- I-XAI iguqula amamodeli ebhokisi elimnyama abe izinhlelo ezisobala, okusivumela ukuthi siqonde i-logic ngemuva kwesinqumo ngasinye se-algorithmic.
- Kukhona umehluko obalulekile phakathi kokuhunyushwa (amamodeli acacile ngaphakathi) kanye nokuchazeka (amasu okucacisa amamodeli ayinkimbinkimbi).
- Ukusetshenziswa kwezindlela ezifana nezindinganiso zeShapley noma zeLIME kusiza ekunciphiseni ukucwasa okubandlululayo futhi kuqinisekisa ukuhambisana nemithethonqubo yezomthetho efana neye-EU.
Cishe uke wabhekana nokusebenzisa ithuluzi lobuhlakani bokwenziwa futhi uzibuza, "Likuthathephi lokhu?" Lowo muzwa wokuba phambi kwebhokisi elimnyama , lapho ufaka khona idatha bese uthola umphumela ngaphandle kokuthi othile azi ngempela ukuthi kwenzekeni phakathi, yilokho kanye i-XAI ezama ukukuxazulula. Akukhona nje ukusebenza komshini, kodwa ngathi, njengabantu, ukukwazi ukuqonda isizathu ngemuva kwempendulo ngayinye ukuze singadingi ukwethemba i-algorithm ngokungacabangi.
Ezweni lapho i-AI isivele inquma konke kusukela ekutheni uthola imali mboleko kuya ekutheni ungasithola kanjani isifo, ikhono lokuvula leso sikhwama bese ubona ukuthi sisebenza kanjani libalulekile. Ukuchaza i-AI akuyona nje into yobuchwepheshe konjiniyela, kodwa kuyisidingo sezenhlalo nezomthetho sokugwema ukucwasa, ukuqinisekisa ubulungisa, futhi, ngaphezu kwakho konke, ukuvimbela abasebenzisi bokugcina ukuthi bazizwe belahlekile. Ake sihlole le ndlela yokusebenza ukuze siqonde ukuthi singasuka kanjani ekungabonakali ngokuphelele siye ekubonakaleni kwangempela.
Iyini i-Explanable AI (XAI)?
Uma sikhuluma nge-XAI, noma i-Explainable Artificial Intelligence, sibhekisela egatsheni elikhethekile elihlose ukwenza izinqubo zangaphakathi zama-algorithms zibe sobala futhi zifinyeleleke kalula . Umqondo ulula: uhlelo akufanele nje luveze umphumela, kodwa kufanele luhambisane nesizathu esingaqondwa yinoma ubani, uchwepheshe noma cha. Lokhu kubalulekile ngoba amamodeli amanje, ikakhulukazi ukufunda okujulile kanye namanethiwekhi e-neural , ayinkimbinkimbi kangangokuthi enza izibalo eziyizigidigidi zegama elilodwa, okwenza kube nzima ngokomuntu ukuphinda inqubo ngesandla.
Ukuhunyushwa kuqhathaniswa Nokuchazeka: Azifani
Zivame ukusetshenziswa ngokushintshana, kodwa kunomehluko obalulekile. Ukuhunyushwa kuyikhwalithi engenzi lutho; okungukuthi, amanye amamodeli abonakala ngokwemvelo, njengezihlahla zezinqumo noma ukuhlehla okuqondile, lapho ungalandela khona indlela yesinqumo cishe ngepensela nephepha nje. Lawa aziwa ngokuthi amabhokisi amhlophe noma amabhokisi engilazi.
Ngakolunye uhlangothi, ukuchazeka kuyinqubo esebenzayo. Isetshenziswa kakhulu kumamodeli angacacile (amabhokisi amnyama), njengamanethiwekhi e-neural, ukuzama ukuthola incazelo eqondakalayo yalokho okwenzekile. Empeleni, ngenkathi ukuchazeka kugxila ekusebenzeni kwangaphakathi, ukuchazeka kugxila ekuqinisekiseni umphumela wokugcina kusetshenziswa amasu angaphandle.
Ukubaluleka kokuvumelanisa umlayezo nomsebenzisi
Awukwazi ukuchaza imodeli yokufunda komshini kusosayensi wedatha ngendlela efanayo ongayenza kumakhasimende asebhange. Ukuvumelanisa ulimi kuyisihluthulelo sempumelelo ye-XAI. Kumuntu wezobuchwepheshe, incazelo izogxila ekulinganisweni kwezilinganiso kanye nesisindo sezinto eziguquguqukayo; kuchwepheshe webhizinisi, into ebalulekile ukuqonda ukuthi yiziphi izici zamakhasimende eziholele imodeli ekuphakamiseni umkhiqizo owodwa hhayi omunye.
- Ochwepheshe be-AI: Bafuna imininingwane yobuchwepheshe ejulile ukuze bathuthukise futhi bathuthukise imodeli.
- Abasebenzisi bedatha: Ochwepheshe (odokotela, abameli) abadinga ukuqinisekisa ukuthembeka ukuze benze izinqumo zezokwelapha noma zomthetho.
- Abantu abasafufusa noma umphakathi jikelele: Abantu abafuna nje ukwazi ukuthi idatha yabo isetshenziswa kahle nokuthi kukhona ukuqapha kwabantu inqubo.
