?url_ver=Z39.88-2004&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Adc&rft.title=+Overlearning+in+marginal+distribution-based+ICA%3A+analysis+and+solutions&rft.creator=S%C3%A4rel%C3%A4%2C+Mr+Jaakko&rft.subject=Statistical+Models&rft.subject=Machine+Learning&rft.subject=Neural+Nets&rft.description=++The+present+paper+is+written+as+a+word+of+caution%2C+with+users+of%0A++independent+component+analysis+(ICA)+in+mind%2C+to+overlearning%0A++phenomena+that+are+often+observed.%5C%5C%0A++We+consider+two+types+of+overlearning%2C+typical+to+high-order%0A++statistics+based+ICA.++These+algorithms+can+be+seen+to+maximise+the%0A++negentropy+of+the+source+estimates.++The+first+kind+of+overlearning%0A++results+in+the+generation+of+spike-like+signals%2C+if+there+are+not%0A++enough+samples+in+the+data+or+there+is+a+considerable+amount+of%0A++noise+present.++It+is+argued+that%2C+if+the+data+has+power+spectrum%0A++characterised+by+%241%2Ff%24+curve%2C+we+face+a+more+severe+problem%2C+which%0A++cannot+be+solved+inside+the+strict+ICA+model.+This+overlearning+is%0A++better+characterised+by+bumps+instead+of+spikes.+Both+overlearning%0A++types+are+demonstrated+in+the+case+of+artificial+signals+as+well+as%0A++magnetoencephalograms+(MEG).+Several+methods+are+suggested+to%0A++circumvent+both+types%2C+either+by+making+the+estimation+of+the+ICA%0A++model+more+robust+or+by+including+further+modelling+of+the+data.%0A&rft.publisher=MIT+press&rft.contributor=Lee%2C+prof.+Te-Won&rft.contributor=Cardoso%2C+prof.+Jean-Francois&rft.contributor=Oja%2C+prof.+Erkki&rft.contributor=Amari%2C+prof.+Shun-Ichi&rft.date=2003-12&rft.type=Journal+(Paginated)&rft.type=PeerReviewed&rft.format=application%2Fpdf&rft.identifier=http%3A%2F%2Fcogprints.org%2F3638%2F1%2Fsarela03.pdf&rft.identifier=++S%C3%A4rel%C3%A4%2C+Mr+Jaakko++(2003)++Overlearning+in+marginal+distribution-based+ICA%3A+analysis+and+solutions.++%5BJournal+(Paginated)%5D+++++&rft.relation=http%3A%2F%2Fcogprints.org%2F3638%2F