Isiqulatho[Fihla][Bonisa]
- 1. Yintoni kanye kanye i-Deep Learning?
- 2. Yintoni eyahlula ukuFunda okuNzulu kuFundo ngoomatshini?
- 3. Ziziphi iingqiqo zakho zangoku zothungelwano lwe-neural?
- 4. Yintoni kanye kanye i-perceptron?
- 5. Yintoni kanye kanye inethiwekhi ye-neural enzulu?
- 6. Yintoni kanye kanye i-Multilayer Perceptron (MLP)?
- 7. Yeyiphi injongo imisebenzi yokuvula idlala kwinethiwekhi ye-neural?
- 8. Yintoni kanye kanye ukwehla kweGradient?
- 9. Yintoni Kanye Kanye Umsebenzi Weendleko?
- 10. Uthungelwano olunzulu lunokubagqwesa njani abo bangekho nzulu?
- 11. Chaza usasazo oluya phambili.
- 12. Yintoni i-backpropagation?
- 13. Kumxholo wokufunda nzulu, ukuqonde njani ukunqunyulwa kokuthambeka?
- 14. Yintoni iSoftmax kunye neReLU Imisebenzi?
- 15. Ngaba imodeli yenethiwekhi ye-neural inokuqeqeshwa kunye nabo bonke ubunzima obusetelwe ku-0?
- 16. Yintoni eyahlula ixesha kwibhetshi nokuphindaphinda?
- 17. Yintoni iBatch Normalization kunye nokuyeka?
- 18. Yintoni eyahlula iStochastic Gradient Descent kwiBatch Gradient Descent?
- 19. Kutheni le nto kubalulekile ukuquka okungeyo-linearities kuthungelwano lwe-neural?
- 20. Yintoni i-tensor ekufundeni nzulu?
- 21. Ungawukhetha njani umsebenzi wokuvula imodeli yokufunda nzulu?
- 22. Uthetha ukuthini nge-CNN?
- 23. Zeziphi iileya ezininzi ze-CNN?
- 24. Iba yintoni imiphumo yokungasebenzi kakuhle, yaye unokuyiphepha njani?
- 25. Kufundo olunzulu, yintoni iRNN?
- 26. Chaza i-Adam Optimizer
- 27. Ii-autoencoders ezinzulu: ziyintoni?
- 28. Ithetha ukuthini iTensor kwiTensorflow?
- 29. Ingcaciso yegrafu yokubala
- 30. Uthungelwano lwe-adversarial networks (GANs): yintoni?
- 31. Uza kukhetha njani inani lee-neuron kunye neeleyile ezifihliweyo ukuba zibandakanywe kwinethiwekhi ye-neural njengoko uyila uyilo loyilo?
- 32. Loluphi uhlobo lothungelwano lwe-neural olusetyenziswa ngokufunda okomeleza?
- isiphelo
Ukufunda nzulu asingombono omtsha. Uthungelwano lwe-neural eyenziweyo lusebenza njengesiseko sokuphela kweseti yokufunda koomatshini eyaziwa njengokufunda nzulu.
Ukufunda nzulu kukulinganisa ubuchopho bomntu, njengokuba kunjalo ngothungelwano lwe-neural, njengoko lwaludalelwe ukuxelisa ingqondo yomntu.
Kukho oku ixesha elide. Kwezi ntsuku, wonke umntu uthetha ngayo kuba asinawo amandla okuqhuba okanye idatha njengoko sisenza ngoku.
Kwiminyaka engama-20 edlulileyo, ukufunda okunzulu kunye nokufunda koomatshini kuye kwavela ngenxa yokunyuka okumangalisayo komthamo wokucubungula.
Ukukunceda ukuba ulungiselele nayiphi na imibuzo onokuthi ujongane nayo xa ufuna umsebenzi wamaphupha akho, esi sithuba siya kukukhokela kwinani lemibuzo enzulu yodliwano-ndlebe yokufunda, ukusuka kokulula ukuya kontsokothileyo.
1. Yintoni kanye kanye i-Deep Learning?
Ukuba uzimasa a ukufunda okunzulu udliwano-ndlebe, ngokungathandabuzekiyo uyayiqonda into enzulu yokufunda. Udliwano-ndlebe, nangona kunjalo, ulindele ukuba unikeze impendulo eneenkcukacha kunye nomzekeliso ekuphenduleni lo mbuzo.
