Okuqukethwe[Fihla][Bonisa]
- 1. Kuyini ngempela Ukufunda Okujulile?
- 2. Yini ehlukanisa i-Deep Learning kusukela ekuFundeni ngomshini?
- 3. Kuyini ukuqonda kwakho kwamanje kwamanethiwekhi e-neural?
- 4. Iyini ngempela i-perceptron?
- 5. Iyini ngempela inethiwekhi ye-neural ejulile?
- 6. Iyini Ngempela I-Multilayer Perceptron (MLP)?
- 7. Iyiphi injongo edlalwa imisebenzi yokuvula kunethiwekhi ye-neural?
- 8. Kuyini Ngempela Ukwehla Kwegradient?
- 9. Iyini Kahle Izindleko Umsebenzi?
- 10. Amanethiwekhi ajulile angawadlula kanjani angajulile?
- 11. Chaza ukusakazwa phambili.
- 12. Iyini i-backpropagation?
- 13. Esimeni sokufunda okujulile, ukuqondisisa kanjani ukunqunywa kwe-gradient?
- 14. Iyini Imisebenzi yeSoftmax ne-ReLU?
- 15. Ingabe imodeli yenethiwekhi ye-neural ingaqeqeshwa nazo zonke izisindo ezibekwe ku-0?
- 16. Yini ehlukanisa inkathi neqoqo nokuphindaphinda?
- 17. Iyini I-Batch Normalization kanye Nokuyeka?
- 18. Yini Ehlukanisa Ukwehla Kwe-Stochastic Gradient kusukela ku-Batch Gradient Descent?
- 19. Kungani kubalulekile ukufaka okungewona olayini kumanethiwekhi emizwa?
- 20. Iyini i-tensor ekufundeni okujulile?
- 21. Ungawukhetha kanjani umsebenzi wokuvula imodeli yokufunda ejulile?
- 22. Usho ukuthini nge-CNN?
- 23. Yiziphi izingqimba eziningi ze-CNN?
- 24. Iyini imiphumela yokusetshenziswa ngokweqile nokungafaneleki, futhi ungayigwema kanjani?
- 25. Ekufundeni okujulile, iyini iRNN?
- 26. Chaza i-Adam Optimizer
- 27. Ama-autoencoder ajulile: ayini?
- 28. Isho ukuthini i-Tensor ku-Tensorflow?
- 29. Incazelo yegrafu yokubala
- 30. Amanethiwekhi ezitha ezikhiqizayo (ama-GAN): ayini?
- 31. Uzolikhetha kanjani inani lama-neurons nezingqimba ezifihliwe ozozifaka kunethiwekhi ye-neural njengoba uklama i-architecture?
- 32. Yiziphi izinhlobo zamanethiwekhi e-neural asetshenziswa ngokufunda okujulile kokuqinisa?
- Isiphetho
Ukufunda ngokujulile akuwona umqondo omusha. Amanethiwekhi e-neural okwenziwa asebenza njengesisekelo sodwa sesethi engaphansi yokufunda yomshini eyaziwa ngokuthi ukufunda okujulile.
Ukufunda okujulile kulingisa ubuchopho bomuntu, njengoba kunjalo ngamanethiwekhi emizwa, njengoba adalwe ukuze alingise ubuchopho bomuntu.
Sekunesikhathi kukhona lokhu. Kulezi zinsuku, wonke umuntu ukhuluma ngakho njengoba asinawo amandla okucubungula amaningi noma idatha njengoba sinawo manje.
Eminyakeni engu-20 edlule, ukufunda okujulile nokufunda ngomshini kuye kwavela ngenxa yokukhuphuka okumangalisayo kwamandla okucubungula.
Ukuze sikusize ulungiselele noma yimiphi imibuzo ongabhekana nayo lapho ufuna umsebenzi wamaphupho akho, lokhu okuthunyelwe kuzokuqondisa emibuzweni eminingi ejulile yenhlolokhono, kusukela kokulula kuye kokuyinkimbinkimbi.
