adapt transfer learn discrimin learn method wide natur languag process method work train test data drawn distribut nlp task confront domain label data scarc exist case seek adapt exist model resourcerich sourc domain resourc poor target domain introduc structur correspond learn automat induc correspond featur domain test techniqu part speech tag show perform gain vari amount sourc target train data improv target domain pars accuraci improv tagger gaussian process latent random field transduct learn adapt graph multi dtask warp gaussian process person ag estim automat ag estim facial imag arous research interest recent year due promis potenti comput vision applic method propos date person ag estim method gener outperform global ag estim method learn separ ag estim person train data set typic ag databas limit train data person train separ ag estim train data person run high risk overfit data predict perform limit paper propos approach ag estim formul problem multi task learn problem base variant gaussian process gp call warp gaussian process wgp propos multi task extens call multi task warp gaussian process mtwgp ag estim formul multi task regress problem learn task refer estim ag function person mtwgp model common featur share task person task specif person specif featur learn automat unlik previou ag estim method form regress function determin paramet function ineffici method cross valid form regress function mtwgp implicitli defin kernel function model paramet learn data automat conduct experi publicli ag databas fg net morph experiment result promis show mtwgp compar favor state ofth art ag estim method sed supervis experiment design applic text classif recent year activ learn method base experiment design achiev state art perform text classif applic method exploit distribut unlabel data support batch select make label data carri inform activ learn paper propos activ learn method text classif call supervis experiment design sed seamlessli incorpor label inform experiment design experiment result show sed outperform counterpart discard label inform fail exploit distribut unlabel data transfer metric learn learn task relationship distanc metric learn plai crucial role data mine algorithm perform algorithm reli heavili choos good metric label data applic scarc metric learn unsatisfactori paper transfer learn set relat sourc task label data learn target task propos convex formul multi task metric learn model task relationship form task covari matrix regard transfer learn special case multitask learn adapt formul multi task metric learn transfer learn set method call transfer metric learn tml tml learn metric task covari sourc task target task unifi convex formul solv convex optim problem altern method subproblem effici solut experiment result commonli transfer learn applic demonstr effect method sed supervis experiment design applic text classif recent year activ learn method base experiment design achiev state art perform text classif applic method exploit distribut unlabel data support batch select make label data carri inform activ learn paper propos activ learn method text classif call supervis experiment design sed seamlessli incorpor label inform experiment design experiment result show sed outperform counterpart discard label inform fail exploit distribut unlabel data mild multipl dinstanc learn disambigu multipl instanc learn mil individu call instanc bag singl multipl instanc class label train set bag instanc bag label posit instanc posit bag label neg posit bag neg instanc addit posit instanc true label instanc posit bag bag label instanc label inher ambigu paper propos effici robust mil method call mild multipl instanc learn disambigu gener mil problem propos disambigu method identifi true posit instanc posit bag propos featur represent scheme instanc level classif bag level classif convert mil problem standard singl instanc learn sil problem solv sil algorithm support vector machin induct semi supervis learn method propos mil evalu method extens challeng mil applic demonstr promis effici robust accuraci regular framework multiclass classif determinist anneal approach local content dbase imag retriev evid region identif heteroscedast probabilist linear discrimin analysi semi dsupervis extens semi dsupervis multi dtask regress human action recognit local spatio dtempor discrimin embed semi dsupervis discrimin analysi robust path dbase similar learn dview stereo match structur visual word arbitrari dview object local semi dsupervis discrimin analysi cccp posterior consist silverman dprior bayesian model choic kernel supervis learn method unifi util tool regular theori dualiti regular prior lead interpret regular method term maximum posteriori estim motiv bayesian interpret kernel method paper pursu bayesian interpret sparsiti kernel set make mixtur point mass distribut prior refer silverman prior provid theoret analysi posterior consist bayesian model choic procedur base prior establish asymptot relationship procedur bayesian inform criterion scalabl kernel dbase semisupervis metric learn algorithm dof dsampl gener abil robust path dbase spectral cluster spectral cluster path base cluster recent develop cluster approach deliv impress result number challeng cluster task robust nois outlier data paper base estim robust statist develop robust path base spectral cluster method defin robust path base similar measur spectral cluster method significantli robust spectral cluster path base cluster perform experi base synthet real world data compar method method color imag berkelei segment dataset benchmark imag segment experi experiment result show method consist outperform method due higher robust imag hallucin neighbor embed visual