Ikamva lilapha. Kwaye, kolu matshini kwixesha elizayo bayaliqonda ihlabathi elibangqongileyo ngendlela efanayo naleyo abantu benza ngayo. Iikhompyutha zinokuqhuba iimoto, zixilonge izifo, zize zichaze ngokuchanekileyo ikamva.
Oku kunokubonakala ngathi yintsomi yesayensi, kodwa iimodeli zokufunda ezinzulu ziyenza ibe yinyani.
Ezi algorithms ezintsonkothileyo zityhila iimfihlo ze kukubhadla okungeyonyani, ukuvumela iikhompyutha ukuba zizifundele kwaye ziphuhlise. Kule posi, siza kungena kummandla weemodeli zokufunda nzulu.
Kwaye, siya kuphanda amandla amakhulu abanawo okuguqula ubomi bethu. Lungiselela ukufunda ngetekhnoloji ephambili etshintsha ikamva lomntu.
Zeziphi kanye kanye iiModeli zokuFunda nzulu?
Ngaba ukhe wadlala umdlalo apho kufuneka uchonge umahluko phakathi kwemifanekiso emibini?
Kumnandi nangona kunjalo, kunokuba nzima, akunjalo? Khawufane ucinge ukwazi ukufundisa ikhompyuter ukudlala loo mdlalo kwaye uphumelele ngalo lonke ixesha. Iimodeli zokufunda nzulu zifezekisa kanye oko!
Iimodeli zokufunda ezinzulu zifana noomatshini abakrelekrele kakhulu abanokuhlola inani elikhulu lemifanekiso kwaye babone ukuba bafana ngantoni. Bakwenza oku ngokuqhaqha imifanekiso baze bafunde ngamnye ngamnye.
Basebenzisa oko bakufundileyo ukuchonga iipateni kwaye benze uqikelelo malunga nemifanekiso emitsha abangazange bayibone ngaphambili.
Iimodeli zokufunda nzulu ziinethiwekhi ze-neural ezenziweyo ezinokufunda kwaye zikhuphe iipateni ezintsonkothileyo kunye neempawu kwiiseti zedatha ezinkulu. Ezi modeli zenziwe ngamanqwanqwa amaninzi adibeneyo, okanye i-neurons, ehlalutya kwaye iguqule idatha engenayo ukuze ivelise imveliso.
Iimodeli zokufunda nzulu zifaneleke ngokukodwa kwimisebenzi efuna ukuchaneka okukhulu kunye nokuchaneka, okufana nokuchongwa kwemifanekiso, ukuqonda intetho, ukusetyenzwa kolwimi lwendalo, kunye neerobhothi.
Baye basetyenziswa kuyo yonke into ukusuka kwiimoto eziziqhubayo ukuya kuxilongo lwezonyango, iinkqubo zokucebisa, kunye Uhlalutyo oluqikelelweyo.
Nalu uguqulelo olwenziwe lula lomboniso ukubonisa ukuhamba kwedatha kwimodeli yokufunda enzulu.
Idatha yegalelo iqukuqela kuluhlu lwegalelo lomfuziselo, othi ke udlulise idatha ngenani leeleya ezifihliweyo ngaphambi kokubonelela ngoqikelelo lwemveliso.
Umaleko ngamnye ofihliweyo wenza uthotho lwemisebenzi yemathematika kwidatha yegalelo phambi kokuba uyidlulisele kuluhlu olulandelayo, olubonelela ngoqikelelo lokugqibela.
Ngoku, makhe sibone ukuba yeyiphi imodeli yokufunda enzulu kwaye singayisebenzisa njani ebomini bethu.
1. Convolutional Neural Networks (CNNs)
Ii-CNN ziyimodeli yokufunda enzulu eye yaguqula indawo yombono wekhompyuter. Ii-CNN zisetyenziselwa ukwahlula imifanekiso, ukuqaphela izinto, kunye nemifanekiso yecandelo. Ubume kunye nomsebenzi we-cortex ebonakalayo yomntu yazisa ukuyilwa kwee-CNNs.
Basebenza njani?
I-CNN yenziwe liqela leeleya zeconvolution, iileya zokudibanisa, kunye neeleya eziqhagamshelwe ngokupheleleyo. Igalelo ngumfanekiso, kwaye imveliso yingqikelelo yeleyibhile yeklasi yomfanekiso.
