Okuqukethwe[Fihla][Bonisa]
I-Artificial Intelligence (AI) ekuqaleni kwakucatshangwa ukuthi iyiphupho elikude, ubuchwepheshe besikhathi esizayo, kodwa lokho akusenjalo.
Okwake kwaba yisihloko socwaningo manje sekuyaqhuma emhlabeni wangempela. I-AI manje isitholakala ezindaweni ezihlukene, okuhlanganisa indawo yakho yokusebenza, isikole, amabhange, izibhedlela, ngisho nefoni yakho.
Amehlo ezimoto ezizishayelayo, amazwi kaSiri kanye ne-Alexa, izingqondo ezilandela ukubikezela isimo sezulu, izandla zokuhlinzwa okusizwa ngamarobhothi, nokunye.
Ukuhlakanipha okungekhona okwangempela (AI) isiba isici esivamile sempilo yesimanje. Eminyakeni embalwa edlule, i-AI iye yavela njengomdlali omkhulu ezinhlobonhlobo eziningi zobuchwepheshe be-IT.
Ekugcineni, inethiwekhi ye-neural isetshenziswa yi-AI ukufunda izinto ezintsha.
Ngakho-ke namuhla sizofunda ngamaNeural Networks, ukuthi asebenza kanjani, izinhlobo zawo, izinhlelo zokusebenza, nokunye okuningi.
Iyini iNeural Network?
In ukufunda imishini, inethiwekhi ye-neural iyinethiwekhi ehlelwe ngesofthiwe yama-neurons okwenziwa. Izama ukulingisa ubuchopho bomuntu ngokuba nezingqimba eziningi “zezinzwa,” ezifana nezinzwa ezisebuchosheni bethu.
Isendlalelo sokuqala sama-neurons sizokwamukela izithombe, ividiyo, umsindo, umbhalo, nokunye okokufaka. Le datha igeleza kuwo wonke amaleveli, okuphumayo kwesendlalelo esisodwa kugelezela kokulandelayo. Lokhu kubalulekile emisebenzini enzima kakhulu, njengokucutshungulwa kolimi lwemvelo lokufunda ngomshini.
Nokho, kwezinye izimo, kuhloswe ukucindezelwa kwesistimu ukunciphisa usayizi wamamodeli kuyilapho kugcinwa ukunemba nokusebenza kahle kungcono. Ukuthena inethiwekhi ye-neural kuyindlela yokucindezela ehlanganisa ukususa izisindo kumodeli efundiwe. Cabangela inethiwekhi ye-artificial intelligence neural eqeqeshelwe ukuhlukanisa abantu ezilwaneni.
Isithombe sizohlukaniswa sibe izingxenye ezikhanyayo nezimnyama ngongqimba lokuqala lwama-neurons. Le datha izodluliselwa kusendlalelo esilandelayo, esizocacisa ukuthi imiphetho ikuphi.
Isendlalelo esilandelayo sizozama ukubona amafomu akhiqizwe yinhlanganisela yamaphethelo. Ngokusho kwedatha eqeqeshwe ngayo, idatha izodlula izendlalelo eziningi ngendlela efanayo ukuze kunqunywe ukuthi isithombe owethule ngesomuntu noma isilwane.
Uma idatha inikezwa kunethiwekhi ye-neural, iqala ukuyicubungula. Ngemuva kwalokho, idatha icutshungulwa ngamaleveli ayo ukuze kutholwe umphumela oyifunayo. Inethiwekhi ye-neural umshini ofunda kokufakwayo okuhleliwe bese ubonisa imiphumela. Kunezinhlobo ezintathu zokufunda ezingenzeka kumanethiwekhi we-neural:
- Ukufunda Okugadiwe - Okufakiwe kanye nemiphumela kunikezwa ama-algorithms kusetshenziswa idatha enelebula. Ngemva kokufundiswa indlela yokuhlaziya idatha, babikezela umphumela ohlosiwe.
- Ukufunda Okungagadiwe - I-ANN ifunda ngaphandle kosizo lomuntu. Ayikho idatha enelebula, futhi okukhiphayo kunqunywa ngamaphethini atholakala kudatha yokuphumayo.
- Ukuqinisa Ukufunda yilapho inethiwekhi ifunda empendulweni eyitholayo.
Asebenza kanjani amanethiwekhi e-Neural?