Ukuze uqeqeshe amanethiwekhi kwimfundo enzulu, amanani abalulekileyo edatha ecwangcisiweyo okanye engacwangciswanga kufuneka isetyenziswe. Ukufumana iipateni ezifihliweyo kunye neempawu, yenza iinkqubo ezinzima (umzekelo, ukwahlula umfanekiso wekati ukusuka kwinja).
2. Yintoni eyahlula ukuFunda okuNzulu kuFundo ngoomatshini?
Njengesebe lobukrelekrele bokwenziwa elaziwa ngokuba kukufunda koomatshini, siqeqesha iikhomputha sisebenzisa idatha kunye neendlela zokwenza izibalo kunye ne-algorithmic ukuze zibengcono ngokuhamba kwexesha.
Njengomba we yokufunda umatshini, ukufunda okunzulu kulinganisa i-neural network architecture ebonwa kwingqondo yomntu.
3. Ziziphi iingqiqo zakho zangoku zothungelwano lwe-neural?
Iinkqubo ezenziweyo ezaziwa ngokuba ziinethiwekhi ze-neural zifana ne-organic neural networks ezifumaneka kumzimba womntu ngokusondeleyo kakhulu.
Ukusebenzisa ubuchule obufana nendlela i Ingqondo yomntu imisebenzi, inethiwekhi ye-neural yingqokelela ye-algorithms ejolise ekuchongeni ulungelelwaniso olusisiseko kwisiqwenga sedatha.
Ezi nkqubo zifumana ulwazi lomsebenzi othile ngokuziveza kuluhlu lweedatha kunye nemizekelo, endaweni yokulandela nayiphi na imigaqo engqamene nomsebenzi.
Ingcamango kukuba endaweni yokuba nokuqonda okucwangcisiweyo kwangaphambili kwezi datha, inkqubo ifunda iimpawu zokwahlula kwidatha eyondliwayo.
Iileya ezintathu zenethiwekhi ezisetyenziswa kakhulu kwiiNeural Networks zezi zilandelayo:
- Umaleko wegalelo
- Umaleko ofihlakeleyo
- Imveliso umaleko
4. Yintoni kanye kanye i-perceptron?
I-biological neuron efumaneka kwingqondo yomntu ithelekiseka ne-perceptron. Amagalelo amaninzi afunyanwa yi-perceptron, ethi emva koko yenza iinguqu ezininzi kunye nemisebenzi kwaye ivelise imveliso.
Imodeli yomgca ebizwa ngokuba yi-perceptron isetyenziswa kuhlelo lokubini. Ilinganisa i-neuron eneentlobo ezahlukeneyo zamagalelo, ngalinye linobunzima obahlukileyo.
I-neuron ibala umsebenzi usebenzisa la magalelo anobunzima kwaye ikhupha iziphumo.
5. Yintoni kanye kanye inethiwekhi ye-neural enzulu?
Inethiwekhi ye-neural enzulu yi-artificial neural network (ANN) enemigangatho emininzi phakathi kwe-input and output layers (DNN).
Uthungelwano olunzulu lwe-neural lubunzulu boyilo lwe-neural networks. Igama elithi "nzulu" libhekisa kwimisebenzi enamanqanaba amaninzi kunye neeyunithi kumaleko omnye. Iimodeli ezichaneke ngakumbi zinokudalwa ngokongeza iileya ezingaphezulu nangaphezulu ukuze ubambe amanqanaba amakhulu eepateni.
6. Yintoni kanye kanye i-Multilayer Perceptron (MLP)?
Ungeniso, olufihliweyo, kunye nolwaleko lwemveliso lukhona kwii-MLPs, kufana nothungelwano lwe-neural. Yakhiwe ngokufana ne-perceptron yoluhlu olulodwa kunye nemigangatho efihliweyo enye okanye ngaphezulu.
Isiphumo sokubini somaleko omnye we-perceptron singahlela kuphela iiklasi ezikwahlulwayo zomgca (0,1), ngelixa i-MLP inokwahlula iiklasi ezingezizo.
7. Yeyiphi injongo imisebenzi yokuvula idlala kwinethiwekhi ye-neural?
Umsebenzi wokuvula umisela ukuba i-neuron kufuneka isebenze na kwelona nqanaba lisisiseko. Nawuphi na umsebenzi wokuvula unokwamkela isixa esikalisiweyo samagalelo kunye nokuthambekela njengegalelo. Imisebenzi yokuvuselela ibandakanya umsebenzi wesinyathelo, i-Sigmoid, i-ReLU, i-Tanh, kunye ne-Softmax.