I-CNN's convolutional layers yakha imephu yesici ngokwenza imveliso yamachaphaza phakathi komfanekiso wegalelo kunye nesethi yezihluzi. Iileya zokudibanisa zehlisa ubungakanani bemephu yesici ngokuyithoba.
Okokugqibela, imephu yesici isetyenziswa ngamaleya aqhagamshelwe ngokupheleleyo ukuqikelela ileyibhile yodidi lomfanekiso.
Kutheni zibalulekile ii-CNNs?
Ii-CNN zibalulekile kuba zinokufunda ukufumanisa iipateni kunye neempawu kwimifanekiso abantu abakufumanisa kunzima ukuyibona. Ii-CNNs zinokufundiswa ukuqaphela iimpawu ezinje ngencam, iikona, kunye nokwakheka kusetyenziswa iiseti zedatha ezinkulu. Emva kokufunda ezi zakhiwo, i-CNN inokuzisebenzisa ukuchonga izinto kwiifoto ezintsha. Ii-CNNs zibonise ukusebenza kweyona ndlela iphambili kwiintlobo ngeentlobo zezicelo zokuchonga umfanekiso.
Sisebenzisa phi ii-CNNs
Ukhathalelo lwempilo, ishishini lemoto, kunye nokuthengisa ngamacandelo ambalwa asebenzisa ii-CNNs. Kwishishini lezempilo, banokuba luncedo ekuxilongeni ukugula, uphuhliso lwamayeza, kunye nohlalutyo lwemifanekiso yezonyango.
Kwicandelo leemoto, bancedisa ekubhaqweni kwendlela, ukubona into, kunye nokuqhuba ngokuzimeleyo. Zikwasetyenziswa kakhulu ekuthengiseni ukukhangela okubonakalayo, isincomo semveliso esekwe kumfanekiso, kunye nolawulo lwempahla.
Umzekelo; UGoogle usebenzisa ii-CNNs kwiinkqubo ezahlukeneyo, kuquka Google Lens, isixhobo sokuchonga umfanekiso othandwayo. Inkqubo isebenzisa ii-CNN ukuvavanya iifoto kunye nokunika abasebenzisi ulwazi.
ILens kaGoogle, umzekelo, inokubona izinto emfanekisweni kwaye inike iinkcukacha ngazo, njengohlobo lweentyatyambo.
Isenokuguqulela umbhalo othathwe kumfanekiso ukuya kwiilwimi ezininzi. ILens kaGoogle iyakwazi ukunika abathengi ulwazi oluluncedo ngenxa yoncedo lweeCNNs ekuchongeni ngokuchanekileyo izinto kunye nokukhupha iimpawu kwiifoto.
2. Uthungelwano lweMemori yeXesha eliFutshane (LSTM).
Uthungelwano lweMemori yeXesha eliMfutshane (LSTM) lwenziwa ukujongana neziphene zothungelwano lwe-neural recurrent recurrent neural (RNNs). Uthungelwano lwe-LSTM lufanelekile kwimisebenzi efuna ukusetyenzwa kolandelelwano lwedatha ngexesha lonke.
Basebenza ngokusebenzisa iseli yememori ethile kunye neendlela ezintathu zokungena.
Balawula ukuhamba kolwazi ukungena nokuphuma kwiseli. Isango lokungena, isango lokulibala kunye nesango lokuphuma ngamasango amathathu.
Isango lokungena lilawula ukuhamba kwedatha kwiseli yememori, isango lokulibala lilawula ukucinywa kwedatha kwiseli, kwaye isango lokuphuma lilawula ukuhamba kwedatha ngaphandle kweseli.
Iyintoni Intsingiselo Yazo?
Uthungelwano lwe-LSTM luluncedo kuba lunokumela ngempumelelo kunye nokuqikelela ulandelelwano lwedatha kunye nobudlelwane bexesha elide. Banokurekhoda kwaye bagcine ulwazi malunga namagalelo angaphambili, okubavumela ukuba benze uqikelelo oluchanekileyo malunga namagalelo exesha elizayo.
Ukuqondwa kwentetho, ukuqondwa kobhalo lwesandla, ukusetyenzwa kolwimi lwendalo, kunye nokuchazwa kwemifanekiso zezinye zezicelo ezithe zasebenzisa uthungelwano lwe-LSM.
