Iyini iNethiwekhi Yezinzwa ku-AI?

Iyini iNethiwekhi Yezinzwa ku-AI?

Amanethiwekhi e-neural azwakala eyimfihlakalo aze angabi njalo. Uma wake wazibuza ukuthi iyini i-Neural Network ku-AI? nokuthi ngabe iyizibalo nje ngesigqoko esihle, usesendaweni efanele. Sizoyigcina isebenza, sisakaze izindlela ezincane, futhi yebo - ama-emoji ambalwa. Uzoshiya wazi ukuthi ziyini lezi zinhlelo, ukuthi kungani zisebenza, ukuthi zehluleka kuphi, nokuthi ungakhuluma kanjani ngazo ngaphandle kokuziphakamisa ngesandla.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Kuyini ukubandlulula kwe-AI
Ukuqonda ukubandlulula ezinhlelweni ze-AI namasu okuqinisekisa ukulingana.

🔗 Kuyini i-AI yokubikezela
Indlela i-AI ebikezelayo esebenzisa ngayo amaphethini ukubikezela imiphumela yesikhathi esizayo.

🔗 Uyini umqeqeshi we-AI
Ukuhlola indima kanye nemithwalo yemfanelo yochwepheshe abaqeqesha i-AI.

🔗 Kuyini umbono wekhompyutha ku-AI
Indlela i-AI ehumusha futhi ihlaziye ngayo idatha ebonakalayo ngokusebenzisa umbono wekhompyutha.

Iyini iNethiwekhi Yezinzwa ku-AI? Impendulo yemizuzwana eyi-10 ⏱️

Inethiwekhi yemizwa iyinqwaba yamayunithi wokubala alula abizwa ngokuthi ama-neurons adlulisa izinombolo phambili, alungise amandla awo okuxhumana ngesikhathi sokuqeqeshwa, futhi kancane kancane afunde amaphethini kudatha. Uma uzwa ukufunda okujulile , lokho kuvame ukusho inethiwekhi yemizwa enezendlalelo eziningi ezihlanganisiwe, ukufunda izici ngokuzenzakalelayo esikhundleni sokuzibhala ngesandla. Ngamanye amazwi: izingcezu eziningi ezincane zezibalo, ezihlelwe ngobuhlakani, eziqeqeshwe kudatha kuze kube yilapho ziwusizo [1].

Yini eyenza iNethiwekhi Yezinzwa ibe usizo? ✅

  • Amandla okumelwa : Ngokwakhiwa nosayizi ofanele, amanethiwekhi angalinganisa imisebenzi eyinkimbinkimbi kakhulu (bheka i-Universal Approximation Theorem) [4].

  • Ukufunda kusukela ekuqaleni kuze kube sekupheleni : Esikhundleni sezici zobunjiniyela bezandla, imodeli iyazithola [1].

  • Ukuhlanganisa : Inethiwekhi ehlelwe kahle ayigcini nje ngokukhumbula - isebenza kudatha entsha, engabonakali [1].

  • Ukusabalala : Amasethi edatha amakhulu kanye namamodeli amakhulu avame ukuqhubeka nokuthuthukisa imiphumela… kuze kufike emikhawulweni esebenzayo njengekhwalithi yokubala kanye nedatha [1].

  • Ukudluliswa : Izici ezifundwe komunye umsebenzi zingasiza komunye (ukudlulisa ukufunda kanye nokulungisa kahle) [1].

Inothi elincane lensimu (isibonelo sesimo): Ithimba elincane lokuhlukanisa umkhiqizo lishintsha izici ezakhiwe ngesandla nge-CNN encane, lengeza ukukhushulwa okulula (ama-flips/crops), bese libuka ukwehla kwephutha lokuqinisekisa - hhayi ngoba inethiwekhi "iyimilingo," kodwa ngoba ifunde izici eziwusizo kakhulu ngqo kumaphikseli.

“Kuyini iNethiwekhi Yezinzwa ku-AI?” ngesiNgisi esilula, nesingathekiso esingaqondakali 🍞

Cabanga ngomugqa wokubhaka. Izithako ziyangena, izisebenzi zilungisa iresiphi, abahloli bokunambitha bakhononda, bese ithimba libuyekeza iresiphi futhi. Kunethiwekhi, okokufaka kugeleza ngezendlalelo, umsebenzi wokulahlekelwa ubeka amazinga okukhiphayo, kanye nama-gradients ashukumisa izisindo ukuze enze kangcono ngesikhathi esizayo. Akuphelele njengesingathekiso - isinkwa asihlukaniseki - kodwa siyanamathela [1].

