Isakhelo esomeleleyo samanani eBayesi sisetyenziswe ngokubanzi kwiinkalo ezininzi, kubandakanywa ukufunda ngomatshini.
Iinkcukacha-manani zaseBayesi zibonelela ngendlela eguquguqukayo nenokwenzeka yokuthelekelela, ngokwahlukileyo kumanani-nkcazo akudala, axhomekeke kwiiparamitha ezibekiweyo kunye noqikelelo lwamanqaku.
Kusinceda ukuba sithathele ingqalelo ulwazi olukhoyo size silungise iimbono zethu xa ulwazi olutsha luvela.
Izibalo zaseBayesi zisinika amandla okwenza izigwebo ezinolwazi ngakumbi kwaye sifikelele kwizigqibo ezithembeke ngakumbi ngokwamkela ukungaqiniseki kunye nokusebenzisa unikezelo olunokwenzeka.
Iindlela zeBayesian zibonelela ngombono owahlukileyo wokulinganisa unxibelelwano olunzima, ukulawula idatha encinci, kunye nokujongana nokugqithisa kumxholo yokufunda umatshini.
Siza kujonga ukusebenza kwangaphakathi kwezibalo zeBayesi kweli nqaku, kunye nokusetyenziswa kwayo kunye neenzuzo kwintsimi yokufunda koomatshini.
Ezinye iikhonsepthi eziphambili kwizibalo zaseBayesi ziqhele ukusetyenziswa kwiSifundo soomatshini. Masijonge eyokuqala; Indlela yaseMonte Carlo.
Indlela yaseMonte Carlo
Kwizibalo zaseBayesi, ubuchule beMonte Carlo buyimfuneko, kwaye banefuthe elibalulekileyo kwizicelo zokufunda ngomatshini.
I-Monte Carlo ibandakanya ukudala iisampulu ezingakhethiyo ukusuka kunikezelo olunokwenzeka ukuya kuqikelelo oluqikelelo lwezibalo ezifana nezidibaniso okanye unikezelo lwangasemva.
Indlela ye-Monte Carlo ibonelela ngendlela esebenzayo yokuqikelela ubuninzi bomdla kunye nokuphonononga izithuba zeparamitha ezinomgangatho ophezulu ngokuphindaphindiweyo ngokuphindaphindiweyo ekusasazweni komdla kunye nokulinganisa iziphumo.
Ngokusekwe kulinganiso lwamanani, obu buchule bunceda abaphandi ukuba benze izigwebo ezinolwazi, balinganise ukungaqiniseki, kwaye bafumane iziphumo eziqinileyo.
Ukusebenzisa i-Monte Carlo ekubaleni okusebenzayo
Ukubala ukusasazwa ngasemva kwizibalo zaseBayesi rhoqo kufuna izidibanisi ezinzima.
Ukuqikelelwa okusebenzayo kwezi zidibanisi ezibonelelwa bubuchule beMonte Carlo kusenza sikwazi ukuphonononga ngokufanelekileyo ukusasazwa ngasemva.
Oku kubalulekile ekufundeni koomatshini, apho iimodeli ezintsonkothileyo kunye nezithuba zeparamitha ezinomgangatho ophezulu ziyinto eqhelekileyo.
Ngokuqikelela ngokufanelekileyo izinto eziguquguqukayo zomdla ezifana namaxabiso okulindelekileyo, i-histograms, kunye ne-marginalizations kusetyenziswa ubuchule be-Monte Carlo, sixhotyiswe ngcono ukuba sihlolisise idatha kwaye senze izigqibo kuyo.
Ukuthatha iSampulu kwi-Postererior Distribution
Kwi-inference ye-Bayesian, isampuli ukusuka kwi-posterior distribution linyathelo elibalulekileyo.
Ukukwazi ukwenza isampuli ukusuka ngasemva kubalulekile kwizicelo zokufunda koomatshini, apho sizama ukufunda kwiidatha kunye nokwenza uqikelelo.
Iindlela ze-Monte Carlo zibonelela ngeendlela ezahlukeneyo zesampulu ukusuka ekuhanjisweni okungekho mthethweni, kubandakanywa ne-posterior.
