Ingabe ibhamuza lokufunda umshini seliqhumile, noma ingabe isiqalo sokuntwela kokusa?

Okusanda kukhishwa indatshana, okubonisa ukuthambekela okuhle ekufundeni komshini eminyakeni yamuva. Ngamafuphi: inani lokuqalisa ukufunda ngomshini lehlile eminyakeni emibili edlule.

Hhayi-ke. Ake sibheke "ukuthi i-bubble iqhume", "indlela yokuqhubeka nokuphila" futhi sikhulume ngokuthi le squiggle ivela kuphi kwasekuqaleni.

Okokuqala, ake sikhulume ngokuthi kwakuyini i-booster yaleli jika. Uvelaphi? Cishe bazokhumbula yonke into ukunqoba ukufunda ngomshini ngo-2012 emqhudelwaneni we-ImageNet. Phela, lesi isenzakalo sokuqala emhlabeni wonke! Kodwa empeleni akunjalo. Futhi ukukhula kwejika kuqala ngaphambi kwesikhathi. Bengingayihlukanisa ibe amaphuzu amaningana.

1. Ngo-2008 kwavela igama elithi "idatha enkulu". Imikhiqizo yangempela iqalile ukuvela kusukela ngo-2010. Idatha enkulu ihlobene ngokuqondile nokufunda komshini. Ngaphandle kwedatha enkulu, ukusebenza okuzinzile kwama-algorithms ayekhona ngaleso sikhathi akunakwenzeka. Futhi lawa akuwona amanethiwekhi e-neural. Kuze kube ngu-2012, amanethiwekhi e-neural ayengawokugcina abantu abambalwa. Kodwa-ke, ama-algorithms ahluke ngokuphelele aqala ukusebenza, ayekhona iminyaka, noma ngisho namashumi eminyaka: I-SVM(1963,1993), Ihlathi Elingahleliwe (1995), I-AdaBoost (2003)

   Okuphuma kuleli gagasi lokuqala isethi yezinhlaka ezifana ne-XGBoost, CatBoost, LightGBM, njll.

2. Ngo-2011-2012 amanethiwekhi we-convolutional neural uwine imincintiswano eminingi yokuqashelwa kwezithombe. Ukusetshenziswa kwabo kwangempela kwabambezeleka ngandlela-thile. Ngingasho ukuthi iziqalo ezinengqondo nezisombululo zaqala ukuvela ngo-2014. Kuthathe iminyaka emibili ukugaya ukuthi ama-neuron asasebenza, ukudala izinhlaka ezikahle ezingafakwa futhi zethulwe ngesikhathi esiphusile, ukuthuthukisa izindlela ezizosimamisa futhi zisheshise isikhathi sokuhlangana.

   Amanethiwekhi e-Convolutional enze kwaba nokwenzeka ukuxazulula izinkinga zombono wekhompyutha: ukuhlukaniswa kwezithombe nezinto ezisesithombeni, ukutholwa kwezinto, ukuqashelwa kwezinto nabantu, ukuthuthukiswa kwesithombe, njll., njll., njll.

3. 2015-2017. I-boom of algorithms namaphrojekthi asekelwe kumanethiwekhi aphindekayo noma ama-analogue awo (LSTM, GRU, TransformerNet, njll.). Ama-algorithms asebenza kahle enkulumo-kuya-umbhalo kanye nezinhlelo zokuhumusha ngomshini sezivele. Ngokwengxenye zisekelwe kumanethiwekhi okuxhumana ukuze kukhishwe izici eziyisisekelo. Ngokwengxenye ngenxa yokuthi sifunde ukuqoqa amasethi edatha amakhulu namahle kakhulu.

“Ingabe liqhumile ibhamuza? Ingabe i-hype ishisa ngokweqile? Ngabe bafa njenge-blockchain?"
Kungenjalo! Kusasa uSiri uzoyeka ukusebenza ocingweni lwakho, futhi ngakusasa uTesla ngeke azi umehluko phakathi kokujika nekhangaru.

