Ukuqeqesha amamodeli e-AI: Usesho lokuthuthukiswa kwe-AI okunezindleko eziphansi

• About the Author:
• Igama lomlobi

  I-Quantumrun Foresight
• Mashi 21, 2023

Ukufunda okujulile (i-DL) kufakazele ukuthi kuyisixazululo esinekhono ezinseleleni ezimbalwa ekuthuthukisweni kobuhlakani bokwenziwa (AI). Kodwa-ke, i-DL nayo ibiza kakhulu. Ukusebenza amanethiwekhi e-neural ajulile kudinga izinsiza eziphezulu zokucubungula, ikakhulukazi ekuqeqeshweni kwangaphambili. Okubi nakakhulu, le nqubo ebamba amandla isho ukuthi lezi zidingo ziholela ekulandeleni kwekhabhoni enkulu, kulimaze izilinganiso ze-ESG zokuthengisa ucwaningo lwe-AI.

Ukuqeqesha amamodeli e-AI umongo

Ukuqeqeshwa kwangaphambili manje kuyindlela ethandwa kakhulu yokwakha amanethiwekhi amakhulu e-neural, futhi kubonise impumelelo enkulu kumbono wekhompyutha (CV) kanye nokucubungula ulimi lwemvelo (NLP). Kodwa-ke, ukwakhiwa kwamamodeli amakhulu e-DL kuye kwabiza kakhulu. Isibonelo, ukuqeqesha i-OpenAI's Generative Pre-trained Transformer 3 (GPT-3), enamapharamitha ayizigidi eziyizinkulungwane eziyi-175 futhi edinga ukufinyelela kumaqoqo amaseva amakhulu anamakhadi ezithombe ezisezingeni eliphezulu, ibe nenani elilinganiselwe le-USD $12 million. Iseva enamandla namakhulu amagigabhayithi ememori yokufinyelela engahleliwe yevidiyo (VRAM) nayo iyadingeka ukuze kuqalise imodeli.

Ngenkathi izinkampani ezinkulu zobuchwepheshe zingakwazi ukukhokhela izindleko ezinjalo zokuqeqesha, kuba nqabela ukuqalisa okuncane nezinhlangano zocwaningo. Kunezici ezintathu ezidala lezi zindleko. 

1. Izindleko zokubala ezibanzi, ezingadinga amaviki ambalwa nezinkulungwane zamayunithi okucubungula izithombe (GPUs).

2. Amamodeli acushwe kahle adinga isitoreji esikhulu, ngokuvamile athatha amakhulukhulu amagigabhayithi (GBs). Ngaphezu kwalokho, amamodeli amaningi emisebenzi ehlukene adinga ukugcinwa.

3. Ukuqeqesha amamodeli amakhulu kudinga amandla okuhlanganisa anembile nehadiwe; ngaphandle kwalokho, imiphumela ingase ingabi mihle.

Ngenxa yezindleko ezingavumelekile, ucwaningo lwe-AI seluya ngokuya luthengiswa, lapho izinkampani zeBig Tech zihola izifundo kulo mkhakha. Lezi zinkampani nazo zizozuza kakhulu kulokho ezikutholile. Khonamanjalo, izikhungo zocwaningo kanye nezinhlangano ezingenzi inzuzo ngokuvamile kufanele zisebenzisane nalawa mabhizinisi uma zifuna ukwenza ukuhlola kwazo kulo mkhakha. 

Umthelela ophazamisayo

Kunobufakazi obuphakamisa ukuthi amanethiwekhi e-neural "angathenwa." Lokhu kusho ukuthi ngaphakathi kwamanethiwekhi amakhulu we-neural, iqembu elincane lingafinyelela izinga elifanayo lokunemba njengemodeli ye-AI yoqobo ngaphandle kwemithelela enzima ekusebenzeni kwayo. Isibonelo, ngo-2020, abacwaningi be-AI e-Swarthmore College nase-Los Alamos National Laboratory babonisa ukuthi noma imodeli eyinkimbinkimbi ye-DL ingafunda ukubikezela izinyathelo zesikhathi esizayo ku-Game of Life yesazi sezibalo uJohn Conway, kuhlala kunenethiwekhi encane ye-neural engafundiswa. ukwenza into efanayo.

