
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 Ababunuli: 
 
 
 
 
 
 Thomas Feltin Léo Marchó Juan-Antonio Cordero-Fuertes Frank Brockners Thomas H. Clausen Isishwankela: Ukubalwa kwe-Deep neural network (DNN) kudatha yokuhamba kufuna izixhobo zokubala ukuhlangabezana neemfuno zokugqitha kokubalwa. Nangona kunjalo, izicelo zokufunda ezibucayi malunga nobude bexesha kunye nemfihlo azinakukwazi ukuthwala ubalele efini elikude ngenxa yeendleko zothumelo kunye nokungcatshwa kubanikezeli abathathu abangabaxhamli. Phakathi kwezisombululo zokwandisa ukusebenza ngelixa ugcina ubalele kwindawo enomlinganiselo, ukukhawulezisa izixhobo kunokuba nzima, kwaye ukulungelelaniswa kwemodeli kufuna imizamo emikhulu yoyilo ngelixa kuthintela ukuchaneka. Ukwahlulwa kwe-DNN yindlela yesithathu ehambelanayo, kwaye iquka ukuhambisa umthwalo wokubala phezu kwezixhobo ezahlukeneyo ezikhoyo, uthathela ingqalelo iipropati zenethiwekhi ye-edge kunye nesakhiwo se-DNN, kunye nenjongo yokwandisa ukugqithiswa kokubala (inani lokubala ngomzuzwana). Eli phepha lingenisa indlela yokuchaza ubude bexesha bokubala kunye nobunye bokubala kweendlela ezininzi ezihanjelweyo ze-DNN, kwaye ichaza inkqubo yokulungelelanisa ukwandisa ukugqithiswa kokubala. Emva koko kuvezwa isixazululi se-branch and bound kwaye sihlalutywa ukubala ukusebenza okufunyenweyo kunye nobunzima. Oku kuhlalutya kukhokelele ekuchazeni indawo yokukhawulezisa, echaza iimeko ezichaziweyo kwiipropati ze-DNN kunye nezenethiwekhi apho ukwahlulwa kwe-DNN kunenzuzo. Okokugqibela, iziphumo zemvavanyo ziqinisekisa iimlinganiselo kwaye zibonisa ukuphucuka kokugqithiswa kokubala kwiimzekelo zokubeka ezikumaphethelo. Ukwahlulwa kwe-DNN kumaphethelo kwenza kube lula ukukhawulezisa ukugqithiswa kokubala ngelixa kunciphisa ukusetyenziswa kwemibhala kunye nobude bexesha kwiiservara ezizingqongileyo, umzekelo, efini. Intshayelelo Izixhobo ezixhunyiweyo ziye zavelisa idatha elinganiswayo yi-2.5 quintillion bytes yonke imihla ngo-2020.  Kulo mba, i-Artificial Intelligence (AI) yenye yezobuchwepheshe ezinokubhekana kakuhle nobuninzi bedatha eyaveliswa, kubude beendawo ezifana nokubonwa kwezinto kwimbono yekhompyutha, ukuqaphela ubuso, ukuqaphela ilizwi, ukucutshungulwa kwelwimi lolwimi, okanye ukuziqhelanisa nokuziqhelanisa nokuziqhelanisa. I-AI, kwaye ngokukodwa i-Deep Learning (DL), ifuna, ngaphezulu kwedatha eninzi, amandla abalulekileyo kwisixhobo sokubala kunye nememori, eyenza ukuba ukubala efini kube yisisombululo sokubamba imisebenzi ye-AI. Ngenxa yoko, ngo-2019, i-96% yimisebenzi ye-AI yenziwa efini. 1 2 Izicelo ze-AI zinokuba nemiqobo enzima malunga nobude bexesha, umzekelo, ekuziqhelaneni okuzenzekelayo, ukubeka iliso kwimveliso, okanye naziphi na izicelo zokubala ixesha lokwenyani ezibandakanya unxibelelwano nehlabathi lokwenyani. Umzekelo, amaxesha empendulo omndilizi ekuziqhelaneni okuzenzekelayo afuna ukuba ngaphantsi kwe-100 yeemilliseconds , kwenza amaxesha ohambo oluya efini abe made kakhulu kwezi sicelo. Ngokufanayo, ezinye izicelo zinokubangela ukusetyenziswa okubalulekileyo kwenethiwekhi, umzekelo, xa kuthathwa umgangatho ophezulu wokucutshungulwa kwevidiyo kwimbono yekhompyutha, nto leyo ethintela ukuba ziqhutywe efini. Imigaqo-nkqubo yemfihlo yedatha okanye imigaqo yokukhusela idatha ingavumeli idatha ukuba ishiye imeko