en
0.25
0.5
0.75
1.25
1.5
1.75
2
Building Sparse Support Vector Machines for Multi-Instance Classification
Published on Oct 03, 20113057 Views
We propose a direct approach to learning sparse Support Vector Machine (SVM) prediction models for Multi-Instance (MI) classification. The proposed sparse SVM is based on a "label-mean" formulation of
Related categories
Chapter list
Building Sparse Support Vector Machines for Multi-Instance Classi cation00:00
Outline00:16
Problem De nition00:54
An Example of MI Classi cation02:01
Applications of MI Classi cation02:49
Prediction Function for Kernel SVM04:29
Why Sparsity is Important for MI Classi cation?05:54
"Label-Max" and Existing MI Formulations07:04
The "Label-Mean" Surrogate (1)08:06
The "Label-Mean" Surrogate (2)08:24
Connections with MI-Kernel08:28
Intuitions08:48
Prediction with "Label-Mean" (1)10:16
Prediction with "Label-Mean" (2)10:34
The Sparse-MI Problem10:47
Proposed Approach11:22
Optimization Scheme (1)11:53
Optimization Scheme (2)12:18
The Optimal Value Function12:19
Computation of derivatives12:31
Algorithm12:42
Remarks12:55
Multi-class Sparse MI Classi er13:42
Synthetic Data 1: Data Set13:58
Synthetic Data 1: Iteration 114:16
Synthetic Data 1: Iteration 1014:23
Synthetic Data 1: Function values over iterations14:27
Synthetic Data 2: Data set14:33
Synthetic Data 2: Iteration 114:40
Synthetic Data 2: Iteration 1014:42
Synthetic Data 2: Function values over iterations14:44
Data Set Descriptions14:47
Performance Comparison14:54
Convergence Results: COREL2015:27
Convergence Results: GENRE15:41
Conclusions15:41
Future Work16:05