Abacwaningi bathole ukuthi uma belahla imingcele eminingi yemodeli ye-DL ngemva kokuba isiqede yonke inqubo yokuqeqeshwa, bangakwazi ukuyehlisa ibe amaphesenti angu-10 osayizi wayo wangempela futhi bazuze umphumela ofanayo. Izinkampani eziningi zobuchwepheshe sezivele zicindezela amamodeli azo e-AI ukuze zonge isikhala kumadivayisi afana namalaptop nama-smartphone. Le ndlela ayigcini nje ngokugcina imali kodwa futhi ivumela isofthiwe ukuthi isebenze ngaphandle koxhumano lwe-inthanethi futhi ithole imiphumela ngesikhathi sangempela. 

Kuphinde kwaba nezimo lapho i-DL ibingenzeka khona kumadivayisi anikwe amandla amabhethri elanga noma amaseli wezinkinobho, ngenxa yamanethiwekhi amancane e-neural. Kodwa-ke, umkhawulo wendlela yokuthena ukuthi imodeli isadinga ukuqeqeshwa ngokuphelele ngaphambi kokuba incitshiswe. Kube nezifundo ezithile zokuqala kuma-neural subsets ezingaqeqeshwa ngokwazo. Kodwa-ke, ukunemba kwawo akufani nalawo amanethiwekhi e- supersized neural.

Imithelela yokuqeqeshwa kwamamodeli e-AI

Imithelela ebanzi yokuqeqeshwa kwamamodeli e-AI ingafaka: 

• Ucwaningo olwandisiwe ngezindlela ezahlukene zokuqeqesha amanethiwekhi e-neural; nokho, inqubekelaphambili ingase ihambe kancane ngenxa yokuntuleka kwezimali.
• Ubuchwepheshe obukhulu buyaqhubeka nokuxhasa ngezimali amalebhu abo ocwaningo lwe-AI, okuholela ekungqubuzaneni okwengeziwe kwezintshisekelo.
• Izindleko zokuthuthukiswa kwe-AI zidala izimo zokuthi ama-monopoly akheke, zikhawulela amandla okuqalisa okusha kwe-AI ukuncintisana ngokuzimela namafemu ezobuchwepheshe amisiwe. Isimo samabhizinisi asafufusa singase sibone idlanzana lamafemu amakhulu ezobuchwepheshe athuthukisa amamodeli amakhulu we-AI obunikazi futhi awaqashele amafemu amancane e-AI njengesevisi/usizo.
• Izikhungo zocwaningo, izinhlangano ezingenzi nzuzo, namanyuvesi axhaswe ngobuchwepheshe obukhulu ukuze enze ezinye izivivinyo ze-AI egameni lazo. Lo mkhuba ungaholela ekuchithekeni kwengqondo okwengeziwe kusuka ezifundweni kuya ezinkampanini.
• Ingcindezi eyengeziwe yokuthi ubuchwepheshe obukhulu bushicilele futhi bubuyekezwe njalo imihlahlandlela yabo yokuziphatha ye-AI ukuze ibenze baphendule ngocwaningo lwabo nemiklamo yabo yentuthuko.
• Ukuqeqesha amamodeli e-AI kuya ebiza kakhulu njengoba amandla aphezulu ekhompiyutha edingeka, okuholela ekukhishweni kwekhabhoni okwengeziwe.
• Ezinye izikhungo zikahulumeni ezizama ukulawula idatha esetshenziswa ekuqeqeshweni kwalawa mamodeli amakhulu we-AI. Futhi, ama-ejensi okuncintisana angakha umthetho ophoqelela amamodeli e-AI osayizi othize ukuba enziwe afinyeleleke ezinkampanini ezincane zasekhaya ngomzamo wokukhuthaza ukusungulwa kwama-SME.

Imibuzo okufanele icatshangelwe

• Uma usebenza emkhakheni we-AI, inhlangano yakho iwathuthukisa kanjani amamodeli e-AI esimeme ngokwemvelo?
• Yimiphi imiphumela yesikhathi eside engaba khona yamamodeli e-AI abizayo?