Sizisebenzisa phi iiNethiwekhi zeLSTM?
Iinkqubo ezininzi zesoftware kunye nethekhnoloji zisebenzisa iinethiwekhi zeLSTM, kubandakanywa iinkqubo zokuqaphela intetho, izixhobo zokucwangcisa iilwimi zendalo ezifana Uhlalutyo lweemvakalelo, iisistim zoomatshini bokuguqulela, kunye neenkqubo zokuvelisa isicatshulwa kunye nemifanekiso.
Zikwasetyenziswe ekudaleni iimoto eziziqhubayo kunye neerobhothi, kunye nakwishishini lezemali ukubona ubuqhophololo kunye nokulindela. imakethi yesitokhwe iintshukumo.
3. IiNethiwekhi zoNxibelelwano lweNkathazo (GANs)
IiGAN zi ukufunda okunzulu ubuchule obusetyenziselwa ukuvelisa iisampuli zedatha entsha efana neseti yedatha enikiweyo. IiGAN zenziwe ngambini uthungelwano lwe-neural: enye efunda ukuvelisa iisampuli ezintsha kunye nenye efunda ukwahlula phakathi kweesampulu zokwenyani kunye nezenziweyo.
Ngendlela efanayo, ezi networks zimbini ziqeqeshwa kunye de ijenereyitha ikwazi ukuvelisa iisampulu ezingaqondakaliyo kwezi zikhoyo.
Kutheni Sisebenzisa ii-GAN
Ii-GAN zibalulekile ngenxa yomthamo wazo wokuvelisa umgangatho ophezulu idatha yokwenziwa enokuthi isetyenziselwe usetyenziso olwahlukeneyo, kubandakanya ukuveliswa kwemifanekiso kunye nevidiyo, ukuveliswa kombhalo, kunye nokuveliswa komculo.
I-GANs nayo isetyenziselwe ukwandiswa kwedatha, eveliswa idatha yokwenziwa ukongeza idatha yehlabathi lokwenyani kunye nokuphucula ukusebenza kweemodeli zokufunda ngomatshini.
Ngaphezu koko, ngokudala idatha yokwenziwa engasetyenziselwa ukuqeqesha iimodeli kunye nokuxelisa izilingo, ii-GAN zinamandla okuguqula amacandelo afana namayeza kunye nophuhliso lwamachiza.
Usetyenziso lwee-GANs
Ii-GAN zinokongeza iiseti zedatha, zenze imifanekiso emitsha okanye iimuvi, kwaye zivelise idatha eyenziweyo yokulinganisa kwezenzululwazi. Ngaphaya koko, ii-GAN zinamandla okuqeshwa kwiintlobo ngeentlobo zezicelo ukusuka kulonwabo ukuya kwezonyango.
iminyaka kunye neevidiyo. Umzekelo weNVIDIA's StyleGAN2, isetyenziselwe ukwenza iifoto ezikumgangatho ophezulu zabantu abadumileyo kunye nomsebenzi wobugcisa.
4. Uthungelwano lweNkolo enzulu (DBNs)
Uthungelwano lweeNkolo eziNzulu (DBNs) zi kukubhadla okungeyonyani iinkqubo ezinokufunda ukubona iipateni kwidatha. Bafeza oku ngokwahlula-hlula idatha ibe ngamacandelwana amancinci nancinci, befumana ukuqonda okucokisekileyo kumgangatho ngamnye.
Ii-DBN zinokufunda kwidatha ngaphandle kokwaziswa ukuba yintoni na (oku kubhekiselwa kuko “njengokufunda okungajongwanga”). Oku kubenza baxabiseke kakhulu ekubhaqeni iipateni kwidatha umntu anokuthi ayifumane inzima okanye ingenakwenzeka ukuyibona.
Yintoni eyenza ukuba ii-DBN zibaluleke?
Ii-DBN zibalulekile ngenxa yesakhono sazo sokufunda ukumelwa kwedatha yemigangatho. Ezi zibonisi zingasetyenziselwa iintlobo ngeentlobo zezicelo ezifana nokuhlelwa, ukufumanisa okungaqhelekanga, kunye nokunciphisa ubukhulu.
Amandla e-DBNs ukwenza uqeqesho lwangaphambili olungajongwanga, olunokwandisa ukusebenza kweemodeli zokufunda ezinzulu kunye nedatha encinci ebhaliweyo, yinzuzo ebalulekileyo.
