Second Order Optimization of Kernel Parameters

We investigate the use of second order optimization approaches for solving the multiple kernel learning (MKL) problem. We show that the hessian of the MKL can be computed efficiently and this information can be used to compute a better descent direction than the gradient (used in the state-of-the-art SimpleMKL algorithm). We then empirically show that our new approaches outperforms SimpleMKL in terms of computational efficiency.