Isanda kukhutshwa , ebonisa umkhwa omhle wokufunda koomatshini kwiminyaka yamuva. Ngamafutshane: inani lokuqalwa komatshini wokufunda liye lehla kule minyaka mibini idlulileyo.
Kulungile. Makhe sijonge "ingaba i-bubble iqhume", "indlela yokuqhubeka uphila" kwaye sithethe malunga nokuba le squiggle ivela phi kwindawo yokuqala.
Okokuqala, makhe sithethe malunga nokuba yayiyintoni i-booster yeli jiko. Wayevela phi? Mhlawumbi baya kukhumbula yonke into ukufunda koomatshini kwi-2012 kukhuphiswano lwe-ImageNet. Ngapha koko, esi sisiganeko sokuqala ehlabathini! Kodwa enyanisweni oku akunjalo. Kwaye ukukhula kwegophe kuqala kancinci ngaphambili. Bendiya kuyiqhekeza ibe ngamanqaku amaninzi.
- I-2008 yabona ukuvela kwegama elithi "idatha enkulu". Iimveliso zokwenyani zaqala ukusukela ngo-2010. Idatha enkulu ihambelana ngqo nokufunda koomatshini. Ngaphandle kwedatha enkulu, ukusebenza okuzinzile kwee-algorithms ezazikho ngelo xesha akunakwenzeka. Kwaye ezi ayizizo iinethiwekhi ze-neural. Ukuza kuthi ga ngo-2012, uthungelwano lwe-neural lwalulondolozwe ligcuntswana le-marginal. Kodwa ke i-algorithms eyahluke ngokupheleleyo yaqala ukusebenza, eyayikho iminyaka, okanye amashumi eminyaka: (1963,1993), (1995) (2003)
I-derivative yale wave yokuqala yiseti yesakhelo ezifana ne-XGBoost, CatBoost, LightGBM, njl.
- Ngo-2011-2012 iphumelele inani lokhuphiswano lokuqondwa komfanekiso. Ukusetyenziswa kwazo okona kwalibaziseka ngandlel’ ithile. Ndingatsho ukuba uqalo olunentsingiselo kunye nezisombululo zaqala ukuvela ngo-2014. Kwathatha iminyaka emibini ukugaya ukuba i-neurons isasebenza, ukudala izikhokelo ezifanelekileyo ezinokufakelwa kwaye ziqaliswe ngexesha elifanelekileyo, ukuphuhlisa iindlela eziza kuzinzisa kwaye zikhawuleze ixesha lokudibanisa.
Iinethiwekhi ze-Convolutional zenze ukuba kube lula ukusombulula iingxaki zombono wekhompyuter: ukuhlelwa kwemifanekiso kunye nezinto ezisemfanekisweni, ukubonwa kwezinto, ukuqatshelwa kwezinto kunye nabantu, ukuphuculwa komfanekiso, njl., njl.
- 2015-2017. I-boom ye-algorithms kunye neeprojekthi ezisekelwe kuthungelwano oluphindaphindiweyo okanye ii-analogues zabo (LSTM, GRU, TransformerNet, njl.). Ukusebenza kakuhle kwentetho-ukuya-kumbhalo algorithms kunye neenkqubo zokuguqulela koomatshini ziye zavela. Ngokuyinxenye zisekwe kuthungelwano lwe-convolution ukukhupha iimpawu ezisisiseko. Ngokuyinxenye ngenxa yokuba sifunde ukuqokelela iiseti zedatha ezinkulu kakhulu nezilungileyo.
“Ingaba iqamza ligqabhukile? Ngaba i-hype ishushu kakhulu? Ngaba bafa njenge-blockchain?"
Kungenjalo! Ngomso uSiri uya kuyeka ukusebenza kwifowuni yakho, kwaye ngosuku olulandelayo uTesla akayi kuwazi umahluko phakathi kokujika kunye nekhangaru.
Uthungelwano lweNeural sele lusebenza. Bakwizixhobo ezininzi. Bakuvumela ngokwenene ukuba ufumane imali, utshintshe imarike kunye nehlabathi elikujikelezayo. I-Hype ibonakala yahluke kancinci:
Kuphela nje iinethiwekhi ze-neural azisekho into entsha. Ewe, abantu abaninzi balindele lukhulu. Kodwa inani elikhulu leenkampani lifunde ukusebenzisa i-neurons kunye nokwenza iimveliso ezisekelwe kuzo. IiNeurons zibonelela ngokusebenza okutsha, zikuvumela ukuba unciphise imisebenzi, kwaye unciphise amaxabiso eenkonzo:
- Iinkampani zokuvelisa zidibanisa i-algorithms ukuhlalutya iziphene kumgca wemveliso.
