Yana da wuya cewa kowa zai yi jayayya da gaskiyar cewa ƙirar LSIs (ASICs) na al'ada ba shi da nisa daga tsari mai sauƙi da sauri. Amma ina so kuma ina buƙatar shi da sauri: a yau na ba da algorithm, kuma bayan mako guda na cire aikin dijital da aka gama. Gaskiyar ita ce, ƙwararrun LSIs kusan samfuri ne na kashe-kashe. Waɗannan ba safai ake buƙata ba a cikin batches na miliyoyin, akan haɓakawa waɗanda zaku iya kashe kuɗi da yawa da albarkatun ɗan adam gwargwadon yadda kuke so, idan ana buƙatar yin hakan cikin kankanin lokaci. ASICs na musamman, sabili da haka mafi tasiri don magance ayyukan su, ya kamata su kasance masu rahusa don haɓakawa, wanda ke zama mega-dace a halin yanzu na ci gaba na ilmantarwa na inji. A wannan gaba, kayan da kasuwar kwamfuta ke tarawa, musamman ma, ci gaban GPU a fagen koyon injin (ML) ba za a iya kauce masa ba.
Don haɓaka ƙirar ASICs don ayyukan ML, DARPA tana kafa sabon shiri - Real Time Machine Learning (RTML). Shirin koyo na inji na ainihi ya ƙunshi haɓaka mai tarawa ko dandamalin software wanda zai iya tsara tsarin gine-gine ta atomatik don takamaiman tsarin ML. Ya kamata dandamali ya bincika ta atomatik na injin koyo algorithm da aka saita don horar da wannan algorithm, bayan haka yakamata ya samar da lamba a cikin Verilog don ƙirƙirar ASIC na musamman. Masu haɓakawa na ML algorithm ba su da ilimin masu ƙirar guntu, kuma masu ƙira ba su saba da ƙa'idodin koyon injin ba. Shirin RTML ya kamata ya taimaka wajen tabbatar da cewa an haɗa fa'idodin duka biyu a cikin tsarin ci gaban ASIC mai sarrafa kansa don koyon inji.
Yayin zagayowar rayuwar shirin RTML, hanyoyin da aka samo za su buƙaci a gwada su a cikin manyan wuraren aikace-aikacen guda biyu: hanyoyin sadarwar 5G da sarrafa hoto. Hakanan, shirin RTML da dandamalin software da aka ƙirƙira don ƙira ta atomatik na masu haɓaka ML za a yi amfani da su don haɓakawa da gwada sabbin algorithms na ML da bayanan bayanai. Don haka, ko da kafin zana siliki, zai yiwu a tantance abubuwan da za a yi na sabon tsarin. Abokin DARPA a cikin shirin RTML zai kasance Gidauniyar Kimiyya ta Kasa (NSF), wacce kuma ke da hannu cikin matsalolin koyon injin da haɓaka algorithms ML. Za a canja wurin mai tarawa da aka haɓaka zuwa NSF, kuma baya DARPA yana tsammanin karɓar mai tarawa da dandamali don tsara algorithms ML. A nan gaba, ƙirar kayan aiki da ƙirƙirar algorithms za su zama haɗin haɗin kai, wanda zai haifar da fitowar tsarin na'ura waɗanda ke koyo da kansu a ainihin lokacin.
source: 3dnews.ru