8. Yintoni kanye kanye ukwehla kweGradient?
Eyona ndlela ilungileyo yokunciphisa umsebenzi weendleko okanye impazamo kukuhla komgangatho. Ukufumana i-minima yendawo-yehlabathi yeyona njongo. Oku kuxela indlela ekufuneka imodeli ilandele ukunciphisa impazamo.
9. Yintoni Kanye Kanye Umsebenzi Weendleko?
Umsebenzi weendleko yimetric yokuvavanya ukuba imodeli yakho isebenza kakuhle kangakanani; ngamanye amaxesha kwaziwa ngokuba "yilahleko" okanye "impazamo." Ngexesha lokusasazwa kwakhona, isetyenziswa ukubala impazamo yomaleko wemveliso.
Sisebenzisa oko kungachanekanga ukuqhubela phambili iinkqubo zoqeqesho lwenethiwekhi ye-neural ngokuyibuyisela umva nge-neural network.
10. Uthungelwano olunzulu lunokubagqwesa njani abo bangekho nzulu?
Iileya ezifihliweyo zongezwa kuthungelwano lwe-neural ukongeza kwigalelo kunye neeleya zemveliso. Phakathi kweeleya zegalelo kunye nemveliso, uthungelwano lwe-neural olunganzulwanga lusebenzisa umaleko omnye ofihliweyo, ngelixa uthungelwano olunzulu lwe-neural lisebenzisa amanqanaba amaninzi.
Umsebenzi womnatha ongekho nzulu ufuna iiparamitha ezininzi ukuze ukwazi ukungena kuwo nawuphi na umsebenzi. Uthungelwano olunzulu lunokulungela imisebenzi ngcono nokuba nenani elincinci leeparamitha kuba zibandakanya iileya ezininzi.
Iinethiwekhi ezinzulu ngoku zikhethwa ngenxa yokuguquguquka kwazo ekusebenzeni kunye naluphi na uhlobo lwemodeli yedatha, nokuba kukuthetha okanye ukuqatshelwa kwemifanekiso.
11. Chaza usasazo oluya phambili.
Amagalelo ahanjiswa kunye kunye neentsimbi kumaleko angcwatywe kwinkqubo eyaziwa ngokuba yi-propagation yokudlulisa.
Imveliso yomsebenzi wokwenza isebenze ibalwa kumaleko ngamnye ongcwatyiweyo phambi kokuba uqwalaselo luqhubele phambili kolu luhlu lulandelayo.
Inkqubo iqala kuluhlu lwegalelo kwaye iqhubela phambili ukuya kwinqanaba lokugqibela lemveliso, ngaloo ndlela igama lokusasaza phambili.
12. Yintoni i-backpropagation?
Xa izisindo kunye ne-biases zihlengahlengiswa kwinethiwekhi ye-neural, i-backpropagation isetyenziselwa ukunciphisa umsebenzi weendleko ngokuqala ngokujonga indlela ixabiso elitshintsha ngayo.
Ukuqonda i-gradient kumaleko ngamnye ofihliweyo kwenza ukubala olu tshintsho lube lula.
Inkqubo, eyaziwa ngokuba yi-backpropagation, iqala kumaleko wemveliso kwaye ibuyela umva kumaleko okufakwayo.
13. Kumxholo wokufunda nzulu, ukuqonde njani ukunqunyulwa kokuthambeka?
I-Gradient Clipping yindlela yokusombulula umba we-gradients eziqhumayo ezivela ngexesha lokubuyisela umva (imeko apho i-gradients engachanekanga iqokelela ngokuhamba kwexesha, ekhokelela kuhlengahlengiso olubalulekileyo kwi-neural network weights model ngexesha loqeqesho).
Ukugqabhuka kwe-gradients ngumba ovela xa i-gradients ikhula kakhulu ngexesha loqeqesho, okwenza imodeli ingazinzi. Ukuba i-gradient inqumle kuluhlu olulindelekileyo, amaxabiso okuthambeka atyhalwa isiqalelo-ngesiqalelo ukuya kwelona nani lichazwe kwangaphambili lisezantsi okanye ixabiso eliphezulu.
Ukunqunyulwa kwegradient konyusa uzinzo lwamanani lwenethiwekhi ye-neural ngexesha loqeqesho, kodwa kunempembelelo encinci ekusebenzeni kwemodeli.
