Isifinyezo: Ucwaningo olusha lusebenzisa ukufunda kuka-opharetha we-Koopman ukuze kufakazele ukuthi amasistimu athile ayinkimbinkimbi, anesiphithiphithi anemikhawulo eyisisekelo yokubikezela engakwazi ukunqotshwa idatha yokuqeqeshwa engapheli. Ngokudizayina amasistimu aphikisanayo ukuze enze imephu lapho amamodeli okufunda emishini ewa khona, ithimba lachaza umsuka wezibalo we-LLM ukubona izinto ezingekho, kuyilapho lethula i-algorithm esebenza kahle kakhulu enemingcele yamaphutha eyakhelwe ngaphakathi efake imephu ngempumelelo amaphethini eqhwa olwandle lwase-Arctic kusetshenziswa ikhompuyutha ephathekayo evamile.
Amaqiniso Abalulekile
- Inganekwane Yedatha Engapheli Ichithwa: Ucwaningo lubonisa ukuthi ifilosofi evamile ye-tech-industry “yedatha eyengeziwe ilingana nokufunda okuqinisekisiwe” ayilungile ngokwezibalo. Izinkinga ezithile eziyinkimbinkimbi kakhulu noma eziyizinxushunxushu zifaka amaphethini anezingqimba afihlekile noma angenzeki ukuhlukaniswa ngobunono, okusho ukuthi okungcono kakhulu okungatholwa yi-algorithm uhlamvu lwemali ($50/50$), okunikeza inkinga ngokungaxazululeki ngokwezibalo ngokunganaki usayizi wedathasethi.
- Kungani ama-Chatbots e-Hallucinate: Ukuntengantenga kwezibalo okwephula ukubikezela ngokomzimba kwesikhathi eside kuchaza ukuthi kungani amamodeli ezilimi ezinkulu (ama-LLM) njenge-ChatGPT noma u-Claude asungula ngokuzethemba ulwazi olungamanga ngokuhamba kwesikhathi. Kumasistimu azwela kakhulu, ukuhluka kwamaminithi esikhumbuzweni sokuqala kubangela amaphutha ahlanganisayo athumela imodeli phansi ezindleleni ezihlukene kakhulu, okugcina ukuhambisana kwesikhashana kuyilapho ulahlekelwa ngokuphelele ukuthintana neqiniso.
- Izinsika ezimbili Zokufeyila Komshini Wokufunda: Ithimba likaDkt. Matthew Colbrook likhombe izizathu ezimbili eziqondile zokuthi kungani ukumodela kwe-AI kubhidlika ngokwemvelo lapho kuhlangana nezimo eziyinkimbinkimbi:
- Ukuhluleka Kokuqinisekisa Ukushoda Kwedatha: I-algorithm yokufunda yomshini ayinakho indlela yezibalo yangaphakathi yokunquma nini idle amasampula okuqeqesha anele ukuze ikhiphe isibikezelo esizinzile, esiqinisekile.
- I-Obfuscation Yephethini Efihliwe: Izixhumanisi zokulandelela okubalulekile ngaphakathi kwezakhiwo eziguquguqukayo zihlala zifihliwe ngokwezibalo noma zibambene ngokujulile, okwenza kungenzeki ngamanethi e-neural ajwayelekile ukuthi ahlukanise.
- Inkinga Ye-Chaos Frequency: Lapho i-AI ihlaziya isistimu enesiphithiphithi (lapho izinguquko ezincane kumapharamitha okuqala ziveza ukwehluka okukhulu), opharetha we-Koopman ukhiqiza ukusabalala okuqhubekayo kwamaza agqagqene kunokuguquguquka okuhlanzekile, okukodwa okulandela umkhondo. Lokhu kuchaza ukuthi kungani izibikezelo zesikhathi esifushane zihlala zinembile, kuyilapho ukuqagela kwesistimu yesikhathi eside kuwa ngokuhlelekile.
- I-algorithm ye-Provably Reliable: Ukuze kuxazululwe lobu sengozini yesakhiwo, abacwaningi benze inoveli, i-algorithm eqinile yezibalo enemingcele yamaphutha eyakhelwe ngaphakathi, engaguquleki. Le khithi yamathuluzi inika abacwaningi imitha yesiqiniseko sesikhathi sangempela, eqinisekisa ngqo ukuthi okukhiphayo kwe-AI kungethenjwa nini ngaphandle kokudinga amakhompiyutha amakhulu ezigidi zamadola.