Ukwakheka kwenethiwekhi yezinzwa 🧩

  • Ama-Neuron : Ama-calculator amancane asebenzisa isamba esilinganisiwe kanye nomsebenzi wokwenza kusebenze.

  • Izisindo kanye nokubandlulula : Izinkinobho ezilungisekayo ezichaza indlela amasignali ahlangana ngayo.

  • Izendlalelo : Isendlalelo sokufaka sithola idatha, izendlalelo ezifihliwe ziyayiguqula, isendlalelo sokukhipha senza isibikezelo.

  • Imisebenzi Yokwenza Kusebenze : Ukushintshashintsha okungekhona okuqondile njenge-ReLU, i-sigmoid, i-tanh, kanye ne-softmax kwenza ukufunda kube lula.

  • Umsebenzi wokulahlekelwa : Isilinganiso sokuthi isibikezelo singalungile kangakanani (i-cross-entropy yokuhlukaniswa, i-MSE yokuhlehla).

  • I-Optimizer : Ama-algorithm afana ne-SGD noma i-Adam asebenzisa ama-gradients ukuvuselela izisindo.

  • Ukuhlelwa kabusha : Amasu anjengokuyeka noma ukubola kwesisindo ukuvimbela imodeli ukuthi ingalingani kakhulu.

Uma ufuna ukwelashwa okusemthethweni (kodwa okusafundeka), incwadi evulekile ethi Deep Learning ihlanganisa yonke inqwaba: izisekelo zezibalo, ukwenza ngcono, kanye nokuhlanganisa [1].

Imisebenzi yokuqalisa, kafushane kodwa iwusizo ⚡

  • I-ReLU : Akukho okungalungile, okuqondile okuhle. Kulula, kuyashesha, futhi kuyasebenza.

  • I-Sigmoid : I-Squashes amanani aphakathi kuka-0 no-1 - awusizo kodwa angagcwala.

  • I-Tanh : Njenge-sigmoid kodwa i-symmetric ezungeze u-zero.

  • I-Softmax : Iguqula amaphuzu angakahlelwa abe amathuba kuzo zonke izigaba.

Akudingeki ukuthi ubambe ngekhanda yonke imilo yejika - vele wazi ama-trade-offs kanye nokuzenzakalelayo okuvamile [1, 2].

Indlela ukufunda okwenzeka ngayo ngempela: i-backprop, kodwa ayesabisi 🔁

  1. Ukudlula phambili : Idatha igeleza ungqimba ngengqimba ukukhiqiza isibikezelo.

  2. Ukulahlekelwa kokubala : Qhathanisa ukubikezela neqiniso.

  3. Ukusabalalisa emuva : Bala ama-gradients okulahlekelwa maqondana nesisindo ngasinye usebenzisa umthetho weketanga.

  4. Isibuyekezo : I-Optimizer ishintsha isisindo kancane.

  5. Phinda : Izinkathi eziningi. Imodeli ifunda kancane kancane.

Ukuze uthole ulwazi olusebenzayo olunezithombe nezincazelo ezihambisana nekhodi, bheka amanothi e-CS231n akudala ku-backprop kanye nokwenza ngcono [2].

Imindeni emikhulu yamanethiwekhi ezinzwa, ngokubuka nje 🏡

  • Amanethiwekhi okudlulisela impendulo (ama-MLP) : Uhlobo olulula kakhulu. Idatha iya phambili kuphela.

  • Amanethiwekhi Emizwa Eguquguqukayo (ama-CNN) : Kuhle kakhulu ezithombeni ngenxa yezihlungi zendawo ezibona imiphetho, ukuthungwa, kanye nezimo [2].

  • Amanethiwekhi Emizwa Aphindaphindayo (ama-RNN) kanye nezinhlobo : Akhelwe ukulandelana okufana nombhalo noma uchungechunge lwesikhathi ngokugcina umuzwa wokuhleleka [1].

  • Ama-Transformers : Sebenzisa ukunaka ekuboniseni ubudlelwano phakathi kwezikhundla ngokulandelana ngasikhathi sinye; obubusayo olimini nangale kwalokho [3].

  • Amanethiwekhi E-Neural Graph (ama-GNN) : Asebenza kuma-node nasemaphethelweni egrafu - awusizo kuma-molecule, kumanethiwekhi omphakathi, izincomo [1].