14. Yintoni iSoftmax kunye neReLU Imisebenzi?
Umsebenzi wokuvula obizwa ngokuba yiSoftmax uvelisa imveliso kuluhlu oluphakathi kwe-0 kunye ne-1. Isiphumo ngasinye sahlulwe ukuze isamba seziphumo zonke sibe sinye. Kumaleko emveliso, iSoftmax isetyenziswa rhoqo.
Iyunithi yomgca elungisiweyo, ngamanye amaxesha eyaziwa ngokuba yi-ReLU, ngowona msebenzi usetyenziswayo wokuvula. Ukuba u-X ulungile, ikhupha u-X, kungenjalo ikhupha ooziro. I-ReLU isetyenziswa rhoqo kumaleko angcwatywayo.
15. Ngaba imodeli yenethiwekhi ye-neural inokuqeqeshwa kunye nabo bonke ubunzima obusetelwe ku-0?
Inethiwekhi ye-neural soze ifunde ukugqiba umsebenzi onikiweyo, kungoko akunakwenzeka ukuqeqesha imodeli ngokuqalisa zonke iintsimbi ukuya ku-0.
I-derivatives iya kuhlala ifana kubo bonke ubunzima kwi-W [1] ukuba zonke iintsimbi ziqaliswa ukuya kwi-zero, okuya kubangela ukuba i-neurons ifunde iimpawu ezifanayo ngokuphindaphindiweyo.
Hayi nje ukuqalisa iintsimbi ukuya ku-0, kodwa nakweyiphi na indlela yokungaguquguquki kunokubangela iziphumo ezisezantsi.
16. Yintoni eyahlula ixesha kwibhetshi nokuphindaphinda?
Iindlela ezahlukeneyo zokusetyenzwa kwedatha kunye nobuchule bokwehla kwegradient bubandakanya ibhetshi, iteration, kunye ne-epoch. I-Epoch ibandakanya kube kanye-nge-neural network ene-dataset epheleleyo, phambili nangasemva.
Ukuze unikeze iziphumo ezithembekileyo, iseti yedatha idla ngokugqithiswa amatyeli amaninzi kuba inkulu kakhulu ukuba ingadlula ngetry enye.
Lo mkhuba wokuqhuba ngokuphindaphindiweyo inani elincinci ledatha ngokusebenzisa inethiwekhi ye-neural ubizwa ngokuba yi-iteration. Ukuqinisekisa ukuba isethi yedatha inqumla ngempumelelo uthungelwano lwe-neural, inokwahlulwa ngokwenani leebhetshi okanye ii-subsets, ezaziwa ngokuba yi-batching.
Ngokuxhomekeke kubungakanani bokuqokelelwa kwedatha, zontathu iindlela-ixesha, ukuphindaphinda, kunye nobungakanani bebhetshi-ziindlela ezingundoqo zokusebenzisa i-algorithm yokwehla kwe-gradient.
17. Yintoni iBatch Normalization kunye nokuyeka?
Ukuyeka kuthintela ukugqithiswa kwedatha ngokususa ngokungaqhelekanga zombini iiyunithi zenethiwekhi ezibonakalayo nezifihliweyo (ngokuqhelekileyo ukulahla i-20 yeepesenti zeendawo). Iphinda kabini inani lophindaphindo olufunekayo ukuze inethiwekhi ihlangane.
Ngokuhlengahlengisa amagalelo kumaleko ngamnye ukuba abe nesiphumo sokuvula i-zero kunye nokutenxa okusemgangathweni komnye, ibhetshi yesiqhelo sisicwangciso sokuphucula ukusebenza kunye nokuzinza kothungelwano lwe-neural.
18. Yintoni eyahlula iStochastic Gradient Descent kwiBatch Gradient Descent?
Ukuhla kweBatch Gradient:
- Iseti yedatha epheleleyo isetyenziselwa ukwakhiwa kwethareyidi yebhetshi yodidi.
- Isixa esikhulu sedatha kunye neentsimbi ezihlaziya ngokucothayo zenza ukudibanisa kube nzima.
Ukuhla kweStochastic Gradient:
- Igradient yestochastic isebenzisa isampulu enye ukubala ukuthambeka.
- Ngenxa yotshintsho oluthe kratya lobunzima, buguquka ngokukhawuleza ngakumbi kune-batch gradient.