Amanethiwekhi e-Neural aseyasebenza. Zikumadivayisi amaningi. Bakuvumela ngempela ukuthi uthole imali, ushintshe imakethe nomhlaba okuzungezile. I-Hype ibukeka ihluke kancane:

Ukuthi nje amanethiwekhi e-neural awaseyona into entsha. Yebo, abantu abaningi balindele lukhulu. Kodwa inani elikhulu lezinkampani lifunde ukusebenzisa ama-neurons nokwenza imikhiqizo ngokusekelwe kuwo. Ama-Neurons ahlinzeka ngokusebenza okusha, akuvumela ukuthi unciphise imisebenzi, futhi wehlise intengo yezinsizakalo:

• Izinkampani ezikhiqizayo zihlanganisa ama-algorithm ukuze zihlaziye amaphutha emugqeni wokukhiqiza.
• Amapulazi emfuyo athenga izinhlelo zokulawula izinkomo.
• Kuhlanganisa okuzenzakalelayo.
• Izikhungo Zezingcingo ezizenzakalelayo.
• Izihlungi ku-SnapChat. (kahle, okungenani okuthile okuwusizo!)

Kodwa into esemqoka, hhayi esobala kakhulu: "Ayisekho imibono emisha, noma ngeke ilethe imali esheshayo." Amanethiwekhi e-Neural axazulule inqwaba yezinkinga. Futhi bazonquma nakakhulu. Yonke imibono esobala eyayikhona yabangela iziqalo eziningi. Kodwa konke okwakuphezu komhlaba kwase kuqoqiwe. Kule minyaka emibili edlule, angikaze ngihlangabezane nombono owodwa omusha wokusetshenziswa kwamanethiwekhi e-neural. Ayikho indlela entsha (kahle, kulungile, kunezinkinga ezimbalwa ngama-GAN).

Futhi ukuqalisa ngakunye okulandelayo kuba nzima nakakhulu. Akusadingi abafana ababili abaqeqesha i-neuron besebenzisa idatha evulekile. Idinga abahleli bezinhlelo, iseva, ithimba labamaka, ukwesekwa okuyinkimbinkimbi, njll.

Ngenxa yalokho, kukhona ukuqalisa okumbalwa. Kodwa kukhona ukukhiqizwa okwengeziwe. Udinga ukungeza ukuqashelwa kwepuleti lelayisense? Kunamakhulu ochwepheshe abanolwazi olufanele emakethe. Ungaqasha othile futhi ezinyangeni ezimbalwa isisebenzi sakho sizokwenza uhlelo. Noma thenga esenziwe ngomumo. Kodwa wenza isiqalo esisha?.. Uyahlanya!

Udinga ukwakha uhlelo lokulandelela izivakashi - kungani ukhokhela inqwaba yamalayisensi kuyilapho ungenza owakho ezinyangeni ezi-3-4, ulolele ibhizinisi lakho.

Manje amanethiwekhi e-neural adlula endleleni efanayo namashumi obunye ubuchwepheshe adlule kuyo.

Uyakhumbula ukuthi umqondo "wonjiniyela wewebhusayithi" ushintshe kanjani kusukela ngo-1995? Imakethe ayikagcwaliswa ngochwepheshe. Kukhona ochwepheshe abambalwa kakhulu. Kodwa ngingabheja ukuthi eminyakeni emi-5-10 ngeke kube nomehluko omkhulu phakathi komklami we-Java kanye nomthuthukisi wenethiwekhi ye-neural. Kuzokwanela zombili ochwepheshe emakethe.

Kuzomane kube nekilasi lezinkinga ezingaxazululwa ngama-neurons. Kuye kwavela umsebenzi - ukuqasha uchwepheshe.

"Yini elandelayo? Buphi ubuhlakani bokwenziwa obuthenjisiwe?”

Kodwa lapha kunokungaqondi okuncane kodwa okuthakazelisayo :)

Inqwaba yobuchwepheshe ekhona namuhla, ngokusobala, ngeke isiholele kubuhlakani bokwenziwa. Imibono kanye nentsha yayo izikhathaze kakhulu. Ake sikhulume ngalokho okuphethe izinga lamanje lentuthuko.

Izithibelo

Ake siqale ngezimoto ezizishayelayo. Kubonakala kusobala ukuthi kungenzeka ukwenza izimoto ezizimele ngokugcwele ngobuchwepheshe banamuhla. Kodwa eminyakeni emingaki lokhu kuzokwenzeka akucaci. U-Tesla ukholelwa ukuthi lokhu kuzokwenzeka eminyakeni embalwa -

Baningi abanye ochwepheshe, abalinganisa iminyaka engu-5-10.