  • Ama-Autoencoders nama-VAE : Funda izethulo ezicindezelwe bese udala ukuhlukahluka [1].

  • Amamodeli akhiqizayo : Kusukela kuma-GAN kuya kumamodeli okusabalalisa, asetshenziselwa izithombe, umsindo, ngisho nekhodi [1].

Amanothi e-CS231n anobungane kakhulu kuma-CNN, kanti iphepha le-Transformer lingumthombo oyinhloko osetshenziswa kakhulu kumamodeli asekelwe ekunakeni [2, 3].

Ithebula lokuqhathanisa: izinhlobo zenethiwekhi ye-neural ezivamile, ukuthi zingobani, ama-vibe ezindleko, nokuthi kungani zisebenza 📊

Ithuluzi / Uhlobo Izithameli Intengo-ngokufanayo Kungani kusebenza
Impendulo (i-MLP) Abaqalayo, abahlaziyi Okuphansi-okuphakathi Izisekelo ezilula, eziguquguqukayo, nezinhle
I-CNN Amaqembu ombono Okuphakathi nendawo Amaphethini endawo + ukwabelana ngamapharamitha
I-RNN / LSTM / GRU Abantu abalandelanayo Okuphakathi nendawo Inkumbulo yesikhashana... ibamba ukuhleleka
I-Transformer I-NLP, i-multimodal Okuphakathi nendawo okuphezulu Ukunaka kugxile ebudlelwaneni obufanele
I-GNN Ososayensi, ama-recsys Okuphakathi nendawo Ukudlulisa imiyalezo kumagrafu kwembula isakhiwo
I-Autoencoder / i-VAE Abacwaningi Okuphansi-okuphakathi Ufunda izethulo ezicindezelwe
I-GAN / Ukusabalalisa Amalebhu okudala Okuphakathi nendawo okuphezulu Umlingo wokuphikisana noma ophindaphindayo oveza umsindo

Amanothi: amanani amayelana nokubala nesikhathi; ibanga lakho liyahlukahluka. Iseli elilodwa noma amabili axoxa ngamabomu ngamabomu.

"Iyini iNethiwekhi Yezinzwa ku-AI?" uma kuqhathaniswa nama-algorithms e-ML akudala ⚖️

  • Ubunjiniyela bezici : I-ML yakudala ivame ukuthembela kuzici ezenziwa ngesandla. Amanethi e-neural afunda izici ngokuzenzakalelayo - ukunqoba okukhulu kwedatha eyinkimbinkimbi [1].

  • Indlala yedatha : Amanethiwekhi avame ukukhanya ngedatha eningi; idatha encane ingase ithande amamodeli alula [1].

  • Ukubala : Amanethiwekhi athanda ama-accelerator afana nama-GPU [1].

  • Umkhawulo wokusebenza : Ngedatha engahlelekile (izithombe, umsindo, umbhalo), ama-deep nets avame ukubusa [1, 2].

Ukuhamba komsebenzi wokuqeqesha okusebenza ngempela 🛠️

  1. Chaza umgomo : Ukuhlela, ukuhlehla, ukukala, ukukhiqiza - khetha ukulahlekelwa okuhambisanayo.

  2. Ukuxazululwa kwedatha : Hlukanisa kube yisitimela/ukuqinisekiswa/ukuhlolwa. Yenza izici zifane. Balance amakilasi. Ukuze uthole izithombe, cabanga ngokwandisa okufana nokuphenduka, ukunqampuna, umsindo omncane.

  3. Ukukhetha ukwakheka : Qala kalula. Engeza umthamo kuphela uma kudingeka.

  4. I-loop yokuqeqesha : Hlanganisa idatha. Dlulisa phambili. Bala ukulahlekelwa. I-Backprop. Buyekeza. Bhala amamethrikhi.

  5. Ukulungisa kabusha : Ukuyeka, ukuwohloka kwesisindo, ukuyeka kusenesikhathi.

  6. Hlola : Sebenzisa isethi yokuqinisekisa yama-hyperparameter. Beka isethi yokuhlola yokuhlola kokugcina.

  7. Thumela ngokucophelela : Qapha ukuzulazula, hlola ubandlululo, hlela ukugoba emuva.

Kwezifundo ezisekelwe ekhodini kusukela ekuqaleni kuya ekugcineni ezinethiyori eqinile, incwadi evulekile kanye namanothi e-CS231n kuyizinto ezithembekile [1, 2].