Ama-neuron okwenziwa asetshenziswa kumanethiwekhi emizwa, okuyizinhlelo eziyinkimbinkimbi. Ama-neuron okwenziwa, aziwa nangokuthi ama-perceptron, akhiwe yizingxenye ezilandelayo:
- Ukufaka
- Isisindo
- I-Bias
- Umsebenzi Wokuqalisa
- Okukhiphayo
Izendlalelo zama-neurons akha amanethiwekhi emizwa. Inethiwekhi ye-neural iqukethe izendlalelo ezintathu:
- Isendlalelo okokufaka
- Isendlalelo esifihliwe
- Isendlalelo sokuphumayo
Idatha esesimweni senani lezinombolo ithunyelwa kungqimba yokufaka. Izendlalelo ezifihliwe zenethiwekhi yizo ezenza izibalo eziningi. Isendlalelo sokuphumayo, okokugcina kodwa okungenani, sibikezela umphumela. Ama-Neurons ayabusa elinye kunethiwekhi ye-neural. Ama-Neurons asetshenziselwa ukwakha isendlalelo ngasinye. Idatha ihanjiswa kusendlalelo esifihliwe ngemuva kokuthi isendlalelo sokufaka siyitholile.
Isisindo sisetshenziswa kokokufaka ngakunye. Ngaphakathi kwezendlalelo ezifihliwe zenethiwekhi ye-neural, isisindo siyinani elihumusha idatha engenayo. Isisindo sisebenza ngokuphindaphinda idatha yokokufaka ngenani lesisindo kusendlalelo yokokufaka.
Bese iqala inani lesendlalelo esifihliwe sokuqala. Idatha yokufaka iyaguqulwa futhi idluliselwe kwesinye isendlalelo ngezendlalelo ezifihliwe. Isendlalelo sokuphumayo sinesibopho sokukhiqiza umphumela wokugcina. Okokufaka nezisindo kuyaphindaphindwa, futhi umphumela ulethwa kuma-neurons ongqimba ofihliwe njengesamba. I-neuron ngayinye inikezwa ukuchema. Ukuze ubale isamba, i-neuron ngayinye yengeza okokufaka ekutholayo.
Ngemva kwalokho, inani lidlula ngomsebenzi wokwenza kusebenze. Umphumela womsebenzi wokwenza kusebenze unquma ukuthi i-neuron yenziwe yasebenza noma cha. Uma i-neuron isebenza, ithumela ulwazi kwezinye izendlalelo. Idatha idalwe kunethiwekhi kuze kube yilapho i-neuron ifinyelela ungqimba oluphumayo isebenzisa le ndlela. Ukusabalalisa phambili kungenye itemu yalokhu.
Isu lokuphakela idatha ku-node yokufaka kanye nokuthola okukhiphayo ngendawo yokukhiphayo yaziwa ngokuthi i-feed-forward propagation. Uma idatha yokufaka yamukelwe isendlalelo esifihliwe, ukusakazwa kokudlulisela phambili kwenzeka. Icutshungulwa ngokuya ngomsebenzi wokwenza kusebenze bese idluliselwa kokuphumayo.
Umphumela uboniswa yi-neuron esendlaleloni okukhiphayo ngamathuba aphezulu kakhulu. I-backpropagation ivela uma okukhiphayo kungalungile. Isisindo siqaliswa kokokufaka ngakunye ngenkathi kwakhiwa inethiwekhi ye-neural. I-Backpropagation iyinqubo yokulungisa izisindo zokufaka ngakunye ukuze kwehliswe amaphutha futhi kunikeze okukhiphayo okunembe kakhudlwana.
Izinhlobo zeNeural Network
1. Perceptron
Imodeli ye-Minsky-Papert perceptron ingenye yamamodeli alula futhi amadala ama-neuron. Iyunithi encane kakhulu yenethiwekhi ye-neural eyenza izibalo ezithile ukuze kutholwe izici noma ubuhlakani bebhizinisi kudatha engenayo. Kudingeka okokufaka okunesisindo futhi kusebenzisa umsebenzi wokuvula ukuze uthole umphumela wokugcina. I-TLU (threshold logic unit) elinye igama le-perceptron.
I-Perceptron iwuhlelo lokufunda olugadiwe oluhlukanisa idatha ngamaqembu amabili. Amasango eLogic okufana nokuthi KANYE, NOMA, kanye ne-NAND kungasetshenziswa ngama-perceptron.