primit manifold local smooth metric learn applic imag retriev paper propos metric learn method base regular move squar unlik previou metric learn method learn global mahalanobi distanc defin local smooth metric local affin transform flexibl data set metric learn preserv origin topolog structur method fairli effici preprocess step subsequ learn task includ classif cluster nonlinear dimension reduct demonstr method boost perform content base imag retriev cbir task experiment result provid empir evid effect approach kernel select forl semi dsupervis kernel machin kernel path algorithm support vector machin choic kernel function determin map input space featur space crucial import kernel method past year effort learn kernel function kernel matrix paper address model select issu learn hyperparamet kernel function support vector machin svm trace solut path respect kernel hyperparamet train model multipl time kernel hyperparamet optim solut obtain find solut neighborhood hyperparamet calcul solut path exhibit piecewis linear extend nonlinearli result breakpoint comput advanc propos method approxim breakpoint method effici gener sens appli kernel function common scalabl kernel dbase algorithm semi dsupervis metric learn boost kernel discrimin analysi applic tissu classif gene express data abstract paper address problem object classif biosonar base mobil robot natur environ boost method present algorithm base gradient boost biosanar base robot recogn object tree reflect sonar echo gradient boost machin learn approach build strong classifi base learner present kind base learner gradient boost ordinari squar ol kernel base base learner compar previou work present time resolv spectrum kernel extract similar echo effici accur result newli propos boost method compar method term sensit specif accuraci matthew correl coeffici runtim train test semi dsupervis cast index featur dlength film kernel approach semisupervis metric learn kernel dbase distanc metric learn content dbase imag retriev surrog maxim fminim algorithm extens learn kernel matrix maxim kfd dbase class separ criterion advantag kernel method depend critic proper choic kernel function promis approach learn kernel data auto matic paper propos method learn kernel matrix base maxim class separ criterion similar linear discrimin analysi lda kernel fisher discrimin kfd terest note optim criterion function requir invert possibl singular class scatter matrix comput problem encount lda kfd method conduct experi synthet data real world data uci feret show method consist outperform previou kernel learn method structur paper explor theori discours structur stress role purpos process discours theori discours structur compos separ interr late compon structur sequenc utter call linguist structur struc ture purpos call intent structur state focu attent call attent state linguist structur consist segment discours utter anc natur aggreg intent structur captur discours relev purpos express linguist segment relationship attent state abstract focu attent particip discours unfold attent state dynam record object properti relat salient point discours distinct compon essenti provid adequ explan discours phenomena cue phrase refer express interrupt theori attent intent aggreg utter illustr paper number discours properti discours explan behavior cue phrase refer express interrupt explor theori framework describ process utter discours discours process requir recogn utter discours aggreg segment recogn intent express discours relationship intent track ing discours oper mechan attent state process descript specifi recognit task role inform discours particip knowledg domain manifold regular approach calibr reduct sensor dnetwork base track abil accur detect locat mobil node sensor network import artifici intellig ai task rang robot context awar comput previou approach locat estim problem assum avail calibr data obtain data requir great effort paper present manifold regular approach leman calibr effort reduct track mobil node wireless sensor network comput subspac map function signal space physic space small amount label data larg amount unlabel data map function onlin determin locat mobil node sensor network base signal receiv crossbow mica setup network usb camera arrai obtain ground truth experiment result show achiev higher accuraci calibr effort compar previou system tensor embed method local linear model face appear manifold applic dual dsubspac base classif graph laplacian kernel object classif singl classif label class challeng problem machin learn pattern recognit attempt address problem context specif applic work problem gener object classif set paper propos graph base approach problem base robust path base similar measur propos recent construct weight graph robust path base similar edg weight kernel matrix call graph laplacian kernel defin base graph laplacian kernel matrix principl kernel base classifi classif demonstr kernel nearest neighbor classifi synthet data real world imag set show method successfulli solv difficult classif task label exampl extend kernel fisher discrimin analysi weight pairwis chernoff criterion abstract linear discrimin analysi lda kernel fisher discrimin analysi kfd method base restrict assumpt data homoscedast paper propos kfd method call heteroscedast kernel weight discrimin analysi hkwda appeal characterist kernel method handl nonlinear effici disciplin manner incorpor weight function captur heteroscedast data distribut discrimin criterion work realist situat enhanc classif accuraci real world applic effect deal small sampl size problem perform face recognit experi compar hkwda linear nonlinear dimension reduct method show hkwda consist result face recognit imag set hierarch