Ziziphi ii-Aplikheshini zee-DBN?
Enye yezona zicelo zibalulekileyo ukubona into, apho ii-DBN zisetyenziselwa ukuqaphela iintlobo ezithile zezinto ezifana neenqwelomoya, iintaka kunye nabantu. Zikwasetyenziselwa ukwenza imifanekiso kunye nokuhlelwa, ukubonwa kwentshukumo kwiifilimu, kunye nokuqonda kolwimi lwendalo ukusetyenzwa kwelizwi.
Ngaphaya koko, ii-DBN ziqhele ukuqeshwa kwiiseti zedatha ukuvavanya ukuma kwabantu. Ii-DBN zisisixhobo esihle kumashishini ahlukeneyo, kubandakanywa ukhathalelo lwezempilo kunye neebhanki, kunye neteknoloji.
5. IiNethiwekhi zokuFunda ngokuNxibelela ngokunzulu (DRLs)
nzulu Ukomeleza ukuFunda Uthungelwano (DRLs) ludibanisa uthungelwano olunzulu lwe-neural kunye neendlela zokufunda okomeleza ukuvumela ii-arhente ukuba zifunde kwindawo enzima ngolingo nangempazamo.
Ii-DRL zisetyenziselwa ukufundisa iiarhente indlela yokwandisa umqondiso womvuzo ngokunxibelelana nendawo ezingqongileyo nokufunda kwiimpazamo zabo.
Yintoni Eyenza Ziphawuleke?
Ziye zasetyenziswa ngempumelelo kwizicelo ezahlukeneyo, kubandakanya imidlalo, iirobhothi, kunye nokuqhuba ukuzimela. Ii-DRL zibalulekile kuba ziyakwazi ukufunda ngokuthe ngqo kwigalelo le-sensory ekrwada, evumela ii-arhente ukuba zenze izigqibo ezisekelwe ekusebenzisaneni kwazo nokusingqongileyo.
Izicelo ezibalulekileyo
Ii-DRL ziqeshwe kwiimeko zangempela zehlabathi kuba ziyakwazi ukusingatha imiba enzima.
Ii-DRLs zibandakanyiwe kwisoftware edumileyo kunye neeplatifti zetekhnoloji, kubandakanya ne-OpenAI's Gym, Ubumbano ML-Arhente, kunye neLebhu ye-DeepMind kaGoogle. AlphaGo, eyakhiwe nguGoogle DeepMind, umzekelo, usebenzisa iDRL ukudlala umdlalo webhodi Yiya kwinqanaba leentshatsheli zehlabathi.
Olunye usetyenziso lwe-DRL lukwirobhothi, apho isetyenziselwa ukulawula ukunyakaza kweengalo zerobhothi ukwenza imisebenzi efana nokubamba izinto okanye iibhloko zokupakisha. Ii-DRLs zinosetyenziso oluninzi kwaye zisisixhobo esiluncedo iiarhente zoqeqesho ukufunda kwaye wenze izigqibo kwiimeko ezinzima.
6. Iikhowudi ezizenzekelayo
Iikhowudi ezizenzekelayo luhlobo olunomdla lwe inethiwekhi yomnatha oko kuye kwabamba umdla wabaphengululi kunye neenzululwazi zedatha. Ngokwesiseko ziyilelwe ukufunda indlela yokucinezela nokubuyisela idatha.
Idatha yegalelo isondliwa ngokulandelelana kweeleya ezinciphisa ngokuthe ngcembe i-dimensionality yedatha ide ixinzelelwe kwi-bottleneck layer kunye ne-nodes ezimbalwa kuneengeniso kunye neengqimba eziphumayo.
Olu phawu lucinezelweyo luthi ke lusetyenziswe ukwenza kwakhona idatha yegalelo loqobo kusetyenziswa ulandelelwano lweemaleko eziphakamisa ngokuthe ngcembe ubume bedatha bubuyele kubume bayo bokuqala.
Kutheni Ibalulekile?
Autoencoders yinxalenye ebalulekileyo ukufunda okunzulu kuba benza utsalo lwesici kunye nokunciphisa idatha kunokwenzeka.