ethile. Oku kukhuthaza izicelo ze-AI ekubekeni kwimicubungulo ye-edge, equka ukubuyisela imisebenzi yokubala ukusuka kwiiservara ukuya kwizixhobo ezikufutshane nemithombo yedatha. [](javascript:void()) Ngokuthelekiswa nee-clouds ezininzi, umphetho yindawo enomlinganiselo enezithintelo ezibalulekileyo , kwaye ukuphuhlisa i-AI kumaphethelo ngoko kuyahluka kakhulu ekuphuhliseni i-AI yefu. Ukuhlalutya idatha ikufutshane nomthombo wayo kunokuba nzima ngenxa yamandla esixhobo aphantsi, ukunganyamezeli kwezixhobo ezikhoyo, kunye nokwahluka kwezithuthi zemiphetho. Ukunciphisa amandla okubala kuthetha ukunciphisa ukugqithiswa kokubala okufumanekayo, okubangela ukuphazamiseka kokuhambisa imithwalo enzima ukusuka efini ukuya kumaphethelo. Kwezinye izicelo, ukugqithiswa kokubala, okuthetha ukugqithiswa okufumanekayo ngomzuzwana we-DNN, kunokudityaniswa nokuchaneka kwe-DNN, umzekelo, kunye nokulandelela izinto kwiintlanganiso zevidiyo, apho ukulahla izinga lokugqithiswa ukusuka kwi-15 ukuya kwi-5 zingomzuzwana kubonakale kunciphisa isikolo se-F1__3__somlandeleli wesixhobo nge-10% . Ngamanye amazwi, izicelo ezinokugqithiswa okuphantsi kokubala zinokubangela ukulahleka kolwazi olubalulekileyo, umzekelo, ukulahleka kwezinto ezifunyenweyo kwizicelo zokubeka iliso kuvidiyo. [](javascript:void()) [](javascript:void()) Zimbini iindlela eziphambili ezikhoyo zokuhlangabezana neemfuno zokusebenza kwiimiphetho enomlinganiselo: ukucudiswa kwemodeli okanye ukukhawulezisa izixhobo.    quka ukunciphisa kunye nokulungelelanisa iinethiwekhi zenaliti ukuya kwinguqulelo elula, ngokufuthi engasebenzi kakuhle, yemodeli. Iindlela ezisemgangathweni zokucudiswa kwemodeli zibandakanya ukusika ,  , okuthetha ukunciphisa imemori ethe see ukugcina imodeli, kunye nokufundisa ulwazi , okuthetha ukuba uqeqeshe inethiwekhi yenaliti encinci kunye nolwazi olukhutshiweyo kwinethiwekhi enkulu. Ukulungelelanisa imodeli kufuna imizamo emikhulu yoyilo kwaye kungaba sisithintelo ekuchanweni kwemodeli.   kumaziko angaphandle kufuna ukufumana izixhobo ezinokuthi zisebenze kakuhle kwaye zibe lula, ngelixa ziseyakwazi ukwenza ubalele ngokusasa okwaneleyo. Kukho iintlobo ezahlukeneyo zezixhobo ezifanelekileyo zokukhawulezisa kumaphethelo, ezahlukileyo ngomsebenzi kunye nokukhetha, kodwa iiyunithi zokucutshungulwa ezisemgangathweni azihlangani nentando kwaye izixhobo ezikhethekileyo ezinje nge-ASICs ziye zafunyanwa zingasebenzi kwaye zibiza kakhulu . Ukucudiswa kwemodeli [](javascript:void()) ukucwina [](javascript:void()) [](javascript:void()) Ukukhawulezisa izixhobo [](javascript:void()) Ukwahlulwa kwe-DNN yindlela ehambelanayo yokukhawulezisa ubalele ngokusebenzisa izixhobo ezininzi ezikhoyo kwimiphetho ukuhambisa ubalele bokubala – obunokuthi busetyenziswe ngokwabo, okanye kunye nezinye izindlela ezimbini. Ukwahlulwa kwe-DNN kubandakanya ukuthatha inethiwekhi yenaliti njengomjikelezo wokuhambisa kwiindawo, kunye nokuhambisa ezo ndawo kumaziko angaphandle. Ukubeka ezo ndawo kusekelwe kwiipropati ze-DNN kunye nezenethiwekhi ezisezantsi. Ukwahlulwa kwe-DNN kuxhomekeke ekuchongeni amanqaku okwahlula, angamanqaku kwigrafu yemodeli apho imodeli yahlulwa kwiindawo. Ngexesha lokusebenza, iindawo ziqhutywa ngokulandelelana, nganye ithumela iziphumo zokubala eziphakathi kwindawo elandelayo. Oku kwenza kube lula kwindawo