Cishe, ngombono wami, eminyakeni engu-15 ingqalasizinda yamadolobha izoshintsha ngokwawo ngendlela yokuthi ukuvela kwezimoto ezizimele kuyoba okungenakugwemeka futhi kuzoba ukuqhubeka kwayo. Kodwa lokhu akukwazi ukubhekwa njengokuhlakanipha. I-Tesla yesimanje iyipayipi eyinkimbinkimbi kakhulu yokuhlunga idatha, ukusesha nokuqeqesha kabusha. Lena imithetho-imithetho-imithetho, ukuqoqwa kwedatha nezihlungi phezu kwayo (lapha lapha Ngibhale okwengeziwe mayelana nalokhu, noma ngibuke kusuka lokhu amamaki).

Inkinga yokuqala

Futhi lapha sibona inkinga yokuqala eyisisekelo. Idatha enkulu. Yilokhu kanye okuzale igagasi lamanje lamanethiwekhi e-neural nokufunda komshini. Kulezi zinsuku, ukwenza into eyinkimbinkimbi futhi ezenzakalelayo, udinga idatha eningi. Hhayi nje okuningi, kodwa kakhulu, kakhulu. Sidinga ama-algorithms azenzakalelayo wokuqoqwa kwawo, ukumakwa, nokusebenzisa. Sifuna ukwenza imoto ibone amaloli ebheke elangeni - kumele siqale siqoqe inani elanele lawo. Sifuna imoto ingahlanyi ibhayisikili eliboshelwe esiqwini - amasampula amaningi.

Ngaphezu kwalokho, isibonelo esisodwa asanele. Amakhulu? Izinkulungwane?

Inkinga yesibili

Inkinga yesibili - ukubona ngeso lokhozi inethiwekhi yethu yezinzwa. Lona umsebenzi ongabalulekile kakhulu. Kuze kube manje, bambalwa abantu abaqondayo ukuthi bangakubona kanjani lokhu ngeso lengqondo. Lezi zihloko ezakamuva kakhulu, lezi yizibonelo ezimbalwa nje, noma zikude:
Ukubonakala ukuthanda izinto zokwakha. Ibonisa kahle lokho i-neuron ejwayele ukulungisa kukho + lokho ekubona njengolwazi lokuqala.

Ukubonakala Ukunaka ku izinguqulo. Eqinisweni, ukukhanga kungasetshenziswa ngokunembile ukukhombisa ukuthi yini ebangele ukusabela kwenethiwekhi okunjalo. Ngizibonile izinto ezinjalo kokubili ukulungisa iphutha nezixazululo zomkhiqizo. Ziningi izindatshana ezikhuluma ngalesi sihloko. Kodwa uma idatha iyinkimbinkimbi, kuba nzima nakakhulu ukuqonda indlela yokufeza ukubonwa okuqinile.

Yebo, yebo, isethi enhle yakudala yokuthi “bheka okungaphakathi kwe-mesh phakathi izihlungi" Lezi zithombe zazidumile eminyakeni engu-3-4 edlule, kodwa wonke umuntu waqaphela ngokushesha ukuthi izithombe zazizinhle, kodwa zazingenayo incazelo enkulu.

Angizange ngikhulume inqwaba yamanye amagajethi, izindlela, ama-hacks, ucwaningo lwendlela yokubonisa okungaphakathi kwenethiwekhi. Ingabe la mathuluzi ayasebenza? Ingabe zikusiza ukuthi uqonde ngokushesha ukuthi yini inkinga futhi ulungise iphutha lenethiwekhi?.. Thola amaphesenti okugcina? Nokho, kucishe kufane:

Ungabuka noma yimuphi umncintiswano ku-Kaggle. Nencazelo yokuthi abantu benza kanjani izinqumo zokugcina. Sibeke amayunithi angama-100-500-800 wamamodeli futhi kwasebenza!

Ngenza ihaba, kunjalo. Kodwa lezi zindlela azinikezi izimpendulo ezisheshayo neziqondile.

Ukuba nesipiliyoni esanele, uhlole izinketho ezahlukene, unganikeza isinqumo sokuthi kungani isistimu yakho yenze isinqumo esinjalo. Kodwa kuzoba nzima ukulungisa ukuziphatha kwesistimu. Faka i-crutch, hambisa i-threshold, engeza idathasethi, thatha enye inethiwekhi ye-backend.

Inkinga yesithathu

Inkinga Yesithathu Eyisisekelo - amagridi afundisa izibalo, hhayi ukucabanga. Ngokwezibalo lokhu ubuso:

Ngokunengqondo, akufani kakhulu. Amanethiwekhi e-Neural awafundi lutho oluyinkimbinkimbi ngaphandle uma ephoqelelwe. Bahlala befundisa izimpawu ezilula ngangokunokwenzeka. Unamehlo, ikhala, ikhanda? Ngakho lobu ubuso! Noma unikeze isibonelo lapho amehlo engasho ubuso. Futhi futhi - izigidi zezibonelo.