19. Kutheni le nto kubalulekile ukuquka okungeyo-linearities kuthungelwano lwe-neural?
Kungakhathaliseki ukuba zininzi kangakanani iileya, inethiwekhi ye-neural iya kuziphatha njenge-perceptron ngokungabikho kwezinto ezingabonakaliyo, okwenza imveliso ixhomekeke kwigalelo.
Ukuyibeka ngenye indlela, inethiwekhi ye-neural ene-n layers kunye ne-m efihliweyo iiyunithi kunye nemisebenzi ye-linear activation ilingana ne-linear neural network ngaphandle kweeleyile ezifihliweyo kunye nokukwazi ukubona imida yokwahlula okukodwa kuphela.
Ngaphandle kwe-non-linearities, inethiwekhi ye-neural ayikwazi ukusombulula imiba enzima kwaye ihlele ngokuchanekileyo igalelo.
20. Yintoni i-tensor ekufundeni nzulu?
I-multidimensional array eyaziwa njenge-tensor isebenza njenge-generalization ye-matrices kunye ne-vectors. Lubume bedatha ebalulekileyo yokufunda nzulu. Uluhlu lwe-N-dimensional lweentlobo zedatha esisiseko zisetyenziselwa ukumela i-tensor.
Yonke icandelo le-tensor inohlobo lwedatha efanayo, kwaye olu hlobo lwedatha luhlala lusaziwa. Kusenokwenzeka ukuba sisiqwenga somlo kuphela—oko kukuthi, mingaphi imilinganiso ekhoyo nokuba mikhulu kangakanani na—eyaziwayo.
Kwiimeko apho amagalelo nawo aziwa ngokupheleleyo, uninzi lwemisebenzi luvelisa i-tensor ezaziwayo ngokupheleleyo; kwezinye iimeko, imo ye-tensor inokusekwa kuphela ngexesha lokwenziwa kwegrafu.
21. Ungawukhetha njani umsebenzi wokuvula imodeli yokufunda nzulu?
- Iyavakala ukusebenzisa umgca wokuvula umsebenzi ukuba isiphumo ekufuneka silindelwe sesokwenene.
- Umsebenzi weSigmoid kufuneka usetyenziswe ukuba imveliso ekufuneka ixelwe kwangaphambili ludidi olunokwenzeka lokubini.
- Umsebenzi weTanh ungasetyenziswa ukuba isiphumo esiqikelelweyo siqulathe iindidi ezimbini.
- Ngenxa yokulula kwayo ukubala, umsebenzi we-ReLU usebenza kuluhlu olubanzi lweemeko.
22. Uthetha ukuthini nge-CNN?
Uthungelwano lwe-neural olunzulu olusebenza ngokukhethekileyo ekuvavanyeni umfanekiso obonakalayo luquka uthungelwano lwe-neural ye-convolutional (CNN, okanye i-ConvNet). Apha, kunokuba kuthungelwano lwe-neural apho i-vector imele igalelo, igalelo ngumfanekiso oneetshaneli ezininzi.
Ii-perceptron ze-Multilayer zisetyenziswa ngendlela ekhethekileyo yi-CNNs efuna ukulungiswa kwangaphambili okuncinci kakhulu.
23. Zeziphi iileya ezininzi ze-CNN?
I-Convolutional Layer: Umaleko oyintloko ngumaleko we-convolution, onezihluzo ezahlukeneyo ezifundekayo kunye nomhlaba owamkelayo. Olu luhlu lokuqala luthatha idatha yegalelo kwaye lukhuphe iimpawu zalo.
Uluhlu lwe-RELU: Ngokwenza uthungelwano lungabikho emgqeni, olu maleko lujika iipikseli ezikhabayo zibe zero.
Umaleko wokudibanisa: Ngokunciphisa ukusetyenzwa kunye noseto lwenethiwekhi, umaleko wokudibanisa ngokuthe ngcembe unciphisa ubungakanani besithuba sokumelwa. Ukudityaniswa okuphezulu yeyona ndlela isetyenziswa kakhulu yokudibanisa.
24. Iba yintoni imiphumo yokungasebenzi kakuhle, yaye unokuyiphepha njani?
Oku kwaziwa njengokugqwesa kakhulu xa imodeli ifunda izinto ezintsonkothileyo kunye nengxolo kwidatha yoqeqesho ukuya kwinqanaba apho ichaphazela kakubi ukusetyenziswa kwemodeli yedatha entsha.
