Namhlanje sibona utshintsho kummandla wokulungiswa kolwimi lwendalo. Kwaye, kuqinisekileyo ukuba akukho kamva ngaphandle kobukrelekrele bokwenziwa. Sele sisebenzisa "abancedisi" be-AI ezahlukeneyo.
Ii-Chatbots ziyimizekelo emihle kwimeko yethu. Bamele ixesha elitsha lonxibelelwano. Kodwa, yintoni ebenza bakhetheke kangaka?
Ii-chatbots zangoku zinokuqonda kwaye ziphendule imibuzo yolwimi lwendalo ngokuchanekileyo kunye neenkcukacha njengeengcali zabantu. Kuyavuyisa ukufunda malunga neendlela ezingena kwinkqubo.
Bopha kwaye masifumanise itekhnoloji esemva kwayo.
Ukuntywila kwiTekhnoloji
I-AI Transformers ligama eliphambili kule ndawo. Banje amanethiwekhi eye yaguqula inkqubo yolwimi lwendalo. Ngokwenyani, kukho ukungqinelana okukhulu koyilo phakathi kweenguqu ze-AI kunye neenethiwekhi ze-neural.
Zombini zenziwe ngamaleko aliqela eeyunithi zokusetyenzwa ezenza uthotho lwezibalo ukuguqula idatha yegalelo ibe yingqikelelo njengemveliso. Kule post, siza kujonga amandla e-AI Transformers kunye nendlela abalitshintsha ngayo ihlabathi elisingqongileyo.
Ukubanakho kokuPhuculwa koLwimi lweNdalo
Masiqale ngezinto ezisisiseko. Siyiva yonke indawo phantse. Kodwa, yintoni kanye kanye ukusetyenzwa kolwimi lwendalo?
Licandelo le kukubhadla okungeyonyani egxininisa kunxibelelwano lwabantu kunye noomatshini ngokusetyenziswa kolwimi lwendalo. Injongo kukuvumela iikhompyutha ukuba zibone, zitolike, kwaye zivelise ulwimi lwabantu ngendlela enentsingiselo nenyanisekileyo.
Ukuqondwa kwentetho, ukuguqulelwa kolwimi, Uhlalutyo lweemvakalelo, kunye nesishwankathelo sombhalo yonke imizekelo yezicelo ze-NLP. Iimodeli zeNLP zesiNtu, kwelinye icala, ziye zasokola ukubamba amakhonkco antsonkothileyo phakathi kwamagama kwibinzana. Oku kwenza amanqanaba aphezulu okuchaneka kwimisebenzi emininzi ye-NLP ayinakwenzeka.
Oku kuxa i-AI Transformers ingena emfanekisweni. Ngenkqubo yokuziqwalasela, abaguquli banokubhala ukuxhomekeka kwexesha elide kunye nokudibanisa phakathi kwamagama kwibinzana. Le ndlela yenza ukuba imodeli ikhethe ukuya kumacandelo ahlukeneyo olandelelwano lwegalelo. Ngoko ke, inokuqonda umongo nentsingiselo yegama ngalinye kwibinzana.
Yintoni kanye kanye iiModeli zeTransformers
Isiguquli se-AI yi ukufunda okunzulu i-architecture eqondayo kunye neenkqubo ezahlukeneyo zolwazi. Iyagqwesa ekumiseleni indlela amasuntswana olwazi anxibelelana ngayo, njengokuba amagama ahlukeneyo kwibinzana adityaniswa njani okanye adityaniswa njani amacandelo omfanekiso.
Isebenza ngokwahlula-hlula ulwazi lube ngamasuntswana amancinci kwaye emva koko ujonge onke la macandelo ngaxeshanye. Kufana nokuba iirobhothi ezininzi ezincinci ziyasebenzisana ukuqonda idatha. Okulandelayo, yakuba iyazi yonke into, iphinda ihlanganise onke amacandelo ukuze inike impendulo okanye isiphumo.
Iziguquli ze-AI zixabiseke kakhulu. Bayakwazi ukuwuqonda umxholo kunye namakhonkco exesha elide phakathi kolwazi olwahlukeneyo. Oku kubalulekile kwimisebenzi efana noguqulo lolwimi, isishwankathelo, kunye nokuphendula imibuzo. Ke, bangobuchopho obusemva kwezinto ezininzi ezinomdla ezinokufezwa yi-AI!
