Design and Analysis of Local Kernel Machines

Ravi Sastry (Mentor Alexander G. Gray) - Local kernel machines work by finding locally linear decision surfaces in the input space. They are built upon the fact that any differentiable function can be locally expressed as a linear function upto a small remainder. Hence by exploiting the smoothness of the underlying decision surface, one can build a globally non-linear decision surface by finding locally linear decision surfaces. This allows us to decompose a single global quadratic optimization problem in the kernel space into many small quadratic optimization problems in the input space..