2. Inethiwekhi ye-Feed-Forward Neural
Inguqulo eyisisekelo kakhulu yamanethiwekhi e-neural, lapho idatha yokokufaka igeleza ngokukhethekile ohlangothini olulodwa, idlula ngamanodi emizwa yokwenziwa futhi iphume ngamanodi okukhiphayo. Izendlalelo zokokufaka nokuphumayo zikhona ezindaweni lapho izingqimba ezifihliwe zingase zibe khona noma zingabi khona. Angabonakala njengenethiwekhi ye-neural enesendlalelo esisodwa noma enezendlalelo eziningi okuphakelayo okuya phambili ngokusekelwe kulokhu.
Inani lezendlalelo ezisetshenzisiwe linqunywa ubunkimbinkimbi bomsebenzi. Isakazeka phambili kuphela ohlangothini olulodwa futhi ayisakazeli emuva. Lapha, izisindo zihlala zingashintshi. Okokufaka kuphindaphindwa ngezisindo ukuze kuphakelwe umsebenzi wokwenza kusebenze. Umsebenzi wokwenza kusebenze ukuhlukanisa noma umsebenzi wokwenza kusebenze isinyathelo usetshenziselwa ukwenza lokhu.
3. I-perceptron enezingqimba eziningi
Isingeniso sokuyinkimbinkimbi amanetha e-neural, lapho idatha yokufaka ihanjiswa ngezingqimba eziningi zama-neurons okwenziwa. Kuyinethiwekhi ye-neural exhunywe ngokuphelele, njengoba yonke i-node ixhunywe kuwo wonke ama-neurons kungqimba olulandelayo. Izendlalelo eziningi ezifihliwe, okungukuthi, okungenani izendlalelo ezintathu noma ngaphezulu, zikhona kuzindlalelo zokufakwayo neziphumayo.
Inokusakazwa okuphindwe kabili, okusho ukuthi ingasakaza kokubili phambili nangemuva. Okokufaka kuphindaphindwa ngezisindo futhi kuthunyelwe kumsebenzi wokwenza kusebenze, lapho ashintshwa khona nge-backpropagation ukuze kuncishiswe ukulahlekelwa.
Izisindo zingamanani afundiwe ngomshini avela ku-Neural Networks, ukukubeka kalula. Kuye ngomehluko phakathi kokuphumayo okulindelekile nokokufaka koqeqesho, bayazilungisa. I-Softmax isetshenziswa njengomsebenzi wokwenza kusebenze isendlalelo esiphumayo ngemva kwemisebenzi yokuvula engaqondile.
4. Convolutional Neural Network
Ngokuhlukile kuhlelo lwendabuko lwezinhlangothi ezimbili, inethiwekhi ye-convolution ye-neural inokumiswa kwezinhlangothi ezintathu zama-neurons. Isendlalelo sokuqala saziwa ngokuthi yi-convolutional layer. I-neuron ngayinye isendlalelo se-convolutional icubungula kuphela ulwazi olusuka engxenyeni elinganiselwe yenkambu yokubuka. Njengesihlungi, izici zokufaka zithathwa ngemodi ye-batch.
Inethiwekhi iqonda izithombe ezigabeni futhi ingenza lezi zenzo izikhathi eziningi ukuze iqedele ukucutshungulwa kwesithombe.
Isithombe siguqulwa sisuka ku-RGB noma i-HSI sibe greyscale ngesikhathi sokucutshungulwa. Ukwehluka okwengeziwe kwevelu yamaphikseli kuzosiza ekutholeni imiphetho, futhi izithombe zingahlelwa ngamaqembu amaningana. Ukusakazwa ngendlela engaqondile kwenzeka lapho i-CNN iqukethe isendlalelo esisodwa noma eziningi ezilandelwayo ezilandelwa ukuhlanganisa, futhi ukusakazwa okuphindwe kabili kwenzeka lapho okukhiphayo kongqimba lwe-convolution kuthunyelwa kunethiwekhi ye-neural exhunywe ngokugcwele ukuze kuhlukaniswe isithombe.
Ukuze kukhishwe izici ezithile zesithombe, izihlungi ziyasetshenziswa. Ku-MLP, okokufaka kukalwa futhi kunikezwa kumsebenzi wokwenza kusebenze. I-RELU isetshenziswa ku-convolution, kuyilapho i-MLP isebenzisa umsebenzi wokuvula ongaqondile olandelwa yi-softmax. Ekubonakalweni kwesithombe nevidiyo, ukucozululwa kwe-semantic, kanye nokutholwa kwamagama-magama, amanethiwekhi e-convolutional neural akhiqiza imiphumela emihle kakhulu.