Ingqalelo kuko konke okufunayo
Isihlokwana esithi "Ingqalelo yiyo yonke into oyifunayo" ibhekisela kupapasho lwe-2017 olucebise imodeli ye-transformer. Itshintshe indlela yokuziphatha kolwimi lwendalo (NLP).
Ababhali bolu phando bathi imodeli ye-transformer yokuziqwalasela yomelele ngokwaneleyo ukuba ithathe indima yesiqhelo kunye uthungelwano lwe-neural convolution isetyenziselwa imisebenzi ye-NLP.
Yintoni Ukuzihoya kanye?
Yindlela evumela imodeli ukuba igxile kumacandelo ahlukeneyo olandelelwano lwegalelo xa isenza uqikelelo.
Ngamanye amazwi, ukuziqwalasela kunceda imodeli ukubala iseti yamanqaku engqwalasela yento nganye malunga nawo onke amanye amacandelo, ivumela imodeli ukuba ilinganise ukubaluleka kwento nganye yokufaka.
Kwindlela esekwe kwi-transformer, ukuzithathela ingqalelo kusebenza ngolu hlobo lulandelayo:
Ulandelelwano lwegalelo kuqala luzinziswe kuluhlu lwee-vectors, enye kwilungu ngalinye lolandelelwano.
Kwinto nganye elandelelanayo, imodeli yenza iiseti ezintathu zevektha: i-vector yombuzo, i-vector engundoqo, kunye nevektha yexabiso.
I-vector yombuzo ithelekiswa nazo zonke ii-vectors eziphambili, kwaye ukufana kubalwa kusetyenziswa imveliso yamachaphaza.
Amanqaku engqwalasela aphumela ekubeni aqheleke kusetyenziswa umsebenzi we-softmax, ovelisa isethi yobunzima obubonisa ukubaluleka kwesiqwenga ngasinye ngokulandelelana.
Ukudala ukubonakaliswa kokugqibela kwemveliso, i-vectors yexabiso iphindaphindwe ngobunzima bengqwalasela kwaye ishwankathelwe.
Imifuziselo esekwe kwiTransformer, esebenzisa ukuzihoya, inokubamba ngempumelelo ubudlelwane bexesha elide kulandelelwano lwegalelo ngaphandle kokuxhomekeka kwiifestile zobude obusisigxina, ezibenza zibe luncedo ngakumbi kwizicelo zokwenziwa kolwimi lwendalo.
umzekelo
Cinga ukuba sinolandelelwano lweempawu ezithandathu: "Ikati yahlala emethini." Ithokheni nganye inokumelwa njengevektha, kwaye ulandelelwano lwegalelo lunokubonwa ngolu hlobo lulandelayo:
Okulandelayo, kwithokheni nganye, siza kwakha iiseti ezintathu zevektha: i-vector yombuzo, i-vector engundoqo, kunye nevektha yexabiso. Ivektha yethokheni ehlanganisiweyo iphindaphindwa ngamatriki amathathu obunzima obufundiweyo ukuvelisa ezi vektha.
Kuphawu lokuqala "I," umzekelo, umbuzo, isitshixo, kunye nevectors zexabiso ziya kuba:
Ivector yombuzo: [0.4, -0.2, 0.1]
Ivector ephambili: [0.2, 0.1, 0.5]
Ivektha yexabiso: [0.1, 0.2, 0.3]
Amanqaku engqwalasela phakathi kwebini ngalinye lamathokheni kulandelelwano lwegalelo lubalwa ngomatshini wokuziqwalasela. Umzekelo, amanqaku okuqwalaselwa phakathi kwamathokheni 1 kunye no-2 "I" aya kubalwa njengemveliso yamachaphaza yombuzo wabo kunye neevektha eziphambili:
Inqaku lengqalelo = dot_product(I-Vector yombuzo we-Token 1, iVector engundoqo ye-Token 2)
= (0.4 * 0.8) + (-0.2 * 0.2) + (0.1 * 0.1)
= 0.31
La manqaku engqalelo abonisa ukuhambelana okunxulumene nomqondiso ngamnye ngokulandelelana kwabanye.