Bayakwazi ukuchonga izinto eziphambili zedatha engenayo kwaye baziguqulele kwifom ecinezelweyo enokuthi isetyenziswe kweminye imisebenzi efana nokuhlela, ukudibanisa, okanye ukudala idatha entsha.
Sizisebenzisa phi iiKhowudi ezizenzekelayo?
Ukufunyaniswa okungaqhelekanga, ukusetyenzwa kolwimi lwendalo, kunye umbono wekhompyutha zimbalwa zezifundo apho kusetyenziswa ii-autoencoders. Ii-autoencoders, umzekelo, zinokusetyenziselwa ucinezelo lomfanekiso, ukukhupha umfanekiso, kunye nokudibanisa imifanekiso kumbono wekhompyuter.
Singasebenzisa ii-Autoencoders kwimisebenzi efana nokudalwa kokubhaliweyo, ukwahlulahlula okubhaliweyo, kunye noshwankathelo lokubhaliweyo kulungiso lolwimi lwendalo. Inokuchonga umsebenzi ongaqhelekanga kwidatha etenxayo kwisiqhelo ekuchongeni okungaqhelekanga.
7. Iinethiwekhi zeCapsule
I-Capsule Networks lulwakhiwo olutsha lokufunda olunzulu olwathi lwaphuhliswa njengotshintsho lweConvolutional Neural Networks (CNNs).
I-Capsule Networks zisekelwe kwingcinga yokuhlanganisa iiyunithi zobuchopho ezibizwa ngokuba zii-capsules ezinoxanduva lokuqaphela ubukho bento ethile emfanekisweni kunye nokufaka ikhowudi yeempawu zayo, ezifana nokuqhelaniswa kunye nokuma, kwii-vectors zazo eziphumayo. I-Capsule Networks ke ngoko inokulawula ukusebenzisana kwendawo kunye nokuguquguquka kwembono engcono kune-CNNs.
Kutheni sikhetha iiNethiwekhi zeCapsule ngaphezu kweCNN's?
IiNethiwekhi ze-Capsule ziluncedo kuba zoyisa ubunzima be-CNN ekubambeni ubudlelwane be-hierarchical phakathi kwezinto ezisemfanekisweni. Ii-CNNs zinokubona izinto ezinobungakanani obahlukeneyo kodwa ziyasokola ukuqonda ukuba ezi zinto zinxibelelana njani enye kwenye.
I-Capsule Networks, kwelinye icala, inokufunda ukuqaphela izinto kunye neziqwenga zazo, kunye nendlela ezibekwe ngayo kwindawo emfanekisweni, zibenze bakhuphisane ngokufanelekileyo kwizicelo zombono wekhompyuter.
Imimandla yezicelo
I-Capsule Networks sele ibonise iziphumo ezithembisayo kwiinkqubo ezahlukeneyo, kubandakanywa ukuhlelwa komfanekiso, ukuchongwa kwezinto, kunye nokwahlulwa kwemifanekiso.
Zisetyenziselwe ukwahlula izinto kwiifoto zonyango, ukuqaphela abantu kwiifilim, kunye nokwenza iimodeli ze-3D kwimifanekiso ye-2D.
Ukwandisa ukusebenza kwabo, iiNethiwekhi zeCapsule ziye zadityaniswa kunye nezinye izakhiwo ezinzulu zokufunda ezifana ne-Generative Adversarial Networks (GANs) kunye ne-Variational Autoencoders (VAEs). IiNethiwekhi zeCapsule ziqikelelwa ukuba zidlala indima ebaluleke kakhulu ekwandiseni itekhnoloji yembono yekhompyuter njengoko isayensi yokufunda nzulu iguquka.
Umzekelo; Nibabel sisixhobo esaziwayo sePython sokufunda nokubhala iintlobo zefayile ze-neuroimaging. Ukwahlulwa kwemifanekiso, isebenzisa iCapsule Networks.
8. Iimodeli ezisekwe kwingqalelo
Iimodeli zokufunda nzulu ezaziwa njengeemodeli ezisekwe kwingqalelo, ezikwabizwa ngokuba ziindlela zokuthathela ingqalelo, zizamela ukunyusa izinga lokuchaneka kwento. iimodeli zokufunda ngomatshini. Le mifuziselo isebenza ngokugxininisa kwiimpawu ezithile zedatha engenayo, ekhokelela ekusebenzeni ngokufanelekileyo nangokusebenzayo.
