SPIDER The Spider Objects

RSC_MP reduced set selectiona = 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