- Iifama zemfuyo zithenga iinkqubo zokulawula iinkomo.
- Ukudibanisa okuzenzekelayo.
- Amaziko oMnxeba azenzekelayo.
- Izihluzi kwi-SnapChat. (Kulungile, ubuncinci into eluncedo!)
Kodwa eyona nto iphambili, kwaye hayi eyona icacileyo: "Akusekho mibono emitsha, okanye abayi kuzisa inkunzi kwangoko." Uthungelwano lweNeural lusombulule inkitha yeengxaki. Kwaye baya kwenza isigqibo ngakumbi. Zonke iimbono ezicacileyo ezazikho zabangela uqalo oluninzi. Kodwa yonke into eyayiphezu komhlaba yayisele iqokelelwe. Kule minyaka mibini idlulileyo, andikhange ndibe nombono omnye omtsha wokusetyenziswa kothungelwano lwe-neural. Ayikho enye indlela entsha (kakuhle, kulungile, kukho iingxaki ezimbalwa ngee-GAN).
Kwaye isiqalo ngasinye esilandelayo sinzima ngakumbi nangakumbi. Ayisafuni abafana ababini abaqeqesha i-neuron besebenzisa idatha evulekileyo. Ifuna abaprogram, iseva, iqela labamakishi, inkxaso enzima, njl.
Ngenxa yoko, zimbalwa iziqalo. Kodwa kukho imveliso engakumbi. Ngaba ufuna ukongeza ulwazi lwepleyiti yelayisensi? Kukho amakhulu eengcali ezinamava afanelekileyo kwimarike. Ungaqesha umntu kwaye kwiinyanga ezimbalwa umqeshwa wakho uya kwenza inkqubo. Okanye uthenge sele selenziwe. Kodwa ukwenza isiqalo esitsha?.. Uphambene!
Kuya kufuneka udale inkqubo yokulandelela iindwendwe - kutheni uhlawulela iqela leelayisenisi xa unokwenza eyakho kwiinyanga ezi-3-4, ulolele ishishini lakho.
Ngoku iinethiwekhi ze-neural zihamba ngendlela efanayo naleyo ininzi yetekhnoloji eye yadlula kuyo.
Ngaba uyakhumbula ukuba ingcamango "yomphuhlisi wewebhusayithi" itshintshile njani ukususela kwi-1995? Imakethi ayikazaliswa ziingcali. Zimbalwa kakhulu iingcali. Kodwa ndingabheja ukuba kwiminyaka emi-5 ukuya kweli-10 akusayi kubakho mahluko ungako phakathi komdwelisi weprogram yeJava kunye nomphuhlisi wenethiwekhi ye-neural. Kuya kubakho ngokwaneleyo kwiingcali zombini kwimarike.
Kuya kubakho udidi lweengxaki ezinokuthi zisonjululwe ngemithambo-luvo. Kuye kwavela umsebenzi - ukuqesha ingcali.
"Yintoni elandelayo? Buphi ubukrelekrele bokuzenzela obuthenjisiweyo?”
Kodwa apha kukho ukungaqondi okuncinci kodwa okunomdla :)
Isitaki setekhnoloji esikhoyo namhlanje, ngokucacileyo, asiyi kusikhokelela kubukrelekrele bokwenziwa. Izimvo kunye nobutsha bazo ubukhulu becala zidiniwe. Makhe sithethe ngento ebambe inqanaba langoku lophuhliso.
Zi thintelo
Masiqale ngeemoto eziziqhubayo. Kubonakala kucacile ukuba kunokwenzeka ukwenza iimoto ezizimeleyo ngokupheleleyo ngeteknoloji yanamhlanje. Kodwa iminyaka emingaphi oku kuya kwenzeka akucaci. UTesla ukholelwa ukuba oku kuya kwenzeka kwiminyaka embalwa-
Kukho nabanye abaninzi , abaqikelela ukuba yiminyaka emi-5-10.