Kunokwenzeka ukuba kwenzeke ngeemodeli ezingezizo ezizilungelelanisa ngakumbi ngelixa ufunda umsebenzi wenjongo. Imodeli inokuqeqeshwa ukuba ibone iimoto kunye neelori, kodwa inokukwazi kuphela ukuchonga izithuthi ezinefomu yebhokisi ethile.
Ngenxa yokuba yayiqeqeshwe kuphela kudidi olunye lwelori, isenokungakwazi ukubona ilori ene-flatbed. Kwidatha yoqeqesho, imodeli isebenza kakuhle, kodwa kungekhona kwihlabathi langempela.
Imodeli engafakwanga ngokwaneleyo ibhekisa kulowo ungaqeqeshwanga ngokwaneleyo kwidatha okanye okwaziyo ukudibanisa ulwazi olutsha. Oku kwenzeka rhoqo xa imodeli iqeqeshwa ngedatha enganelanga okanye engachanekanga.
Ukuchaneka kunye nokusebenza zombini zisengozini ngenxa yokungafihli.
Ukuphinda kusetyenziswe idatha ukuqikelela ukuchaneka kwemodeli (ukuqinisekiswa kwe-K-fold cross-validation) kunye nokusebenzisa i-dataset yokuqinisekisa ukuvavanya imodeli ziindlela ezimbini zokuphepha ukugqithisa kunye nokunciphisa.
25. Kufundo olunzulu, yintoni iRNN?
Uthungelwano lwe-neural oluqhelekileyo (RNNs), iindidi eziqhelekileyo zothungelwano lwe-neural eyenziweyo, hamba ngesishunqulelo se-RNN. Baqeshelwe ukucubungula i-genomes, ukubhala ngesandla, isicatshulwa, kunye nolandelelwano lwedatha, phakathi kwezinye izinto. Kuqeqesho oluyimfuneko, iiRNN zisebenzisa i-backpropagation.
26. Chaza i-Adam Optimizer
I-Adam optimizer, ekwabizwa ngokuba yi-adaptive momentum, bubuchule bokuphucula obuphuhliswe ukujongana neemeko zengxolo ngeegradient ezinqabileyo.
Ukongeza ekuboneleleni ngohlaziyo lweparamitha nganye yokudibanisa okukhawulezileyo, i-Adam optimizer yongeza ukuhlangana ngesantya, iqinisekisa ukuba imodeli ayibanjiswa kwindawo yesali.
27. Ii-autoencoders ezinzulu: ziyintoni?
I-autoencoder enzulu ligama elidityanisiweyo lothungelwano lwenkolelo enzulu elinganayo ebandakanya ngokubanzi iileya ezine okanye ezintlanu ezingenzulwanga zesiqingatha sothungelwano lwekhowudi kunye nenye iseti yezine okanye ezintlanu umaleko wesiqingatha sekhowudi.
Ezi maleko zenza isiseko sothungelwano lweenkolelo ezinzulu kwaye zinyanzelwa ngoomatshini baseBoltzmann. Emva kwe-RBM nganye, i-autoencoder enzulu isebenzisa utshintsho lokubini kwi-dataset ye-MNIST.
Zisenokusetyenziswa kwezinye iiseti zedatha apho uguqulo olulungisiweyo lwe-Gaussian luya kukhethwa kune-RBM.
28. Ithetha ukuthini iTensor kwiTensorflow?
Lo ngomnye umbuzo wodliwano-ndlebe onzulu obuzwa rhoqo. I-tensor yingqiqo yemathematika ebonwa njenge-arrays ezinomlinganiselo ophezulu.
I-tensors zezi zintlu zedatha ezibonelelwa njengegalelo kuthungelwano lwe-neural kwaye zinemilinganiselo eyahlukeneyo kunye nokuhlelwa.
29. Ingcaciso yegrafu yokubala
Isiseko seTensorFlow kukwakhiwa kwegrafu yokubala. Indawo nganye isebenza kuthungelwano lweenodi, apho iindawo zimele imisebenzi yemathematika kunye nemiphetho yeetensor.
Ngamanye amaxesha ibizwa ngokuba yi "DataFlow Graph" kuba idatha ihamba ngendlela yegrafu.
30. Uthungelwano lwe-adversarial networks (GANs): yintoni?
Kwisifundo esinzulu, umfuziselo ovelisayo uphunyezwa kusetyenziswa uthungelwano oluchasayo oluvelisayo. Kungumsebenzi ongajongwanga apho umphumo uveliswa ngokuchonga iipateni kwidatha yegalelo.
