Proximal Methods for Sparse Hierarchical Dictionary Learning

advertisement
Proximal Methods for Sparse
Hierarchical Dictionary Learning
Rodolphe Jenatton, Julien Mairal,
Guillaume Obozinski, Francis Bach
Presented by Bo Chen, 2010, 6.11
Outline
•
•
•
•
1. Structured Sparsity
2. Dictionary Learning
3. Sparse Hierarchical Dictionary Learning
4. Experimental Results
Structured Sparsity
• Lasso (R. Tibshirani.,1996)
• Group Lasso (M. Yuan & Y. Lin, 2006)
• Tree-Guided Group Lasso (Kim & Xing, 2009)
Tree-Guided Structure Example
Multi-task:
Tree Regularization Definition:
Kim & Xing, 2009
Tree-Guided Structure Penalty
Introduce two parameters:
Rewrite the penalty term, if the number of tasks is 2. (K=2):
Generally:
Kim & Xing, 2009
In Detail
Kim & Xing, 2009
Some Definitions about Hierarchical Groups
Hierarchical Sparsity-Inducing Norms
Dictionary Learning
 If the structure information is introduced, the difference
between dictionary learning and group lasso:
1. Group Lasso is a regression problem. Each feature has its own physical
meaning. The structure information should be meaningful and correct.
Otherwise, the ‘structure’ will hurt the method.
2. In dictionary learning, the dictionary is unknown. So the structure information
will be a guide to help learn the structured dictionary.
Optimization
• Proximal Operator for Structure Norm
Fix the dictionary D, the objective function:
=
Transformed to a proximal problem:
Proximal operator with the structure penalty:
Learning the Dictionary
Updating D 5 times in each iteration,
Updating A,
Experiments : Natural Image Patches
• Use the learned dictionary from training set to impute the
missing values in testing samples. Each sample is a 8x8
patch.
• Training set: 50000; Testing set: 25000
• Test 21 balanced tree structures of depth 3 and 4. Also
set the number of the nodes in each layer.
Learned Hierarchical Dictionary
Experiments : Text Documents
Key points:
Visualization of NIPS proceedings
Documents: 1714
Words: 8274
Postings Classification
Training set: 1000; Testing set: 425; Documents: 1425; Words:13312
Goal: classify the postings from the two newsgroups, alt.atheism and talk.religion.misc.
Download