Ezi ndlela, ezibandakanya indlela yokuguqula, indlela yokuqulunqa, indlela yokugatya, kunye nokubaluleka kwesampulu, kusenza sikwazi ukukhupha iisampulu ezimele ukusuka ngasemva, okusivumela ukuba sihlolisise kwaye siqonde ukungaqiniseki okuhambelana neemodeli zethu.
Monte Carlo kwiSifundo soomatshini
Ii-algorithms ze-Monte Carlo zisetyenziswa ngokubanzi ekufundeni komatshini ukuqikelela ukuhanjiswa kwangasemva, okubandakanya ukungaqiniseki kweeparamitha zemodeli ezinikwe idatha ephawulweyo.
Ubuchwephesha be-Monte Carlo buvumela umlinganiselo wokungaqiniseki kunye noqikelelo lobungakanani bomdla, njengamaxabiso okulindelekileyo kunye nezalathi zentsebenzo yemodeli, ngokuthatha iisampulu ukusuka kulwabiwo lwangasemva.
Ezi sampuli zisetyenziswa kwiindlela ezahlukeneyo zokufunda ukuvelisa uqikelelo, ukwenza ukhetho lwemodeli, ukulinganisa ukuntsokotha kwemodeli, kunye nokwenza inference yeBayesian.
Ngaphaya koko, ubuchule beMonte Carlo bubonelela ngesakhelo esiguquguqukayo sokujongana neendawo ezinomgangatho ophezulu weparamitha kunye neemodeli ezintsonkothileyo, ezivumela ukuphononongwa kokusasazwa kwangasemva kunye nokwenziwa kwezigqibo ezinamandla.
Ukuqukumbela, iindlela ze-Monte Carlo zibalulekile ekufundeni komatshini kuba ziququzelela ukulinganisa ukungaqiniseki, ukwenza izigqibo, kunye nokuziphatha okusekelwe kwi-posterior distribution.
Markov Chains
Iikhonkco zeMarkov ziyimizekelo yeemathematika ezisetyenziselwa ukuchaza iinkqubo zestochastic apho imeko yenkqubo ngexesha elithile imiselwe kuphela yimeko yangaphambili.
Ikhonkco le-Markov, ngamagama alula, lulandelelwano lweziganeko ezingalindelekanga okanye iindawo apho amathuba okutshintsha ukusuka kwelinye ilizwe ukuya kwelinye kuchazwa liqela lezinto ezinokwenzeka ezaziwa ngokuba yinguquko enokwenzeka.
Amatyathanga eMarkov asetyenziswa kwifiziksi, ezoqoqosho, nakwisayensi yekhompyuter, kwaye abonelela ngesiseko esiqinileyo sokufunda kunye nokulinganisa iinkqubo ezintsonkothileyo ezinokuziphatha okunokwenzeka.
Amakhonkco eMarkov aqhagamshelwe ngokusondeleyo kumatshini wokufunda kuba akuvumela ukuba wenze imodeli kwaye uvavanye ubudlelwane obuguquguqukayo kwaye wenze iisampulu ukusuka kulwabiwo olunzima olunokwenzeka.
Amakhonkco eMarkov aqeshwe ekufundeni koomatshini kwizicelo ezifana nokwandiswa kwedatha, ulandelelwano lwemodeli, kunye nemodeli yokuvelisa.
Ubuchule bokufunda ngoomatshini bunokubamba iipatheni ezisisiseko kunye nobudlelwane ngokwakha kunye nokuqeqesha imodeli yekhonkco yeMarkov kwidatha eqatshelweyo, ibenze ibe luncedo kwizicelo ezifana nokuqondwa kwentetho, ukusetyenzwa kolwimi lwendalo, kunye nohlalutyo lothotho lwexesha.
Amakhonkco eMarkov abaluleke kakhulu kwiindlela ze-Monte Carlo, ezivumela ukuba kwenziwe isampulu esebenzayo kunye noqikelelo lokuqikelelwa kwiBayesian umatshini wokufunda, ojolise ekuqikeleleni usasazo lwangasemva olunikwe idatha ephawulweyo.