Umcalu-calulo usetyenziselwa ukucalula iimeko eziveliswa ngumenzi, ngelixa ijenereyitha isetyenziselwa ukuvelisa imizekelo emitsha.
31. Uza kukhetha njani inani lee-neuron kunye neeleyile ezifihliweyo ukuba zibandakanywe kwinethiwekhi ye-neural njengoko uyila uyilo loyilo?
Ukunikezelwa komngeni wezoshishino, inani elichanekileyo lee-neurons kunye neengqimba ezifihliweyo ezifunekayo ukwakha i-neural network architecture ayikwazi ukumiselwa nayiphi na imithetho enzima kwaye ekhawulezayo.
Kuthungelwano lwe-neural, ubungakanani bomaleko ofihliweyo kufuneka siwe kwindawo ethile embindini wobungakanani begalelo kunye neeleya zemveliso.
Ukuqala kwentloko ekudaleni uyilo lwenethiwekhi ye-neural kunokufezekiswa ngeendlela ezimbalwa ezithe ngqo, nangona:
Ukuqala ngovavanyo oluthile olusisiseko olucwangcisiweyo ukubona ukuba yintoni enokwenza ngcono kuyo nayiphi na isethi yedatha ethile esekwe kumava angaphambili kunye neenethiwekhi ze-neural kwiindawo ezifanayo zehlabathi lokwenyani yeyona ndlela ilungileyo yokujongana nayo yonke imiceli mngeni yoqikelelo lwehlabathi lokwenyani.
Ubumbeko lwenethiwekhi lunokukhethwa ngokusekelwe kulwazi lomntu malunga nomba wesizinda kunye namava angaphambili enethiwekhi ye-neural. Xa uvavanya ukusetwa kwenethiwekhi ye-neural, inani leeleya kunye ne-neuron ezisetyenziswa kwiingxaki ezinxulumeneyo yindawo elungileyo yokuqala.
Ukuntsokotha kwenethiwekhi ye-neural kufuneka kwandiswe ngokuthe chu ngokusekwe kwimveliso eqikelelweyo kunye nokuchaneka, kuqalwa ngoyilo olulula lwenethwekhi ye-neural.
32. Loluphi uhlobo lothungelwano lwe-neural olusetyenziswa ngokufunda okomeleza?
- Kwiparadigm yokufunda koomatshini ebizwa ngokuba kukufunda ukomeleza, imodeli isebenza ukwandisa umbono womvuzo owongezelekayo, kanye njengokuba zisenza izinto eziphilayo.
- Imidlalo kunye nezithuthi eziziqhubayo zombini zichazwa njengeengxaki ezibandakanya nokuqiniswa ukufunda.
- Isikrini sisetyenziswa njengegalelo ukuba ingxaki eza kumelwa ngumdlalo. Ukuze kuveliswe isiphumo sezigaba ezilandelayo, i-algorithm ithatha iipixels njengegalelo kwaye iqhubeke kusetyenziswa iileya ezininzi ze-convolutional neural network.
- Iziphumo zezenzo zemodeli, nokuba zilungile okanye zimbi, zisebenza njengokuqiniswa.
isiphelo
UkuFunda okunzulu kuye kwakhula ekuthandeni ukutyhubela iminyaka, kunye nezicelo kuwo wonke ummandla woshishino.
Iinkampani ziya zikhangela ngakumbi iingcali ezinobuchule ezinokuyila iimodeli eziphindaphinda indlela yokuziphatha kwabantu zisebenzisa ukufunda okunzulu kunye neendlela zokufunda koomatshini.
Abaviwa abonyusa isethi yesakhono sabo kwaye bagcine ulwazi lwabo kwezi teknoloji zokusika banokufumana uluhlu olubanzi lwamathuba omsebenzi kunye nomvuzo onomtsalane.
Ungaqala ngodliwano-ndlebe ngoku ukuba ubambe ngamandla malunga nendlela yokuphendula eminye yemibuzo edla ngokucelwa ngokunzulu yodliwano-ndlebe. Thatha inyathelo elilandelayo ngokusekelwe kwiinjongo zakho.
Ndwendwela iHashdork's Uluhlu lodliwano-ndlebe ukulungiselela udliwano-ndlebe.
Shiya iMpendulo