- Isakhiwo esiyisisekelo samanethiwekhi e-neural esisekelwe ezingqimbeni zokufaka, ezifihliwe kanye neziphumayo ezicubungula idatha ngemisebenzi yokwenza kusebenze.
- Izindlela zokufunda eziqondisiwe zigxile ekusakazeni phambili kanye nokwenza ngcono amaphutha ngokusebenzisa ukusakazwa kwangemuva kanye nokwehla kwe-gradient.
- Ukuvela kwezakhiwo ezisebenzayo ezifana namanethiwekhi alula kanye namamodeli angenasisindo aqeda ukuphindaphinda okubizayo ukuze kuncishiswe ukusetshenziswa kwamandla.
Uma wake wazibuza ukuthi yini ebangela umlingo wobuhlakani bokwenziwa, impendulo emfushane iwukuthi sizama ukulingisa ukusebenza kobuchopho bomuntu . Cabanga ngenethiwekhi eyinkimbinkimbi kakhulu yamaseli abizwa ngokuthi ama-neurons athumelana ama-electrical impulses ukuze siqonde umhlaba; kahle, ososayensi bekhompyutha benze inguqulo yesofthiwe, enama-node nama-algorithms, ukuxazulula izinkinga zezibalo ezingasithatha amahora amaningi.
Ezweni le-AI, akuzona zonke izinto ezifanayo. Sisuke kumamodeli alula saya kumanethiwekhi ajulile e-neural aphatha izigidi zokuxhumeka, futhi manje sibheke ezakhiweni ezihlose ukuba ezizinzile kakhulu futhi ezisheshayo, ziphule amapharadigm esiwasebenzise amashumi eminyaka. Ake sihlukanise konke lokhu ukuze ungabi nokungabaza.
Isakhiwo esiyisisekelo senethiwekhi ye-neural
Ukuze inethiwekhi isebenze, ngokuvamile ihlelwe ngezinhlobo ezintathu zezingqimba. Okokuqala, kukhona ungqimba lokufaka , lapho idatha evela ngaphandle ingena khona; lapha, ama-node ahlaziya futhi ahlukanise ulwazi ngaphambi kokuluthumela ezingeni elilandelayo. Bese kufika izingqimba ezifihliwe , ezingaba ungqimba olulodwa noma izinkulungwane uma kuziwa ekufundeni okujulile. Lezi zingqimba zenza umsebenzi wokuphakamisa osindayo, zicubungula umphumela wongqimba odlule ukuze kukhishwe amaphethini.
Ekugcineni, sifika ku -output layer , esinika umphumela wokugcina. Kuye ngokuthi sifunani, kungase kube ne-node eyodwa (njengombuzo othi yebo/cha) noma eziningana uma sibhekene nenkinga yokuhlukaniswa kwamakilasi amaningi . Kumanethiwekhi ajulile, isihluthulelo sisezisindweni , izinombolo ezibonisa ukuthi i-neuron eyodwa iyavuselela noma iyayivimbela enye, ngaleyo ndlela kunqunywe ithonya lokuxhumana ngakunye emphumeleni wokugcina.
Inqubo yokufunda: Kusukela ku-theory kuya ekusebenzeni
Ukufunda kuwukulungisa lezi zisindo ukuze kugwenywe amaphutha. Isinyathelo sokuqala ukusabalala phambili , okuyidatha ehamba kusukela kwesobunxele kuya kwesokudla senethiwekhi. Ama-neurons enza isamba esilinganisiwe sezinto ezifakiwe , engeza ipharamitha ebizwa ngokuthi i-bias (ukuze anikeze ukuguquguquka okwengeziwe kwe-algebra), bese edlulisa umphumela ngomsebenzi wokwenza kusebenze.
Ake sixoxe ngalezi zinhlelo, ngoba yizo ezivimbela inethiwekhi ekubeni yi-linear regression elula. Ezivame kakhulu umsebenzi we-Sigmoid , obuyisela amanani aphakathi kuka-0 no-1 (afanele amathuba), kanye nomsebenzi we-ReLu , oyindlovukazi yokufunda okujulile ngoba kulula kakhulu futhi uvimbela i-gradient ukuthi inyamalale ngesikhathi sokuqeqeshwa.
