Kwenzekeni
Abacwaningi be-MIT bathuthukise i-HardFlow, i- yesikhathi sokuthunyelwa yamamodeli e-AI akhiqizayo asebenza ngaphansi kwezidingo zokuphepha okungaxoxiswana ngazo, ngokomzimba, noma okuqondene nomsebenzi othize. Indlela ivumela izinyathelo zesizukulwane esimaphakathi inkululeko eyengeziwe, bese iqondisa okukhiphayo kokugcina ukwanelisa izithiyo ezinzima. I-MIT ibika izivivinyo ekwenziweni kwe-robotic, ukuzulazula kwe-maze, kanye nokuhlelwa kwesithombe esiqondiswa umbhalo, lapho i-HardFlow ihlangabezane nezingqinamba ezidingekayo kanye nezindlela eziyisisekelo ezisebenza kahle kakhulu ezilinganisweni zekhwalithi yesixazululo.
Indlela yokwenza kabusha isampula ebambeke kanzima njengenkinga yokwenza kahle i-trajectory isebenzisa izindlela zokulawula kahle. Kunokuba iphoqelele yonke isampula emaphakathi ukuthi ihlangabezane nezingqinamba, i-HardFlow iqondisa umzila wesizukulwane futhi idinga okukhiphayo kokugcina ukuthi kwenzeke. Abacwaningi bathi lokhu kugcina inkululeko eyengeziwe ngesikhathi sokukhiqiza futhi kuvumela izinjongo ezengeziwe, njengokunciphisa ubude bendlela yerobhothi.
Umthombo uthi ithimba lisebenzise ukwakheka kwamamodeli afana nokugeleza ukunqamula inkinga enkulu yokwenza kahle ibe yizinkinga ezincane zesinyathelo esisodwa. Ibika ukuthi i-HardFlow ithole ukugculiseka okuphelele okukhawulelwe kuzo zonke izivivinyo zokukhohlisa irobhothi, ukuzulazula kwe-maze, kanye nokuhlelwa kwesithombe esiqondiswa umbhalo, kuyilapho iqhubeka nokusebenza ngendlela edlula izindlela eziyisisekelo ezilinganisweni zekhwalithi yesixazululo. Umthombo uphinde uthi isikhathi sayo sokubala besiqhathaniseka noma siphansi kunesezindlela eziningi eziqhudelanayo. Ayikho imiphumela yezinombolo, izisekelo ezishiwo, ukukhishwa kwekhodi, noma imininingwane yokufinyelela eyimodeli enikeziwe.
Imininingwane yomthombo: news.mit.edu ↗
Kungani kubalulekile
Amamodeli akhiqizayo acishe alungile asengangaphephile kuzilungiselelo ezifana nokuzulazula kwerobhothi eduze kwabantu noma imishini. I-HardFlow ibhekana nenkinga yokuthunyelwa emaphakathi: ukuphoqelela izithiyo ngaphandle kokunciphisa ngokungadingekile ukucinga kwemodeli kwezixazululo ezingcono. Uma imiphumela ebikiwe iba jikelele, indlela ingasiza amaqembu ukuthi asebenzise amamodeli akhiqizayo aqeqeshwe kusengaphambili ezinhlelweni ezidinga kokubili ukuba nokwenzeka nokulungiselelwa, njengokugwema ukungqubuzana okuhlanganiswe nama-trajectories amarobhothi amafushane. Umthombo awusunguli ukusebenza kokukhiqiza komhlaba wangempela, ukuqinisekiswa okuzimele, noma ukutholakala okubanzi.
Inkinga engokoqobo ukuthi ukwanelisa izidingo zokuphepha kukodwa kungase kuveze umphumela ongasebenzi kahle noma ongemuhle, kuyilapho ukukhulisa ikhwalithi ngaphandle kokuphoqelela umthetho oqinile kungaveza ukwehluleka okungamukeleki. Izibonelo ze-MIT zihlela i-tradeoff ngokuhlela indlela yamarobhothi: indlela ecishe igweme isithiyo isengadala ukushayisana, kuyilapho indlela ebambeke kakhulu ingathola umzila omude kakhulu.
I-HardFlow ichazwa njenge-plug-and-play futhi isebenziseka ngesikhathi sokuphakelwa ngamamodeli akhiqiza aqeqeshwe kusengaphambili, angase anciphise isidingo sokuphinda aqeqeshwe lapho umkhawulo omusha wethulwa. Lokho kuyisimangalo esivela emthonjeni, hhayi umphumela ozimele. I-athikili ayicacisi izidingo zokusetshenziswa, ukulayisensa, izingxenyekazi zekhompuyutha ezisekelwayo, noma ukuthi izinhlangano ezingaphandle kweqembu locwaningo zingasebenzisa yini indlela manje.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Imibuzo eyinhloko evulekile eyokuthi i-HardFlow iyatholakala yini esidlangalaleni, yiziphi izinhlobo eziqeqeshwe kusengaphambili nezinhlobo ezivimbelayo ezisekelayo ekusebenzeni, nokuthi izindleko zayo zokubala zishintsha kanjani ngamasistimu amakhulu noma izindawo eziyinkimbinkimbi. Umsebenzi oqhubekayo ungavumela imodeli eyisisekelo ukuthi ibuyekezwe kanye nokuphoqelela ukuphoqelelwa. Ukuhlolwa okuzimele nobufakazi bokuthunyelwa kuzodingeka ukuze kunqunywe ukuthi imiphumela ebikiwe idlulisa yini ngaphezu kokuhlolwa komthombo.
Ucwaningo luvela kokuthi I-IEEE Transactions on Pattern Analysis and Machine Intelligence, ngokusho komthombo. Umsebenzi wokulandelela kulindeleke ukuthi uhlole izilungiselelo lapho imodeli ye-AI ngokwayo ingabuyekezwa, okungenzeka ibe ngcono kokubili ukwaneliseka okunezingqinamba kanye nekhwalithi yokuphuma ngokuhamba kwesikhathi.
Okubalulekile okungaziwa kufaka phakathi ukusebenza kumarobhothi aphathekayo ngaphandle kokuhlolwa okulawulwayo, ukuziphatha ngaphansi kokungaphelele noma okushintsha ulwazi lwezemvelo, izinhlobo zezithiyo ezinzima ezingabhalwa ngekhodi ngendlela enokwethenjelwa, kanye nokukala kumanethiwekhi amakhulu emizwa. Umthombo awuhlinzeki ngamanani noma ulwazi lokufinyelela kwezohwebo futhi awubiki ukuphindaphinda okuzimele noma isitifiketi sokuphepha.