KuneMakamelo Amaningi Ngezansi

Ngingasho ukuthi yilezi zinkinga ezintathu zomhlaba wonke okwamanje ezinciphisa ukuthuthukiswa kwamanethiwekhi e-neural nokufunda komshini. Futhi lapho lezi zinkinga zingazange zikhawule khona, isivele isetshenziswa ngenkuthalo.

Lesi ukuphela? Ingabe amanethiwekhi e-neural aphakanyisiwe?

Akwaziwa. Kodwa-ke, wonke umuntu unethemba lokuthi akunjalo.

Ziningi izindlela nezikhombisi-ndlela zokuxazulula izinkinga eziyisisekelo engizigqamisile ngenhla. Kodwa kuze kube manje, ayikho kulezi zindlela eziye zenza kwaba nokwenzeka ukwenza okuthile okusha ngokuyisisekelo, ukuxazulula okuthile okungakaxazululwa. Kuze kube manje, wonke amaphrojekthi abalulekile enziwa ngesisekelo sezindlela ezizinzile (Tesla), noma aqhubeke amaphrojekthi okuhlola ezikhungo noma izinkampani (Google Brain, OpenAI).

Uma sikhuluma nje, isiqondiso esiyinhloko siwukudala ukumelwa kwezinga eliphezulu kwedatha yokufaka. Ngomqondo othile, “inkumbulo”. Isibonelo esilula kakhulu senkumbulo “Ukushumeka” okuhlukahlukene - izethulo zezithombe. Nokho, isibonelo, zonke izinhlelo zokuqaphela ubuso. Inethiwekhi ifunda ukuthola ebusweni ukumelela okuthile okungaxhomekile ekujikelezeni, ekukhanyeni, noma ekulungisweni. Empeleni, inethiwekhi inciphisa imethrikhi "ubuso obuhlukile bukude" futhi "ubuso obufanayo buseduze."

Ukuze uthole ukuqeqeshwa okunjalo, kudingeka amashumi nezinkulungwane zezibonelo. Kodwa umphumela uqukethe ezinye zeziqalo ze-"One-shot Learning". Manje asisadingi amakhulu obuso ukuze sikhumbule umuntu. Ubuso obubodwa nje futhi yilokho kuphela esiyikho ake sithole!
Kunenkinga eyodwa... Igridi ingafunda kuphela izinto ezilula. Uma uzama ukuhlukanisa hhayi ubuso, kodwa, isibonelo, "abantu ngezingubo" (umsebenzi Ukuphinda kukhonjwe) - izinga liwela ngama-oda amaningi wobukhulu. Futhi inethiwekhi ngeke isakwazi ukufunda izinguquko ezisobala kuma-engeli.

Futhi ukufunda ezigidini zezibonelo nakho kuwuhlobo olujabulisayo.

Kunomsebenzi wokunciphisa kakhulu ukhetho. Isibonelo, umuntu angakhumbula ngokushesha omunye wemisebenzi yokuqala I-OneShot Learning kusuka ku-Google:

Kunemisebenzi eminingi enjalo, isibonelo 1 noma 2 noma 3.

Kukhona ukususa okukodwa - ngokuvamile ukuqeqeshwa kusebenza kahle kwezinye izibonelo ezilula, "MNIST". Futhi lapho udlulela emisebenzini eyinkimbinkimbi, udinga isizindalwazi esikhulu, imodeli yezinto, noma uhlobo oluthile lomlingo.
Ngokuvamile, ukusebenza ekuqeqeshweni kwe-One-Shot kuyisihloko esithakazelisa kakhulu. Uthola imibono eminingi. Kodwa ingxenye enkulu, izinkinga ezimbili engizibalile (ukuqeqeshwa kusengaphambili kudathasethi enkulu / ukungazinzi kudatha eyinkimbinkimbi) ziphazamisa kakhulu ukufunda.

Ngakolunye uhlangothi, ama-GAN—amanethiwekhi ezitha akhiqizayo—asondela esihlokweni sokushumeka. Cishe ufunde inqwaba yezindatshana ezikhuluma ngo-Habré ngalesi sihloko. (1, 2,3)
Isici se-GAN ukwakheka kwesikhala sombuso sangaphakathi (empeleni Ukushumeka okufanayo), okukuvumela ukuthi udwebe isithombe. Kungaba njalo ubuso, kungaba njalo isenzo.