Okunokwenzeka, ngokombono wam, kwiminyaka eyi-15 iziseko zezixeko ziya kutshintsha ngokwazo ngendlela yokuba ukuvela kweemoto ezizimeleyo kuya kuba yinto engenakuphepheka kwaye iya kuba ukuqhubeka kwayo. Kodwa oku akunakuthathwa njengobukrelekrele. I-Tesla yanamhlanje yipayipi enzima kakhulu yokucoca idatha, ukukhangela kunye nokuqeqeshwa kwakhona. Le yimithetho-imithetho-imithetho, ukuqokelelwa kwedatha kunye nezihluzo phezu kwayo (apha Ndibhale kancinci malunga noku, okanye ndibukele ukusuka amanqaku).
Ingxaki yokuqala
Kwaye kulapho sibona khona ingxaki yokuqala esisiseko. Idatha enkulu. Yile nto kanye eyazala amaza akhoyo othungelwano lwe-neural kunye nokufunda koomatshini. Kule mihla, ukwenza into entsonkothileyo kwaye ngokuzenzekelayo, udinga idatha eninzi. Hayi nje kakhulu, kodwa kakhulu, kakhulu. Sidinga iialgorithms ezizenzekelayo zokuqokelela kwazo, ukumakisha, kunye nokusetyenziswa. Sifuna ukwenza imoto ibone iilori ezijonge elangeni - kufuneka kuqala siqokelele inani elaneleyo lazo. Sifuna ukuba imoto ingaphambani ngebhayisekile ebotshelelwe esiqwini - iisampuli ezininzi.
Ngaphezu koko, umzekelo omnye awanele. Amakhulu? Ngamawaka?
Ingxaki yesibini
Ingxaki yesibini -Ukubona into eqondwayo yinethiwekhi yethu ye-neural. Lo ngumsebenzi ongengomsebenzi omncinane kakhulu. Ukuza kuthi ga ngoku, bambalwa abantu abayiqondayo indlela yokubona oku. La manqaku ngawamva nje, le yimizekelo nje embalwa, nokuba ekude:
Ukuthanda ukuthungwa. Ibonisa kakuhle ukuba i-neuron ithanda ukulungisa ntoni + kwinto eyibona njengokuqala kolwazi.
Ukuqwalaselwa kwi . Enyanisweni, umtsalane unokusetyenziselwa ngokuchanekileyo ukubonisa ukuba yintoni ebangele ukusabela kwenethiwekhi. Ndizibonile izinto ezinjalo zombini ukulungisa iimpazamo kunye nezisombululo zemveliso. Kukho amanqaku amaninzi kulo mbandela. Kodwa okukhona kuntsonkothe ngakumbi idatha, kokukhona kunzima ukuqonda indlela yokufezekisa umbono owomeleleyo.
Ewe, ewe, iseti endala elungileyo “jonga into engaphakathi kumnatha ngaphakathi " Le mifanekiso yayidumile kwiminyaka eyi-3-4 edlulileyo, kodwa wonke umntu waqaphela ngokukhawuleza ukuba imifanekiso yayintle, kodwa yayingenayo intsingiselo eninzi.
Khange ndikhankanye ininzi yezinye izixhobo, iindlela, iihacks, uphando malunga nendlela yokubonisa i-insides yenethiwekhi. Ngaba ezi zixhobo ziyasebenza? Ngaba bayakunceda ukuba uqonde ngokukhawuleza ukuba yintoni ingxaki kwaye ulungise inethiwekhi? .. Fumana ipesenti yokugqibela? Ewe, iphantse ifane:
Ungabukela naluphi na ukhuphiswano kwi Kaggle. Kwaye inkcazo yendlela abantu abazenza ngayo izigqibo zokugqibela. Sipakishe iiyunithi eziyi-100-500-800 zeemodeli kwaye zasebenza!
Ndiyayibaxa, kunjalo. Kodwa ezi ndlela azinikezeli iimpendulo ezikhawulezayo nezithe ngqo.
Ukuba namava aneleyo, uphonononge iinketho ezahlukeneyo, unokunika isigwebo malunga nokuba kutheni inkqubo yakho yenze eso sigqibo. Kodwa kuya kuba nzima ukulungisa indlela yokuziphatha kwenkqubo. Faka i-crutch, hambisa umda, yongeza isethi yedatha, thatha enye inethiwekhi ye-backend.
