sparse Co-clustering via Proximal Alternating Linearized Minimization(CVPR15)

• Single view rank-one approximation.

Page 4: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Multi-view co-clustering

• Objective Function.

• Optimization: PALM.

Page 5: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Proximal Alternating Linearized Minimization for Nonconvex and Nonsmooth Problems(Bolte, Mathematical Programming14)

• General nonconvex nonsmooth function

• Example:

Page 6: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Coordinate descent methods

• Gauss-Seidel iteration scheme:

– Drawback: strict convexity assumption.

• Proximal regularization of the Gauss-Seidel scheme:

– Existing problems: Hard to optimize each step exactly.

Page 7: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Proposed PALM method

• Idea: linear expansion

Page 8: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Proposed PALM method

• Algorithm.

Page 9: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

• An example: Sparse NMF

• Generate

• Generate

Page 10: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Entropic Graph-based Posterior Regularization.MaxwellW Libbrecht(ICML15)

• Objective Function.

• R: regularization term over ,

• Existing ways : L0, L1, L2 ...

Page 11: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Posterior Regularization

• Posterior regularization:– Nearby variables have simple posterior distributions.

– More nature.

• Formulation.

Page 12: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Optimation & Simulations

• EM-like algorithm.

• has closed form.

Page 13: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Unsupervised Simultaneous Orthogonal Basis Clustering Feature Selection.(CVPR15)

• Regularized regression model:

• Proposed Model.

• B: latent cluster center.

• E: encoding matrix.each row of E has only one non-zero term.

Page 14: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.
Page 15: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Just for Insights

Page 16: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Optimal Graph Learning with Partial Tags and Multiple Features for Image and Video Annotation.(CVPR15)

• Extensions: out of sample annotation; noise label information.

Page 17: BrainStorming 樊艳波 2015-7-20. Outline Several papers on icml15 & cvpr15 PALM Information Theory Learning.

Information-Theoretic Dictionary Learning for Image Classification(TPAMI14)

• Traditional Dictionary learning.

• ITL based Dictionary learning.

– D^{0}: initialization by k-svd or others.

– Dictionary compactness:

– Dictionary discrimination:

– Dictionary representation:

• Optimization: gauss process, kernel density estimates...