Amasu okuvula ibhokisi elimnyama
Ukuze kwenziwe i-AI ingabi yimfihlakalo, kusetshenziswa izindlela ezahlukene, ezihlukaniswe kakhulu ngezindlela ezimbili: zomhlaba wonke kanye nezendawo.
Ukuhlaziywa Komhlaba Wonke: Imephu iyonke
Lezi zindlela zihlose ukuqonda ukuthi imodeli iziphatha kanjani iyonke. Isibonelo, inqubo ye -Permutation Importances ilinganisa ukuthi iphutha likhuphuka kangakanani uma sishintsha amanani e-variable, iveza ukuthi yiziphi izici ezinomthelela omkhulu ohlelweni. Kukhona futhi ne-Partial Dependence Plots , ezihlaziya ukuthi ukubikezela kushintsha kanjani lapho inani le-variable ethile lishintshwa, yize zombili zivame ukucabanga ukuthi i-variables ayizimele.
Ukuhlaziywa Kwendawo: Icala Eliqondile
Yilapho sifuna ukuqonda khona ukuthi kungani lesi siguli sithole isexwayiso sezokwelapha. Amanani e-Shapley (asekelwe kumbono womdlalo) abalulekile lapha, njengoba enikeza umnikelo othize emphumeleni wokugcina wenguquko ngayinye. Lokhu kuvumela udokotela ukuthi abone ukuthi isexwayiso sibangelwe kakhulu yizinga le-cholesterol , hhayi iminyaka yesiguli, okusenza sikwazi ukwenza ngokushesha nangokunembile.
Izingozi zokungacaci kanye nezinzuzo zokucaca
Ukuthembela ebhokisini elimnyama ngokungazi kuyingozi. Ingozi enkulu kakhulu ukucwasa nokubandlulula ; uma i-AI iqeqeshwe ngedatha yomlando lapho kwenzeka khona ubandlululo ngokobulili noma ngokobuhlanga, i-algorithm izofunda lawo maphethini bese iwaphinda, ijezise ama-resume noma yenqabe amakhredithi ngaphandle kwesizathu esinengqondo. Ngaphezu kwalokho, ukuntuleka kokubonakala kudala isikhala sokuzibophezela ngokomthetho : uma imoto ezishayela yona iphahlazeka, sidinga ukwazi ukuthi kungani sinquma ukuthi ubani onephutha.
Ukusebenzisa i-XAI akugcini nje ngokuvimbela izinhlekelele, kodwa futhi kunikeza inzuzo yokuncintisana . Kuvumela ukutholakala kwamaphutha ngokushesha, ukuhambisana nemithetho eqinile njenge- European Union's General Data Protection Regulation (GDPR) , futhi, ngaphezu kwakho konke, kwakha ukwethembana. Uma umsebenzisi eqonda isizathu sesinqumo, cishe uzokwamukela ubuchwepheshe futhi abuhlanganise empilweni yakhe yansuku zonke.
Izinselele kanye nenkinga yokunemba
Akuyona yonke into ehamba kahle. Kukhona ukungezwani okuqhubekayo okwaziwa ngokuthi ukuhwebelana ngokunembile nokuchaza . Ngokuvamile, amamodeli anembe kakhulu (njengokufunda okujulile) awacaci kakhulu, kuyilapho okulula ukuwaqonda evame ukungabi namandla kakhulu. Inselele yamanje ukuthola indawo ephakathi lapho uhlelo lusebenza khona ngokwanele ukuba luwusizo, kodwa lucace ngokwanele ukuba luhlolwe futhi luqondwe.
Ngaphezu kwalokho, kunezindleko zokubala. Ukubala amanani e-Shapley ezigidini zezibikezelo ngomzuzwana kubiza kakhulu maqondana namandla okucubungula. Ngakho-ke, izindlela zokuhumusha ze-mechanistic (ukuqonda ukusebenza kwangaphakathi) kanye ne-post-hoc (ukuchaza ngemuva kokuqeqeshwa) ziyathuthukiswa ukuze kwenziwe inqubo isebenze kahle kakhulu.
I-AI echazekayo emikhakheni ebalulekile
Kwezokunakekelwa kwempilo , i-XAI ingaba umehluko phakathi kokuphila nokufa, okuvumela odokotela ukuthi baqinisekise iziphakamiso ze-AI ngaphambi kokuhlinzwa. Kwezemfundo , kubalulekile ukuthi othisha balondoloze ukuzimela kwabo futhi bagweme ukuba ngabaphathi nje bemiyalelo ye-algorithmic, okukhuthaza ukucabanga okubucayi kubafundi. Kwezokuthutha, kusiza ukuphatha impahla ngokuchaza ukuthi kungani kungcono ukuhambisa iphalethi ngokusekelwe ekuvameni kokukhetha kanye nokusondelana nama-load docks.
Ukucaca okuphelele, kusukela emthonjeni wedatha kuya ku-logic yokukhipha, kuguqula ubuhlakani bokwenziwa bube ithuluzi lokusebenzisana phakathi kwabantu nomshini . Ngokuhlanganisa ukuqapha nokuqonda komuntu, ubuchwepheshe buqinisekisiwe ukuthi buyisisekelo sokuziphatha nokuphephile, bugwema ukukhohlisa kokuqonda nokuqinisekisa ukuthi zonke izinqumo zilungile, ziyaqinisekiswa, futhi, ngaphezu kwakho konke, ziwusizo ekuthuthukiseni umphakathi.