5. Inethiwekhi ye-Radial Bias
Ivektha yokufaka ilandelwa ungqimba lwama-neuron e-RBF kanye nongqimba oluphumayo olunenodi eyodwa esigabeni ngasinye ku-Radial Basis Function Network. Okokufaka kuhlukaniswa ngokuqhathanisa namaphoyinti edatha kusukela kusethi yokuqeqeshwa, lapho neuron ngayinye igcina i-prototype. Lesi ngesinye sezibonelo zesethi yokuqeqeshwa.
I-neuron ngayinye ibala ibanga le-Euclidean phakathi kokokufaka kanye ne-prototype yayo lapho ivektha yokufaka entsha [ivektha engu-n-dimensional ozama ukuyibeka ngokwezigaba] kufanele ihlukaniswe. Uma sinezigaba ezimbili, Ikilasi A kanye Nekilasi B, okokufaka okusha okuzohlukaniswa kufana kakhulu nama-prototypes ekilasi A kunezibonelo zekilasi B.
Ngenxa yalokho, ingase ilebulwe noma ihlukaniswe njengesigaba A.
6. Inethiwekhi Yezinzwa Ezivamile
I-Recurrent Neural Networks yakhelwe ukulondoloza okuphumayo kwesendlalelo bese iwubuyisela kokokufaka ukusiza ukubikezela umphumela wesendlalelo. Ukudlulisela phambili inethiwekhi ye-neural ngokuvamile isendlalelo sokuqala, esilandelwa isendlalelo senethiwekhi ye-neural esiphindelelayo, lapho umsebenzi wenkumbulo ukhumbula ingxenye yolwazi ebinalo esinyathelweni sesikhathi sangaphambilini.
Lesi simo sisebenzisa ukusabalalisa phambili. Ilondoloza idatha ezodingeka esikhathini esizayo. Esimeni lapho ukubikezela kungalungile, izinga lokufunda lisetshenziselwa ukwenza izinguquko ezincane. Ngenxa yalokho, njengoba i-backpropagation iqhubeka, izoba nenembile ngokwandayo.
Izicelo
Amanethiwekhi e-Neural asetshenziselwa ukusingatha izinkinga zedatha emikhakheni eyahlukene; ezinye izibonelo zikhonjisiwe ngezansi.
- Ukuqashelwa Ubuso - Izixazululo Zokubona Ubuso zisebenza njengezinhlelo zokugada ezisebenzayo. Amasistimu wokuqaphela ahlobanisa izithombe zedijithali nobuso babantu. Asetshenziswa emahhovisi uma kungenwa ngokukhetha. Ngakho, amasistimu aqinisekisa ubuso bomuntu futhi abuqhathanise nohlu lwama-ID agcinwe kusizindalwazi sakhe.
- I-Stock Prediction - Ukutshalwa kwezimali kuvezwe ezingozini zemakethe. Kunzima ukubona kusengaphambili intuthuko yesikhathi esizayo emakethe yamasheya eshintshashintsha kakhulu. Ngaphambi kwamanethiwekhi e-neural, izigaba ze-bullish ne-bearish ezishintsha njalo zazingenakubikezelwa. Kodwa, yini eyashintsha konke? Yebo, sikhuluma ngamanethiwekhi emizwa… I-Multilayer Perceptron MLP (uhlobo lwesistimu ye-feedforward intelligence yokwenziwa) isetshenziselwa ukwakha isibikezelo sezulu esiyimpumelelo ngesikhathi sangempela.
- Social Media - Kungakhathaliseki ukuthi i-corny ingazwakala kanjani, inkundla yezokuxhumana ishintshile indlela yokuphila. Ukuziphatha kwabasebenzisi benkundla yezokuxhumana kufundwa kusetshenziswa i-Artificial Neural Networks. Ukuze kuhlaziywe ngokuncintisana, idatha enikezwa nsuku zonke ngokusebenzisana okubonakalayo iyanqwabelana futhi ihlolwe. Izenzo zabasebenzisi benkundla yezokuxhumana ziphindaphindwa amanethiwekhi we-neural. Ukuziphatha komuntu ngamunye kungaxhunywa kundlela abantu abasebenzisa ngayo imali uma idatha isihlaziywe kusetshenziswa amanethiwekhi ezokuxhumana. Idatha evela ezinhlelweni zokuxhumana zenziwa kusetshenziswa i-Multilayer Perceptron ANN.