- I-Laptop vs. Supercomputer Benchmark: Lapho ingcindezi ivivinywa kuqhathaniswa neminyaka engu-40 yamarekhodi esimo sezulu e-Arctic, i-algorithm yangokwezifiso yeqembu yakhomba amaphethini okubola esakhiwo alahleka kudala kumashidi eqhwa. Ihlale ingcono kunezinhlelo ze-AI ezihamba phambili zomhlaba zezentengiselwano ngenkathi isebenza ngokuphelele kukhompyutha ephathekayo eyisisekelo, yezinga lomthengi ngengxenyana yezindleko zokubala.
Umthombo: Inyuvesi yaseCambridge
Singayethemba nini imiphumela esiyithola ku-AI, futhi kunini lapho ukufunda kungenakwenzeka? Abacwaningi babonise ukuthi kunezinkinga ezithile ngisho ne-AI enamandla kakhulu engazixazulula ngokuthembekile, kungakhathaliseki ukuthi ingakanani idatha enikeziwe.
Abacwaningi, abavela eNyuvesi yaseCambridge kanye naseNyuvesi yaseCalifornia Santa Barbara, baklame izinhlelo zezibalo 'eziphikisanayo' eziklanyelwe ukukhohlisa noma iyiphi i-algorithm ye-AI. Njengabaduni abaqotho abahlola ukuphepha kwenethiwekhi, lezi zinhlelo eziphikisanayo zaziklanyelwe ukukhomba kahle ukuthi ukubikezela kwe-AI kwephuka kuphi futhi kungani.
Izinhlelo eziningi zomhlaba wangempela – njengalezo ezisolwandle, ubuchopho bomuntu, noma amarobhothi – ziyinkimbinkimbi kakhulu ukuthi zingachazwa ngobunono ngezibalo, ngakho abacwaningi bavame ukufunda indlela abaziphatha ngayo ngokusebenzisa ukufunda ngomshini. Kodwa lezi zindlela ze-AI azisebenzi kahle ngaso sonke isikhathi, zibuyisela imiphumela engathembekile noma izibikezelo ezimbi.
Ngezinye izikhathi, nokho, ukunikeza izixazululo ezinokwethenjelwa akunakwenzeka, ngisho nangedatha engapheli. Izinhlelo eziphikisanayo ezakhiwe abacwaningi zingasiza abathuthukisi nabasebenzisi bezinhlelo ze-AI bazi ukuthi basebenza enkingeni engaxazululeka noma engaxazululeki, bakhe izindlela ezisebenzayo, futhi bagweme ukuchitha isikhathi, umzamo noma amathokheni e-AI lapho inkinga ingaphezu kwemingcele okungenzeka.
Imiphumela yabo, kubikwe kujenali Ukuxhumana Kwemveloingasiza futhi ukuchaza ukuthi kungani ama-chatbots e-AI adumile njenge-ChatGPT noma i-Claude ekwazi ukunemba esikhathini esifushane, kodwa angakhukhuleka noma akhohlise ngokuhamba kwesikhathi.
“Siphenya imingcele yalokho ongakwazi ukukwenza nongakwazi ukukwenza nge-AI,” kusho umbhali oholayo uDkt Matthew Colbrook, woMnyango Wezibalo Ezisetshenziswayo kanye neTheory Physics yaseCambridge. “Kubaluleke kakhulu ukuqonda ukuthi yiziphi izinkinga ezingaxazululeki ngalezi zindlela, ngoba uma kungenjalo ugcina uchithe isikhathi nemali eningi.”
U-Colbrook nababhali abakanye naye basebenzisa indlela ebizwa ngokuthi i-Koopman operator learning, eshintsha ukuziphatha okuyinkimbinkimbi okungahambisani nomugqa kube indlela yomugqa okulula ukuyihlaziya.
“Esasikwenza ngalezi ‘zitha’ kwakuwukuzama ukuthola izinhlobo zezinhlelo okunzima noma okungenakwenzeka ukuzibikezela, kanye nezinhlobo zezinhlelo ezingase zishintshwe ukuze kubuyiselwe imiphumela ethembekile,” kusho uColbrook.
Abacwaningi bahlonze izizathu ezimbili eziyinhloko zokuthi kungani ukufunda komshini kuphuka lapho kuhlaziywa amasistimu ayinkimbinkimbi: noma i-algorithm ayikwazi ukusho uma ibonwa idatha eyanele ukubuyisela umphumela othembekile, noma amaphethini kusistimu angafihlwa noma kube nzima ukuwahlukanisa.
“Ocwaningweni oluningi lwe-AI, umcabango ojwayelekile ukuthi uma siqoqa idatha eyengeziwe, ukufunda kuzogcina kusebenze,” kusho uColbrook. “Kodwa sithole ukuthi lokhu akulungile. Ukufunda kuvame ukuhlukaniswa, futhi kudinga izinyathelo eziningi ngendlela efanele ukuze kusebenze.”