extract exemplar appear manifold unsupervis nonparametr approach propos automat extract repres face sampl exemplar video sequenc imag set multipleshot face recognit motiv nonlinear dimension reduct algorithm call isomap local neighborhood inform approxim geodes distanc face imag hierarch agglom cluster hac algorithm appli group similar face base estim geodes distanc approxim locat appear manifold defin exemplar cluster center templat match subsequ test stage final recognit outcom major vote scheme combin decis individu frame test set experiment result subject video databas demonstr effect flexibl propos method solut path semi dsupervis classif manifold regular facial express recognit advanc local binari pattern paper propos facial express recognit approach base set featur extract face imag textur featur global appear featur set obtain extend local binari pattern intens gradient map comput tsalli entropi gabor filter respons set featur obtain perform nullspac base linear discrimin analysi train face imag propos method evalu extens experi jaff databas compar wide facial express recognit approach experiment result show propos approach maintain high recognit rate wide rang resolut level outperform altern method index term face recognit machin vision local discrimin embed tensor represent ddimension solut path support vector regress recent appeal approach propos comput entir solut path support vector classif svc low extra comput cost approach extend support vector regress svr model call svr method requir error paramet set priori desir accuraci approxim advanc paper show solut path svr piecewis linear respect propos effici algorithm explor dimension solut space defin regular error paramet oppos algorithm svc propos algorithm svr initi number support vector increas gradual algorithm proce good regress function possess spars properti obtain iter defend tcp syn flood attack type ip spoof tcp base flood attack common form distribut denial servic ddo attack abus network resourc bring threat internet incorpor ip spoof make difficult defend attack ip spoof techniqu includ random spoof subnet spoof fix spoof subnet spoof difficult type fight paper propos simpl effici method detect defend tcp syn flood attack ip spoof type includ subnet spoof method make storag effici data structur chang point detect method distinguish complet tcp handshak incomplet simul experi consist show method effici effect defend tcp base flood attack ip spoof type intrus detect router design model dbase transduct learn kernel matrix paper address problem transduct learn kernel matrix probabilist spectiv defin kernel matrix wishart process prior construct hierarch gener model kernel matrix learn specif target kernel matrix random matrix wishart distribut posit definit paramet matrix degre freedom paramet matrix turn invert wishart distribut posit definit hyperpa ramet matrix conjug prior degre freedom equal dimension featur space induc target kernel resort miss data problem devis expect maxim em algorithm infer miss data paramet matrix featur dimension maximum posteriori map manner set target kernel hyperparamet matric model appli type learn problem applic semi supervis learn set present classif method classif experi report benchmark data set encourag result addit devis em algorithm kernel matrix complet relax metric adapt applic semi dsupervis cluster content dbase imag retriev perform supervis unsupervis learn algorithm sensit choic distanc metric previou work metric learn adapt focus classif task make class label inform standard cluster task class label format order adapt metric improv cluster result background knowledg side inform need type side inform form pairwis similar dissimilar inform recent method parametr method propos xing al learn global metric base pairwis side inform shown demonstr promis result paper propos nonparametr method call relax metric adapt rma metric adapt prob lem rma local sens local adapt metric global pattern vicin long rang effect metric adapt process experiment result semi supervis cluster base simul real world data set show rma outperform xing preprint submit elsevi scienc juli al method situat appli rma semi supervis learn improv perform content base imag retriev system metric adapt experiment result base real world imag databas show rma significantli outperform meth od improv imag retriev perform extend relev compon analysi algorithm metric learn posit neg equival constraint relev compon analysi rca recent propos metric learn method semi supervis learn applic simpl effici method appli successfulli give impress result rca make supervisori inform form posit equival constraint paper propos extens rca posit neg equival constraint incorpor experiment result show extend rca algorithm effect kei word metric learn mahalanobi metric semi supervis learn robust local linear embed past year nonlinear dimension reduct nldr nonlinear manifold learn method arous great deal interest machin learn commun method promis automat discov low dimension nonlinear manifold high dimension data space emb data point low dimension embed space tractabl linear algebra techniqu easi implement prone local minima appeal properti nldr method robust outlier data address robust problem paper address problem context nldr method call local linear embed lle base robust estim techniqu propos approach make lle robust refer approach robust local linear embed rlle present specif method realiz gener rlle approach experiment result synthet realworld data show rlle robust outlier local linear metric adapt applic semi dsupervis cluster imag retriev throttl spoof syn flood traffic sourc robust path dbase spectral cluster applic imag segment spectral cluster path base cluster recent develop