Ukufaka ngokweqile, ukuhlanganisa, kanye namanye ama-gremlins 👀

  • Ukufaka ngokweqile : Imodeli ikhumbula ngekhanda izici zokuqeqesha. Lungisa ngedatha eyengeziwe, ukuhlelwa kabusha okuqinile, noma ukwakheka okulula.

  • Ukufaneleka : Imodeli ilula kakhulu noma ukuqeqeshwa kuyathusa kakhulu. Khulisa umthamo noma ukuqeqesha isikhathi eside.

  • Ukuvuza kwedatha : Ulwazi oluvela kusethi yokuhlola lungena ekuqeqeshweni. Hlola kathathu ukuhlukaniswa kwakho.

  • Ukulinganisa okungekuhle : Imodeli eqinisekayo kodwa engalungile iyingozi. Cabanga ngokulinganisa noma okunye ukwehlisa isisindo.

  • Ukushintsha kokusabalalisa : Ukushintsha kwedatha yangempela. Qapha futhi uvumelane nezimo.

Ukuze uthole umbono ongemuva kokwenza izinto zibe zijwayelekile kanye nokuhlelwa kabusha, ncika ezinkombeni ezijwayelekile [1, 2].

Ukuphepha, ukuhunyushwa kalula, kanye nokuthunyelwa okunesibopho 🧭

Amanethiwekhi e-neural angenza izinqumo ezibalulekile. Akwanele ukuthi asebenze kahle ebhodini labaphambili. Udinga izinyathelo zokuphatha, zokulinganisa, kanye nezokuncishiswa kulo lonke umjikelezo wokuphila. Uhlaka Lokuphathwa Kwengozi lwe-NIST AI luchaza imisebenzi ewusizo - UKULAWULA, UKUPHATHA, UKULINGANISA, UKUPHATHA - ukusiza amaqembu ukuhlanganisa ukuphathwa kwengozi ekwakhiweni nasekusetshenzisweni [5].

Ukugudluza okumbalwa okusheshayo:

  • Ukuhlolwa kokukhetha : Hlola kuzo zonke izingxenye zabantu lapho kufaneleka khona futhi kusemthethweni.

  • Ukuhunyushwa : Sebenzisa amasu anjengokuqonda noma izici. Aziphelele, kodwa ziwusizo.

  • Ukuqapha : Setha izexwayiso zokwehla kwe-metric noma ukuzulazula kwedatha ngokuzumayo.

  • Ukuqapha kwabantu : Gcina abantu benolwazi ngezinqumo ezinzima. Akukho maqhawe, inhlanzeko nje kuphela.

Imibuzo evame ukubuzwa owawunayo ngasese 🙋

Ingabe inethiwekhi yemizwa empeleni iwubuchopho?

Kuphefumulelwe ubuchopho, yebo - kodwa kwenziwe lula. Ama-neurons kumanethiwekhi ayimisebenzi yezibalo; ama-neurons ebhayoloji angamaseli aphilayo ane-dynamics eyinkimbinkimbi. Ama-vibes afanayo, i-physics ehlukene kakhulu [1].

Zingaki izendlalelo engizidingayo?

Qala kancane. Uma ungenazo izimbobo, engeza ububanzi noma ukujula. Uma ungenazo izimbobo eziningi, hlela kabusha noma wehlise umthamo. Ayikho inombolo yomlingo; kukhona nje ama-curve okuqinisekisa kanye nokubekezela [1].

Ingabe ngihlala ngiyidinga i-GPU?

Akunjalo ngaso sonke isikhathi. Amamodeli amancane anedatha encane angaqeqeshwa kuma-CPU, kodwa ezithombeni, kumamodeli amakhulu ombhalo, noma kumasethi amakhulu edatha, ama-accelerator agcina isikhathi esiningi [1].

Kungani abantu bethi ukunaka kunamandla?

Ngoba ukunaka kuvumela amamodeli ukuthi agxile ezingxenyeni ezifanele kakhulu zokufaka ngaphandle kokuhamba ngokulandelana okuqinile. Kubamba ubudlelwano bomhlaba wonke, okuyinto enkulu emisebenzini yolimi kanye neyokusebenzisa izindlela eziningi [3].

Ingabe "Kuyini iNethiwekhi Yezinzwa ku-AI?" kuhlukile "kuyini ukufunda okujulile"?

