Automatic Kernel Machine Learning
Penalized linear regression (Ridge, Lasso, ..) and most of the kernel methods (SVM, SVR, OCSVM) included multiple kernel learning can be formulated as a Quadratic programming (QP) optimization problem. However, standard QP solvers do not take into account the sparse nature of the solution. That is why, we have internally developed a dedicated solver called monQP for matlab. This solver has been ported to CUDA thanks to mex and cuBLAS. It includes an adapted version of the one rank update for the Cholesky factorization.