Abstract
We study the problem of structured output prediction. Some methods such as structured output support vector machines (SSVM) and conditional random fields (CRFs) are state-of-the-art in dealing with the structured data. However, these classical methods have some limits in scalability because of high memory requirements and slow training speed. Recently, the method joint kernel support estimation (JKSE) has been proposed based on one-class SVM which can be trained efficiently. However, JKSE is not as powerful as those classical methods from the point of prediction performance. To improve the performance of JKSE, we introduce privileged information into it. Learning using privileged information (LUPI) is an advanced machine learning paradigm by taking advantage of some elements of human teaching that are only available at the training stage, not at testing. Motivated by the LUPI, we propose three new models based on JKSE by considering three forms of privileged information. The resulting optimization problems are convex quadratic programming that can be easily solved. Our new models not only persist the advantage of JKSE but also improve its performance. The experimental results show the superiority of new models over the JKSE when solving object detection and multi-class classification problems.
Highlights
This paper deals with the structured learning problems which learn function: f : X → Y, where the elements of X and Y are structured objects such as sequences, trees, bounding boxes, strings
Conditional random fields [4], [5], maximum margin markov networks [6] and structured output support vector machines(SSVM) [7] and their extensions have been developed as powerful tools to predict the structured data
1) We first incorporate the Learning using privileged information (LUPI) paradigm into joint kernel support estimation (JKSE) and propose three new models: JKSE+, JKSE+part and JKSE+different which deal with three cases of privileged information respectively; 2) For multi-class problems, we propose the taxonomy-based JKSE+ approach that depends on the joint feature map
Summary
This paper deals with the structured learning problems which learn function: f : X → Y, where the elements of X and Y are structured objects such as sequences, trees, bounding boxes, strings. It is difficult or infeasible to solve large scale problems except for some special output structures To overcome these drawbacks, a method called Joint Kernel Support Estimation(JKSE) has been proposed in [8]. In this paper, motivated by the LUPI paradigm, we propose three novel algorithms base on JKSE by considering three specific forms of privileged information. We conduct some experiments on object detection and multiclass classification problems to show the superiority of our new methods over JKSE. 1) We first incorporate the LUPI paradigm into JKSE and propose three new models: JKSE+, JKSE+part and JKSE+different which deal with three cases of privileged information respectively; 2) For multi-class problems, we propose the taxonomy-based JKSE+ approach that depends on the joint feature map.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