Ekugqibeleni, kwithokheni nganye, ukubonakaliswa kwemveliso kudalwa ngokuthatha isixa esinexabiso leevetha zexabiso, kunye nemilinganiselo enqunywe ngamanqaku okuqwalaselwa. Umboniso wemveliso wethokheni yokuqala "I," umzekelo, iya kuba:
Ivektha yemveliso yoMqondiso 1 = (Inqaku lokuqwalaselwa ngoMqondiso 1) * Ivektha yexabiso yoMqondiso 2
+ (Amanqaku okuqwalaselwa ngoMqondiso wesi-3) * Ivektha yexabiso yoMqondiso wesi-3
+ (Amanqaku okuqwalaselwa ngoMqondiso wesi-4) * Ivektha yexabiso yoMqondiso wesi-4
+ (Amanqaku okuqwalaselwa ngoMqondiso wesi-5) * Ivektha yexabiso yoMqondiso wesi-5
+ (Amanqaku okuqwalaselwa ngoMqondiso wesi-6) * Ivektha yexabiso yoMqondiso wesi-6
= (0.31 * [0.1, 0.2, 0.3]) + (0.25 * [0.2, -0.1, 0.7]) + (0.08 * [0.3, 0.5, -0.1]) + (0.14 * [0.1, 0.3, -0.2] ) + (0.22 * [0.6, -0.3, 0.4])
= [0.2669, 0.1533, 0.2715]
Njengomphumo wokuziqwalasela, imodeli esekelwe kwi-transformer inokukhetha ukuya kumacandelo ahlukeneyo okulandelelana kwegalelo xa udala ukulandelelana kwemveliso.
Izicelo zingaphezulu kunoko ucinga
Ngenxa yokuguquguquka kwabo kunye nokukwazi ukuphatha uluhlu olubanzi lwemisebenzi ye-NLP, njengokuguqulelwa komatshini, uhlalutyo lweemvakalelo, ukushwankathela okubhaliweyo, kunye nokunye, abaguquli be-AI baye bakhula ekuthandeni kwiminyaka yamuva.
Iziguquli ze-AI zisetyenziswe kwiinkalo ezahlukeneyo, kubandakanywa ukuqatshelwa kwemifanekiso, iinkqubo zokucebisa, kunye nokufunyanwa kweziyobisi, ngaphezu kwezicelo ezisekelwe kulwimi lwakudala.
Iziguquli ze-AI zinosetyenziso olungenamda kuba zinokulungelelaniswa kwiindawo ezininzi zeengxaki kunye neentlobo zedatha. Abaguquli be-AI, kunye namandla abo okuhlalutya ulandelelwano lwedatha olunzima kunye nokubamba ubudlelwane bexesha elide, lubekwe ukuba lube yinto ebalulekileyo yokuqhubela phambili ekuphuhlisweni kwezicelo ze-AI kwiminyaka ezayo.
Ukuthelekiswa nezinye iiNeural Network Architectures
Njengoko bekwazi ukuhlalutya ulandelelwano lwegalelo kwaye babambe ubudlelwane obude kumbhalo, abaguquli be-AI bafaneleke ngakumbi ukusetyenzwa kolwimi lwendalo xa kuthelekiswa nezinye izicelo zenethiwekhi ye-neural.
Olunye ulwakhiwo lwenethiwekhi ye-neural, efana ne-convolutional neural networks (CNNs) kunye ne-recurrent neural networks (RNNs), kwelinye icala, ifaneleka ngcono imisebenzi ebandakanya ukusetyenzwa kwegalelo elicwangcisiweyo, njengemifanekiso okanye idatha yothotho lwexesha.
Ikamva liJonga Liqaqambile
Ikamva labaguquli be-AI libonakala liqaqambile. Enye inkalo yophononongo oluqhubekayo kuphuhliso lweemodeli ezinamandla ngokuqhubekayo ezikwaziyo ukusingatha imisebenzi entsonkothileyo.
Ngaphezu koko, iinzame ziyenziwa ukudibanisa i-AI transformers kunye nobunye ubuchwepheshe be-AI, obufana nokuqiniswa ukufunda, ukubonelela ngezakhono zokuthatha izigqibo eziphucukileyo.
Ishishini ngalinye lizama ukusebenzisa amandla e-AI ukuqhubela phambili izinto ezintsha kunye nokufezekisa ukhuphiswano. Ke, abaguquli be-AI kunokwenzeka ukuba badityaniswe ngokuqhubekayo kwizicelo ezahlukeneyo, kubandakanya ukhathalelo lwempilo, imali, kunye nezinye.
Ngophuculo oluqhubekayo kwitekhnoloji yokuguqula i-AI kunye nokubanakho kwezi zixhobo zinamandla ze-AI ukuguqula indlela abantu abaqhuba ngayo kunye nokuqonda ulwimi, ikamva libonakala liqaqambile.
Shiya iMpendulo