LCSL / GURLS

GURLS: a Least Squares Library for Supervised Learning
http://lcsl.mit.edu/#/downloads/gurls
63 stars 37 forks source link

SVM Subgradient Method #29

Open hanssusilo opened 8 years ago

hanssusilo commented 8 years ago

Wrote SVM Subgradient script with demo function Follows the kernel formulation Demo runs both RBF and Linear kernels to show off SVM Sub functionality and compares to dual RLS formulation. The other methods can be easily configured in the demo script.

Requires addition of two parameters for defopt: 1) 'Niter' - the total number of iterations of subgradient descent 2) 'gammafunc' - the function that maps the iteration number to the step size

Notes: Reuses the Dual RLS prediction code - might want to create a new one, if that suits convention better Can be adapted to use Matlab sparse matrices to take even better advantage of SVM sparsity.