nganye ukuba iqale ukucutshungulwa kwedatha elandelayo ngelixa ezinye izixhobo ziqhubeka nokucutshungulwa kwedatha ethunyelweyo, ngoko ke kuphuculwe ukugqithiswa kokubala, njengoko kubonisiwe kumfanekiso  . Kulo lonke iphepha, ukwahlulwa kwe-DNN kuvezwa ngomzekelo wokubala ixesha lokwenyani kwiintlanganiso zevidiyo. we-1 A. Imisebenzi ehambelanayo Iindlela zokwahlulwa kwe-DNN, eziveliswe ukusuka ekuthwaleni kusuka kwimicubungulo ye-edge ukuya efini, zifuna ukulungelelanisa izinto ezahlukeneyo zokusebenza, njenge-computing latency, ukusetyenziswa kwamandla, ukusetyenziswa kwezixhobo, iindleko, okanye ukugqithiswa. Ukwahlulwa kwe-DNN kuxhomekeke kudityaniso lwezinto ezinye okanye ngaphezulu zezi zinto ukuchaza injongo yokulungelelanisa, kunye nendlela esebenzisa le nto ukufumana izikimu zokwahlula. Umzekelo, iNeurosurgeon  ifuna indawo enye yokwahlula, igcina indawo yokuqala kumaphethelo kwaye ithwala indawo yesibini efini, ukunciphisa ubude bexesha kunye nokusetyenziswa kwamandla. Izicelo kwi-IoT ziye zathatha ukwahlulwa okudibeneyo kunye nokuthwala ii-DNN ezininzi ukulungelelanisa amandla, ukubuyela emva, kunye/okanye iindleko, kusetyenziswa izixazululi ezahlukeneyo ezifana ne-SPSO-GA  idibanisa i-Particle Swarm Optimization (PSO) kunye neGenetics Algorithm (GA), okanye i-DDPQN  esebenzisa i-Deep Reinforcement Learning ukufumana izikimu zokwahlula. Ezinye izibonelo zibandakanya i-DINA , echaza ingxaki ye-Integer Non Linear Programming (INLP), kwaye isebenzisa umxhasi osengqiqweni wesayensi yothutho, ukulungelelanisa ubuyela emva kunye nokusetyenziswa kwezixhobo kwiithutho zemifuno. Iindlela zokulungelelanisa ukugqithiswa kokubala zibandakanya i-DNN surgery , esebenzisa ubukhulu obuncinci bebunye bokubala kwe-DNN ukwahlula ubalele bokubala phakathi komphetho kunye nefini. Zombini iialgorithm zidala iindawo ezimbini, enye kumaphethelo kunye nenye efini. Ezinye izifundo kwibala le-mobile edge-cloud  zikwabandakanya ubalele obuninzi kwifini ukuze kwakhona kukhawuleziswe ukugqithiswa kokubala konke. Ukuxhomekeka kwifini ngobalele besibini bokubala kufuna unxibelelwano lwenethiwekhi oluhle, olunokuthi lungaqiniseki kunxibelelwano olubi lwezicelo zangaphandle ezithile ezikude. I-Edgent  iyazikhupha indlela yangaphambili kwiizixhobo eziphathwayo kunye nee-servara ezikhoyo, ibuyisela isithintelo sokuthwala efini. Le ndlela idityaniswe noqeqesho olukhethekileyo kunye nesakhiwo semodeli, kwaye isaqulela ukwahlula kubini. Ezinye iindlela ziye zathatha amanqaku amaninzi okwahlula, umzekelo, i-IONN , echaza ingxaki yokwahlula ukunciphisa ubude bexesha kunye nokusetyenziswa kwamandla, ikwayithathela ingqalelo ixesha lokulayisha iimodeli kwii-servara ezikude. [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) Ezi ndlela zihlala zidibanisa nee-clouds ezikude, okanye zikhethwe ngendlela yokwahlula, umzekelo, kufuneka zohlulwe zibe zizicucu ezimbini, kunye okanye ngaphandle kobunye. Akukho nalunye ukufundiswa kokwahlulwa kwe-DNN okudweliswe koluphononongo luthathela ingqalelo zombini iindawo ezininzi zokwahlula kunye nobuninzi. B. Igalelo Eli phepha lichaza inkqubo yokubala ehlanjiweyo yokwandisa ukugqithiswa kokubala kokubala kwe-DNN ngokubala ixesha lokwenyani kwiidatha ezihamba phambili, kunye nendawo zokwahlula ezininzi kunye nobahluliweyo. Igalelo leli phepha linje: 
 