1. Kuyini ngempela Ukufunda Okujulile?
Uma uhambela a ukufunda okujulile interview, ngokungangabazeki uyaqonda ukuthi kuyini ukufunda okujulile. Obuza imibuzo, nokho, ulindele ukuthi unikeze impendulo enemininingwane kanye nomfanekiso ekuphenduleni lo mbuzo.
Ukuze uqeqeshe amanethiwekhi we-neural ekufundeni okujulile, amanani abalulekile edatha ehleliwe noma engahlelekile kufanele isetshenziswe. Ukuthola amaphethini nezici ezifihliwe, yenza izinqubo eziyinkimbinkimbi (isibonelo, ukuhlukanisa isithombe sekati kuleso senja).
2. Yini ehlukanisa i-Deep Learning kusukela ekuFundeni ngomshini?
Njengegatsha lobuhlakani bokwenziwa elaziwa ngokuthi ukufunda komshini, siqeqesha amakhompyutha sisebenzisa idatha namasu ezibalo kanye ne-algorithmic ukuze abe ngcono ngokuhamba kwesikhathi.
Njengengxenye ye ukufunda imishini, ukufunda okujulile kulingisa i-neural network architecture ebonwa ebuchosheni bomuntu.
3. Kuyini ukuqonda kwakho kwamanje kwamanethiwekhi e-neural?
Amasistimu okwenziwa aziwa ngokuthi amanethiwekhi e-neural afana namanethiwekhi e-organic neural atholakala emzimbeni womuntu eduze kakhulu.
Ukusebenzisa inqubo efana nendlela i- ubuchopho bomuntu function, inethiwekhi ye-neural iqoqo lama-algorithms ahlose ukukhomba ukuhlobana okuyisisekelo ocezwini lwedatha.
Lawa masistimu athola ulwazi oluqondene nomsebenzi othile ngokuziveza kuhlu lwamadathasethi nezibonelo, kunokulandela noma yimiphi imithetho eqondene nomsebenzi othile.
Umqondo uwukuthi esikhundleni sokuba nokuqonda okuhlelwe ngaphambilini kwalawa madathasethi, isistimu ifunda izici ezihlukanisayo kudatha ephakelwayo.
Izendlalelo ezintathu zenethiwekhi ezivame ukusetshenziswa kakhulu ku-Neural Networks zimi kanje:
- Isendlalelo okokufaka
- Isendlalelo esifihliwe
- Isendlalelo sokuphumayo
4. Iyini ngempela i-perceptron?
I-biological neuron etholakala ebuchosheni bomuntu iqhathaniswa ne-perceptron. Okokufaka okuningi kwamukelwa yi-perceptron, ebese yenza izinguquko nemisebenzi eminingi futhi ikhiqize okukhiphayo.
Imodeli yomugqa ebizwa ngokuthi i-perceptron isetshenziswa ekuhlukaniseni kanambambili. Ilingisa i-neuron enokufakwa okuhlukahlukene, ngayinye enesisindo esihlukile.
I-neuron ibala umsebenzi isebenzisa lokhu okokufaka okunesisindo futhi ikhiphe imiphumela.
5. Iyini ngempela inethiwekhi ye-neural ejulile?
Inethiwekhi ye-neural ejulile iyinethiwekhi ye-neural yokwenziwa (ANN) enezendlalelo ezimbalwa phakathi kwezendlalelo zokufaka nokuphumayo (DNN).
Amanethiwekhi e-neural ajulile angamanethiwekhi e-neural ezakhiwo ezijulile. Igama elithi “deep” libhekisela emisebenzini enamazinga amaningi namayunithi kusendlalelo esisodwa. Amamodeli anembe kakhudlwana angadalwa ngokungeza izendlalelo ezengeziwe nezikhudlwana ukuze kuthwebule amaleveli amakhulu amaphethini.
