Urusobe rw'imitsi mu buhanga bw'ubwenge (AI) ni iki?

Urusobe rw'imitsi mu buhanga bw'ubwenge (AI) ni iki?

Imiyoboro y'imitsi isa n'amayobera kugeza ubwo itabikora. Niba wigeze wibaza icyo ikoranabuhanga rya Neural Network muri AI ari cyo? kandi niba ari imibare gusa hamwe n'ingofero nziza, uri ahantu heza. Tuzayikomeza mu buryo bw'ingirakamaro, duce duto duto, kandi yego - emoji nke. Uzasiga uzi izi sisitemu, impamvu zikora, aho zinanirwa, n'uburyo bwo kuzivugaho udakoresheje ukuboko.

Ingingo ushobora gukunda gusoma nyuma y'iyi:

🔗 Ubwiyongere bw'ubukorano (AI) ni iki?
Gusobanukirwa uruhande rubogamye muri sisitemu za AI n'ingamba zo kwemeza ko ubutabera bungana.

🔗 Ni iki gisobanura ko umuntu akoresha ubwenge bwo gutekereza (AI) mu gihe cy'ibiza?
Uburyo AI iteganya guhanura ikoresha imiterere mu guhanura ibisubizo by'ejo hazaza.

🔗 Umutoza w'ubuhanga bwo gukora imibonano mpuzabitsina (AI) ni iki?
Gusuzuma uruhare n'inshingano z'abahanga batoza ubuhanga bwo gukora imibonano mpuzabitsina hakoreshejwe ubuhanga bwa “AI”.

🔗 Kureba kwa mudasobwa muri AI ni iki?
Uburyo AI isobanura kandi igasesengura amakuru aboneka binyuze mu kureba kuri mudasobwa.

Urusobe rw'imitsi mu buhanga bw'ubwenge (AI) ni iki? Igisubizo cy'amasegonda 10 ⏱️

Urusobe rw'imitsi ni urusobe rw'ibice byoroshye byo kubara byitwa neurone zishyira imibare imbere, zigahindura imbaraga zazo mu gihe cy'imyitozo, kandi ziga buhoro buhoro imiterere y'amakuru. Iyo wumvise ubumenyi bwimbitse , ibyo akenshi bivuze urusobe rw'imitsi rufite ibice byinshi byashyizwe hamwe, imikorere yo kwiga mu buryo bwikora aho kubishyira mu nyandiko n'intoki. Mu yandi magambo: uduce twinshi tw'imibare, dutondetse neza, twatojwe ku makuru kugeza igihe bigiriye akamaro [1].

Ni iki gituma umuyoboro w'imitsi (neural network) ugira akamaro? ✅

  • Imbaraga zo guhagararira : Hamwe n'imiterere n'ingano bikwiye, imiyoboro ishobora kugera ku mikorere igoye cyane (reba Inyigisho ya Universal Approximation Theorem) [4].

  • Kwiga kuva ku iherezo kugeza ku iherezo : Aho kugira ngo habeho imikorere y'intoki, icyitegererezo kirabivumbura [1].

  • Guhuza : Umuyoboro ugenzurwa neza ntufata mu mutwe gusa - ukora ku makuru mashya kandi atagaragara [1].

  • Uburyo bwo kwagura amakuru : Amakuru manini hamwe n'amakuru manini akenshi atuma umusaruro urushaho kuba mwiza… kugeza ku nzitizi zifatika nko kubara no kunoza amakuru [1].

  • Kwimura : Ibintu byigwa mu murimo umwe bishobora gufasha undi (kwimura imyigire no gutunganya neza) [1].

Icyitonderwa gito (urugero rw'ibyabaye): Itsinda rito rishinzwe gushyira mu byiciro ibicuruzwa rihinduranya ibikoresho byubatswe n'intoki kuri CNN nto, rikongeraho uburyo bworoshye bwo kongeramo amakuru (flips/crops), kandi rikareba igabanuka ry'amakosa yo kwemeza - atari ukubera ko umuyoboro ari "magic", ahubwo ni ukubera ko ryize ibintu by'ingirakamaro byinshi kuri pixels.

"Urusobe rw'imitsi mu buhanga bw'ubwenge ni iki?" mu Cyongereza gisanzwe, hamwe n'ikigereranyo cy'iffy 🍞

Tekereza umurongo w’imigati. Ibikoresho bishyirwamo, abakozi bahindura resept, abapima uburyohe bakinubira, maze itsinda rikongera kuvugurura resept. Mu muyoboro, ibintu byinjira binyura mu byiciro, imikorere y’igihombo ishyira amanota ku musaruro, kandi imiterekero itera uburemere gukora neza ubutaha. Ntabwo ari ikigereranyo cyiza - umugati ntushobora gutandukana - ariko urafatana [1].

