Power SVM: Generalization with Exemplar Classification Uncertainty

TitlePower SVM: Generalization with Exemplar Classification Uncertainty
Publication TypeConference Paper
Year of Publication2012
AuthorsZhang, W., Yu S. X., & Teng S-H.
Page(s)2144-2151
Other Numbers3316
Abstract

The human vision tends to recognize more variants ofa distinctive exemplar. This observation suggests that discriminativepower of training exemplars could be utilizedfor shaping a desirable global classifier that generalizesmaximally from a few exemplars. We propose to derive classificationuncertainty for each exemplar, using a local classificationtask to separate the exemplar from those in othercategories. We then design a global classifier by incorporatingthese uncertainties into constraints on the classifiermargins. We show through the dual form that the classificationcriterion can be interpreted as finding closest points betweenconvex hulls in the feature space augmented by classificationuncertainty. We call this scheme Power SVM (asin Power Diagram), since each exemplar is no longer a singularpoint in the feature space, but a super-point with itsown governing power in the classifier space. We test PowerSVM on digit recognition, indoor-outdoor categorization,and large-scale scene classification tasks. It shows consistentimprovement over SVM and uncertainty weighted SVM,especially when the number of training exemplars is small.

Acknowledgment

This work was partially supported by funding provided to Stella Yu through National Science Foundation Career Award IIS-0644204. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors or originators and do not necessarily reflect the views of the National Science Foundation.

URLhttp://www.icsi.berkeley.edu/pubs/vision/ICSI_powersvm12.pdf
Bibliographic Notes

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2012), Providence, Rhode Island, pp. 2144-2151

Abbreviated Authors

W. Zhang, S. X. Yu, and S.-H. Teng

ICSI Research Group

Vision

ICSI Publication Type

Article in conference proceedings