Ngoku, kukho enye ingcamango ebalulekileyo kwiBayesian Statistics ivelisa amanani angaqhelekanga osasazo olungenasizathu. Makhe sibone ukuba kunceda njani ukufunda ngomatshini.
Isizukulwana seNani esiRandom sokusasazwa ngokungenamkhethe
Kwiintlobo ngeentlobo zemisebenzi ekufundeni koomatshini, amandla okuvelisa amanani angaqhelekanga ukusuka kunikezelo olungenasizathu kubalulekile.
Iindlela ezimbini ezidumileyo zokufezekisa le njongo yi-algorithm ye-inversion kunye ne-algorithm yokwamkelwa-yokwala.
Inversion Algorithm
Sinokufumana amanani angaqhelekanga ukusuka kunikezelo kunye nomsebenzi owaziwayo wokusasaza oqokelelweyo (CDF) usebenzisa i-algorithm yokuguqula.
Singakwazi ukuguqula amanani afanayo aqhelekileyo abe ngamanani angaqhelekanga kunye nosasazo olufanelekileyo ngokubuyisela umva i-CDF.
Le ndlela ifanelekile kwizicelo zokufunda zoomatshini ezibiza iisampulu ezivela kunikezelo olwaziwayo kuba lusebenza kwaye lusebenza ngokubanzi.
Ulwamkelo-Ukwala Algorithm
Xa i-algorithm yesiqhelo ingafumaneki, i-algorithm yokumkelwa-i-rejection iyindlela eguquguqukayo kunye nesebenzayo yokuvelisa amanani angaqhelekanga.
Ngale ndlela, amanani apheleleyo ayamkelwa okanye ayaliwe ngokusekwe kuthelekiso lomsebenzi wemvulophu. Isebenza njengokwandiswa kwenkqubo yokwakheka kwaye ibalulekile ekuveliseni iisampulu ukusuka kulwabiwo oluntsonkothileyo.
Ekufundeni komatshini, i-algorithm yokwala ukwamkelwa ibaluleke kakhulu xa kujongwana nemiba emininzi okanye iimeko apho indlela yokuguqula uhlalutyo oluthe tye ayinakwenzeka.
Ukusetyenziswa kuBomi bokwenyani kunye nemingeni
Ukufumana imisebenzi yemvulophu efanelekileyo okanye uqikelelo olwenza kakhulu unikezelo ekujoliswe kulo luyimfuneko ukuze zombini iindlela zokwenza oko.
Oku kufuna rhoqo ukuqondwa ngokucokisekileyo kweempawu zonikezelo.
Enye into ebalulekileyo ekufuneka ithathelwe ingqalelo ngumlinganiselo wokwamkelwa, olinganisa ukusebenza kwe-algorithm.
Ngenxa yobunzima bokusabalalisa kunye nesiqalekiso se-dimensionality, indlela yokwamkelwa-yokwaliwa inokuthi, nangona kunjalo, ibe yingxaki kwimiba ephezulu. Iindlela ezizezinye ziyafuneka ukujongana nezi ngxaki.
Ukuphucula ukuFunda koomatshini
Kwimisebenzi efana nokwandisa idatha, ukuseta imodeli, kunye noqikelelo lokungaqiniseki, ukufunda ngomatshini kufuna ukuveliswa kwee-integers ezingenamkhethe ukusuka kulwabiwo olungenamkhethe.
Ubuchule bokufunda ngomatshini ingakhetha iisampulu kwiintlobo ezahlukeneyo zonikezelo ngokusebenzisa i-inversion kunye neendlela zokwala ukwamkelwa, ukuvumela ukumodareyitha okuguquguqukayo kunye nokusebenza okwandisiweyo.
Ekufundeni koomatshini baseBayesi, apho ukusasazwa ngasemva rhoqo kufuna ukuqikelelwa ngesampulu, ezi ndlela ziluncedo kakhulu.
Ngoku, masiqhubele phambili kwenye ingqiqo.
Intshayelelo kwi-ABC (I-Computation ye-Bayesi esondeleyo)
I-Approximate Bayesian Computation (ABC) yindlela yamanani esetyenziswa xa kubalwa umsebenzi onokwenzeka, omisela ukuba nokwenzeka kobungqina bedata enikwe imodeli yeeparamitha, ingumceli mngeni.
Endaweni yokubala umsebenzi onokwenzeka, i-ABC isebenzisa ukulinganisa ukuvelisa idatha ukusuka kumzekelo kunye namaxabiso eparameter.
Idatha efanisiweyo kunye nebonwayo iyathelekiswa ke, kwaye imimiselo yeparameter eyenza ukulinganisa okuthelekisayo kuyagcinwa.
Uqikelelo olurhabaxa lokusasazwa ngasemva kweeparameters lunokuveliswa ngokuphinda le nkqubo kunye nenani elikhulu lokulinganisa, ukuvumela ukuba iBayesian inference.
Ingcamango ye-ABC
Ingqikelelo engundoqo ye-ABC kukuthelekisa idatha eyenziweyo eyenziwe yimodeli ukujongwa kwedatha ngaphandle kokubala ngokucacileyo umsebenzi wokunokwenzeka.
I-ABC isebenza ngokuseka umgama okanye i-metric yokungafani phakathi kwedatha eqatshelweyo neyenziweyo.
Ukuba umgama ungaphantsi komda othile, amaxabiso eparameter asetyenziselwa ukwakha ukulinganisa okuhambelanayo kucingelwa ukuba kunengqiqo.
I-ABC idala uqikelelo lokusasazwa kwangasemva ngokuphinda le nkqubo yokwamkelwa-kwaliwa ngamaxabiso ahlukeneyo eparameter, ebonisa amaxabiso eparamitha abambekayo anikwe idatha ephawulweyo.
Ii-ABC zeZifundo zoomatshini
I-ABC isetyenziswa ekufundeni ngoomatshini, ngakumbi xa ingqikelelo esekwe ekucingeni inzima ngenxa yeemodeli ezintsonkothileyo okanye ezixabisa kakhulu. I-ABC ingasetyenziselwa usetyenziso olwahlukeneyo olubandakanya ukhetho lwemodeli, uqikelelo lweparamitha, kunye nemodeli yokuvelisa.
I-ABC ekufundeni koomatshini ivumela abaphandi ukuba bazobe uqikelelo malunga neeparamitha zemodeli kwaye bakhethe ezona modeli zingcono ngokuthelekisa idatha eyenziweyo kunye neyonyani.
Ubuchule bokufunda ngomatshini bangafumana ulwazi kwimodeli yokungaqiniseki, benze uthelekiso lwemodeli, kwaye benze uqikelelo olusekwe kwidatha eqatshelweyo ngokuqikelela usasazo lwangasemva nge-ABC, naxa uvavanyo lokunokwenzeka lubiza okanye lungenakwenzeka.
isiphelo
Ekugqibeleni, izibalo ze-Bayesian zibonelela ngesakhelo esomeleleyo sokubethelela kunye nomzekelo wokufunda ngomatshini, okusivumela ukuba sibandakanye ulwazi lwangaphambili, ukujongana nokungaqiniseki, kunye nokufikelela kwiziphumo ezithembekileyo.
Iindlela ze-Monte Carlo zibalulekile kwizibalo zaseBayesi kunye nokufunda komatshini kuba zivumela ukuphononongwa ngokufanelekileyo kweendawo eziyinkimbinkimbi zeparameter, ukuqikelelwa kwamaxabiso omdla, kunye nesampuli ukusuka kwi-posterior distribution.
Amatyathanga eMarkov anyusa umthamo wethu wokuchaza kunye nokulinganisa iinkqubo ezinokwenzeka, kunye nokuvelisa amanani angaqhelekanga osasazo olwahlukeneyo luvumela umfuziselo oguquguqukayo kunye nokusebenza ngcono.
Okokugqibela, i-Approximate Bayesian Computation (ABC) bubuchule obuluncedo bokwenza ukubala okunokwenzeka okunzima kunye nokuvelisa izigwebo zeBayesian ekufundeni koomatshini.
Sinokuphuhlisa ukuqonda kwethu, siphucule iimodeli, kwaye senze izigwebo ezifundisiweyo kwicandelo lokufunda koomatshini ngokusebenzisa le migaqo.
Shiya iMpendulo