Imiterere y'imiyoboro y'ubwonko 🧩

  • Uturemangingo : Utwuma duto two kubara dukoresha igipimo cy’ingano n’imikorere yo gukora.

  • Ibiremereye n'ibirangamimerere : Udupfundo dushobora guhindurwa tugaragaza uburyo ibimenyetso bihurira.

  • Ibyiciro : Ibyinjira byakira amakuru, ibyiciro byihishe birayahindura, ibyiciro bisohoka bigakora isuzuma.

  • Imikorere yo gukora : Impinduka zitari umurongo umwe nka ReLU, sigmoid, tanh, na softmax zituma kwiga byoroha.

  • Umumaro w'igihombo : Amanota y'uburyo iteganyagihe ritari ryo (interropy yo gushyira mu byiciro, MSE yo gusubira inyuma).

  • Optimizer : Algorithms nka SGD cyangwa Adam zikoresha gradients kugira ngo zivugurure uburemere.

  • Gutunganya : Uburyo bwo gusiga imyenda cyangwa kwangirika kw'ibiro kugira ngo wirinde ko imashini ishyirwa mu mwanya urenze urugero.

Niba ushaka uburyo bwo kuvura (ariko buracyasomwa), igitabo gifunguye cyitwa Deep Learning gikubiyemo urutonde rwuzuye: ishingiro ry'imibare, gutunganya, no gukwirakwiza [1].

Imikorere yo gukora, mu buryo bw'akanya gato ariko bufasha ⚡

  • ReLU : Zero ku bintu bibi, umurongo ugororotse ku byiza. Byoroshye, byihuse, kandi bigira ingaruka nziza.

  • Sigmoid : Ipima agaciro ka squashes hagati ya 0 na 1 - ni ingirakamaro ariko ishobora gukabya.

  • Tanh : Nk'ikimenyetso cya sigmoid ariko gihwanye hafi ya zeru.

  • Softmax : Ihindura amanota y’ibanze mo amahirwe mu byiciro bitandukanye.

Ntabwo ukeneye gufata mu mutwe buri shusho y'umugozi - menya gusa impinduka n'uburyo busanzwe bwo gukoresha [1, 2].

Uko kwiga bibaho mu by'ukuri: inyuma, ariko ntabwo biteye ubwoba 🔁

  1. Inzira yo gusohora amakuru: Amakuru agenda ku rundi ruhande kugira ngo habeho iteganyagihe.

  2. Igihombo cyo kubara : Gereranya ubuhanuzi n'ukuri.

  3. Gukwirakwiza inyuma : Kubara imivuduko y'igihombo ukurikije buri buremere ukoresheje itegeko ry'uruhererekane.

  4. Amakuru mashya : Optimizer ihindura ibiro gato.

  5. Subiramo : Ibihe byinshi. Icyitegererezo kigenda cyiga buhoro buhoro.

Kugira ngo ubone uburyo bwo gusobanukirwa ibintu hamwe n'amashusho n'ibisobanuro bijyanye n'inyandiko, reba inyandiko za kera za CS231n kuri backprop na optimization [2].

Imiryango ikomeye y'imiyoboro y'ubwonko, mu magambo make 🏡

  • Imiyoboro y'amakuru (MLPs) : Ubwoko bworoshye cyane. Amakuru arushaho gutera imbere.

  • Imiyoboro y'imitsi ya Convolutional Neural Networks (CNNs) : Ni nziza ku mashusho bitewe n'amashusho ayungurura imiterere y'ahantu areba impande, imiterere, imiterere [2].

  • Imiyoboro y'imitsi isubiramo (RNNs) n'ibindi bitandukanye : Byubakiwe urukurikirane nk'inyandiko cyangwa urukurikirane rw'igihe binyuze mu kubungabunga uburyo ibintu bigenda neza [1].

  • Transformers : Shyira imbaraga mu mibanire y’abantu mu buryo bw’icyitegererezo mu buryo bukurikiranye icyarimwe; iganza mu rurimi no hanze yarwo [3].

  • Imiyoboro y'imitsi ya Graph (GNNs) : Ikora ku ngingo no ku nkengero z'imbonerahamwe - ingirakamaro ku matsinda, ku mbuga nkoranyambaga, inama [1].

