Inkinga yakudala kubuhlakani bokwenziwa ukuphishekela umshini okwazi ukuqonda ulimi lwabantu.
Isibonelo, lapho usesha "izindawo zokudlela eziseduze zesi-Italian" enjinini yakho yokusesha eyintandokazi, i-algorithm kufanele ihlaziye igama ngalinye embuzweni wakho bese ikhipha imiphumela efanele. Uhlelo lokusebenza lokuhumusha oluhloniphekile kuzodingeka luqonde umongo wegama elithile ngesiNgisi futhi ngandlela thize lubhekele umehluko wohlelo lolimi phakathi kwezilimi.
Yonke le misebenzi nokunye okuningi kuwela ngaphansi kwesizinda sesayensi yekhompyutha eyaziwa ngokuthi Ukucutshungulwa kolimi ngokwemvelo noma i-NLP. Intuthuko ku-NLP iholele ekusetshenzisweni okubanzi okusebenzayo okuvela kubasizi ababonakalayo njenge-Amazon's Alexa kuya kuzihlungi zogaxekile ezithola ama-imeyili anonya.
Ukuphumelela kwakamuva ku-NLP umqondo wokuthi a imodeli yolimi olukhulu noma i-LLM. Ama-LLM afana ne-GPT-3 abe namandla kangangokuthi abonakala ephumelela cishe kunoma yimuphi umsebenzi we-NLP noma icala lokusebenzisa.
Kulesi sihloko, sizobheka ukuthi ayini ngempela ama-LLM, ukuthi la mamodeli aqeqeshwa kanjani, kanye nemikhawulo yamanje anayo.
Ithini imodeli yolimi enkulu?
Emgogodleni wayo, imodeli yolimi imane iyi-algorithm eyazi ukuthi mangakanani amathuba okuthi ukulandelana kwamagama kube umusho ovumelekile.
Imodeli yolimi elula kakhulu eqeqeshelwe ezincwadini ezingamakhulu ambalwa kufanele ikwazi ukusho ukuthi "Wagoduka" iyasebenza kakhulu kunokuthi "Waya Ekhaya".
Uma sishintsha idathasethi encane ngokuqhathaniswa nedathasethi enkulu ekhishwe ku-inthanethi, siqala ukusondela embonweni we imodeli yolimi olukhulu.
Ukusebenzisa amanethiwekhi we-neural, abacwaningi bangaqeqesha ama-LLM ngenani elikhulu ledatha yombhalo. Ngenxa yenani ledatha yombhalo imodeli elibonile, i-LLM iba yinhle kakhulu ekubikezeleni igama elilandelayo ngokulandelana.
Imodeli iba yinkimbinkimbi kakhulu, ingenza imisebenzi eminingi ye-NLP. Le misebenzi ihlanganisa ukufinyeza umbhalo, ukwakha okuqukethwe kwenoveli, ngisho nokulingisa ingxoxo efana neyomuntu.
Isibonelo, imodeli yolimi edume kakhulu ye-GPT-3 iqeqeshwe ngamapharamitha angaphezu kwezigidigidi ezingu-175 futhi ibhekwa njengemodeli yolimi ethuthuke kakhulu kuze kube manje.
Iyakwazi ukukhiqiza ikhodi yokusebenza, ibhale zonke izindatshana, futhi ingathatha isibhamu ekuphenduleni imibuzo nganoma yisiphi isihloko.
Aqeqeshwa Kanjani Ama-LLM?
Sithinte kafushane iqiniso lokuthi ama-LLM akweleta amandla awo amaningi ngosayizi wedatha yawo yokuqeqeshwa. Kunesizathu esenza sizibize ngezibonelo zezilimi “ezinkulu” phela.
Ukuqeqeshwa kwangaphambili nge-Transformer Architecture
Phakathi nesigaba sangaphambi kokuqeqeshwa, ama-LLM ethulwa kudatha yombhalo ekhona ukuze kufundwe isakhiwo nemithetho ejwayelekile yolimi.
Eminyakeni embalwa edlule, ama-LLM aqeqeshwe kusengaphambili kumadathasethi afaka ingxenye enkulu ye-inthanethi yomphakathi. Isibonelo, imodeli yolimi ye-GPT-3 yaqeqeshwa kudatha evela ku- Ukukhasa Okuvamile Idathasethi, iqoqo lokuthunyelwe kwewebhu, amakhasi ewebhu, nezincwadi ezenziwe ngedijithali ezikhishwe ezizindeni ezingaphezu kwezigidi ezingu-50.
Idathasethi enkulu ibe isifakwa kumodeli eyaziwa ngokuthi a I-transformer. Ama-Transformer awuhlobo lwe inethiwekhi ye-neural ejulile okusebenza kangcono kudatha elandelanayo.
Ama-Transformer asebenzisa i- isakhiwo se-encoder-decoder ngokubamba okokufaka nokuphumayo. Empeleni, i-transformer iqukethe amanethiwekhi amabili e-neural: i-encoder ne-decoder. Isishumeki singakhipha incazelo yombhalo ofakiwe futhi siwugcine njengevekhtha. Idekhoda ibe isithola i-vector bese ikhiqiza incazelo yayo yombhalo.