cluster approach deliv impress result number challeng cluster task robust nois outlier data paper base estim robust statist develop robust path base spectral cluster method defin robust path base similar measur spectral cluster method significantli robust spectral cluster path base cluster perform experi base synthet real world data compar method method color imag berkelei segment dataset benchmark imag segment experi experiment result show method consist outperform method due higher robust nonlinear dimension reduct classif kernel weight subspac method abstract studi kernel subspac method learn low dimension subspac represent classif task propos method call kernel weight nonlinear discrimin analysi kwnda possess appeal properti kernel method handl nonlinear disciplin manner computation attract introduc weight function discrimin criterion outperform exist kernel discrimin analysi method term classif accuraci effect deal small sampl size problem empir compar subspac method respect classif perform facial imag base simpl nearest neighbor rule experiment result show kwnda substanti outperform compet linear nonlinear subspac method kernel dbase metric adapt pairwis constraint abstract supervis unsupervis learn algorithm depend choic distanc metric metric learn supervis learn task long histori extend learn task weaker supervisori inform studi recent method propos semi supervis metric learn base pairwis di similar inform paper propos kernel base approach nonlinear metric learn perform local linear translat kernelinduc featur space formul metric learn problem kernel learn problem solv effici kernel matrix adapt experiment result base synthet real world data set show approach promis semi supervis metric learn bayesian infer princip compon analysi revers jump markov chain mont carlo super dresolut neighbor embed abstract biometr person authent face fingerprint palmprint signatur depend qualiti imag process low resolut imag accuraci impair recov lost inform downsampl imag import authent preprocess base super resolut neighbor embed algorithm histogram match propos improv super resolut approach choos reason train imag train imag select histogram match neighbor embed algorithm emploi recov high resolut imag experi imag show improv super resolut approach promis potenti applic low resolut mobil phone cctv close circuit televis imag person authent reinforc learn episod dstationari markovian environ svc intern signatur verif competit abstract handwritten signatur wide accept biometr ident verif facilit object evalu comparison algorithm field automat handwritten signatur verif organ intern signatur verif competit svc recent step establish common benchmark databas benchmark rule task competit signatur databas involv set signatur data creat genuin signatur skill forgeri set eventu team compet task team compet task evalu data skill forgeri team task equal error rate eer task eer svc successfulli achiev goal experi gain svc similar activ futur bayesian infer transduct learn kernel matrix tanner dwong data augment algorithm kernel method interest recent develop seek learn good kernel empir data automat paper transduct learn kernel matrix miss data problem propos bayesian hierarch model problem devis tanner wong data augment algorithm make infer model tanner wong algorithm close relat gibb sampl bear strong resembl expect maxim em algorithm cient implement propos simplifi bayesian hierarch model tannerwong algorithm express relationship kernel input space kernel output space symmetr definit gener eigenproblem local linear metric adapt semi dsupervis cluster supervis unsupervis learn algorithm sensit choic distanc metric classif task make class label inform metric learn inform gener unavail convent cluster task recent research sought address variant convent cluster problem call semi supervis cluster perform cluster presenc background knowledg supervisori inform express pairwis similar dissimilar constraint exist metric learn method semisupervis cluster perform global metric learn linear transform paper propos metric learn method perform nonlinear transform global linear transform local formul learn problem optim problem present method solv toi data set show empir local linear metric adapt llma method handl difficult case handl satisfactorili previou method demonstr effect method real data set idr intrus detect router defend distribut denial dof dservic ddo attack surrog maxim fminim algorithm adaboost logist regress model surrog maxim minim sm algorithm famili algorithm regard gener expect maxim em algorithm idr intrus detect router defend distribut denial dof dservic ddo attack time seri cluster arma mixtur parametr distanc metric learn label inform distanc base method pattern recognit machin learn reli similar dissimilar measur pattern input space applic euclidean distanc input space good choic complic distanc metric paper propos parametr method distanc metric learn base class label inform defin dissimilar measur prove metric favor properti class dissimilar larger class dissimilar perform parametr learn find regress map input space euclidean featur space dissimilar pattern input space approxim euclidean distanc point featur space parametr learn perform iter major algorithm method test synthet real world benchmark dataset experiment result show approach promis host dbase intrus detect dynam static behavior model intrus detect emerg importantapproach network secur paper adopt anomali detect approach detect intrus base program user pro le built normal usag data program pro le base unix system call user pro le base unix shell command model di erent type behavior model data mine dynam model approach isbas hidden markov model hmm principl maximum likelihood static model