6. Iyini Ngempela I-Multilayer Perceptron (MLP)?
Izendlalelo okokufaka, ezifihliwe, neziphumayo zikhona kuma-MLP, njengakunethiwekhi ye-neural. Yakhiwe ngokufana ne-perceptron yongqimba olulodwa enezingqimba ezifihliwe eyodwa noma ngaphezulu.
Okukhiphayo kanambambili kwe-perceptron yesendlalelo esisodwa kungahlukanisa kuphela amakilasi ahlukanisekayo ngomugqa (0,1), kanti i-MLP ingahlukanisa amakilasi angaqondile.
7. Iyiphi injongo edlalwa imisebenzi yokuvula kunethiwekhi ye-neural?
Umsebenzi wokwenza kusebenze unquma ukuthi i-neuron kufanele isebenze ezingeni elibaluleke kakhulu. Noma yimuphi umsebenzi wokwenza kusebenze ungamukela isamba esinesisindo sokufakwayo kanye nokuchema njengokufakiwe. Imisebenzi yokuvula ihlanganisa umsebenzi wesinyathelo, i-Sigmoid, i-ReLU, i-Tanh, ne-Softmax.
8. Kuyini Ngempela Ukwehla Kwegradient?
Indlela engcono kakhulu yokunciphisa umsebenzi wezindleko noma iphutha ukwehla kwe-gradient. Ukuthola i-minima yendawo yomhlaba wonke yomsebenzi kuwumgomo. Lokhu kucacisa indlela imodeli okufanele iyilandele ukuze kuncishiswe iphutha.
9. Iyini Kahle Izindleko Umsebenzi?
Umsebenzi wezindleko uyimethrikhi yokuhlola ukuthi imodeli yakho isebenza kahle kangakanani; ngezinye izikhathi kwaziwa ngokuthi “ukulahlekelwa” noma “iphutha.” Ngesikhathi sokusatshalaliswa kabusha, isetshenziselwa ukubala iphutha lesendlalelo sokuphumayo.
Sisebenzisa lokho kungalungile ukuze siqhubekisele phambili izinqubo zokuqeqeshwa zenethiwekhi ye-neural ngokuyibuyisela emuva ngenethiwekhi ye-neural.
10. Amanethiwekhi ajulile angawadlula kanjani angajulile?
Izendlalelo ezifihliwe zengezwa kumanethiwekhi e-neural ngaphezu kwezingqimba zokufakwayo neziphumayo. Phakathi kwezingqimba zokufakwayo neziphumayo, amanethiwekhi we-neural angashoni asebenzisa isendlalelo esisodwa esifihliwe, kanti amanethiwekhi ajulile e-neural asebenzisa amaleveli amaningi.
Inethiwekhi engajulile idinga amapharamitha ambalwa ukuze ikwazi ukungena kunoma yimuphi umsebenzi. Amanethiwekhi ajulile angafanela imisebenzi kangcono ngisho nenani elincane lamapharamitha njengoba ahlanganisa izendlalelo ezimbalwa.
Amanethiwekhi ajulile manje ayathandwa ngenxa yokuguquguquka kwawo ekusebenzeni nanoma yiluphi uhlobo lokumodela idatha, noma ngabe okokukhuluma noma ukubonwa kwesithombe.
11. Chaza ukusakazwa phambili.
Okokufaka kudluliselwa kanye nezisindo kungqimba engcwatshiwe ngenqubo eyaziwa ngokuthi ukusabalalisa ukudlulisa.
Okukhiphayo komsebenzi wokwenza kusebenze kubalwa kusendlalelo ngasinye esingcwatshiwe ngaphambi kokuthi ukucubungula kudlulele kusendlalelo esilandelayo.
Inqubo iqala kusendlalelo okokufaka futhi iqhubekele kwesendlalelo sokugcina, ngaleyo ndlela igama liya phambili.