Ingxaki yesithathu
Ingxaki Yesithathu Esisiseko - iigridi zifundisa izibalo, hayi ingqiqo. Ngokwezibalo oku :
Ngokwengqiqo, akufani kakhulu. Uthungelwano lwe-Neural alufundi nantoni na entsokothileyo ngaphandle kokuba lunyanzeliswe. Basoloko befundisa eyona miqondiso ilula. Unamehlo, impumlo, intloko? Ke lo ubuso! Okanye unike umzekelo apho amehlo engathethi ubuso. Kwaye kwakhona - izigidi zemizekelo.
Kukho Ininzi yeGumbi emazantsi
Ndingatsho ukuba zezi ngxaki zintathu zehlabathi ezinciphisa uphuhliso lwenethiwekhi ye-neural kunye nokufunda koomatshini. Kwaye apho ezi ngxaki zingayikhawulelanga, sele isetyenziswa ngokusebenzayo.
Esi sisiphelo? Ngaba iinethiwekhi ze-neural ziphezulu?
Akwaziwa. Kodwa, ngokuqinisekileyo, wonke umntu akanalo ithemba.
Zininzi iindlela kunye nezalathiso zokusombulula iingxaki ezingundoqo endizibalule ngasentla. Kodwa ukuza kuthi ga ngoku, akukho nanye kwezi ndlela yenze ukuba kwenzeke into entsha ngokusisiseko, ukusombulula into engekasonjululwa. Ukuza kuthi ga ngoku, zonke iiprojekthi ezisisiseko zenziwa ngokwesiseko seendlela ezizinzileyo (Tesla), okanye zihlale iiprojekthi zovavanyo lwamaziko okanye amaqumrhu (Google Brain, OpenAI).
Xa sithetha, ulwalathiso oluphambili kukudala ukumelwa komgangatho ophezulu wedatha yokufaka. Ngandlela-thile, “inkumbulo”. Owona mzekelo ulula wenkumbulo yohlukile "Ufakelo" - ukuboniswa kwemifanekiso. Ewe, umzekelo, zonke iinkqubo zokuqaphela ubuso. Uthungelwano lufunda ukufumana ebusweni ukumelwa okuzinzile okungaxhomekeke ekujikelezweni, ekukhanyeni, okanye ekusonjululweni. Ngokusisiseko, inethiwekhi inciphisa imetriki "ubuso obahlukeneyo bukude" kwaye "ubuso obufanayo busondele."
Kuqeqesho olunjalo, kufuneka amashumi namakhulu amawaka emizekelo. Kodwa isiphumo sithwala ezinye zeziseko ze "One-shot Learning". Ngoku asidingi makhulu obuso ukuze sikhumbule umntu. Ubuso nje obunye kwaye yiyo yonke into esiyiyo !
Inye nje ingxaki... Igridi inokufunda kuphela izinto ezilula ngokufanelekileyo. Xa uzama ukwahlula ubuso, kodwa, umzekelo, "abantu ngeempahla" (umsebenzi ) - umgangatho wehla ngemiyalelo emininzi yobukhulu. Kwaye inethiwekhi ayisakwazi ukufunda utshintsho olucacileyo kwii-engile.
Kwaye ukufunda kwizigidi zemizekelo kukwayinto eyonwabisayo.
Kukho umsebenzi wokunciphisa kakhulu unyulo. Ngokomzekelo, umntu unokukhumbula ngokukhawuleza enye yemisebenzi yokuqala Isifundo seOneShot :
Mininzi imisebenzi enjalo, umzekelo okanye okanye .
Kukho uthabatha omnye - uqeqesho luqhele ukusebenza kakuhle kwimizekelo elula, "MNIST". Kwaye xa uqhubela phambili kwimisebenzi enzima, udinga i-database enkulu, imodeli yezinto, okanye uhlobo oluthile lomlingo.
Ngokubanzi, umsebenzi kuqeqesho lwe-One-Shot sisihloko esinomdla kakhulu. Ufumana iingcamango ezininzi. Kodwa ubukhulu becala, iingxaki ezimbini endizidwelisileyo (ukuqeqeshwa kwangaphambili kwidathasethi enkulu / ukungazinzi kwidatha entsonkothileyo) iphazamisana kakhulu nokufunda.
Ngakolunye uhlangothi, ii-GANs-uthungelwano lwe-adversarial networks-lusondela kwisihloko se-Embedding. Ukhe wafunda iqela lamanqaku malunga noHabré ngalo mxholo. (, ,)
Uphawu lwe-GAN kukubunjwa kwendawo ethile yangaphakathi yelizwe (eyona nto ifana nokuFakela), ekuvumela ukuba uzobe umfanekiso. Inokuba njalo , ingayi .