Uma isistimu inesiphithiphithi – okusho ukuthi umehluko omncane ezimeni zokuqala uholela emigudwini ehluke kakhulu, njengokukhetha indaba yakho yokuzidela – u-opharetha we-Koopman ngokuvamile ugcina nge okuqhubekayo ukusabalala kwamafrikhwensi esikhundleni sezindlela ezihlanzekile, ezihlukile. Ukubikezela isikhathi esifushane kwakunembile, kodwa ukubikezela kwesikhathi eside akuzange kuthembeke, ngoba ukuzwela kwezimo zokuqala kuyahlangana ngokuhamba kwesikhathi.
Ukungazinzi okufanayo kwezibalo okwehlula ama-algorithms wokubikezela kungase futhi kuchaze ukuthi kungani ama-chatbot e-AI enza amaqiniso ngokuzethemba: izinguquko ezincane embuzweni zingathumela i-chatbot phansi ngendlela ehluke ngokuphelele, ebukeka izwakala izwi negama kodwa ilahlekelwe ukubamba kwayo iqiniso ngenxa yemiphumela emide.
Abacwaningi bathuthukise indlela yokuhlukanisa lezi zinkinga ngokusekelwe ekutheni zingaki izinyathelo ezidingekayo ukuze zixazululwe. Lapho idatha ingafakwanga ngokwanele noma ilandelana ngendlela efanele, okungcono kakhulu i-algorithm engayenza – ngisho nedatha engapheli – ingu-50/50, okuhlukanisa inkinga njengengaxazululeki.
Ithimba liphinde lakhiqiza i-algorithm entsha, ethembeke futhi esebenza kahle kakhulu enemingcele yamaphutha eyakhelwe ngaphakathi: empeleni inika abacwaningi be-AI indlela yokwazi lapho bekwazi ukuyethemba impendulo, ngengxenyana yezindleko zamakhompiyutha amaningi amakhulu.
Abacwaningi bahlole indlela yabo eminyakeni engaphezu kwengu-40 yedatha yeqhwa lasolwandle i-Arctic. Besebenzisa i-algorithm yabo bathole amaphethini afihliwe endleleni iqhwa elincipha ngayo, futhi bakwazi ukudlula amamodeli ahamba phambili e-AI amanje ngenani elincane lezindleko, kukhompyutha ephathekayo evamile.
“Sisesigabeni manje lapho kube nezibonelo eziningi eziwubukhazikhazi nezindaba zempumelelo ku-AI, kodwa kubalulekile ukuthi siphinde sibuze ukuthi amamodeli aqiniseke kangakanani, nokuthi sazi kanjani ukuthi aqinisekile,” kusho uColbrook. Ngaphandle kwalokho, sakhela phezu kwezisekelo ezintengantengayo.
Imibuzo Ebalulekile Iyaphendulwa:
A: Zibone ngeso lengqondo uzama ukubikezela indlela yentuthu ephuma emlilweni. Intuthu iyasonteka, iyagoqa, futhi ihlukana phakathi ngendlela eyinkimbinkimbi kakhulu, engeyona yomugqa cishe okungenakwenzeka ukuyilandelela ngezibalo eziyisisekelo. Umsebenzisi we-Koopman uyindlela yezibalo ethatha lokhu kuziphatha okungcolile, okungaqondile futhi ikufakisele kwenye indawo, indawo engabonakali lapho ukunyakaza kusebenza njengomugqa oqondile, oqondile. Ngokuguqula amasistimu ayinkimbinkimbi abe yilolu hlobo lomugqa, ithimba likaDkt. Matthew Colbrook lingakwazi ukuhlola izibalo, lidale “amanethiwekhi aphikisanayo” ukuze likhombe ngqo lapho izibalo zehla khona futhi kungenzeki ukuthi i-AI ixazulule.
A: I-Chatbots icubungula umbhalo kufana nesimo sezulu esinesiphithiphithi sicubungula isimo, igama negama, indlela incike ngokuphelele endaweni yokuqala. Uma isistimu inesiphithiphithi, amashifu amancane kokokufaka kokuqala athumela indlela ezansi kuma-trajectories ahluke kakhulu. Ku-chatbot, ukushintsha uhlamvu olulodwa noma igama ngokushesha kungabangela i-AI ukuthi iphambuke endleleni yayo yeqiniso. Igama negama, impendulo izwakala izwakala ngokuphelele, kodwa ngenxa yomphumela omude, ukuzwela okuhlanganisiwe kulolo shintsho oluncane lwasekuqaleni kubangela ukuba imodeli isuke eqinisweni, okuholela ekuboneni izinto ezingekho.