12. Iyini i-backpropagation?
Uma izisindo nokuchema kulungiswa kunethiwekhi ye-neural, i-backpropagation isetshenziselwa ukunciphisa umsebenzi wezindleko ngokuqala ngokubheka ukuthi inani lishintsha kanjani.
Ukuqonda i-gradient kusendlalelo ngasinye esifihliwe kwenza ukubala lolu shintsho lube lula.
Inqubo, eyaziwa ngokuthi i-backpropagation, iqala kusendlalelo sokuphumayo bese ihlehlela emuva izendlalelo zokufakwayo.
13. Esimeni sokufunda okujulile, ukuqondisisa kanjani ukunqunywa kwe-gradient?
I-Gradient Clipping iyindlela yokuxazulula inkinga yama-gradients aqhumayo avela phakathi no-backpropagation (isimo lapho ama-gradient angalungile aqoqwa khona ngokuhamba kwesikhathi, okuholela ekulungiseni okubalulekile ezisindweni zemodeli yenethiwekhi ye-neural phakathi nokuqeqeshwa).
Ama-gradient aqhumayo yinkinga ephakamayo lapho ama-gradient eba mkhulu kakhulu ngesikhathi sokuqeqeshwa, okwenza imodeli ingazinzi. Uma i-gradient yeqe ububanzi obulindelekile, amanani egrediyenti aphushwa i-elementi-by-elementi enani elichazwe ngaphambilini eliphansi noma eliphakeme.
Ukusika i-Gradient kuthuthukisa ukuzinza kwezinombolo kwenethiwekhi ye-neural phakathi nokuqeqeshwa, kodwa kunomthelela omncane ekusebenzeni kwemodeli.
14. Iyini Imisebenzi yeSoftmax ne-ReLU?
Umsebenzi wokwenza kusebenze obizwa ngokuthi i-Softmax ukhiqiza okukhiphayo kububanzi obuphakathi kuka-0 no-1. Okukhiphayo ngakunye kuhlukaniswa ukuze isamba sayo yonke imiphumela ibe yinye. Okwezendlalelo eziphumayo, iSoftmax ivamise ukusetshenziswa.
Iyunithi Yolayini Elungisiwe, kwesinye isikhathi eyaziwa ngokuthi i-ReLU, iwumsebenzi wokwenza kusebenze osetshenziswa kakhulu. Uma u-X ethi phozithivu, ukhipha u-X, uma kungenjalo ukhipha amaziro. I-ReLU isetshenziswa njalo ezingxenyeni ezingcwatshiwe.
15. Ingabe imodeli yenethiwekhi ye-neural ingaqeqeshwa nazo zonke izisindo ezibekwe ku-0?
Inethiwekhi ye-neural ayisoze yafunda ukuqedela umsebenzi othile, ngakho-ke akwenzeki ukuqeqesha imodeli ngokuqalisa zonke izisindo zibe ngu-0.
Okuphuma kokunye kuzohlala kufana kuso sonke isisindo ku-W [1] uma zonke izisindo ziqaliswa zibe ziro, okuzoholela ekutheni ama-neurons afunde izici ezifanayo ngokuphindaphindiwe.
Hhayi nje ukuqalisa izisindo ziye ku-0, kodwa kunoma yiluphi uhlobo lokungaguquki kungenzeka kuphumele kumphumela ophansi.
16. Yini ehlukanisa inkathi neqoqo nokuphindaphinda?
Izinhlobo ezihlukene zokucubungula amasethi edatha kanye namasu okwehla kwe-gradient afaka inqwaba, i-iteration, ne-epoch. I-Epoch ihlanganisa kanye-ngenethiwekhi ye-neural enedathasethi egcwele, kokubili phambili nangemuva.
Ukuze kuhlinzekwe imiphumela ethembekile, idathasethi ivamise ukudlula izikhathi ezimbalwa njengoba inkulu kakhulu ukuthi ingadlula ngokuzama okukodwa.