  • Autoencoders na VAEs : Menya uburyo bwo gushushanya butonze kandi utange impinduka [1].

  • Imiterere ibyara : Kuva kuri GANs kugeza kuri moderi zo gukwirakwiza, zikoreshwa mu mashusho, amajwi, ndetse no kuri kode [1].

Inyandiko za CS231n ni nziza cyane kuri CNN, mu gihe inyandiko ya Transformer ari yo soko nyamukuru y’uburyo bwo kwita ku bantu [2, 3].

Imbonerahamwe yo kugereranya: ubwoko busanzwe bw'imiyoboro y'ubwonko, abo ishingiyeho, ibiciro, n'impamvu ikora 📊

Igikoresho / Ubwoko Abareba Igiciro gikwiye Impamvu bikora
Ibisubizo (MLP) Abatangizi, abasesenguzi Hagati cyane Imiterere y'ibanze yoroshye, ihindagurika kandi ikwiye
CNN Amakipe y'icyerekezo Hagati Imiterere yo mu gace + gusangira ibipimo
RNN / LSTM / GRU Abantu bakurikiranye Hagati Kwibuka by'igihe gito… bifata gahunda
Transformateur NLP, uburyo bwinshi bwo gukoresha Hagati-hejuru Icyitonderwa cyibanda ku mibanire ijyanye n'ibyo
GNN Abahanga mu bya siyansi, abashinzwe ikoranabuhanga Hagati Gutanga ubutumwa ku mbonerahamwe bigaragaza imiterere
Autoencoder / VAE Abashakashatsi Hagati cyane Yiga uburyo bwo kwerekana ibintu mu buryo bufunze
GAN / Gukwirakwiza Laboratwari z'ubuhanzi Hagati-hejuru Ubumaji burwanya cyangwa busubiramo bugaragaza urusaku

Icyitonderwa: ibiciro bireba imibare n'igihe; uburebure bw'urugendo rwawe buratandukanye. Imodoka imwe cyangwa ebyiri zivuga nkana ku bushake.

"Urusobe rw'imitsi muri AI ni iki?" ugereranije na algorithms za ML za kera ⚖️

  • Ubwubatsi bw'imikorere : ML ya kera ikunze kwishingikiriza ku mikorere ikorwa n'intoki. Neti z'imitsi ziga imikorere mu buryo bwikora - intsinzi ikomeye ku makuru akomeye [1].

  • Gushaka amakuru : Imiyoboro ikunze kugaragara iyo amakuru menshi; amakuru make ashobora gukunda moderi zoroshye [1].

  • Kubara : Imiyoboro ikunda accelerators nka GPUs [1].

  • Igipimo cy'imikorere : Ku makuru adakozwe neza (amashusho, amajwi, inyandiko), imiyoboro miremire ikunze kwiganza [1, 2].

Imikorere y'amahugurwa ikora mu buryo bw'imyitozo 🛠️

  1. Sobanura intego : Gushyira mu byiciro, gusubira inyuma, gushyira mu byiciro, guteranya - guhitamo igihombo gihuye.

  2. Guhangana kw'amakuru : Gabanyijemo mo ibice bya gari ya moshi/igenzura/igerageza. Hindura imiterere. Kuringaniza amasomo. Ku mashusho, tekereza ku kongeramo amakuru nko kuzunguruka, gukata, urusaku ruto.

  3. Amahitamo y'inyubako : Tangira byoroshye. Ongeraho ubushobozi gusa igihe bikenewe.

  4. Urukurikirane rw'amahugurwa : Gushyira amakuru mu byiciro. Gutanga amakuru mbere. Kubara igihombo. Gusubiza inyuma. Kuvugurura. Kwandika ibipimo.

  5. Guhindura imikorere : Kureka ishuri, kwangirika kw'ibiro, guhagarika imyitozo hakiri kare.

  6. Isuzuma : Koresha seti yo kwemeza kuri hyperparameter. Tegura seti y'igerageza kugira ngo urebe neza.

  7. Kohereza witonze : Reba uko ibintu bigenda, reba niba hari aho bibogamiye, teganya ko ibintu bisubira inyuma.

Ku nyigisho zishingiye ku iherezo kugeza ku iherezo, zishingiye ku mibare y’amakuru, zikubiyemo inyigisho zifatika, igitabo cy’amasomo gifunguye n’inyandiko za CS231n ni imfashanyigisho zizewe [1, 2].

