| |
RSC_MP reduced set selection
|
a = rss_mp(alg,hyper)
generates a rss object, using the matching pursuit selection method
hyperparameters:
child=svm algorithm worked on
max_loops=1e5 maximum number of basis functions
tolerance=1e-5 tolerated loss in ||w-w*||^2
backfit=1 backfit on every nth iteration
backfit_at_start=100 always backfit for first e.g. 100 iterations
dont_revisit=1 dont return to old basis function optimization, always get new one
reoptimize_b=1 recalculate the threshold b0
alpha_cutoff=0 throw away svs with abs(alpha) bal_w=0 treat multiple w's as equal by normalizing by length
optimizer='iterative' iterative update of matrix inverse
model:
alpha new alphas for rs-vectors
Xsv rs vectors
b0 the threshold
stats:
w2=0 final value of ||w-w*||^2
res=[] results on a separate test set
dtst=[] separate test set
test_on=0 iterations to test on
methods:
train constructs a reduced set, returns trained rs-machine
test tests new rs-machine on supplied data
example:
d=gen(toy2d('2circles','l=100'));
[r,a]=train(svm({kernel('rbf',1),'C=10000','alpha_cutoff=1e-2'}),d);
[r,a2]=train(rss_mp(a,'tolerance=1e-2'),d);
test(a2,d,loss)
author: goekhan bakir, jason weston
reference: fast binary and multi-output rss, 2004