Version 3.0 release of the fast additive kernel classifiers NOTE: added a bug fix for sparse feature vectors. Code does NOT support sparse features, directly. Sparse features are implicitly converted to dense in the approximate model computation and prediction functions. If your features are very sparse, you might want to use an earlier release of the code. Contains: libsvm-mat-3.0-1 : latest version of libsvm at the time of release. Updates: 1. Support for learning models using the intersection, chisquared, and JS kernels 2. Fast approximate classification for these kernels using piecewise linear approximations. 3. ONLY piecewise linear approximations are supported. The binary search based exact classification for intersection kernel is no longer supported. 4. mex code for picewise linear interpolation, weighted kernel sampling and a fast binary search based version for the intersection kernel. Usage : Run the demo.m : It trains a libsvm model for the three kernels and compares the predictions using the exact classifier of LIBSVM and the picewise linear approximations using 100 uniform samples for each dimension. matlab functions: compute_approximate_classifier : computes the approximate classifier svmpredict_approx : performs predictions using piecewise linear approximation. Cite these papers if you use this code: Efficient classification for Additive Kernel SVMs, Subhransu Maji, Alexander C. Berg, Jitendra Malik, IEEE PAMI, 2012 Classification using Intersection Kernel SVMs is efficient, Subhransu Maji, Alexander C. Berg, Jitendra Malik, CVPR 2008. -- Subhransu (email me to report bugs.)