Inkinga nge-GAN ukuthi uma into ekhiqiziwe iyinkimbinkimbi kakhulu, kuba nzima kakhulu ukuyichaza ngomqondo “wokucwasa igenerator”. Njengomphumela, okuwukuphela kwezicelo zangempela ze-GAN ezizwakalayo yi-DeepFake, ephinda futhi, isebenzise izethulo zobuso (okunesisekelo esikhulu sazo).

Ngibone okunye ukusetshenziswa okumbalwa okuwusizo. Ngokuvamile uhlobo oluthile lobuqili oluhlanganisa ukuqedela imidwebo yezithombe.

Futhi futhi. Akekho onombono wokuthi lokhu kuzosivumela kanjani ukuthi sidlulele ekusasa eliqhakazile. Ukumela i-logic/space kunethiwekhi ye-neural kuhle. Kodwa sidinga inani elikhulu lezibonelo, asiqondi ukuthi i-neuron imelela kanjani lokhu ngokwayo, asiqondi ukuthi yenziwa kanjani i-neuron ikhumbule umqondo othile oyinkimbinkimbi ngempela.

Ukuqiniswa kokufunda - lena indlela esuka endaweni ehluke ngokuphelele. Impela uyakhumbula ukuthi i-Google yahlula kanjani wonke umuntu ku-Go. Ukunqoba kwakamuva ku-Starcraft ne-Dota. Kodwa lapha yonke into ikude kakhulu nenhle futhi iyathembisa. Ukhuluma kahle nge-RL nezinkimbinkimbi zayo Lesi sihloko.

Ukufingqa kafushane lokho umbhali akulobile:

• Amamodeli angaphandle kwebhokisi awalingani / asebenza kabi ezimweni eziningi
• Izinkinga ezingokoqobo kulula ukuzixazulula ngezinye izindlela. I-Boston Dynamics ayisebenzisi i-RL ngenxa yobunkimbinkimbi bayo/ukungakwazi ukubikezela/ukuba yinkimbinkimbi kwekhompyutha
• Ukuze i-RL isebenze, udinga umsebenzi oyinkimbinkimbi. Kuvame ukuba nzima ukwakha/ukubhala
• Kunzima ukuqeqesha amamodeli. Kufanele uchithe isikhathi esiningi ukupompa futhi uphume endaweni enhle kakhulu
• Ngenxa yalokho, kunzima ukuphinda imodeli, imodeli ayizinzile ngezinguquko ezincane
• Ivamise ukugcwala ngokweqile uhlobo oluthile lwamaphethini angakwesokunxele, kuze kufike kwijeneretha yenombolo engahleliwe

Iphuzu elibalulekile ukuthi i-RL ayikasebenzi ekukhiqizeni. I-Google inokuhlola okuthile ( 1, 2 ). Kodwa angikaze ngibone uhlelo olulodwa lomkhiqizo.

Memory. Ububi bakho konke okuchazwe ngenhla ukuntuleka kwesakhiwo. Enye yezindlela zokuzama ukuqoqa konke lokhu ukunikeza inethiwekhi ye-neural ukufinyelela kwimemori ehlukene. Ukuze akwazi ukuqopha futhi abhale kabusha imiphumela yezinyathelo zakhe lapho. Bese inethiwekhi ye-neural inganqunywa isimo sememori samanje. Lokhu kufana kakhulu namaphrosesa namakhompyutha wakudala.

Edume kakhulu futhi ethandwa kakhulu indatshana — kusuka ku-DeepMind:

Kubonakala sengathi lesi isihluthulelo sokuqonda ubuhlakani? Kodwa mhlawumbe akunjalo. Uhlelo lusadinga inani elikhulu ledatha yokuqeqeshwa. Futhi isebenza kakhulu ngedatha yethebula ehleliwe. Ngaphezu kwalokho, lapho Facebook yanquma inkinga efanayo, base bethatha indlela "inkumbulo eyisikulufa, yenza nje i-neuron ibe yinkimbinkimbi, futhi ube nezibonelo eziningi - futhi izofunda ngokwayo."