Lo mkhuba wokusebenzisa ngokuphindaphindiwe inani elincane ledatha ngenethiwekhi ye-neural ubizwa ngokuthi i-iteration. Ukuqinisekisa ukuthi isethi yedatha inqamula ngempumelelo amanethiwekhi e-neural, ingahlukaniswa ngenani lamaqoqo noma amasethi angaphansi, aziwa ngokuthi i-batching.
Ngokuya ngosayizi wokuqoqwa kwedatha, zontathu izindlela—inkathi, ukuphindaphinda, nosayizi weqoqo—izindlela zokusebenzisa i-algorithm yokwehla kwe-gradient.
17. Iyini I-Batch Normalization kanye Nokuyeka?
Ukuyeka kuvimbela ukufakwa ngokweqile kwedatha ngokukhipha ngokungahleliwe kokubili amayunithi enethiwekhi abonakalayo nafihlekile (ngokuvamile kwehla amaphesenti angu-20 amanodi). Iphinda kabili inombolo yokuphindaphinda okudingekayo ukuze inethiwekhi ihlangane.
Ngokwenza okokufaka kujwayelekile kusendlalelo ngasinye ukuze kwenziwe kusebenze isilinganiso esimaphakathi sikaziro kanye nokuchezuka okujwayelekile kokukodwa, ukwenziwa kwenqwaba kujwayelekile isu lokuthuthukisa ukusebenza nokuzinza kwamanethiwekhi emizwa.
18. Yini Ehlukanisa Ukwehla Kwe-Stochastic Gradient kusukela ku-Batch Gradient Descent?
Ukwehla kwe-Batch Gradient:
- Idathasethi ephelele isetshenziselwa ukwakha i-gradient ye-batch gradient.
- Inani elikhulu ledatha kanye nezisindo ezibuyekezwa kancane kancane kwenza ukuhlangana kube nzima.
Ukwehla kwe-Stochastic Gradient:
- Igradient ye-stochastic isebenzisa isampula eyodwa ukuze ibale ukuthambekela.
- Ngenxa yezinguquko zesisindo ezivamile, ihlangana ngokushesha kakhulu kune-batch gradient.
19. Kungani kubalulekile ukufaka okungewona olayini kumanethiwekhi emizwa?
Kungakhathaliseki ukuthi zingaki izendlalelo ezikhona, inethiwekhi ye-neural izoziphatha njenge-perceptron lapho kungekho okungekona komugqa, okwenza okukhiphayo kuncike ngokomugqa ngokokufaka.
Ukukubeka ngenye indlela, inethiwekhi ye-neural enezingqimba ezingu-n kanye namayunithi afihliwe m kanye nemisebenzi yokuvula imigqa ilingana nenethiwekhi ye-neural eqondile ngaphandle kwezendlalelo ezifihliwe kanye nekhono lokubona imingcele yokuhlukaniswa komugqa kuphela.
Ngaphandle kokungewona olayini, inethiwekhi ye-neural ayikwazi ukuxazulula izinkinga eziyinkimbinkimbi futhi ihlukanise ngokunembile okokufaka.
20. Iyini i-tensor ekufundeni okujulile?
I-multidimensional array eyaziwa ngokuthi i-tensor isebenza njengokujwayelekile kwamatrices nama-vector. Kuwuhlaka lwedatha olubalulekile lokufunda ngokujulile. Ukuhlelwa kwe-N-dimensional kwezinhlobo zedatha eyisisekelo kusetshenziselwa ukumela ama-tensor.
Yonke ingxenye ye-tensor inohlobo lwedatha olufanayo, futhi lolu hlobo lwedatha luhlale lwaziwa. Kungenzeka ukuthi ucezu lwesimo kuphela—okungukuthi, ukuthi zingaki izilinganiso ezikhona nokuthi ngasinye sikhulu kangakanani—saziwayo.
Ezimeni lapho okokufaka nakho kwaziwa ngokuphelele, iningi lemisebenzi likhiqiza ama-tensor awaziwa ngokugcwele; kwezinye izimo, uhlobo lwe-tensor lungasungulwa kuphela ngesikhathi sokukhishwa kwegrafu.
21. Ungawukhetha kanjani umsebenzi wokuvula imodeli yokufunda ejulile?
- Kunengqondo ukusebenzisa umsebenzi womugqa wokwenza kusebenze uma umphumela okufanele ulindelwe ungokoqobo.
- Umsebenzi we-Sigmoid kufanele usetshenziswe uma okukhiphayo okufanele kubikezelwe kungamathuba ekilasi kanambambili.
- Umsebenzi we-Tanh ungasetshenziswa uma okukhiphayo okubonisiwe kuqukethe izigaba ezimbili.
- Ngenxa yobulula bawo bokubala, umsebenzi we-ReLU usebenza ezimeni eziningi ezahlukene.
22. Usho ukuthini nge-CNN?
Amanethiwekhi e-neural ajulile asebenza ngokukhethekile ekuhloleni izithombe ezibukwayo afaka amanethiwekhi e-convolutional neural (CNN, noma ConvNet). Lapha, kunokuba kumanethiwekhi e-neural lapho ivekhtha imele okokufaka, okokufaka kuyisithombe esineziteshi eziningi.
I-Multilayer perceptrons isetshenziswa ngendlela ekhethekile ngama-CNN adinga ukucubungula ngaphambilini okuncane kakhulu.
23. Yiziphi izingqimba eziningi ze-CNN?
Isendlalelo se-Convolutional: Isendlalelo esiyinhloko ungqimba lwe-convolutional, olunezinhlobonhlobo zezihlungi ezifundekayo kanye nenkambu eyamukelayo. Lesi sendlalelo sokuqala sithatha idatha yokufaka futhi sikhiphe izici zayo.
Isendlalelo se-ReLU: Ngokwenza amanethiwekhi angabi-layini, lesi sendlalelo sishintsha amaphikseli anegethivu abe yiziro.
Isendlalelo sokuhlanganisa: Ngokunciphisa ukucubungula nezilungiselelo zenethiwekhi, isendlalelo sokuhlanganisa kancane kancane sinciphisa usayizi wendawo wokumelela. I-Max pooling iyindlela esetshenziswa kakhulu yokuhlanganisa.
24. Iyini imiphumela yokusetshenziswa ngokweqile nokungafaneleki, futhi ungayigwema kanjani?
Lokhu kwaziwa njengokugcwalisa ngokweqile lapho imodeli ifunda ubunkimbinkimbi nomsindo kudatha yokuqeqeshwa kuze kube yilapho kuthinta kabi ukusetshenziswa kwemodeli kwedatha entsha.
Kungenzeka kakhulu ukuthi kwenzeke ngamamodeli angaqondile avumelana nezimo ngenkathi ufunda umsebenzi wegoli. Imodeli ingaqeqeshelwa ukuthola izimoto namaloli, kodwa ingakwazi kuphela ukuhlonza izimoto ezinefomu lebhokisi elithile.
Uma kubhekwa ukuthi iqeqeshwe ohlotsheni olulodwa lweloli, ingase ingakwazi ukubona iloli eline-flatbed. Kudatha yokuqeqeshwa, imodeli isebenza kahle, kodwa hhayi emhlabeni wangempela.
Imodeli engafakwanga ngokwanele isho leyo engaqeqeshwanga ngokwanele kudatha noma ekwazi ukuhlanganisa ulwazi olusha. Lokhu kuvame ukwenzeka lapho imodeli iqeqeshwa ngedatha enganele noma engalungile.
Ukunemba nokusebenza kokubili kuphazanyiswa ukufakwa ngaphansi kokufaneleka.