 
 
 
 
 Imodeli yokubala kunye nokubala kokuhambisa kobalele obuhlanjiweyo, apho ukugqithiswa okulindelekileyo kokubala kwezicelo zenkqubo yokwahlula kunokuchazwa. Ukubunjwa kwengxaki yokulungelelanisa yokwahlula kwe-DNN, ukusetyenziswa komthwali we-branch and bound kule ngxaki, kunye nokuphathwa kobunzima bayo. Iimlinganiselo zokuhlola ukufumana ukuphucuka okufumanekayo okusebenzisa iipropati zenethiwekhi kunye neze-DNN ezahlukeneyo. Ukuchongwa kweendawo ezichaziweyo kwiipropati zenethiwekhi kunye neze-DNN ezikhokelela kubukho bezicelo zokwahlula ezilungileyo kunye neendleko zokubala ezi zisombululo, oko kukuthi, iimeko apho ukwahlulwa kwe-DNN kunenzuzo. Izilulo zemvavanyo ezibonisa iindawo ezichazwe kwiimlinganiselo, kunye nokuchaneka kokuchaza kunye nokuqhubela phambili ukukhawulezisa kokubala kobuninzi kuzo zonke iindawo ezifanayo nezingafaniyo. Ngalo lonke iphepha, ukuthintela ukudideka, i-deep neural network ehlulweyo iya kuthathwa njenge  , kwaye igama   libonisa inethiwekhi yothumelo yezothutho esezantsi. DNN network C. Isikhokelo sePhepha Ukusala kweli phepha kuhlelwe ngolu hlobo:    sibonisa isiseko sokufunda okunzulu kunye nendlela ii-DNN ezimelwe ngayo koluphononongo.    sichaza iindlela zokuchaza ubude bexesha bokubala kunye nobunye bokuhambisa ezifuneka kwisahluko   esichaza ingxaki yokulungelelanisa, kunye ne-algorithm ye-branch and bound yokwenza izigqibo zokwahlula.    sibonisa iziphumo zemilinganiselo ukuze kubalwe ukusebenza kunye nobunzima be-algorithm ye-branch and bound, kunye neemeko ezikhokelela kubukho besicelo sokwahlula esiphucula ukugqithiswa kokubala kwi-network efanayo.    sibonisa izilulo zemvavanyo eziqinisekisa iimilinganiselo ze  , kwaye ibonisa ukuphuculwa kokugqithiswa kokubala kwisiseko sovavanyo esingafaniyo. Okokugqibela,    siyaxoxa ngeziphumo kwaye siwandise kusebenzekeni okubanzi kwale ndlela. Isahluko II Isahluko III IV Isahluko V Isahluko VI Isahluko V Isahluko VII ISahlulo II. Isiseko: Ukufunda Okunzulu Ngaphakathi kwibala le-Machine Learning, i-Deep Learning (DL) ibhekisa kwiindlela ezixhomekeke kusetyenziso lwe-DNNs, oko kukuthi, iinethiwekhi zenaliti zobugcisa (okanye iindlela ezifanayo zokufunda komatshini) ezinolunye olunye umaleko obufihliweyo. Enye yezona zinto zangaphambili malunga ne-deep learning architectures yapapashwa ngo-1967 , kodwa kwakungekho de kwiminyaka yama-2010 apho i-DL yafumana amandla, kuqala ekuqaphelweni kwelizwi, emva koko ekuboneni izinto kwimifanekiso, ngenxa yobukho bokubamba ubudlelwane obuqatha kwidatha enobuninzi obuphezulu. I-Deep Learning kulindeleke ukuba ibeyinxalenye enkulu yezicelo ze-AI ngo-2022, kunye nemibutho engaphezu kwe-75% esebenzisa ii-DNNs kwizicelo ezinokusebenzisa iindlela zakudala. Iindlela ze-DL ziye zachazwa njenge "   " . Ezi sekhiti