Umlingo wokusabalala kwe-backpropagation kanye nokwehla kwe-gradient
Uma inethiwekhi yehluleka, i-backpropagation iyaqala ukusebenza . Lapha, inqubo iyaguqulwa: sisebenza kusukela kokukhiphayo kuya kokufakayo ukuze sibale iphutha. Sisebenzisa imisebenzi efana ne- Mean Squared Error (MSE) ukuze sinqume ukuthi siphambuke kangakanani eqinisweni. Ukusuka lapho, sisebenzisa i-gradient descent , esitshela ukuthi kufanele silungise izisindo ngakuphi ukuze sinciphise lelo phutha.
Iphuzu elibalulekile lapha izinga lokufunda . Uma liphezulu kakhulu, inethiwekhi ifunda ngokushesha kodwa ingadlula iphuzu elifanele futhi ingalungi; uma liphansi kakhulu, inqubo ayinamkhawulo futhi ingase ingaqedi ukufunda. Kuyibhalansi ebucayi echaza ikhwalithi yokuqeqeshwa.
Ukuthuthuka ekuqondeni i-AI esimeme: Amanethiwekhi alula namamodeli angenasisindo
Inkinga ngokufunda okujulile kwamanje ukuthi kudla amandla amaningi ngenxa yezigidigidi zokuphindaphinda ekwenzayo. Yingakho kuye kwavela amanethiwekhi e-neural alula , esebenzisa amasu afana nokusika ukuze kuqedwe ukuxhumana okungadingekile futhi kuncishiswe amandla awo ngaphandle kokulahlekelwa amandla amaningi okucubungula.
Kodwa igxathu langempela eliphazamisayo liza namanethiwekhi e-neural angenasisindo , acwaningwe ochwepheshe abanjengoLizy K. John. Esikhundleni sokuphindaphinda okufakwayo ngezisindo, basebenzisa amathebula okubheka axhumene anedatha ye-binary. Lolu ushintsho oluphelele lwe-paradigm: sisuka ekwenzeni izibalo ezibizayo zezibalo siye ekwenzeni imibuzo esheshayo, okuvumela inethiwekhi ukuthi ibe ncane ngokuphindwe kayi-1.000 futhi isebenze kahle kakhudlwana.
Izinhlelo zokusebenza zangempela kanye nekusasa le-Edge Computing
Lezi zakhiwo ezisebenza kahle kakhulu ziyigolide elicwengekile lokubala okusemaphethelweni , lapho ukucubungula kwenzeka khona ngqo kudivayisi, hhayi efwini. Cabanga ngezinzwa zezokwelapha eziqapha i-ECG noma umfutho wegazi ngesikhathi sangempela ngaphandle kokukhipha ibhethri ngemizuzu, noma izinhlelo zokuthola amagama angukhiye (njenge-Alexa's) ezidla ingxenye encane yama-nanojoule anamuhla.
Ngaphezu kwalokho, emkhakheni wezimboni, ukwenza ngcono ukupholisa ezindaweni zedatha kusetshenziswa amamodeli alula kunciphise ukusetshenziswa kwamandla ngamaphesenti afinyelela ku-30. Umkhuba ucacile: asisafuni nje amamodeli amakhulu, kodwa sifuna i-AI enomthwalo wemfanelo engaba ngcono ngaphandle kokubhubhisa iplanethi.
Uhambo olusuka kumodeli kaMcCulloch-Pitts ka-1943 kuya kuma-transformer namanethiwekhi angenasisindo lubonisa ukuthi ukusebenza kahle kuyindawo entsha. Ngenkathi ukufunda okujulile kwakudala kusinika amandla okukhiqiza umbhalo nezithombe, ukwenza ngcono ngokusebenzisa izakhiwo ezilula kanye namathebula okubheka kuzosivumela ukuthi sihlanganise ubuhlakani bokwenziwa kunoma iyiphi into yansuku zonke, sibeka phambili ubumfihlo kanye nokonga amandla ngaphezu kwamandla okusebenzisa ikhompyutha.
