Ngaphansi kweminyaka emibili edlule, i-DARPA yethula uhlelo lwe-Lifelong Learning Machines (L2M) ukuze idale ngokuqhubekayo amasistimu erobhothi anezici zobuhlakani bokwenziwa. Uhlelo lwe-L2M bekufanele luholele ekuveleni kwezinkundla zokuzifundela ezikwazi ukuzivumelanisa nendawo entsha ngaphandle kokuhlela noma ukuqeqeshwa kwangaphambili. Kalula nje, amarobhothi kwakufanele afunde emaphutheni awo, futhi angafundi ngokumpompa amasethi edatha yesifanekiso endaweni yaselabhorethri.
Uhlelo lwe-L2M lubandakanya amaqembu ocwaningo angama-30 anenani elihlukahlukene loxhaso. Muva nje, elinye lamaqembu avela eNyuvesi yaseSouthern California likhombise inqubekelaphambili egculisayo ekudaleni amapulatifomu erobhothi okuzifundela wona, njengoba kubikwe kumagazini kaMashi we-Nature Machine Intelligence.
Ithimba labacwaningi abavela enyuvesi liholwa nguFrancisco J. Valero-Cuevas, uprofesa wobunjiniyela bezinto eziphilayo, i-biokinesiology kanye nokwelapha ngokomzimba. Ngokusekelwe ku-algorithm eyakhiwe yiqembu, esekelwe ezindleleni ezithile zokusebenza kwezinto eziphilayo, ukulandelana kwezenzo zobuhlakani bokwenziwa kwadalwa ukufundisa ukunyakaza kwamarobhothi ezithweni ezine. Kubikwa ukuthi izitho zokwenziwa ngendlela yokulingisa imisipha, imisipha namathambo bakwazi ukufunda ukuhamba phakathi nemizuzu emihlanu ngemva kokugijima i-algorithm.
Ngemva kokwethulwa kokuqala, inqubo yayingahlelekile futhi inesiphithiphithi, kodwa i-AI yaqala ukuzivumelanisa ngokushesha namaqiniso futhi yaqala ngempumelelo ukuhamba ngaphandle kokuhlela kwangaphambili. Ngokuzayo, indlela edaliwe yokuqeqeshwa impilo yonke yamarobhothi ngaphandle kokuqeqeshwa kwe-ML kokuqala namasethi wedatha ingashintshwa ukuze kuhlonyiswe izimoto zabantu abashayela izimoto ezizenzakalelayo kanye nezimoto zamarobhothi ezempi. Kodwa-ke, lobu buchwepheshe bunamathemba amaningi nezindawo zokusebenzisa. Into eyinhloko ukuthi i-algorithm ayiboni umuntu njengenye yezithiyo ekuthuthukisweni futhi ayifundi lutho olubi.
Source: 3dnews.ru