- Ukunakekelwa Kwezempilo - Abantu emhlabeni wanamuhla basebenzisa izinzuzo zobuchwepheshe embonini yezokunakekelwa kwempilo. Ebhizinisini lokunakekelwa kwezempilo, i-Convolutional Neural Networks isetshenziselwa ukutholwa kwe-X-ray, ama-CT scan, kanye ne-ultrasound. Idatha yokucabanga yezokwelapha etholwe ezivivinyweni ezishiwo ngenhla iyahlolwa futhi ihlolwe kusetshenziswa amamodeli enethiwekhi ye-neural, njengoba i-CNN isetshenziswa ekucubunguleni izithombe. Ekuthuthukisweni kwezinhlelo zokuqaphela izwi, inethiwekhi ye-neural recurrent (RNN) nayo iyasetshenziswa.
- Umbiko Wesimo Sezulu – Ngaphambi kokusetshenziswa kobuhlakani bokwenziwa, ukuqagela komnyango wesimo sezulu akuzange kunembile. Ukubikezela isimo sezulu kwenziwa kakhulu ukubikezela izimo zezulu ezizokwenzeka esikhathini esizayo. Izibikezelo zesimo sezulu zisetshenziselwa ukulindela amathuba ezinhlekelele zemvelo esikhathini samanje. Ukubikezela isimo sezulu kwenziwa kusetshenziswa i-multilayer perceptron (MLP), i-convolutional neural networks (CNN), kanye namanethiwekhi e-neural aphindaphindiwe (RNN).
- Defense - Logistics, ukuhlaziywa kokuhlasela kuhlonyiwe, kanye nendawo yento konke kusebenzisa amanethiwekhi we-neural. Baphinde baqashwe ekuqapheni emoyeni nasolwandle, kanye nokuphatha ama-drones azimele. Ubuhlakani bokwenziwa bunikeza imboni yezokuvikela amandla adingekayo ukuze ikhulise ubuchwepheshe bayo. Ukuze kutholwe ubukhona bezimayini ezingaphansi kwamanzi, i-Convolutional Neural Networks (CNN) iyasetshenziswa.
Izinzuzo
- Ngisho noma ama-neuron ambalwa kunethiwekhi ye-neural engasebenzi kahle, amanethiwekhi e-neural asazokhiqiza okuphumayo.
- Amanethiwekhi e-Neural anekhono lokufunda ngesikhathi sangempela futhi azivumelanise nezilungiselelo zawo eziguqukayo.
- Amanethiwekhi e-Neural angafunda ukwenza imisebenzi eyahlukene. Ukuhlinzeka ngomphumela olungile ngokusekelwe kudatha enikeziwe.
- Amanethiwekhi e-Neural anamandla nekhono lokusingatha imisebenzi eminingana ngesikhathi esisodwa.
Okumbi
- Amanethiwekhi e-Neural asetshenziselwa ukuxazulula izinkinga. Ayidaluli incazelo ngemuva kokuthi "kungani futhi kanjani" yenze izinqumo ezenzile ngenxa yokuxaka kwamanethiwekhi. Ngenxa yalokho, ukwethenjwa kwenethiwekhi kungase kucekeleke phansi.
- Izingxenye zenethiwekhi ye-neural zincike kwenye. Okusho ukuthi, amanethiwekhi e-neural afuna (noma athembele kakhulu) amakhompyutha anamandla anele okwenza ikhompuyutha.
- Inqubo yenethiwekhi ye-neural ayinawo umthetho othize (noma umthetho wesithupha). Ngendlela yokuhlola nephutha, isakhiwo senethiwekhi esilungile sisungulwa ngokuzama inethiwekhi efanele. Kuyinqubo edinga ukulungiswa kahle kakhulu.
Isiphetho
Inkambu ye amanethiwekhi we-neural ikhula ngokushesha. Kubalulekile ukufunda nokuqonda imiqondo kulo mkhakha ukuze ukwazi ukubhekana nayo.
Izinhlobo eziningi zamanethiwekhi e-neural zimboziwe kulesi sihloko. Ungasebenzisa amanethiwekhi e-neural ukuze ubhekane nezinkinga zedatha kwezinye izinkambu uma ufunda kabanzi ngalesi siyalo.
shiya impendulo