Kwimisebenzi yokulungisa ulwimi lwendalo efana nokuguqulelwa komatshini kunye nohlalutyo lweemvakalelo, iindlela zokuqwalaselwa zibonise ukuba ziphumelele kakhulu.
Iyintoni Intsingiselo Yazo?
Iimodeli ezisekwe kwingqalelo ziluncedo kuba zenza kube lula ukusetyenzwa kwedatha entsonkothileyo.
Uthungelwano lweNeural lwesiqhelo Vavanya yonke idatha yegalelo njengokubaluleka ngokulinganayo, okukhokelela ekuqhubeni kancinane kunye nokuncipha kokuchaneka. Iinkqubo zokuthathela ingqalelo zigxininisa kwimiba ebalulekileyo yedatha yokufaka, evumela uqikelelo olukhawulezayo noluchane ngakumbi.
Iindawo zokuSebenzisa
Kwinkalo yobukrelekrele bokwenziwa, iindlela zokuqwalaselwa zinoluhlu olubanzi lwezicelo, kubandakanywa ukusetyenzwa kolwimi lwendalo, ukuqondwa kwemifanekiso kunye nesandi, kunye nezithuthi ezingaqhubi.
Iindlela zokuthathela ingqalelo, umzekelo, zingasetyenziselwa ukuphucula uguqulelo lomatshini kulungiso lolwimi lwendalo ngokuvumela inkqubo ukuba igxile kumagama athile okanye amabinzana abalulekileyo kumxholo.
Iindlela zokuqwalaselwa kwiimoto ezizimeleyo zingaqeshwa ukunceda inkqubo ekugxininiseni kwizinto ezithile okanye imingeni kwindawo yayo.
9. Uthungelwano lweTransformer
Uthungelwano lweTransformer ziimodeli zokufunda ezinzulu eziphonononga kwaye zivelise ulandelelwano lwedatha. Basebenza ngokucwangcisa ulandelelwano lwegalelo into enye ngexesha kwaye bavelise ulandelelwano lwemveliso yobude obufanayo okanye obahlukeneyo.
Uthungelwano lweTransformer, ngokungafaniyo nemifuziselo esemgangathweni yokulandelelana ukuya kulandelelwano, musa ukwenza ulandelelwano usebenzisa i-neural networks (RNNs). Endaweni yoko, basebenzisa iinkqubo zokuzihoya ukuze bafunde amakhonkco phakathi kwamaqhekeza olandelelwano.
Yintoni Ukubaluleka Kothungelwano LweTransformer?
Uthungelwano lweTransformer lukhule ngokuthandwa kwiminyaka yakutshanje ngenxa yokusebenza kwabo ngcono kwimisebenzi yokulungisa ulwimi lwendalo.
Ziyifanele kakhulu imisebenzi yokudala imibhalo efana nokuguqulela ulwimi, ukushwankathela okubhaliweyo, kunye nokuveliswa kwencoko.
Uthungelwano lweTransformer lusebenza ngokubonakalayo ngakumbi xa kuthelekiswa neemodeli ezisekwe kwi-RNN, nto leyo ezenza zibe lukhetho olukhethwayo kwizicelo ezinkulu.
Ungazifumana phi iiNethiwekhi zeTransformer?
Uthungelwano lweTransformer lusetyenziswa ngokubanzi kuluhlu olubanzi lwezicelo, ngakumbi ukusetyenzwa kolwimi lwendalo.
Uthotho lwe-GPT (Generative Pre-trained Transformer) ngumzekelo obalaseleyo osekwe kwi-transformer esetyenziselwe imisebenzi efana nokuguqulela ulwimi, ukushwankathela okubhaliweyo, kunye nokuveliswa kwe-chatbot.
I-BERT (i-Bidirectional Encoder Representations ezivela kwiTransformers) yenye imodeli eqhelekileyo esekwe kwi-transformer esetyenziselwe izicelo zokuqonda ulwimi lwendalo ezifana nokuphendula imibuzo kunye nohlalutyo lweemvakalelo.
bobabini GPT kunye ne-BERT zadalwa kunye I-PyTorch, isikhokelo sokufunda esinzulu esivulelekileyo esiye sathandwa kakhulu ekuphuhliseni iimodeli ezisekelwe kwi-transformer.