Gushyiramo ibintu byinshi, gukoresha ibintu byose hamwe, n'ibindi bikoresho by'ubwubatsi 👀

  • Gushyira ibintu mu buryo burenze urugero : Iyi moderi yibuka ibintu bisanzwe mu myitozo. Kosora ukoresheje amakuru menshi, gutunganya ibintu neza, cyangwa imiterere yoroshye.

  • Igipimo cyo hasi : Icyitegererezo cyoroshye cyane cyangwa imyitozo iragoye cyane. Ongera ubushobozi cyangwa imyitozo igihe kirekire.

  • Gusohora amakuru : Amakuru aturutse mu itsinda ry'ibizamini yinjira mu myitozo. Reba gatatu uko wagabanyijemo ibice.

  • Gupima nabi : Icyitegererezo cyizewe ariko kitari cyo ni akaga. Tekereza ku gupima cyangwa kugabanya uburemere butandukanye.

  • Ihindagurika ry'ikwirakwizwa : Ihindagurika ry'amakuru mu isi nyayo. Kugenzura no kuyahindura.

Ku bijyanye n'inyigisho zishingiye ku gukwirakwiza no kugena uko ibintu bimeze, shingira ku bipimo ngenderwaho [1, 2].

Umutekano, gusobanuka neza, no gushyira abantu mu bikorwa neza 🧭

Imiyoboro y'imitsi ishobora gufata ibyemezo bikomeye. Ntabwo bihagije ko ikora neza ku rubuga rw'ubuyobozi. Ukeneye imiyoborere, gupima, no kugabanya ingaruka mu buzima bwose. Urwego rwa NIST AI rwo gucunga ibyago rugaragaza imirimo ifatika - GUVERN, MAP, GUPIMA, GUCUNGA - kugira ngo bifashe amatsinda kwinjiza imicungire y'ibyago mu gushushanya no gushyiraho ibikorwa [5].

Gusunika gato vuba:

  • Igenzura ry'ubusumbane : Suzuma ibice by'abaturage aho bikwiye kandi byemewe n'amategeko.

  • Gusobanura : Koresha ubuhanga nk'ubuhanga cyangwa imiterere y'ibintu. Ntibitunganye, ariko bifite akamaro.

  • Gukurikirana : Shyiraho amatangazo ku birebana n'igabanuka ritunguranye ry'ibipimo cyangwa ihindagurika ry'amakuru.

  • Ubugenzuzi bw'abantu : Guma abantu mu murongo w'ibyemezo bikomeye. Nta butwari, isuku gusa.

Ibibazo bikunze kubazwa wagize mu ibanga 🙋

Ese urusobe rw'imitsi muri rusange ni ubwonko?

Yego, byahumetswe n'ubwonko - ariko byoroshye. Uturemangingo (neurons) mu miyoboro ni imikorere y'imibare; uturemangingo (neurons) tw'ibinyabuzima ni uturemangingo tuzima dufite imiterere igoye. Imiterere isa, fiziki itandukanye cyane [1].

Nkeneye urwego rungana iki?

Tangira gato. Niba urimo gushyiramo ibintu bike, ongeramo ubugari cyangwa ubujyakuzimu. Niba urimo gushyiramo ibintu byinshi, shyiramo ibintu mu buryo busanzwe cyangwa ugabanye ubushobozi. Nta mubare w'ubumaji uhari; hari uburyo bwo kwemeza ibintu gusa no kwihangana [1].

Ese buri gihe nkeneye GPU?

Si buri gihe. Modele nto ku makuru make zishobora kwigishwa kuri CPU, ariko ku mashusho, modele nini z'inyandiko, cyangwa seti nini z'amakuru, accelerators zigabanya igihe kinini [1].

Kuki abantu bavuga ko kwita ku bandi ari ibintu bikomeye?

Kubera ko kwita ku bintu bituma abanyamideli bibanda ku bice by'ingenzi by'inyongera batitaye ku murongo ukwiye. Bifata imibanire mpuzamahanga, ikaba ari ikintu gikomeye ku ndimi no ku mirimo myinshi [3].

Ese "Umuyoboro w'imitsi muri AI ni iki?" bitandukanye na "kwiga byimbitse ni iki"?

Kwiga mu buryo bwimbitse ni uburyo bwagutse bukoresha imiyoboro y'ubwonko yimbitse. Rero kubaza uti "Neural Network" muri "AI" ni iki? ni nko kubaza umukinnyi w'ingenzi; kwiga mu buryo bwimbitse ni filime yose [1].