Ukuhlukanisa. Enye indlela yokudala inkumbulo enengqondo ukuthatha ukushumeka okufanayo, kodwa ngesikhathi sokuqeqeshwa, ngenisa indlela eyengeziwe engakuvumela ukuthi ugqamise “izincazelo” kuzo. Isibonelo, sifuna ukuqeqesha inethiwekhi ye-neural ukuhlukanisa phakathi kokuziphatha komuntu esitolo. Uma silandela indlela evamile, kuzodingeka senze amanethiwekhi ayishumi nambili. Omunye ufuna umuntu, owesibili ubheka ukuthi wenzani, owesithathu iminyaka yakhe, owesine ubulili bakhe. I-logic ehlukene ibheka ingxenye yesitolo lapho yenza khona/iqeqeshelwa ukwenza lokhu. Owesithathu unquma umkhondo wayo, njll.

Noma, uma bekunenani elingapheli ledatha, bekuzokwaziwa ukuqeqesha inethiwekhi eyodwa kuyo yonke imiphumela engaba khona (ngokusobala, uhlu olunjalo lwedatha alukwazi ukuqoqwa).

Indlela yokuhlukanisa iyasitshela - masiqeqeshe inethiwekhi ukuze yona ikwazi ukuhlukanisa phakathi kwemiqondo. Ukuze yakhe ukushumeka ngokusekelwe kuvidiyo, lapho indawo eyodwa izonquma isenzo, umuntu uzonquma indawo phansi ngesikhathi, omunye uzonquma ubude bomuntu, futhi omunye uzonquma ubulili bomuntu. Ngesikhathi esifanayo, lapho ngiqeqeshwa, ngingathanda cishe ukungabambisi inethiwekhi ngemiqondo eyinhloko enjalo, kodwa kunalokho ukuze igqamise futhi iqoqe izindawo. Kukhona izindatshana ezinjalo ezimbalwa (ezinye zazo 1, 2, 3) futhi ngokujwayelekile ziyithiyori impela.

Kodwa lesi siqondiso, okungenani ngokwethiyori, kufanele sihlanganise izinkinga ezibalwe ekuqaleni.

Ukubola kwesithombe ngokuya ngamapharamitha “umbala wodonga/umbala wesitezi/umumo wento/umbala wento/njll.”

Ukubola kobuso ngokuya ngemingcele "usayizi, amashiya, umumo, umbala wesikhumba, njll."

Ezenye

Kunezinye izindawo eziningi, hhayi zomhlaba jikelele, ezikuvumela ukuthi ngandlela thize unciphise isizindalwazi, usebenze ngedatha ehlukahlukene, njll.

Attention. Cishe akunangqondo ukwehlukanisa lokhu njengendlela ehlukile. Indlela nje ethuthukisa abanye. Izihloko eziningi zinikezelwe kuye (1,2,3). Iphuzu lokunakwa ukuthuthukisa impendulo yenethiwekhi ikakhulukazi ezintweni ezibalulekile ngesikhathi sokuqeqeshwa. Ngokuvamile ngohlobo oluthile lokuqokwa kwangaphandle, noma inethiwekhi encane yangaphandle.

Ukulingiswa kwe-3D. Uma wenza injini ye-3D enhle, ungakwazi ukumboza u-90% wedatha yokuqeqeshwa ngayo (ngize ngabona isibonelo lapho cishe u-99% wedatha embozwe injini enhle). Kunemibono eminingi kanye nama-hacks wokuthi ungayenza kanjani inethiwekhi eqeqeshelwe ukusebenza kwenjini ye-3D usebenzisa idatha yangempela (Ukushuna okuhle, ukudluliswa kwesitayela, njll.). Kodwa ngokuvamile ukwenza injini enhle kuyimiyalo eminingana yobukhulu enzima kakhulu kunokuqoqa idatha. Izibonelo ngenkathi kwenziwa izinjini:
Ukuqeqeshwa kwerobhothi (google, i-braingarden)
Izifundo zokuqeqesha ukuqashelwa izimpahla esitolo (kodwa kumaphrojekthi amabili esiwenzile, singenza kalula ngaphandle kwawo).
Ukuqeqeshwa eTesla (futhi, ividiyo engenhla).

okutholakele

Isihloko sonke, ngomqondo othile, siyiziphetho. Mhlawumbe umlayezo oyinhloko engangifuna ukuwenza ukuthi "imali yamahhala isiphelile, ama-neurons awasanikezi izixazululo ezilula." Manje kudingeka sisebenze kanzima ukuze senze izinqumo eziyinkimbinkimbi. Noma sebenza kanzima ukwenza ucwaningo lwesayensi oluyinkimbinkimbi.

Ngokuvamile, isihloko kuyaphikiswana ngaso. Mhlawumbe abafundi banezibonelo ezithakazelisayo?

Source: www.habr.com