approach isbasedonev occurr frequenc distribut principl minimum cross entropi novelti detect approach adopt estim model paramet normal train data oppos classi cation approach normal intrus data train determin behavior similar normal model classi ed normal scheme justi ed perspect hypothesi test experiment result show dynam model approach static model approach system call dataset dynam model approach wors shell command dataset static model approach similar perform instanc base learn report previous shell command databas higher comput storag requir method mixtur arma model model dbase time seri cluster cluster problem central knowledg discoveri data mine task exist cluster method work fix dimension represent data pattern paper studi cluster data pattern repres sequenc time seri possibl di erent length propos model base approach problem mixtur autoregress move averag arma model deriv expect maxim em algorithm learn mix coe cient paramet compon model algorithm determin number cluster data automat experi conduct number simul real dataset result experi show method compar favor method recent propos similar time seri cluster problem parzen dwindow network intrus detector network intrus detect problem detect anomal network connect caus intrus activ intrus detect system propos normal intrus data build classifi intrus data scarc difficult collect propos solv problem novelti detect approach propos nonparametr densiti estim approach base parzen window estim gaussian kernel build intrus detect system normal data facilit comparison test system kdd cup dataset system compar favor kdd cup winner base ensembl decis tree bag boost system intrus data normal data train user profil intrus detect dynam static behavior model bidirect deform match applic handwritten charact extract achiev integr segment recognit complex scene model base approach wide accept promis paradigm perform satisfactori target object highli deform level outlier contamin high paper describ bayesian framework classifi input pattern detect target pattern complex scene deform model show framework similar forward revers set hausdorff match match discrimin properti complementari properli combin framework propos match scheme call bidirect match combin approach inherit advantag bayesian framework obtain encourag empir result shape base pattern extract subset cedar handwrit databas handwritten word highli vari shape index term model base segment deform model bayesian infer bidirect match hausdorff match abridg version paper publish dr kwok wai cheung correspond author deform model visual pattern recognit paper review model base method fornon rig shape recoglj method model match andclassif rigg shape aregefixq problemat conventationalalgentati rigg model pencalc applic dline mathemat express recognit technolog calcul program found exist penbas mobil comput devic person digit assist pda handheld devic full advantag pen technolog offer devic input express virtual keypad shown screen stylu electron pen simpli point devic paper propos intellig handwrit base calcul program user enter express simpli write screen stylu addit variabl defin store intermedi result subsequ calcul ordinari algebra calcul propos softwar result applic line mathemat express recognit technolog mathemat express editor program hidden dmode markov decis process nonstationari sequenti decis make paper address condit condit briefli discuss interest special type nonstationari environ repeat dynam manner propos formal model environ develop algorithm learn model paramet comput optim polici proce briefli review categori problem shown figur defin terminolog paper error detect error conceal techniqu video transmiss channel introduc bit error effici result error conceal techniqu reli error detect capabl video decod error detect techniqu emploi fragil watermark propos paper embed fragil watermark quantiz dct coeffici examin integr decod side error detect capabl video decod significantli increas compar wide syntax base error detect scheme mathemat express recognit survei automat recognit mathemat express kei vehicl drive transcrib document scientif engin disciplin electron form problem typic consist major stage symbol recognit structur analysi survei paper review exist work respect major stage recognit process put emphasi similar differ system import issu mathemat express recognit address depth serv provid clear pictur research area develop date kei word error detect correct mathemat express recognit perform evalu structur analysi symbol recognit effici syntact approach structur analysi dline handwritten mathemat express machin recognit mathemat express trivial individu charact symbol express recogn correctli paper propos definit claus grammar dcg formal de ne set replac rule pars mathemat express dcg defin replac rule concis definit readili execut form dcg parser potenti ineffici due frequent backtrack propos method increas effici pars process experi commonli mathemat express show propos method achiev satisfactori speedup make mathemat express recognit feasibl real world applic bidirect match algorithm deform pattern detect applic handwritten word retriev bayesian framework deform pattern classif propos promis result isol handwritten charact recognit perform degrad significantli appli detect deform pattern complex scene amount outlier due neighbor object background larg fact evid measur penal model rest white space result high fals alarm rate paper bayesian framework deform pattern detect propos framework possess intrins properti match part imag segment evid measur penal white space implicitli limit data explor capabl major trade properli combin framework match algorithm call bidirect match propos combin approach possess advantag environ model nonstationari reinforc learn reinforc learn nonstationari environ gener regard import difficult problem paper partial address problem formal subclass nonstationari environ environ model call hidden mode markov decis process hm mdp assum environment confin small number hidden mode mode basic index markov decis process mdp evolv time markov chain hm mdp special case partial observ markov decis process pomdp model hm mdp environ gener pomdp model unnecessarili increas problem complex variant baum welch algorithm develop model learn requir data time introduct reinforc learn rl learn paradigm base framework markov decis process mdp tradit rl research assum environ dynam recogn dline handwritten alphanumer charact flexibl structur match speed accuraci flexibl crucial practic line handwrit recognit extens import concern move domain requir charact set extend paper propos simpl robust structur approach recogn line handwrit approach design achiev reason speed fairli high accuraci suffici toler variat time maintain high degre reusabl facilit extens experiment result show recognit rate digit uppercas letter lowercas letter combin set reject case exclud calcul rate increas averag recognit speed charact run prolog sun sparc unix workstat effici structur analysi mathemat express machin recognit mathemat express trivial individu charact symbol express recogn correctli paper propos definit claus grammar dcg formal defin set replac rule pars mathemat express dcg defin replac rule concis definit readili execut form backtrack parser prolog interpret execut dcg directli natur ineffici propos method increas effici pars process experi typic mathemat express show propos method achiev speedup rang time make mathemat express recognit feasibl real world applic bayesian framework deform pattern recognit applic handwritten charact recognit deform model recent propos pattern recognit applic due abil handl larg shape variat propos approach repres pattern shape deform model deform match input imag subsequ feed extract inform classifi compon model match classif treat independ task paper studi integr deform model bayesian framework unifi approach model match classifi shape handwritten charact recognit serv testb evalu approach system recognit invari affin transform handwrit variat addit preprocess manual set hyperparamet regular paramet charact width requir issu incorpor constraint model fl dline handwritten alphanumer charact recognit domin point stroke competit mixtur deform model pattern classif success appli deform model featur extract natur step appli model pattern classif recent cast deform model bayesian framework classif give promis result deform model method computation expens due requir iter optim process problem sever larg number model charact recognit deform match input data final classif deriv paper propos combin deform model mixtur individu model compet surviv match process classif model compet elimin earli allow substanti save comput process competit elimin appli handwritten digit recognit signific speedup achiev sacrif recognit accuraci bayesian regular construct neural network paper studi incorpor bayesian regular construct neural network degre regular automat control bayesian infer framework requir manual set simul show regular input train full bayesian approach produc network gener perform lower suscept fit network size increas regular input train mackai evid framework produc signific improv problem test introduct multi layer feedforward network popularli pattern classif regress problem standard back propag perform gradient descent weight space network fix topolog construct procedur start small network grow addit hidden unit weight satisfactori solut found parti grammat infer approach dline handwrit model recognit pilot studi dline handwritten alphanumer charact recognit featur sequenc predict drout memori dbase reinforc learn approach adapt traffic control paper propos memori base learn algorithm call predict rout pq rout adapt traffic control attempt address problem encount rout boyan amp littman inabl fine tune rout polici low network load inabl learn optim polici decreas load condit unlik memori base reinforc learn algorithm memori past experi increas learn speed pq rout experi learn reus predict traffic trend effect pq rout verifi network topolog traffic condit simul result show pq rout superior rout term learn speed adapt introduct adapt traffic control problem devis rout polici control router oper stationari environ minim construct neural network estim bayesian discrimin function crucial import success artifici neural network pattern classif problem network size automat determin address issu formul process automat search space function correspond subclass multilay feedforward network learn dynam network construct process involv adjust network weight topolog ad hidden unit correspond extract higher level featur origin input featur reduc residu classif error shown result network approxim bayesian classifi implement bayesian decis rule classif paper report empir result pattern classif experi keyword pattern classif discrimin function bayesian classifi feedforward network construct network transfer learni neural network approach construct induct watanab know guess formal qualit studi wilei york ny weiss indurkhya weiss indurkhya small sampl decis tree prune intern machin learn confer brunswick nj wilson wilson classifi system animat problem machin learn wittgenstein wittgenstein ascomb editor philosoph investig oxford blackwel london uk wnek michalski wnek michalski hypothesi driven construct induct aq hci method experi machin learn bibliographi intern workshop machin learn page evanston il morgan kaufmann slotta chi slotta chi understandin 