zobalo zihlala zihlelwa kwiingqeqe, ezakha amanyathelo okubala. Igama   libhekisa kwiialgorithm ezibandakanya ngaphezu komaleko omnye. [](javascript:void()) iindlela zokufunda komatshini apho izimvo zithatha uhlobo lweesekethi zobalo ezintsonkothileyo ezineendlela zokuqiniswa ezilungisekayo [](javascript:void()) deep Zimbini iintlobo eziphambili ze-DNN: iinethiwekhi zenaliti ezingapheliyo kunye neenethiwekhi zenaliti ezithunyelweyo. Zombini iintlobo ze-DNN zinokumelwa njengeegrafu zokubala, kunye neyona nto ingafaniyo kukuba iinethiwekhi ezingapheliyo zingamela njengeegrafu ezingapheliyo (DAGs), ngelixa iinethiwekhi ezingaphambili zingabandakanya imijikelezo. Iinethiwekhi ezingaphambili zakhelwe imijikelezo yedatha, apho yonke idatha inokuxhomekeka kwidatha yangaphambili kumjikelezo, ngelixa iinethiwekhi ezingapheliyo zingathatheli ingqalelo unxibelelwano phakathi kwamanqaku kumjikelezo wedatha. Ukusala kweli phepha siya kuthatha isigqibo kwisigqibo se-feed-forward DNNs. 5 ISahlulo III. Imodeli Yokubala Ebuhanjelweyo Eli Candelo lichaza imodeli yokumelwa kweempawu ze-DNN kunye nezenethiwekhi ukuze zichaze ubude bexesha bokubala, ubude bexesha bokuhambisa, kunye nokugqithiswa kokubala kokugqibela kwesisombululo esithile sokwahlula. A. Imifanekiso ye-DNN kunye neNethiwekhi I-DNN engapheliyo ye-N iilanga imelwe njenge-DAG GA=(L,E) kunye ne L=(L1,…,LL) ziilanga ze-DNN. Imiphetho (Li,Lj)∈E zizityalo phakathi kweelanga Li kunye noLj. Wonke umaleko Li unokubala okubalayo okubalayo ci , ekalinganiswa kwinani lemisebenzi yokudada yebhoma efunekayo ukubala inyathelo eliphambili ngeyona ilanga. Imiphetho (Li,Lj) inikwe ubunzima si,j obufana nobukhulu bedatha ehamba phakathi kweelanga Li kunye noLj ngama-bytes. Inethiwekhi yenyama imelwe njengegrafu egqityiweyo G=(N,V) apho N=(N1,…,NN) yiseti yee-nodes zokubala kunye no-V yiseti yezitya phakathi kwee-nodes. Kuthathwa ukuba ii-nodes zokubala N1,…,NN zinemilinganiselo yokubala η1,…,ηN, ngokulandelanayo, ekalinganiswe kwi-floating-point operations ngomzuzwana. Ekugqibeni, ukugqithiswa kwekhonkco phakathi kwezimbini i-nodes ezikufuphi Na kunye no Nb kuthathelwa ingqalelo θa,b , ekalinganiswe ngama-bytes ngomzuzwana. Wonke umNode Ni uxhumene naye ngokwakhe kunye nokugqithiswa okungapheliyo ukumelana nekhonkco lokubuyela umva, kwaye izitya zithathwa njengezilinganisi, oko kukuthi, θa,b=θb,a . Izicelo zokwahlula zichazwa njengeemephu P:L→N , oko kukuthi, ukwahlulwa kunika inombolo ye-node kulo naliphi na ilanga kwi-DNN. Isicelo sokwahlula sinokuchazwa njenge-matrix P yobukhulu (N×L) , N ibala i-nodes kwinethiwekhi kunye no-L inani lee-langa ze-DNN kwaye, kunye nomzekelo onikwe kumfanekiso  , L=7 iilanga kunye no-N=2 i-nodes zokubala, ukwahlula okubonisiweyo kunye neelanga {L1,L2,L3,L6,L7} kwi-node N1 , kunye