ayini amamodeli e-AI

Ayini Amamodeli e-AI? Ukucwila Okujulile.

Wake wazithola uskrola ngo-2 ekuseni ubuza ukuthi yiziphi izinhlobo ze-AI emhlabeni nokuthi kungani wonke umuntu ekhuluma ngazo sengathi ziyiziphonso zomlingo? Kunjalo. Le ngxenye iyindlela yami engahlelekile kakhulu, ngezinye izikhathi ekhethayo ukuze ikususe “eh, angazi” uye “ekuthembeni okuyingozi emaphathini esidlweni sakusihlwa.” Sizobona ukuthi ziyini, yini ezenza zibe usizo ngempela (hhayi nje kuphela), ukuthi ziqeqeshwa kanjani, ukuthi zikhethwa kanjani ngaphandle kokungabaza, kanye nezingibe ezimbalwa ozifunda kuphela ngemva kobuhlungu.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Kuyini i-AI arbitrage: Iqiniso ngemuva kwegama elidumile
Kuchaza i-arbitrage ye-AI, ukuduma kwayo, kanye namathuba angempela.

🔗 Kuyini i-AI engokomfanekiso: Konke okudingeka ukwazi
Ihlanganisa i-AI engokomfanekiso, izindlela zayo, kanye nezinhlelo zokusebenza zesimanje.

🔗 Izidingo zokugcina idatha ze-AI: Okudingeka ukwazi
Ihlukanisa izidingo zokugcina idatha ye-AI kanye nokucatshangelwa okusebenzayo.

Ngakho-ke... ayini amamodeli e-AI, ngempela? 🧠

Uma ihlungiwe kakhulu: imodeli ye-AI imane nje iwumsebenzi ofundiwe .Uyinika okokufaka, ikhipha okuphumayo. Okubalulekile ukuthi, ithola ukuthi kanjani ngokubhekabheka izibonelo eziningi bese izilungisa ukuze "ingabi nephutha kangako" isikhathi ngasinye. Phinda lokho ngokwanele bese iqala ukubona amaphethini obungaqapheli nokuthi akhona lapho.

Uma uke wezwa amagama afana nokubuyela emuva okuqondile, izihlahla zezinqumo, amanethiwekhi e-neural, ama-transformer, amamodeli okusabalalisa, noma ngisho nomakhelwane abaseduze kakhulu - yebo, bonke bangama-riff kule ndaba efanayo: idatha iyangena, imodeli ifunda imephu, umphumela uyaphuma. Izingubo ezahlukene, umbukiso ofanayo.

Yini ehlukanisa amathoyizi namathuluzi angempela ✅

Amamodeli amaningi abukeka kahle kakhulu ku-demo kodwa ancipha ekukhiqizweni. Lawo anamathelayo avame ukuba nohlu olufushane lwezimpawu zabantu abadala:

  • Ukuhlanganisa - kuphatha idatha engakaze ibonwe ngaphandle kokuwa.

  • Ukuthembeka - akusebenzi njengokuphonsa uhlamvu lwemali lapho okufakwayo kuba okungajwayelekile.

  • Ukuphepha Nokuphepha - kunzima kakhulu ukudlala noma ukusebenzisa kabi.

  • Ukuchazeka - akucaci kahle ngaso sonke isikhathi, kodwa okungenani kuyalungiswa.

  • Ubumfihlo Nokulunga - ihlonipha imingcele yedatha futhi ayihlangani nobandlululo.

  • Ukusebenza kahle - okungabizi ngokwanele ukuthi kusebenze kahle kakhulu.

Lokho ngokuyisisekelo abalawuli bohlu lwezingubo kanye nezinhlaka zezingozi kanye nothando - ukufaneleka, ukuphepha, ukuzibophezela, ukucaca, ubulungisa, konke okuhle kakhulu. Kodwa ngokweqiniso, lokhu akuzona izinto ezinhle ongaba nazo; uma abantu bethembele ohlelweni lwakho, bayizisulu zetafula.

Ukuhlolwa okusheshayo kokuqonda: amamodeli vs ama-algorithm vs idatha 🤷

Nasi isiqephu esinezingxenye ezintathu:

  • Imodeli - "into" efundiwe eguqula okokufaka kube okuphumayo.

  • I-Algorithm - iresiphi eqeqesha noma esebenzisa imodeli (cabanga ngokuhla kwe-gradient, usesho lwe-beam).

  • Idatha - izibonelo ezingavuthiwe ezifundisa imodeli ukuthi kufanele iphathe kanjani.

Isifaniso esingacacile kancane: idatha iyizithako zakho, i-algorithm iyiresiphi, kanti imodeli iyikhekhe. Ngezinye izikhathi imnandi, ngezinye izikhathi ingena phakathi ngoba ubheke ngokushesha kakhulu.

Imindeni yamamodeli e-AI ozohlangana nawo ngempela 🧩

Kunezigaba ezingenamkhawulo, kodwa nansi uhlu olusebenzayo:

  1. Amamodeli aqondile nahlelekile - alula, asheshayo, angahunyushwa. Izisekelo eziyisisekelo ezisalokhu zinganqotshwa zedatha yethebula.

  2. Izihlahla kanye neqembu - izihlahla ezinqumayo ziyahlukana; hlanganisa ihlathi noma uzikhulise bese ziqine ngendlela emangalisayo.

  3. Amanethi e-Convolutional neural (ama-CNN) - umgogodla wokuqashelwa kwesithombe/ividiyo. Izihlungi → imiphetho → izimo → izinto.

  4. Amamodeli okulandelana: Ama-RNN nama-transformer - ombhalo, inkulumo, amaprotheni, ikhodi. Ukuzinaka kwama-transformer kwaba yinto eyashintsha umdlalo [3].

  5. Amamodeli okusabalalisa - akhiqizayo, aguqula umsindo ongahleliwe ube izithombe ezihambisanayo isinyathelo ngesinyathelo [4].

  6. Amanethi e-neural egrafu (ama-GNN) - akhelwe amanethiwekhi nobudlelwano: ama-molecule, amagrafu omphakathi, izindandatho zokukhwabanisa.

  7. Ukufunda okuqiniswayo (RL) - ama-ejenti okuzama namaphutha athuthukisa umvuzo. Cabanga ngamarobhothi, imidlalo, izinqumo ezilandelanayo.

  8. Okwethembekile kwakudala: kNN, Naive Bayes - izisekelo ezisheshayo, ikakhulukazi zombhalo, uma udinga izimpendulo izolo.

Inothi eliseceleni: kudatha yethebula, ungayicindezeli kakhulu. Ukuhlehla kwe-logistic noma izihlahla ezikhuliswe nge-boost zivame ukugoba amanetha ajulile. Ama-transformer amahle, kodwa awakho yonke indawo.

Ukuthi ukuqeqeshwa kubukeka kanjani ngaphansi kwe-hood 🔧

Amamodeli amaningi anamuhla afunda ngokunciphisa umsebenzi wokulahlekelwa ngohlobo oluthile lwe -gradient desccent. I-Backpropagation isunduza ukulungiswa emuva ukuze ipharamitha ngayinye yazi ukuthi kufanele ihambe kanjani. Fafaza amaqhinga afana nokumisa kusenesikhathi, ukuqondisa kabusha, noma ama-optimizer ahlakaniphile ukuze angangeni esiphithiphithini.

Ukuhlolwa kwamaqiniso okufanele kuqoshwe ngaphezu kwedeski lakho:

  • Ikhwalithi yedatha > ukukhetha imodeli. Ngempela.

  • Hlala uhlela into elula. Uma imodeli eqondile inamathela, ipayipi lakho ledatha cishe nalo liyanamathela.

  • Bukela ukuqinisekiswa. Uma ukulahlekelwa kokuqeqeshwa kwehla kodwa ukulahlekelwa kokuqinisekiswa kunyuka - sawubona, kufanele ukwenze ngokweqile.

Ukuhlola amamodeli: ukunemba kulele 📏

Ukunemba kuzwakala kukuhle, kodwa kuyinombolo eyodwa embi kakhulu. Kuye ngomsebenzi wakho:

  • Ukunemba - uma uthi okuhle, usuke uqinisile kangaki?

  • Khumbula - kuzo zonke izinto ezinhle zangempela, zingaki ozitholile?

  • F1 - ibhalansisa ukunemba kanye nokukhumbula.

  • Ama-curve e-PR - ikakhulukazi kudatha engalingani, athembeke kakhulu kune-ROC [5].

Ibhonasi: hlola ukulinganiswa (ingabe amathuba asho okuthile?) bese uzulazula (ingabe idatha yakho yokufaka iyashintsha ngaphansi kwezinyawo zakho?). Ngisho nemodeli "enhle" iyaphela.

Ukubusa, ingozi, imithetho yomgwaqo 🧭

Uma imodeli yakho ithinta abantu, ukuthobela imithetho kubalulekile. Izikhombisi ezimbili ezinkulu:

  • I-AI RMF ye-NIST - yokuzithandela kodwa ewusizo, enezinyathelo zomjikelezo wokuphila (ukubusa, imephu, ukukala, ukuphatha) kanye namabhakede okuthembeka [1].

  • Umthetho we-EU AI - umthetho osekelwe engcupheni, osuvele ungumthetho kusukela ngoJulayi 2024, obeka imithwalo eqinile yezinhlelo ezinobungozi obukhulu kanye namamodeli athile ajwayelekile [2].

Iphuzu elibalulekile: bhala phansi lokho okwakhile, ukuthi ukuvivinye kanjani, kanye nezingozi ozihlolile. Kukusindisa izingcingo eziphuthumayo zaphakathi kwamabili kamuva.

Ukukhetha imodeli ngaphandle kokulahlekelwa ingqondo 🧭➡️

Inqubo ephindaphindwayo:

  1. Chaza isinqumo - liyini iphutha elihle uma liqhathaniswa nephutha elibi?

  2. Idatha yokuhlola - usayizi, ibhalansi, ukuhlanzeka.

  3. Beka imingcele - ukuchazeka, ukubambezeleka, isabelomali.

  4. Sebenzisa ama-baseliners - qala ngomugqa/i-logistic noma umuthi omncane.

  5. Phindaphinda ngokuhlakanipha - engeza izici, lungisa, bese ushintsha imindeni uma ithuthuka.

Kuyadina, kodwa kuyadina kuhle lapha.

Isithombe sokuqhathanisa 📋

Uhlobo lwemodeli Izithameli Intengo-ngokufanayo Kungani kusebenza
I-Linear & Logistic abahlaziyi, ososayensi okuphansi–okuphakathi amandla ahunyushwayo, asheshayo, atholakala kuthebula
Izihlahla Zezinqumo amaqembu axubile phansi ukuhlukaniswa okufundeka njengomuntu, ukuphathwa okungekho emgqeni
Ihlathi Elingahleliwe amaqembu omkhiqizo okuphakathi nendawo amaqembu anciphisa ukuhlukahluka, ama-generalist aqinile
Izihlahla Ezikhuliswe Nge-Gradient ososayensi bedatha okuphakathi nendawo I-SOTA kuthebula, iqinile ngezici ezingcolile
Ama-CNN abantu abanombono okuphakathi nendawo - okuphezulu ukuhlangana → amazinga endawo
Ama-Transformers I-NLP + i-multimodal phezulu ukuzinaka kulinganisa kahle [3]
Amamodeli Okusabalalisa amaqembu okudala phezulu ukususa umsindo kuveza umlingo wokukhiqiza [4]
Ama-GNN ama-graph nerd okuphakathi nendawo - okuphezulu ukudlulisa umlayezo kuhlanganisa ubudlelwano
kNN / Naive Bayes abaphangi abasheshayo kuphansi kakhulu izisekelo ezilula, ukuthunyelwa okusheshayo
Ukufunda Kokuqinisa ucwaningo oluningi okuphakathi nendawo - okuphezulu kuthuthukisa izenzo ezilandelanayo, kodwa kunzima ukuzilawula

"Izinto ezikhethekile" ezisebenzayo 🧪

  • Izithombe → Ama-CNN ayazigqaja ngokufaka amaphethini endawo abe makhulu.

  • Ulimi → Ama-Transformers, ngokuzinaka, aphatha umongo omude [3].

  • Amagrafu → Ama-GNN ayakhanya lapho ukuxhumana kubalulekile.

  • Imidiya ekhiqizayo → Amamodeli okusabalalisa, asusa umsindo kancane kancane [4].

Idatha: i-MVP ethule 🧰

Amamodeli awakwazi ukulondoloza idatha embi. Izisekelo:

  • Hlukanisa amasethi edatha ngakwesokudla (akukho ukuvuza, hlonipha isikhathi).

  • Ukungalingani kokuphatha (ukuphinda kuthathwe isampula, izisindo, imikhawulo).