10. Oomatshini beBoltzmann abathintelweyo(RBMs)
Oomatshini beBoltzmann abathintelweyo (RBMs) luhlobo lothungelwano lwe-neural olungagadwanga olufunda ngendlela yokuvelisa. Ngenxa yobuchule babo bokufunda kunye nokukhupha iimpawu ezibalulekileyo kwiidatha ezinomgangatho ophezulu, ziqeshwe ngokubanzi kwiinkalo zokufunda koomatshini kunye nokufunda nzulu.
Ii-RBM zenziwe ngamacandelo amabini, abonakalayo kwaye afihliweyo, kunye noluhlu ngalunye olubandakanya iqela lee-neurons ezidityaniswe ngamaphethelo anzima. Ii-RBMs ziyilelwe ukuba zifunde usasazo olunokwenzeka oluchaza idatha yegalelo.
Bathini oomatshini baseBoltzmann abathintelweyo?
Ii-RBM zisebenzisa isicwangciso sokufunda esizivelisayo. Kwi-RBMs, umaleko obonakalayo ubonisa idatha yegalelo, ngelixa umqolo ongcwatyiweyo ufaka iimpawu zedatha yegalelo. Ubunzima bemigangatho ebonakalayo kunye nefihliweyo ibonisa amandla ekhonkco yabo.
Ii-RBM zilungelelanisa iintsimbi kunye nokuthambekela phakathi kweeleya ngexesha loqeqesho kusetyenziswa ubuchule olwaziwa ngokuba yiyantlukwano eyahlukileyo. Ukwahluka okuchaseneyo sisicwangciso sokufunda esingajongwanga esonyusa ukubakhona kwemodeli yokuqikelela.
Yintoni intsingiselo yeeMishini zeBoltzmann eziThintelweyo?
Ii-RBM zibalulekile kwi yokufunda umatshini kunye nokufunda nzulu kuba banokufunda kwaye bakhuphe iimpawu ezifanelekileyo kwiimali ezinkulu zedatha.
Zisebenza kakhulu kumfanekiso kunye nokuqondwa kwentetho, kwaye ziqeshwe kwiintlobo ngeentlobo zezicelo ezifana neenkqubo zokuncoma, ukufumanisa okungaqhelekanga, kunye nokunciphisa ubukhulu. Ii-RBM zinokufumana iipateni kwiiseti zedatha ezininzi, ezikhokelela kuqikelelo oluphezulu kunye nokuqonda.
Banokusetyenziswa phi oomatshini baseBoltzmann abaMiselweyo?
Izicelo zee-RBMs ziquka ukuncitshiswa kobukhulu, ukubhaqwa okungaqhelekanga, kunye neenkqubo zokucebisa. Ii-RBMs ziluncedo ngakumbi kuhlalutyo lweemvakalelo kunye imodeli yesihloko kumxholo wokulungiswa kolwimi lwendalo.
Uthungelwano lweenkolelo ezinzulu, uhlobo lothungelwano lwe-neural olusetyenziselwa ukuqondwa kwelizwi kunye nomfanekiso, lukwasebenzisa ii-RBMs. I-Deep Belief Network Toolbox, TensorFlow, yaye Theano yeminye imizekelo ethile yesoftware okanye iteknoloji esebenzisa iiRBMs.
Songa
Iimodeli zokuFunda okuNzululwazi ziya zibaluleke ngakumbi nangakumbi kumashishini ahlukeneyo, kubandakanywa ukuqondwa kwentetho, ukusetyenzwa kolwimi lwendalo, kunye nombono wekhompyutha.
I-Convolutional Neural Networks (CNNs) kunye ne-Recurrent Neural Networks (RNNs) zibonise esona sithembiso kwaye zisetyenziswa kakhulu kwizicelo ezininzi, nangona kunjalo, zonke iimodeli ze-Deep Learning zineenzuzo kunye nokungalunganga.
Nangona kunjalo, abaphandi basajonge koomatshini beBoltzmann abaMiselweyo (RBMs) kunye nezinye iindidi zeemodeli zokuFunda okuNzululwazi kuba nazo zinezibonelelo ezikhethekileyo.
Iimodeli ezintsha kunye nezobuchule zilindeleke ukuba zidalwe njengoko indawo yokufunda enzulu iqhubela phambili ukujongana neengxaki ezinzima.
Shiya iMpendulo