A: Umkhakha wezobuchwepheshe udinga ukuyeka ukujikijela ngokungaboni amandla amakhulu ekhompiyutha kanye namasethi edatha angacushiwe ezinkingeni eziyinkimbinkimbi. Esikhundleni salokho, abathuthukisi bangasebenzisa i-algorithm entsha yeqembu le-Cambridge-UCSB, efaka imingcele yamaphutha eyakhelwe ngaphakathi. Kucabange njengegeji yedeshibhodi eyakhelwe ngaphakathi ekutshela ukuthi imodeli ye-AI iqiniseke kangakanani mayelana nokuphuma kwayo. Ngokusebenzisa le ndlela, ososayensi bangakwazi ukuhlukanisa ngokushesha ukuthi inkinga eyinkimbinkimbi iyaxazululeka noma ayinakwenzeka ngempela, konga izigidi zamaRandi, ukunqamula isikhathi esichithiwe se-supercomputing, nokugqamisa amaphethini afihliwe kusetshenziswa amakhompyutha aphathekayo ayisisekelo, athengekayo.
Amanothi Omhleli:
- Lesi sihloko sihlelwe umhleli weNeuroscience News.
- Iphepha lejenali libuyekezwe ngokugcwele.
- Ingqikithi eyengeziwe yengezwe abasebenzi bethu.
Mayelana nalezi zindaba zocwaningo lwe-AI
Umbhali: Sarah Collins
Umthombo: Inyuvesi yaseCambridge
Othintana naye: USarah Collins – Inyuvesi yaseCambridge
Isithombe: Isithombe sifakwe ku-Neuroscience News
Ucwaningo lwangempela: Vula ukufinyelela.
“Izinhlelo ze-Adversarial dynamical zibonakala lapho ukufunda okuqhutshwa yidatha kuphumelela noma kwehluleka” ngu-Matthew J. Colbrook, Igor Mezić & Alexei Stepanenko. Ukuxhumana Kwemvelo
I-DOI:10.1038/s41467-026-74220-8
Abstract
Amasistimu aguqukayo aphikisanayo abonakala lapho ukufunda okuqhutshwa yidatha kuphumelela noma kwehluleka
Amasistimu amaningi amelana nokumodela kokuhlaziya, okwenza ukuthi ukuguquguquka okuqhutshwa yidatha kubaluleke kakhulu. Nokho izindlela eziqhutshwa yidatha zingehluleka ukuhlangana noma ukujwayela, okushiya kuvuleke umbuzo omaphakathi: Kunini lapho ukuziphatha kwesistimu kungafundwa ngokwethembeka kudatha, futhi kunini lapho ukufunda okunjalo kungenakwenzeka?
Siphendula lo mbuzo sisebenzisa amasistimu we-adversarial dynamical ukuhlonza umngcele phakathi kwemibuso efinyelelekayo nengafinyeleleki. Ekufundeni komsebenzisi we-Koopman, uhlaka oluhamba phambili lokumela okuguquguqukayo okungaqondile ngokusebenzisa izinto ze-spectral ezinomugqa, siklama ama-algorithms e-spectral aqhutshwa yidatha afaneleka ngokuhlangana neziqinisekiso zesitifiketi ngaphansi kwezimo ezivela ngokubanzi kumasistimu aphathekayo.
Lokhu kuveza ithiyori yokuhlangana yokulinganisa kwe-Koopman-opharetha futhi kuxazulule inkinga evulekile ende ekuhlaziyweni kwe-spectral ye-Koopman. Ngokuphambene, ngokwakha izinhlelo eziphikisanayo, sifakazela imiphumela engenzeki efanayo: ngaphandle kwale mibandela, ayikho inqubo yokukhawulela yokulandelana okukodwa engaqinisekisa ukufunda, kungakhathaliseki ikhwalithi yedatha. Le miphumela igqama kakhulu lapho ukufundwa kwe-spectral okuqhutshwa yidatha kungaphumelela nalapho kufanele kuhluleke. Siqinisekisa uhlaka lwama-oscillator, ukugeleza koketshezi oluyisiphithiphithi kanye nokubikezela ukugcwala kweqhwa lolwandle lwe-Arctic.
Ekugcineni, sembula izindlela ezifihliwe zokuwohloka kweqhwa olwandle lwe-Arctic, siletha izibikezelo zebanga elide ezinemingcele yamaphutha ezindawo, futhi sidlula ukusebenza kahle kwamamodeli ashukumisayo futhi ajulile okufunda ngezindleko eziphansi kakhulu zokubala, okuvumela ukusetshenziswa kwesikhathi sangempela kuma-CPU ajwayelekile.
