🤖 Sannu, mu ne CLEVI.
Mun halarci taron 「Makon Baje-kolin Bincike na JBNU na 2025 (Research Fair Week)」 wanda Jami'ar Jeonbuk ta shirya daga ranar 20 zuwa 22 ga Janairu.
Wannan taron wuri ne mai muhimmanci da ya mayar da hankali kan kyawawan sakamakon bincike na Jami'ar Jeonbuk, ya gabatar da alkiblar fasahohin gaba masu alaƙa da dabarun masana'antu da na ƙasa, tare da samar da damar musayar ra'ayi tsakanin masana'antu, jami'o'i, cibiyoyin bincike da hukumomi.
Musamman ma, an gayyaci Shugaba Hwanho Lee na CLEVI a matsayin mai gabatar da jawabin VIP a bikin buɗe taron da aka gudanar a ranar 20 ga wata! 👏
Bari mu sake jin yanayin wannan ranar!
Ga Shugaba Hwanho Lee yana shirin gabatar da jawabi kafin taron ya fara; ana iya ganin yadda yake tunanin yadda zai gabatar da jawabin cikin cikakken bayani kuma cikin sauƙin fahimta.
An fara taron, kuma aka gudanar da bikin buɗewa.
An buɗe taron cikin ban sha'awa da wasan kwaikwayo na fasahar kai tsaye wanda ɗalibin Jami'ar Jeonbuk ya kammala nan take.
Bayan jawabin buɗe taron na farko, Shugaba Hwanho Lee na CLEVI ya gabatar da jawabin buɗe taron na biyu mai taken “Me ya sa AI ke buƙatar matsakaicin zahiri?” Bari mu duba a taƙaice.
Ta yaya AI ke fahimtar duniya,
ya ƙirƙiri jimloli, har ma ya yi ayyuka?
Muna son mu duba tare cikin 'kwakwalwar' manyan samfuran harshe (LLM) kamar ChatGPT da Claude da muke amfani da su kowace rana, don fahimtar ƙa'idar da suke amfani da ita wajen fahimta da ƙirƙirar jimloli.
- Sanya lambar suna (Token ID): Da farko, AI tana sanya kowace kalma lamba ta musamman kamar '5' ko '8' domin sarrafa ta.
- Sanya adireshi a taswira (Vector): Amma ba za a iya fahimtar ma'ana daga lambobi kaɗai ba, saboda haka ana ba kowace kalma 'cikakken adireshi (coordinates)' da ya ƙunshi ɗaruruwan lambobi.
- Neman maƙwabta: Kalmomin da suke da ma'ana iri ɗaya (misali, 'the' da 'a') suna zama kusa da juna a wannan taswira kamar maƙwabta.
To, ta yaya AI ke hasashen kalma ta gaba?
Yanzu da muka fahimci sauya bayanai na asali, bari mu ga cikakken tsarin yadda waɗannan lambobi ke zama jimloli na ainihi. Bari mu ɗauka cewa akwai wannan shigarwar "Paris is the city...".
Tokenizer (rarrabawa): Kamar yadda muka gani a baya, jimlar da muka shigar tana canzawa zuwa lambobi (Token ID) ta hanyar tokenizer. "Paris" lamba ce ta 1652, "is" kuma ta 5... haka dai.
LLM (babban na'urar lissafi): Waɗannan tarin lambobin suna shiga cikin wata babbar na'urar lissafi mai suna LLM (samfurin Bloom da ke cikin hoton). A nan ne yake fahimtar mahallin ta hanyar aikin 'vector' da muka koya a baya. Yana cewa, 'Ah, yanzu ana magana ne game da birnin Paris!'
Logits & Softmax (lissafin yiwuwa): Bayan ya kammala lissafi, AI yana ba wa dubban kalmomin da za su iya zuwa na gaba maki. Sannan ta hanyar wani tacewa mai suna 'softmax', yana mayar da makiyan zuwa **'yiwuwa'** da za mu iya fahimta cikin sauƙi.
Zaɓi na ƙarshe: Ku kalli jadawalin da ke hannun dama.
- of → 33.7%
- that → 15.3%
- where → 9.4%
A ƙarshe, AI yana zaɓar **'of'**, wadda take da mafi girman yiwuwa, sannan yana ci gaba da kammala jimlar "Paris is the city of...". Yana maimaita wannan tsari dubban sau, ta haka ne yake samar da doguwar amsar da muke gani.
A taƙaice, tuna kalma ga AI ba yana nufin haddace haruffa ba ne, sai dai tuna 'adireshin' wurin da kalmar take a kan babbar taswirar ma'ana.
Ana iya faɗaɗa LLM zuwa gani, mutum-mutumi, da koyon aiki.
AI mai fahimtar ma'ana, mai gani da fahimta, kuma mai aiki.
Ba wa AI hannaye da ƙafafu yana nufin faɗaɗa hankali daga harshe zuwa ganewa, sannan ya wuce zuwa aiki.
AI ba ya cikin fagen ra'ayi ko hasashe kawai; ya shiga matakin aiwatarwa da zahiri.
Mataki na gaba na CLEVI
CLEVI za ta ci gaba da faɗaɗa damammaki iri-iri, kamar tarukan bita a cikin makarantu da wajensu, ayyukan koyarwa na aikace, da rumfunan gwaji, domin mutane da yawa su samu damar ganin kai tsaye yadda fasahar AI ke aiki a zahirin duniya.
Kamar yadda laccar Jami'ar Jeonbuk ta zama gada mai daraja da ta haɗa ɗalibai da mu, za mu ci gaba da taka rawar gada tsakanin masu fasaha da masu amfani, da kuma tsakanin fagen ilimi da masana'antu.