Vokenization: Ulimi i-AI engalubona

ISIKWELETU SESITHOMBE:
Isikweletu sezithombe
iStock

Vokenization: Ulimi i-AI engalubona

Vokenization: Ulimi i-AI engalubona

Umbhalo wesihlokwana
Njengoba izithombe manje sezihlanganiswa nokuqeqeshwa kwezinhlelo ze-artificial intelligence (AI), amarobhothi angase akwazi “ukubona” imiyalo maduze.
    • About the Author:
    • Igama lomlobi
      I-Quantumrun Foresight
    • Kwangathi 9, 2023

    Ukucutshungulwa kolimi lwemvelo (NLP) kunikeze amandla amasistimu obuhlakani bokwenziwa (AI) ukuthi afunde inkulumo yomuntu ngokuqonda amagama nokuqhathanisa umongo nomuzwa. Okubi kuphela ukuthi lezi zinhlelo ze-NLP zisekelwe emibhalweni kuphela. I-Vokenization isizokushintsha konke lokho.

    Umongo we-Vokenization

    Izinhlelo ezimbili zokufunda ngomshini (ML) ezisekelwe embhalweni zivame ukusetshenziselwa ukuqeqesha i-AI ukuze icubungule futhi iqonde ulimi lwabantu: I-OpenAI's Generative Pre-trained Transformer 3 (GPT-3) kanye ne-BERT ye-Google (Ukumela Okubhalwe Ngekhodi Okubili okuvela ku-Transformers). Kumagama e-AI, amagama asetshenziswe ekuqeqeshweni kwe-NLP abizwa ngokuthi amathokheni. Abacwaningi abavela eNyuvesi yaseNorth Carolina (UNC) babone ukuthi izinhlelo zokuqeqesha ezisekelwe emibhalweni zilinganiselwe ngoba abakwazi "ukubona," okusho ukuthi azikwazi ukuthatha ulwazi olubonakalayo nokuxhumana. 

    Isibonelo, uma othile ebuza i-GPT-3 ukuthi uyini umbala wezimvu, uhlelo luvame ukuphendula "olumnyama" ngisho noma lumhlophe ngokusobala. Le mpendulo ingenxa yokuthi uhlelo olusekelwe emibhalweni luzoyihlanganisa negama elithi "izimvu ezimnyama" esikhundleni sokukhomba umbala ofanele. Ngokuhlanganisa okubonwayo namathokheni (voken), amasistimu e-AI angaba nokuqonda okuphelele kwamagama. I-Vokenization ihlanganisa ama-voken ezinhlelweni ze-NLP ezigadiwe, ezivumela ukuthi zithuthukise "umqondo ojwayelekile."

    Ukuhlanganisa amamodeli olimi nombono wekhompyutha akuwona umqondo omusha, futhi kuyinkambu ekhula ngokushesha ocwaningweni lwe-AI. Inhlanganisela yalezi zinhlobo ezimbili ze-AI isebenzisa amandla azo ngamanye. Amamodeli olimi afana ne-GPT-3 aqeqeshwa ngokufunda okungagadiwe, okubavumela ukuba balinganise kalula. Ngokuphambene, amamodeli ezithombe njengamasistimu okubona into angafunda ngokuqondile kokungokoqobo futhi awanciki ekunqanyulweni okunikezwa umbhalo. Isibonelo, amamodeli ezithombe angabona ukuthi imvu imhlophe ngokubuka isithombe.

    Umthelela ophazamisayo

    Inqubo ye-vokenization ilula kakhulu. Ama-Vokens adalwa ngokunikeza izithombe ezihambisanayo noma ezifanele kumathokheni olimi. Bese, ama-algorithms (i-vokenizer) aklanyelwe ukukhiqiza ama-vokens ngokufunda okungagadiwe (ayikho imingcele ecacile/imithetho). I-Common AI eqeqeshwe ngokusebenzisa i-vokenization ingaxhumana futhi ixazulule izinkinga kangcono ngoba inokuqonda okujulile komongo. Le ndlela iyingqayizivele ngoba ayibikezeli kuphela amathokheni olimi kodwa futhi ibikezela amathokheni ezithombe, okuyinto amamodeli we-BERT wendabuko angakwazi ukuyenza.

    Isibonelo, abasizi be-robotic bazokwazi ukubona izithombe futhi bazulazule izinqubo kangcono ngoba bayakwazi "ukubona" ​​okudingekayo kubo. Amasistimu obuhlakani bokwenziwa aqeqeshelwe ukubhala okuqukethwe azokwazi ukwenza ama-athikili azwakala njengomuntu owengeziwe, anemibono egeleza kangcono, esikhundleni semisho ehlukene. Uma kucatshangelwa ukufinyelela okubanzi kwezinhlelo zokusebenza ze-NLP, ukwenziwa kwe-vokenization kungaholela kuma-chatbot asebenza kangcono, abasizi ababonakalayo, ukuxilongwa kwezokwelapha okuku-inthanethi, abahumushi bedijithali, nokuningi.

    Ukwengeza, inhlanganisela yombono nokufunda ulimi ithola ukuduma ezinhlelweni zezithombe zezokwelapha, ikakhulukazi ekuxilongweni kwesithombe sezokwelapha okuzenzakalelayo. Isibonelo, abanye abacwaningi bazama le ndlela ezithombeni ze-radiograph ezinezincazelo zombhalo ezihambisanayo, lapho ukuhlukaniswa kwe-semantic kungase kudle isikhathi. Inqubo ye-vokenization ingathuthukisa lezi izethulo futhi ithuthukise ukuthwebula kwezithombe okuzenzakalelayo ngokusebenzisa ulwazi lombhalo.

    Izicelo ze-vokenization

    Ezinye izinhlelo zokusebenza zokuthi vokenization zingabandakanya:

    • Ama-chatbot anembile angacubungula izithombe-skrini, izithombe, nokuqukethwe kwewebhusayithi. Ama-chatbots osekelo lwamakhasimende, ikakhulukazi, angakwazi ukuncoma ngokunembile imikhiqizo namasevisi.
    • Abahumushi bedijithali abakwazi ukucubungula izithombe namavidiyo futhi banikeze ukuhumusha okunembile okucabangela umongo wamasiko nesimo.
    • Izikena ze-bot zenkundla yezokuxhumana zikwazi ukwenza ukuhlaziya kwemizwa okuphelele ngokuhlanganisa izithombe, amagama-ncazo, namazwana. Lolu hlelo lokusebenza lungaba usizo ekulinganisweni kokuqukethwe okudinga ukuhlaziywa kwezithombe eziyingozi.
    • Ukwandisa amathuba omsebenzi wombono wekhompyutha kanye nonjiniyela bokufunda ngomshini we-NLP nososayensi bedatha.
    • Ama-Startups akha kulezi zinhlelo ze-AI ukuze athengise ngazo noma ahlinzeke ngezixazululo ezenziwe ngokwezifiso zamabhizinisi.

    Imibuzo ongaphawula ngayo

    • Ucabanga ukuthi i-vokenization izoshintsha kanjani indlela esisebenzisana ngayo namarobhothi?
    • I-vokenization ingashintsha kanjani indlela esiqhuba ngayo ibhizinisi kanye nokuxhumana namagajethi ethu (ama-smartphone nezinto zikagesi ezihlakaniphile)?

    Izinkomba zokuqonda

    Izixhumanisi ezilandelayo ezidumile nezikhungo zibhekiselwe kulo mbono: