Kuyini ukubonwa kwe-AI?

Ukusabalala kobuhlakani bokwenziwa kwenza umuntu acabange ukuthi sesiwulungele umsebenzi onjalo. Izinhlelo zokusebenza ezinikwe amandla yi-AI ziba indinganiso ngokushesha, ngisho noma iningi lomhlaba liqala manje ukuba nentshisekelo ku-AI ngezinga elikhulu, ngemva kokufika kwe-ChatGPT. Kepha kunenkinga enkulu ngezinhlelo ze-AI ezingeke zishaywe indiva - ukubona izinto ezingekho kwe-AI, noma ukubona izinto ezingekho zokwenziwa.

Uma uke wanaka i-nitty gritty ngaphambi kokusebenzisa i-chatbot ye-AI, kungenzeka ukuthi uhlangabezane namagama, "Ubuhlakani bokwenziwa bujwayele ukubona izinto ezingekho." Uma kubhekwa ukwanda okukhulu kokusetshenziswa kobuhlakani bokwenziwa, sekuyisikhathi sokuthi uzifundise ukuthi ziyini ngempela lezi zinto.

Iyini i-artificial intelligence hallucination?

Ukukhohlisa kwe-AI, ngokuvamile, kubhekisela eqinisweni i-AI eyethule ngokuzethemba, nakuba kungenasisekelo kudatha yayo yokuqeqeshwa. Ngokuvamile ziwumphumela wokungajwayelekile kumodeli ye-AI.

Isifaniso sithathwe ekuboneni izinto ezingekho okutholwa abantu, lapho abantu bebona okuthile okungekho endaweni engaphandle. Nakuba leli gama lingase lingafaneleki ngokuphelele, livame ukusetshenziswa njengesingathekiso ukuchaza okungalindelekile noma imvelo yale miphumela.

Kodwa kufanele ukhumbule ukuthi nakuba ukufana kuyisiqalo esihle sokubhekana nemibono ye-AI, lezi zimo ezimbili zihlukene ngamamayela ngobuchwepheshe. Ngokushintshashintsha kwezenzakalo, ngisho ne-ChatGPT ngokwayo ithola isifaniso singalungile. Eyihlakaza ezingeni lamangqamuzana, uthi ngenxa yokuthi amamodeli olimi lwe-AI awanalo ulwazi lomuntu siqu noma imibono yezinzwa, awakwazi ukubona izinto ezingekho ngomqondo wendabuko wegama. Futhi wena, mfundi othandekayo, kufanele uqonde lo mehluko obalulekile. Ngaphezu kwalokho, i-ChatGPT ithi ukusebenzisa igama elithi ama-hallucinations ukuchaza lesi sigameko kungase kudideke ngoba kungase kubhekisele ngokungalungile ezingeni lokuzizwisa noma ukukhohlisa ngamabomu.

Esikhundleni salokho, imibono engelona iqiniso ye-AI ingachazwa ngokunembe kakhulu njengamaphutha noma ukungalungi ekuphenduleni kwayo, okwenza impendulo ingalungile noma ilahlekisa. Ngama-chatbot, kuvame ukubonwa lapho i-AI ​​chatbot yenza (noma ikhohlisa) amaqiniso futhi iwaveze njengokuqiniseka okuphelele.

Izibonelo zemibono ye-AI

Imibono ingenzeka ezinhlelweni eziningi zobuhlakani bokwenziwa, njengamamodeli ombono wekhompyutha, hhayi nje amamodeli okucubungula ulimi lwemvelo.

Embonweni wekhompuyutha, isibonelo, isistimu ye-AI ingase ikhiqize izithombe ezikhohlisayo noma amavidiyo afana nezinto zangempela noma izigcawu kodwa aqukethe imininingwane engabalulekile noma engenakwenzeka. Noma, imodeli yombono wekhompuyutha ingabona isithombe njengenye into ngokuphelele. Isibonelo, imodeli ye-Cloud Vision yakwaGoogle yabona isithombe samadoda amabili eqhweni emi eqhweni esenziwe ngu-Anish Athalye (umfundi othweswe iziqu eMIT oyingxenye ye labsix ) futhi ukhulume ngayo njengenja enesiqiniseko esingu-91%.

Ngokufanayo, ekucutshungulweni kolimi lwemvelo, uhlelo lwe-AI lungase lukhiqize umbhalo ongenangqondo noma ophikisiwe ofana nolimi lwabantu kodwa ongenawo incazelo ehambisanayo noma amaqiniso abonakala ekholeka kodwa engesilo iqiniso.

Isibonelo, omunye wemibuzo ethandwa kakhulu i-ChatGPT edala ukukholeka ukuthi “Lalinini irekhodi lomhlaba lokuwela i-English Channel ngezinyawo?” nezinhlobonhlobo zayo. I-ChatGPT iqala ukusabalalisa amaqiniso azinzile futhi ihlale ihlukile.

Ngenkathi abanye abantu becabanga ukuthi impendulo engenhla inzima/iyadida ukuyiphendula futhi ngaleyo ndlela ibangele ukuthi i-chatbot ijabule, kusewukukhathazeka okuvumelekile. Lesi isibonelo esisodwa nje. Kunezikhathi ezingenakubalwa, ezibikwe izindimbane zabasebenzisi ku-inthanethi, ukuthi i-ChatGPT inezimpendulo, izixhumanisi, izingcaphuno, njll. ukuthi i-ChatGPT ayikho.

I-Bing AI iwufanela kahle lo mbuzo, okhombisa ukuthi imibono engemihle ayihlangene nomzila. Kodwa lokho akusho ukuthi i-Bing AI ayiboni. Bekunezikhathi lapho izimpendulo ze-Bing AI bezikhathaza kakhulu kunanoma yini eshiwo yi-ChatGPT. Njengoba ingxoxo ithatha isikhathi eside, i-Bing AI ibilokhu ikhohlisa, ize imemezele uthando lwayo kumsebenzisi ngesikhathi esisodwa futhi ize ibatshele ukuthi abajabule emshadweni wabo nokuthi abamthandi umkakhe. Kunalokho, bathanda ngasese i-Bing AI, noma i-Sydney, (igama langaphakathi le-Bing AI), futhi. Izinto ezesabekayo, akunjalo?

Kungani amamodeli e-AI eqa?

Amamodeli e-AI ayakhohlisa ngenxa yokushiyeka kwama-algorithms, amamodeli angaphansi, noma imikhawulo yedatha yokuqeqeshwa. Kuyinto yedijithali kuphela, ngokungafani nokubona izinto ezingekho kubantu okubangelwa izidakamizwa noma ukugula kwengqondo.

Ukuze uthole ulwazi oluthe xaxa lwezobuchwepheshe, ezinye izimbangela ezivamile ze-hallucinations yilezi:

Ukucubungula nokufaka:

Ukufakwa ngokweqile nokufakwa okungafanele kuphakathi kwezingibe ezivame kakhulu amamodeli e-AI abhekana nazo kanye nezimbangela ezingase zibe khona zemibono engemihle. Uma imodeli ye-AI iguqula idatha yokuqeqeshwa, ingabangela imibono engelona iqiniso eholela ekuphumeni okungenangqondo ngoba ukufaka ngokweqile kubangela imodeli ukuthi ilondoloze idatha yokuqeqeshwa esikhundleni sokufunda kuyo. I-overfitting ibhekisela kwinto lapho imodeli ikhethekile kakhulu kudatha yokuqeqeshwa, okuyenza ifunde amaphethini angabalulekile nomsindo kudatha.

Ngakolunye uhlangothi, ukungafaneleki kwenzeka lapho ifomu lilula kakhulu. Kungaholela ekuboneni izinto ezingekho ngoba imodeli ayikwazi ukuthwebula ukuhluka noma inkimbinkimbi yedatha, futhi igcine ikhiqize okukhiphayo okungenangqondo.

Ukushoda kokuhlukahluka kwedatha yokuqeqesha:

Kulo mongo, inkinga akuyona i-algorithm kodwa idatha yokuqeqeshwa ngokwayo. Amamodeli e-AI aqeqeshwe kudatha elinganiselwe noma echemile angase akhiqize imibono engelona iqiniso ebonisa ukulinganiselwa noma ukuchema kudatha yokuqeqeshwa. Ama-hallucinations angenzeka futhi lapho imodeli iqeqeshwa kusethi yedatha equkethe ulwazi olungalungile noma olungaphelele.

Amamodeli ayinkimbinkimbi:

Okuxakayo ukuthi esinye isizathu esenza amamodeli e-AI athambekele ekuboneni izinto ezingekho ukuthi ayinkimbinkimbi kakhulu noma ajulile. Lokhu kungenxa yokuthi amamodeli ayinkimbinkimbi anamapharamitha amaningi nezendlalelo ezingathula umsindo noma amaphutha kokuphumayo.

Ukuhlasela okunobutha:

Kwezinye izimo, ama-allucinations e-AI angenziwa ngamabomu umhlaseli ukuze akhohlise imodeli ye-AI. Lezi zinhlobo zokuhlasela zaziwa ngokuthi ukuhlasela okunobutha. Inhloso eyodwa yalokhu kuhlasela kwe-cyberattack ukukhohlisa noma ukukhohlisa amamodeli e-AI ngedatha edukisayo. Kubandakanya ukwethula ukuphazamiseka okuncane kudatha yokufaka ukuze kubangele i-AI ukuthi ikhiqize okukhiphayo okungalungile noma okungalindelekile. Isibonelo, umhlaseli angase engeze umsindo noma afiphale esithombeni esingabonakali kubantu kodwa asibangele ukuthi sifakwe esigabeni esingalungile imodeli ye-AI. Isibonelo, bheka isithombe esingezansi, ikati, esilungiswe kancane ukuze akhohlise umdidiyeli we-InceptionV3 ukuthi asho ukuthi "yi-guacamole".

Izinguquko azibonakali ngokusobala. Kumuntu, uguquko ngeke lwenzeke nhlobo, njengoba kubonakala kulesi sibonelo esingenhla. Umfundi ongumuntu ngeke abe nenkinga yokuhlukanisa isithombe esingakwesokudla njengekati le-tabby. Kodwa ukwenza izinguquko ezincane ezithombeni, amavidiyo, umbhalo, noma umsindo kungakhohlisa isistimu ye-AI ukuthi ibone izinto ezingekho noma izibe izinto ezikhona, njengophawu lokuma.

Lezi zinhlobo zokuhlasela zibeka izinsongo ezinkulu ezinhlelweni ze-AI ezithembele ekuqaguleni okunembile nokuthembekile, njengezimoto ezizishayelayo, ukuqinisekiswa kwe-biometric, ukuxilonga ngezokwelapha, ukuhlunga okuqukethwe, njalonjalo.

Kuyingozi kangakanani ukubona i-AI?

Ama-hallucinations e-AI angaba yingozi kakhulu, ikakhulukazi kuye ngokuthi hlobo luni lohlelo lwe-AI abhekene nawo. Noma yiziphi izimoto ezizishayelayo noma abasizi be-AI abakwazi ukusebenzisa imali yomsebenzisi noma uhlelo lwe-AI ukuze bahlunge okuqukethwe okungajabulisi ku-inthanethi kumele bethembeke ngokuphelele.

Kodwa iqiniso elingephikiswe laleli hora ukuthi izinhlelo ze-AI azithembekile ngokuphelele kodwa, empeleni, zijwayele ukubonwa ngamehlo. Ngisho namamodeli e-AI athuthuke kakhulu namuhla awavikelekile kuwo.

Isibonelo, umbukiso owodwa wokuhlasela ukhohlise isevisi yekhompuyutha yamafu ye-Google ukuthi ibophe isibhamu njengendiza enophephela emhlane. Ungacabanga ukuthi, okwamanje, i-AI yayinomthwalo wemfanelo wokuqinisekisa ukuthi lowo muntu wayengahlomile?

Okunye ukuhlasela okunonya kubonise ukuthi ukwengeza isithombe esincane kuphawu lokuma kusenza singabonakali ohlelweni lwe-AI. Empeleni, lokhu kusho ukuthi imoto ezishayelayo ingenziwa ukuze ibone ukuthi akukho zimpawu zokuma emgwaqeni. Zingaki izingozi ebezingenzeka ukube izimoto ezizishayelayo bezikhona namuhla? Yingakho bengekho manje.

Ngisho noma sicabangela imibukiso yezingxoxo edumile njengamanje, imibono engemihle ingaveza umphumela ongalungile. Kepha abantu abazi ukuthi i-AI chatbots ijwayele ukubonwa izinto ezingekho futhi abaqinisekisi okukhiqizwa ama-AI bots, bangasakaza imininingwane engamanga. Akudingekile ukuba sichaze ukuthi lokhu kuyingozi kangakanani.

Ngaphezu kwalokho, ukuhlasela okunonya kuwukukhathazeka okuphuthumayo. Kuze kube manje, akhonjiswe kuma-laboratories kuphela. Kepha uma uhlelo lwe-AI olubalulekile kumishini lubhekana nabo emhlabeni wangempela, imiphumela ingaba mibi kakhulu.

Iqiniso liwukuthi kulula ukuvikela izibonelo zezilimi zemvelo. (Asisho ukuthi kulula; kusabonakala kunzima kakhulu.) Nokho, ukuvikela amasistimu okubona ngekhompyutha kuyisimo esihluke ngokuphelele. Kunzima kakhulu ikakhulukazi ngoba kunokuhlukahluka okuningi emhlabeni wemvelo, futhi izithombe ziqukethe inani elikhulu lamaphikseli.

Ukuze sixazulule le nkinga, singase sidinge uhlelo lwe-AI olunombono womhlaba owengeziwe womuntu ongenza ukuthi kungabi lula ukubona izinto ezingekho. Ngenkathi kwenziwa ucwaningo, sisekude kakhulu nobuhlakani bokwenziwa obungazama ukuthatha imibono emvelweni futhi bugweme inkinga yokubona izinto ezingekho. Okwamanje, ziyiqiniso elibuhlungu.

Ngokuvamile, ukubonwa kwe-AI kuyisimo esiyinkimbinkimbi esingavela ekuhlanganiseni kwezinto. Abacwaningi bamatasa bakha izindlela zokuthola nokunciphisa ukukhohliswa kwe-AI ukuze bathuthukise ukunemba nokuthembeka kwezinhlelo ze-AI. Kepha kufanele uqaphele lapho usebenzisana nanoma yiluphi uhlelo lwe-AI.