  • Izici zonjiniyela ngokucophelela - ngisho namamodeli ajulile ayazuza.

  • Qinisekisa ngokuphambene ukuze uthole ingqondo ehluzekile.

Ukulinganisa impumelelo ngaphandle kokuzikhohlisa 🎯

Qondanisa izilinganiso nezindleko zangempela. Isibonelo: i-triage yamathikithi okusekela.

  • Ukubuyiselwa kwemali kukhulisa izinga lokubamba amathikithi ngokushesha.

  • Ukunemba kuvimbela ama-ejenti ukuthi angaminzi emsindweni.

  • I-F1 ilinganisa kokubili.

  • Landelela ukuzulazula nokulinganisa ukuze uhlelo lungaboli buthule.

Ingozi, ubulungisa, amadokhumenti - kwenze kusenesikhathi 📝

Cabanga ngemibhalo hhayi njenge-red tape kodwa njengomshwalense. Ukuhlolwa kokukhetha, ukuhlolwa kokuqina, imithombo yedatha - kubhale phansi. Izinhlaka ezifana ne-AI RMF [1] kanye nemithetho efana ne-EU AI Act [2] seziba yizingqinamba noma kunjalo.

Imephu yokuqalisa esheshayo 🚀

  1. Qinisekisa isinqumo kanye nesilinganiso.

  2. Qoqa isethi yedatha ehlanzekile.

  3. Isisekelo esinomugqa/isihlahla.

  4. Yeqa uye emndenini ofanele ukuze uthole indlela.

  5. Hlola ngezilinganiso ezifanele.

  6. Izingozi zedokhumenti ngaphambi kokuthunyelwa.

Imibuzo Evame Ukubuzwa mayelana nombani ⚡

  • Linda, ngakho-ke futhi - iyini imodeli ye-AI?
    Umsebenzi oqeqeshwe kudatha ukuze uhlanganise okokufaka kokukhiphayo. Umlingo uwukwenza konke kube ngokoqobo, hhayi ukukhumbula.

  • Ingabe amamodeli amakhulu ahlala ewina?
    Hhayi kumathebula - izihlahla zisabusa. Kumbhalo/izithombe, yebo, usayizi uvame ukusiza [3][4].

  • Ukuchazeka uma kuqhathaniswa nokunemba?
    Ngezinye izikhathi kuyashintshana. Sebenzisa amasu ahlanganisiwe.

  • Ukulungisa kahle noma ubunjiniyela obusheshayo?
    Kuncike - isabelomali kanye nobubanzi bomsebenzi kuyalawula. Kokubili kunendawo yakho.

TL;DR 🌯

Amamodeli e-AI = imisebenzi efunda kudatha. Okuwenza abe usizo akukhona nje ukunemba kodwa ukwethembana, ukuphathwa kwezingozi, kanye nokusetshenziswa okucatshangelwe kahle. Qala ngokulula, linganisa okubalulekile, bhala phansi izingxenye ezimbi, bese (futhi kuphela lapho) ziqala ukuthandwa.

Uma ugcina umusho owodwa kuphela: amamodeli e-AI ayimisebenzi efundwayo, aqeqeshwe ngokwenziwa ngcono, ahlulelwe ngezilinganiso ezithile zomongo, futhi asetshenziswa ngezivikelo. Yilokho kuphela.

Izinkomba

  1. I-NIST - Uhlaka Lokuphathwa Kwengozi Yobuhlakani Bokwenziwa (AI RMF 1.0)
    I-NIST AI RMF 1.0 (PDF)

  2. Umthetho Wezobunhloli Bokwenziwa we-EU - Ijenali Esemthethweni (2024/1689, Julayi 12 2024)
    I-EUR-Lex: Umthetho We-AI (I-PDF Esemthethweni)

  3. Ama-Transformers / Ukuzinaka - Vaswani et al., Ukunaka Yikho Konke Okudingayo (2017).
    arXiv:1706.03762 (PDF)

  4. Amamodeli Okusabalalisa - Ho, Jain, Abbeel, Amamodeli Angalindelekile Okusabalalisa (2020).
    arXiv:2006.11239 (PDF)

  5. I-PR vs i-ROC mayelana nokungalingani - uSaito noRehmsmeier, i-PLOS ONE (2015).
    I-DOI: 10.1371/journal.pone.0118432

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI

Buyela kubhulogi