Inama z'ingirakamaro, zifite ibitekerezo bicye 💡

  • uhitemo imirongo yoroheje . Ndetse n'agace gato k'inyuguti nyinshi gashobora kukubwira niba amakuru ashobora kwigwa.

  • Komeza ko umurongo wawe w'amakuru ushobora kongera gukoreshwa . Niba udashobora kongera kuwukoresha, ntushobora kuwizera.

  • Igipimo cy'imyigire ni ingenzi kurusha uko ubitekereza. Gerageza gahunda. Gushyushya bishobora gufasha.

  • ihindagurika ry'ingano y'amatsinda . Amatsinda manini atuma imiterere y'amatsinda ikomera ariko ashobora kuba atandukanye.

  • Iyo urimo urujijo, igabanuka ry'ibice by'ibishushanyo n'ibipimo by'uburemere birahinduka . Watangazwa n'uburyo igisubizo kiba kenshi mu bice by'ibishushanyo.

  • Andika ibitekerezo byawe. Ejo hazaza - wibagirwa ibintu - vuba [1, 2].

Gucengera mu mazi abira: uruhare rw'amakuru, cyangwa impamvu imyanda ikiri mu mazi isobanura imyanda hanze 🗑️➡️✨

Imiyoboro y'imitsi ntabwo ikosora amakuru afite inenge mu buryo bw'amayobera. Ibirango byangiritse, amakosa yo kwandika, cyangwa ingero ntoya byose bizagaragara muri iyo moderi. Kora urutonde, genzura, kandi wongere. Kandi niba utazi neza niba ukeneye amakuru menshi cyangwa moderi nziza, igisubizo akenshi kiba cyoroshye cyane: byombi - ariko tangira ukoresheje ireme ry'amakuru [1].

"Umuyoboro w'imitsi ni iki muri AI?" - ibisobanuro bigufi ushobora kongera gukoresha 🧾

  • Urusobe rw'imitsi ni uburyo bwo gusuzuma imikorere y'ingingo (layered function approximator) bwiga imiterere igoye binyuze mu guhindura uburemere hakoreshejwe ibimenyetso bya gradient [1, 2].

  • Ni sisitemu ihindura ibikubiye mu bisohoka binyuze mu ntambwe zikurikirana zitari umurongo, yatojwe kugira ngo igabanye igihombo [1].

  • Ni uburyo bworoshye bwo gukoresha icyitegererezo, bukoresha amakuru menshi, bukoreshwa mu gushyira mu bikorwa ibintu bitari byitezwe nk'amashusho, inyandiko, n'amajwi [1, 2, 3].

Igihe kirekire cyane, sinasomye n'amagambo ya nyuma 🎯

Iyo umuntu akubajije ati: "Neural Network" muri AI ni iki? Dore uko amajwi ateye: "neural network" ni uruhurirane rw'ibice byoroshye bihindura amakuru intambwe ku yindi, biga impinduka mu kugabanya igihombo no gukurikiza gradients. Bifite imbaraga kuko bipima, biga imikorere mu buryo bwikora, kandi bishobora kugaragaza imikorere igoye cyane [1, 4]. Bitera ibyago iyo wirengagije ireme ry'amakuru, imiyoborere, cyangwa igenzura [5]. Kandi si ubumaji. Imibare gusa, kubara, n'ubuhanga bwiza - hamwe n'uburyohe buto.

Gusoma byinshi, byatoranijwe neza (inyongera zitari iz'incamake)

Amareferensi

[1] Goodfellow, I., Bengio, Y., & Courville, A. Deep Learning . MIT Press. Verisiyo y'ubuntu kuri interineti: soma byinshi

[2] Stanford CS231n. Imiyoboro y'imitsi ya Convolutional Neural yo Kumenya Ibigaragara (inyandiko z'amasomo): soma byinshi

[3] Vaswani, A., Shazeer, N., Parmar, N., n'abandi (2017). Kwita ku bandi ni byo ukeneye gusa . NeurIPS. arXiv: soma byinshi

[4] Cybenko, G. (1989). Isesengura rishingiye ku miterere y'imikorere ya sigmoidal . Imibare yo kugenzura, ibimenyetso na sisitemu , 2, 303–314. Springer: soma byinshi

[5] NIST. Urwego rw'Imicungire y'Ibibazo bya AI (AI RMF) : soma byinshi

Shaka ubuhanga bwa AI bugezweho mu iduka ryemewe rya AI Assistant Store

Ku bijyanye natwe

Garuka kuri blog