Kodwa-ke, umqondo oyinhloko ovumele ukwakhiwa kwe-transformer ukuthi kusebenze kahle kakhulu ukungezwa kwe-a indlela yokuzinaka. Umqondo wokuzinaka uvumele imodeli ukuthi inake amagama abaluleke kakhulu emshweni othile. Lo mshini ucabangela ngisho nesisindo phakathi kwamagama aqhelelene ngokulandelana.
Enye inzuzo yokuzinaka ukuthi inqubo ingahambisana. Esikhundleni sokucubungula idatha elandelanayo, amamodeli we-transformer angacubungula konke okokufaka ngesikhathi esisodwa. Lokhu kwenza ama-transformer akwazi ukuqeqesha ngenani elikhulu ledatha ngokushesha uma kuqhathaniswa nezinye izindlela.
Ukuhleleka Kwayo
Ngemva kwesigaba sokuqeqeshwa kwangaphambili, ungakhetha ukwethula umbhalo omusha we-LLM eyisisekelo ozoqeqeshelwa kuwo. Le nqubo siyibiza ukuhleleka Kwayo futhi kuvame ukusetshenziselwa ukuthuthukisa umkhiqizo we-LLM emsebenzini othile.
Isibonelo, ungase ufune ukusebenzisa i-LLM ukuze ukhiqize okuqukethwe kwe-akhawunti yakho ye-Twitter. Singahlinzeka imodeli ngezibonelo ezimbalwa zama-tweets akho adlule ukuze siyinikeze umbono wokukhiphayo okufunayo.
Kunezinhlobo ezimbalwa ezahlukene zokulungisa kahle.
Ukufunda okumbalwa isho inqubo yokunikeza imodeli inombolo encane yezibonelo kulindelekile ukuthi imodeli yolimi ithole ukuthi ingenziwa kanjani into efanayo. Ukufunda okukodwa kuyinqubo efanayo ngaphandle kwesibonelo esisodwa kuphela esinikeziwe.
Imikhawulo Yamamodeli Olimi Olukhulu
Ama-LLM afana ne-GPT-3 ayakwazi ukwenza inombolo enkulu yezimo zokusebenzisa ngisho nangaphandle kokucushwa kahle. Kodwa-ke, lawa mamodeli asafika nesethi yawo yokulinganiselwa.
Ukuntuleka Kokuqonda KweSemantic Komhlaba
Phezulu, ama-LLM abonakala ebonisa ubuhlakani. Kodwa-ke, lawa mamodeli awasebenzi ngendlela efanayo ubuchopho bomuntu kwenza. Ama-LLM athembele kuphela ekubalweni kwezibalo ukuze akhiqize okukhiphayo. Abanawo amandla okucabanga imibono nemibono bebodwa.
Ngenxa yalokhu, i-LLM ingakhipha izimpendulo ezingenangqondo ngoba nje amagama abonakala “elungile” noma “ngokwezibalo kungenzeka” uma ebekwe ngalolo hlelo.
ukusangana
Amamodeli afana ne-GPT-3 nawo ahlushwa yizimpendulo ezingalungile. Ama-LLM angahlushwa into eyaziwa ngokuthi ukukhanyiselwa lapho amamodeli akhipha impendulo engalungile ngokweqiniso ngaphandle kokuqaphela ukuthi impendulo ayinasisekelo eqinisweni.
Isibonelo, umsebenzisi angacela imodeli ukuthi ichaze imicabango kaSteve Jobs ku-iPhone yakamuva. Imodeli ingase ikhiqize isilinganiso esivela emoyeni omncane ngokusekelwe kudatha yayo yokuqeqeshwa.
Ukuchema kanye Nolwazi Olulinganiselwe
Njengamanye ama-algorithms amaningi, amamodeli ezilimi amakhulu athambekele ekuzuzeni ifa lokuchema elikhona kudatha yokuqeqeshwa. Njengoba siqala ukuthembela kakhulu kuma-LLM ukuze sibuyise ulwazi, abathuthukisi balawa mamodeli kufanele bathole izindlela zokunciphisa imiphumela engaba yingozi yezimpendulo ezichemile.
Ngendlela efanayo, ama-blindpots edatha yokuqeqeshwa kwemodeli izophinde ivimbele imodeli ngokwayo. Njengamanje, amamodeli ezilimi amakhulu athatha izinyanga ukuqeqeshwa. Lawa mamodeli aphinde ancike kumasethi edatha anomkhawulo kububanzi. Yingakho i-ChatGPT inolwazi olulinganiselwe lwezehlakalo ezenzeke ngo-2021 odlule.
Isiphetho
Amamodeli ezilimi amakhulu anamandla okushintsha ngempela indlela esisebenzisana ngayo nobuchwepheshe kanye nomhlaba wethu uwonke.
Inani elikhulu ledatha elitholakala ku-inthanethi linikeze abacwaningi indlela yokubonisa ubunkimbinkimbi bolimi. Kodwa-ke, endleleni, lezi zibonelo zezilimi zibonakala zithatha ukuqonda okufana nomuntu mayelana nomhlaba njengoba unjalo.
Njengoba umphakathi uqala ukwethemba lezi zinhlobo zezilimi ukuze zinikeze okukhiphayo okunembile, abacwaningi nabathuthukisi sebevele bathola izindlela zokwengeza ama-guardrails ukuze ubuchwepheshe buhlale buhambisana nezimiso zokuziphatha.
Ucabanga ukuthi liyini ikusasa lama-LLM?
shiya impendulo