neelanga {L4,L5} kwi-node N2 , kuboniswe njenge: we-2 Ukunikwa kwisicelo, i-  ichazwa njengeqela leelanga ezilandelelanayo phakathi kweendawo ezimbini zokwahlula, eziqhutywa ngokulandelelana kwi-node enye. Kumzekelo ongentla, i-node N1 iya kuqhuba imisonto emibini, yokuqala equka iilanga {L1,L2,L3} kunye neyesibini equka {L6,L7} . thread Kunye nale miqobo ichaziweyo, ukusala kweli Candelo kuya kuchaza imbeko evaliweyo yokugqithiswa kokubala okufunyenweyo kwisicelo esithile, njengomsebenzi we-inference (isahluko III-B) kunye nobude bexesha bokuhambisa (isahluko III-C). B. Ukuchaza Ubude bexesha bokubala Ukunikwa kwisicelo, eli Candelo sijonga ubude bexesha bokubala kwisicelo esikwisixhobo esizimeleyo. Ubude bexesha obubangelwa kukubala kwelanga Li kunye nokubala okubalayo ci , ukuba kufuneka kubalwe kwi-node Na , kunye nomlinganiselo wokubala ηa ubalwa njenge Tci(Na)=ci/ηa . Kunikwa iqela leelanga L′⊂L kulo lonke imisonto eqhutywa kwi-node Na kunye nomlinganiselo wokubala ηa , ubude bexesha bokubala bunokuchazwa njenge: Lo mboniso yinto yokuqala yokulinganisa ubude bexesha bokubala esekelwe kwinani lemisebenzi yokudada yebhoma (FLOPs) efunekayo ukucutshungulwa kwe-DNN, oko kukuthi, inani lemisebenzi yokubuyabuyela-yongeza kwisigqokele. Ukuchaza ubude bexesha bokubala kumaziko angaphandle kubangela ubunzima ngenxa ye-run-time optimizations. Izisombululo ezikhoyo ezifana ne-nn-Meter  zichaza ubude bexesha bokubala ngokusekelwe kwi-FLOPs. Ezinye izisombululo zixhomekeke kwitshati yokujonga enexesha elibaliweyo lokubala ukuze kubalwe ubude bexesha bokubala. Umzekelo, i-BRP-NAS  isebenzisa inethiwekhi yenaliti egrafu eqeqeshwe kwangaphambili ukuchaza ubude bexesha bokubala, ngelixa ithathela ingqalelo ii-run-time model optimizations. [](javascript:void()) [](javascript:void()) Kwi- , kuphinda kuthathelwe ingqalelo ukuba abaculi basebenzisa amandla abo apheleleyo, nto leyo eyenza ukuba abaculi abaninzi bazibonele njengabaculi abanobunye kunye nomlinganiselo wokubala obulingana nesixa semilinganiselo yabo yomgbebe. Kwaye kuthathelwa ingqalelo ukuba isixhobo sokubala sabelwana ngokulinganayo phakathi kwemimiselo, oko kukuthi, umculi uchitha inxalenye efanayo yexesha lakhe kulo lonke usasazo. incoko 1 Kuzo zonke i-nodes, i-thread enobude bexesha obude ijongana nokugqithiswa kokubala okuhanjelweyo kwe-DNN. Ukushiya ubude bexesha bokuhambisa, oku kuthetha ukuba xa le thread icokozelayo igqibezela ukubala if

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Read My Stories

Cisco Systems

Le audio iveliswe ngolwimi lokuqala lwebali!

Kung Kutheni i-AI Yakho Ayisadingi i-Cloud Ukuba Isebenze Ngokukhawuleza Ngoku

About Author

IZIMVO

ZIJONGE IIMPAWU

ELI NQAKU LINIKEZELWE KU

Related Stories

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps