《Local Deep Kernel Learning for Efficient Non-linear SVM Prediction 2016微软 Cijo》.pdfVIP

Local Deep Kernel Learning for Efficient Non-linear SVM Prediction Cijo Jose, Prasoon Goyal, Parv Aggrwal {josevancijo,prasoongoyal13,parv92}@ Indian Institute of Technology Delhi, Hauz Khas, New Delhi, India 110 016 Manik Varma manik@ Microsoft Research India, Bangalore, India 560 001 Abstract Our objective is to ameliorate the situation by re- ducing the cost of non-linear SVM prediction while Our objective is to speed up non-linear SVM maintaining classification accuracy above an accept- prediction while maintaining classification able threshold. Speeding up SVM prediction is an im- accuracy above an acceptable limit. We gen- portant research problem which has been approached eralize Localized Multiple Kernel Learning from multiple perspectives including approximating so as to learn a tree-based primal feature the kernel function or matrix (B˘az˘avan et al., 2012; embedding which is high dimensional and Kar Karnick, 2012; Maji et al., 2013; Rahimi sparse. Primal based classification decouples Recht, 2007; 2008; Vedaldi Zisserman, 2011; 2012; prediction costs from the number of support Williams Seeger, 2001; Yang et al., 2012), reduc- vectors and our tree-structured features effi- ing the number of support vectors (Cossalter et al., ciently encode non-linearities while speeding 2011; Joachims Yu, 2009; Keerthi et al., 2006) and up prediction exponentially over the state-of