Ukwenza isampula kabusha idatha ukuze ulinganisele ukunemba kwemodeli (ukuqinisekiswa kwe-K-fold cross-validation) nokusebenzisa idathasethi yokuqinisekisa ukuze kuhlolwe imodeli kuyizindlela ezimbili zokugwema ukucwiliswa ngokweqile nokungafaneleki.
25. Ekufundeni okujulile, iyini iRNN?
Amanethiwekhi e-neural avamile (ama-RNN), izinhlobonhlobo ezivamile zamanethiwekhi emizwa yokwenziwa, ahamba ngesifinyezo esithi RNN. Baqashelwe ukucubungula ama-genomes, ukubhala ngesandla, umbhalo, nokulandelana kwedatha, phakathi kwezinye izinto. Ukuze uthole ukuqeqeshwa okudingekayo, ama-RNN asebenzisa i-backpropagation.
26. Chaza i-Adam Optimizer
I-Adam optimizer, eyaziwa nangokuthi i-adaptive momentum, iyindlela yokwenza kahle ethuthukiswe ukuze isingathe izimo ezinomsindo ngama-gradients angenalutho.
Ngokungeziwe ekuhlinzekeni ngezibuyekezo zepharamitha ngayinye ukuze kuhlangane ngokushesha, isilungiseleli sika-Adam sithuthukisa ukuhlangana ngomfutho, siqinisekisa ukuthi imodeli ayibambeki endaweni yesihlalo.
27. Ama-autoencoder ajulile: ayini?
I-autoencoder ejulile igama eliyiqoqo lamanethiwekhi amabili enkolelo ejulile elinganayo ngokuvamile ahlanganisa izendlalelo ezine noma ezinhlanu ezingajulile zengxenye yombhalo wekhodi wenethiwekhi kanye nenye isethi yezendlalelo ezine noma ezinhlanu zengxenye yokukhipha amakhodi.
Lezi zingqimba zakha isisekelo samanethiwekhi enkolelo ejulile futhi ziphonswa imishini ye-Boltzmann. Ngemva kwe-RBM ngayinye, isifaki khodi esijulile sisebenzisa izinguquko kanambambili kudathasethi ye-MNIST.
Angasetshenziswa nakwamanye amadathasethi lapho izinguquko ezilungisiwe ze-Gaussian zingakhethwa kune-RBM.
28. Isho ukuthini i-Tensor ku-Tensorflow?
Lona omunye umbuzo ojulile wenhlolokhono yokufunda ebuzwa njalo. I-tensor umqondo wezibalo obonakala njengama-arrays anobukhulu obuphezulu.
Ama-tensor yilawa mafayela afanayo edatha anikezwa njengokufakwayo kunethiwekhi ye-neural futhi anobukhulu obuhlukahlukene namazinga.
29. Incazelo yegrafu yokubala
Isisekelo se-TensorFlow ukwakhiwa kwegrafu yokubala. I-node ngayinye isebenza kunethiwekhi yamanodi, lapho amanodi amelela ukusebenza kwezibalo kanye nemiphetho yama-tensor.
Ngezinye izikhathi ibizwa ngokuthi "Igrafu ye-DataFlow" njengoba idatha igeleza ngesimo segrafu.
30. Amanethiwekhi ezitha ezikhiqizayo (ama-GAN): ayini?
Ekufundeni Okujulile, ukumodela okukhiqizayo kwenziwa kusetshenziswa amanethiwekhi akhiqizayo aphikisanayo. Kungumsebenzi ongagadiwe lapho umphumela ukhiqizwa ngokuhlonza amaphethini kudatha yokufaka.
Umbandlululi usetshenziselwa ukuhlukanisa izimo ezikhiqizwa ijeneretha, kuyilapho ijeneretha isetshenziselwa ukukhiqiza izibonelo ezintsha.
31. Uzolikhetha kanjani inani lama-neurons nezingqimba ezifihliwe ozozifaka kunethiwekhi ye-neural njengoba uklama i-architecture?