Ukufunda okujulile kuyindlela ebanzi esebenzisa amanethiwekhi ajulile e-neural. Ngakho-ke ukubuza ukuthi Iyini i-Neural Network ku-AI? kufana nokubuza ngomlingiswa oyinhloko; ukufunda okujulile kuyifilimu yonke [1].

Amathiphu awusizo, anemibono emincane 💡

  • Khetha ama-baseline alula kuqala. Ngisho ne-perceptron encane enezingqimba eziningi ingakutshela ukuthi idatha ingafundwa yini.

  • Gcina ipayipi lakho ledatha liphinde likhiqizwe . Uma ungakwazi ukuliphinda ulisebenzise, ​​awukwazi ukulithemba.

  • Izinga lokufunda libaluleke kakhulu kunalokho okucabangayo. Zama ishejuli. Ukuzifudumeza kungasiza.

  • ukuhwebelana ngosayizi weqembu . Amaqoqo amakhulu aqinisa ama-gradient kodwa angase ahluke kakhulu.

  • Uma kudideka, amajika okulahlekelwa kanye nezindinganiso zesisindo . Ungamangala ukuthi impendulo ivame kangakanani kumabali.

  • Bhala phansi izibikezelo. Ikusasa - ukhohlwa izinto - ngokushesha [1, 2].

Ukuchezuka okujulile: indima yedatha, noma ukuthi kungani udoti ungena ngaphakathi usasho ukuthi udoti awusekho 🗑️➡️✨

Amanethiwekhi e-neural awalungisi ngomlingo idatha enephutha. Amalebula aphikisiwe, amaphutha okuchaza, noma ukusampula okuncane konke kuzozwakala kumodeli. Hlela, uhlole, futhi ukhulise. Futhi uma ungaqiniseki ukuthi udinga idatha eyengeziwe noma imodeli engcono, impendulo ivame ukuba lula ngendlela ecasulayo: kokubili - kodwa qala ngekhwalithi yedatha [1].

"Iyini iNethiwekhi Yezinzwa ku-AI?" - izincazelo ezimfushane ongazisebenzisa kabusha 🧾

  • Inethiwekhi ye-neural iyi-approximator yomsebenzi enezingqimba efunda amaphethini ayinkimbinkimbi ngokulungisa izisindo kusetshenziswa izimpawu ze-gradient [1, 2].

  • Luhlelo oluguqula okokufaka kube yimiphumela ngezinyathelo ezilandelanayo ezingezona eziqondile, oluqeqeshwe ukunciphisa ukulahleka [1].

  • Kuyindlela yokumodela eguquguqukayo, edinga idatha echumayo kokufakwayo okungahlelekile njengezithombe, umbhalo, nomsindo [1, 2, 3].

Kude Kakhulu, Angifundanga kanye namazwi okugcina 🎯

Uma othile ekubuza ukuthi Iyini iNeural Network ku-AI? nansi i-sound bite: inethiwekhi ye-neural iyinqwaba yamayunithi alula aguqula idatha isinyathelo ngesinyathelo, afunda ukuguqulwa ngokunciphisa ukulahleka nokulandela ama-gradient. Anamandla ngoba ayakala, afunda izici ngokuzenzakalelayo, futhi angamela imisebenzi eyinkimbinkimbi kakhulu [1, 4]. Ayingozi uma ungayinaki ikhwalithi yedatha, ukubusa, noma ukuqapha [5]. Futhi awawona umlingo. Izibalo nje, ukubala, kanye nobunjiniyela obuhle - ngokunambitha okuncane.

Ukufunda okwengeziwe, okukhethwe ngokucophelela (okungeziwe okungacashuniwe)

Izinkomba

[1] Goodfellow, I., Bengio, Y., & Courville, A. Deep Learning . I-MIT Press. Inguqulo yamahhala eku-inthanethi: funda kabanzi

[2] I-Stanford CS231n. Amanethiwekhi E-Convolutional Neural for Visual Recognition (amanothi ezifundo): funda kabanzi

[3] Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Ukunaka Yikho Konke Okudingayo . NeurIPS. arXiv: funda kabanzi

[4] Cybenko, G. (1989). Ukuqagela ngokufaka izinto eziningi esikhundleni somsebenzi we-sigmoidal . Izibalo Zokulawula, Izimpawu Nezinhlelo , 2, 303–314. I-Springer: funda kabanzi

[5] I-NIST. Uhlaka Lokuphathwa Kwengozi ye-AI (i-AI RMF) : funda kabanzi

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI

Buyela kubhulogi