Ingxaki nge-GAN kukuba eyona nto intsonkothileyo into eyenziweyo, kokukhona kunzima ukuyichaza “kwi-generator-discriminator” logic. Ngenxa yoko, ekuphela kwezicelo zokwenyani ze-GAN eziviwayo yi-DeepFake, ethi, kwakhona, isebenzise ukumelwa kobuso (ekukho isiseko esikhulu sayo).
Ndibone ezinye izinto ezimbalwa eziluncedo. Ngokuqhelekileyo uhlobo oluthile lobuqhetseba olubandakanya ukugqiba imizobo yemifanekiso.
Kwaye kwakhona. Akukho mntu unombono wokuba oku kuya kusivumela njani ukuba sihambe siye kwikamva eliqaqambileyo. Ukumela ingqiqo / indawo kwinethiwekhi ye-neural kulungile. Kodwa sidinga inani elikhulu lemizekelo, asiyiqondi indlela i-neuron emele ngayo le nto ngokwayo, asiqondi indlela yokwenza i-neuron ikhumbule ingcamango enzima ngokwenene.
Ukuqinisa ukufunda - le yindlela esuka kwicala elahluke ngokupheleleyo. Ngokuqinisekileyo uyakhumbula ukuba uGoogle wabetha wonke umntu kwiGo. Uloyiso lwakutsha nje kwiStarcraft kunye neDota. Kodwa apha yonke into ikude kakhulu kwaye iyathembisa. Uthetha kakuhle nge-RL kunye nobunzima bayo .
Ukushwankathela ngokufutshane oko umbhali akubhalileyo:
- Iimodeli ezingaphandle kwebhokisi azilingani / zisebenza kakubi kwiimeko ezininzi
- Iingxaki ezisebenzayo kulula ukuzicombulula ngezinye iindlela. IBoston Dynamics ayisebenzisi iRL ngenxa yokuntsokotha kwayo/ukungacingeleki/ukuntsokotha kwayo
- Ukuze i-RL isebenze, udinga umsebenzi onzima. Kuhlala kunzima ukuyila/ukubhala
- Kunzima ukuqeqesha iimodeli. Kuya kufuneka uchithe ixesha elininzi ukumpompa kwaye uphume kwindawo efanelekileyo yendawo
- Ngenxa yoko, kunzima ukuphinda imodeli, imodeli ayizinzile kunye notshintsho oluncinci
- Ikholisa ukugqwesa iipateni ezithile ezingalandelekiyo, kunye nenani elingaqhelekanga lejenereyitha
Inqaku eliphambili kukuba iRL ayikasebenzi kwimveliso. UGoogle uneminye imifuniselo ( , ). Kodwa andizange ndibone inkqubo enye yemveliso.
Imemori. I-downside yayo yonke into echazwe ngasentla kukungabikho kwesakhiwo. Enye yeendlela zokuzama ukuqoqosha konke oku kukubonelela ngenethiwekhi ye-neural ukufikelela kwimemori eyahlukileyo. Ukuze akwazi ukurekhoda aze aphinde abhale iziphumo zamanyathelo akhe apho. Emva koko inethiwekhi ye-neural inokumiselwa yimeko yememori yangoku. Oku kufana kakhulu neeprosesa zakudala kunye neekhompyuter.
Eyona idumileyo nethandwayo — ukusuka kwi-DeepMind:
Kubonakala ngathi le ngundoqo ekuqondeni ubukrelekrele? Kodwa mhlawumbi akunjalo. Inkqubo isafuna isixa esikhulu sedatha yoqeqesho. Kwaye isebenza ikakhulu ngedatha yetheyibhile ecwangcisiweyo. Ngaphezu koko, xa Facebook Ingxaki efanayo, emva koko bathatha indlela "imemori ekrazulweyo, yenza nje i-neuron ibe nzima ngakumbi, kwaye ube nemizekelo emininzi - kwaye iya kufunda ngokwayo."
Ukwahlukana. Enye indlela yokudala inkumbulo enentsingiselo kukuthatha uthungelwano olufanayo, kodwa ngexesha loqeqesho, yazisa iindlela ezongezelelweyo eziya kukuvumela ukuba uqaqambise “iintsingiselo” kuzo. Umzekelo, sifuna ukuqeqesha inethiwekhi ye-neural ukwahlula phakathi kokuziphatha komntu evenkileni. Ukuba silandele indlela eqhelekileyo, kuya kufuneka senze iinethiwekhi ezilishumi elinesibini. Omnye ukhangela umntu, owesibini umisela into ayenzayo, owesithathu yiminyaka yakhe, owesine sisini sakhe. Ingqiqo eyahlukileyo ijonga indawo yevenkile apho yenza / iqeqeshelwa ukwenza oku. Owesithathu umisela umkhondo wayo, njl.
Okanye, ukuba bekukho inani elingenasiphelo ledatha, ngoko kuya kwenzeka ukuqeqesha inethiwekhi enye kuzo zonke iziphumo ezinokwenzeka (ngokucacileyo, uluhlu olunjalo lwedatha alunakuqokelelwa).
Indlela yokwahlulahlula iyasixelela - masiqeqeshe inethiwekhi ukuze yona ngokwayo ikwazi ukwahlula phakathi kweengqikelelo. Ukuze ibe nokubethelela ngokusekwe kwividiyo, apho ummandla othile uya kumisela isenzo, omnye uya kumisela indawo emgangathweni kwangethuba, omnye uya kumisela ubude bomntu, aze omnye agqibe ngesini somntu. Kwangaxeshanye, xa ndiqeqeshwa, ndingathanda phantse ukuba ndingawukhawulezisi uthungelwano ngeekhonsepthi ezingundoqo ezinjalo, kodwa endaweni yokuba iqaqambise kunye neendawo zeqela. Kukho amanqaku ambalwa anjalo (amanye awo , , ) kwaye ngokubanzi ziyithiyori.
Kodwa olu lwalathiso, ubuncinci ngokwethiyori, kufuneka lugubungele iingxaki ezidweliswe ekuqaleni.
Ukubola komfanekiso ngokweparameters "umbala wodonga / umbala womgangatho / imilo yento / umbala wento / njl."
Ukubola kobuso ngokweparameters "ubungakanani, amashiya, ukuqhelaniswa, umbala wolusu, njl.
Прочее
Zininzi ezinye iindawo, ezingezizo zehlabathi jikelele, ezikuvumela ukuba ngandlel’ ithile unciphise isiseko sedatha, usebenze ngedatha eyongezelelekileyo, njl.
ingqalelo. Mhlawumbi akukho ngqiqweni ukwahlula oku njengendlela eyahlukileyo. Indlela nje eyomeleza abanye. Amanqaku amaninzi anikezelwe kuye (,,). Inqaku lokuqwalaselwa kukuphucula impendulo yenethiwekhi ngokukodwa kwizinto ezibalulekileyo ngexesha loqeqesho. Rhoqo ngohlobo oluthile lwenjongo ekujoliswe kuyo yangaphandle, okanye inethiwekhi encinci yangaphandle.
Ukulinganisa kwe-3D. Ukuba wenza i-injini ye-3D elungileyo, unako ukugubungela i-90% yedatha yoqeqesho nayo (ndade ndabona umzekelo apho phantse i-99% yedatha igutyungelwe yi-injini elungileyo). Kukho iingcamango ezininzi kunye nee-hacks malunga nendlela yokwenza inethiwekhi eqeqeshwe kwi-injini ye-3D isebenze usebenzisa idatha yokwenyani (Ukulungiswa kakuhle, ukuhanjiswa kwesitayela, njl.). Kodwa ngokufuthi ukwenza i-injini elungileyo ziiodolo ezininzi zobukhulu ezinzima kunokuqokelela idatha. Imizekelo xa iinjini zenziwe:
Uqeqesho lwerobhothi (, )
Iiseshoni zoqeqesho iimpahla evenkileni (kodwa kwiiprojekthi ezimbini esizenzileyo, sinokwenza lula ngaphandle kwayo).
Uqeqesho eTesla (kwakhona, ividiyo engentla).
ezifunyanisiweyo
Inqaku lilonke, ngandlel’ ithile, linezigqibo. Mhlawumbi owona myalezo uphambili ebendifuna ukuwenza ibikukuba "izinto zasimahla ziphelile, ii-neuron azisaniki zisombululo ezilula." Ngoku kufuneka sisebenze nzima ukuze senze izigqibo ezinzima. Okanye usebenze nzima ukwenza uphando oluntsonkothileyo lwezenzululwazi.
Ngokubanzi, isihloko siyaxoxwa. Mhlawumbi abafundi banemizekelo enomdla ngakumbi?
umthombo: www.habr.com