Njengoba kunikezwe inselele yebhizinisi, inani elinembile lama-neurons nezingqimba ezifihliwe ezidingekayo ukuze kwakhiwe i-neural network architecture ayikwazi ukunqunywa nganoma yimiphi imithetho eqinile nesheshayo.
Kunethiwekhi ye-neural, usayizi wesendlalelo esifihliwe kufanele siwe ndawana thize phakathi nosayizi wezendlalelo zokufakwayo neziphumayo.
Isiqalo sokuqala ekudaleni ukwakheka kwenethiwekhi ye-neural singatholakala ngezindlela ezimbalwa eziqondile, noma:
Ukuqala ngokuhlolwa okuhlelekile okuyisisekelo ukuze ubone ukuthi yini engenza kangcono kunoma iyiphi idathasethi ethile ngokusekelwe kokuhlangenwe nakho kwangaphambilini namanethiwekhi e-neural kuzilungiselelo ezifanayo zomhlaba wangempela kuyindlela engcono kakhulu yokubhekana nayo yonke inselele eyingqayizivele yokubikezela umhlaba wangempela.
Ukulungiselelwa kwenethiwekhi kungakhethwa ngokusekelwe olwazini lomuntu lwesizinda senkinga kanye nolwazi lwangaphambili lwenethiwekhi ye-neural. Lapho kuhlolwa ukusethwa kwenethiwekhi ye-neural, inani lezendlalelo nama-neurons asetshenziswa ezinkingeni ezihlobene liyindawo enhle yokuqala.
Ubunkimbinkimbi benethiwekhi ye-neural kufanele bukhuliswe kancane kancane ngokusekelwe ekuphumeni okucatshangelwayo nokunemba, kuqalwe ngomklamo olula wenethiwekhi ye-neural.
32. Yiziphi izinhlobo zamanethiwekhi e-neural asetshenziswa ngokufunda okujulile kokuqinisa?
- Ku-paradigm yokufunda komshini ebizwa ngokuthi ukufunda kokuqinisa, imodeli isebenza ukuze ikhulise umqondo womvuzo oqongelelwe, njengoba kwenza izinto eziphilayo.
- Imidlalo nezimoto ezizishayelayo zombili zichazwa njengezinkinga ezibandakanya ukuqinisa ukufunda.
- Isikrini sisetshenziswa njengokufakwayo uma inkinga ezomelwa kuwumdlalo. Ukuze kukhiqizwe okuphumayo kwezigaba ezilandelayo, i-algorithm ithatha amaphikseli njengokufakwayo futhi iwacubungule kusetshenziswa izendlalelo eziningi zamanethiwekhi e-convolutional neural.
- Imiphumela yezenzo zemodeli, evumayo noma embi, isebenza njengokuqiniswa.
Isiphetho
I-Deep Learning ikhule ngokuduma eminyakeni edlule, futhi isetshenziswa cishe kuyo yonke indawo yezimboni.
Izinkampani ziya ngokuya zifuna ochwepheshe abanekhono abangaklama amamodeli aphindaphinda ukuziphatha kwabantu kusetshenziswa ukufunda okujulile nezindlela zokufunda ngomshini.
Abafundi abakhulisa amakhono abo futhi balondoloze ulwazi lwabo lwalobu buchwepheshe obusezingeni eliphezulu bangathola inhlobonhlobo yamathuba omsebenzi aneholo elikhangayo.
Ungaqala ngezingxoxo manje njengoba usuqonda ngokuqinile ukuthi ungayiphendula kanjani eminye yemibuzo evame ukucelwa kakhulu yenhlolokhono yokufunda. Thatha isinyathelo esilandelayo ngokusekelwe ezinhlosweni zakho.
Vakashela i-Hashdork's Uchungechunge Lwezingxoxo ukulungiselela izinhlolokhono.
shiya impendulo