Kernel Perceptron with Budget by Cramer et al.
A=BUDGET_PERCEPTRON(H) returns a budget_perceptron object initialized with hyperparameters H.
The budget_perceptron object trains a potentially kernelized perceptron.
Hyperparameters (with defaults)
cache_size= -- if set to a scalar, this is the maximum cache size used
after this the algorithm throws away "support vectors"
if left empty, an adaptive cache is used
max_loops=1 -- Maximum number of sweeps through the data
margin=0.01 -- potential margin with which to train on
alpha_cutoff=-1 -- keep alphas with abs(a_i)>max(a)/alpha_cutoff
default keeps all alphas, another
reasonable choice is e.g alpha_cutoff=1e5 to remove
zero alphas (i.e non-SVs) to speed up computations.
child=kernel -- the kernel is stored as a member called "child"
alpha -- the weights
Xsv -- the "Support Vectors"
Reference : Online Classification on a Budget |
Author : Crammer, Kandola and Singer |