Kwa zaka zambiri, kuphunzira mozama kwakhala kukupanga mitu muukadaulo. Ndipo, n'zosavuta kumvetsa chifukwa chake.
Nthambi iyi ya intelligence intelligence ikusintha magawo kuyambira chisamaliro chaumoyo mpaka kubanki kupita kumayendedwe, zomwe zimathandizira kupita patsogolo komwe sikunali koyenera.
Kuphunzira mwakuya kumapangidwa ndi ma aligorivimu apamwamba kwambiri omwe amaphunzira kuchotsa ndi kulosera zapakatikati pazambiri zambiri.
Tiwona njira 15 zabwino kwambiri zophunzirira mozama mu positiyi, kuchokera ku Convolutional Neural Networks kupita ku Generative Adversarial Networks mpaka maukonde a Memory Yanthawi Yaifupi.
Positi iyi ikupatsani chidziwitso chofunikira ngati ndinu a woyamba kapena katswiri pa kuphunzira mozama.
1. Transformer Networks
Ma network a Transformer asintha masomphenya a makompyuta ndi kugwiritsa ntchito zilankhulo zachilengedwe (NLP). Amasanthula deta yomwe ikubwera ndikugwiritsa ntchito njira zowunikira kuti agwire maubwenzi atalitali. Izi zimawapangitsa kukhala ofulumira kuposa ma motere otsatizana.
Ma network a Transformer adafotokozedwa koyamba m'buku lakuti "Kusamala Ndi Zonse Zomwe Mukufunikira" ndi Vaswani et al.
Amakhala ndi encoder ndi decoder (2017). Mtundu wa transformer wawonetsa magwiridwe antchito osiyanasiyana a NLP, kuphatikiza kusanthula malingaliro, kugawa malemba, ndi kumasulira kwa makina.
Zitsanzo zokhazikitsidwa ndi ma Transformer zitha kugwiritsidwanso ntchito pakuwona makompyuta pamapulogalamu. Amatha kuzindikiritsa chinthu ndikulemba mawu.
2. Ma Netiweki a Memory akanthawi yayitali (LSTMs)
Ma Networks Memory Memory Networks (LSTMs) ndi mtundu wa neural network makamaka yomangidwa kuti igwire zolowetsa motsatizana. Amatchedwa "kanthawi kochepa" chifukwa amatha kukumbukira chidziwitso cha nthawi yayitali komanso kuiwala zambiri zosafunikira.
Ma LSTM amagwira ntchito kudzera mu "zipata" zina zomwe zimayendetsa kayendedwe ka mauthenga mkati mwa netiweki. Kutengera ngati chidziwitsocho chikuwoneka ngati chofunikira kapena ayi, zipatazi zitha kuloleza kapena kuziletsa.
Njirayi imathandizira ma LSTM kukumbukira kapena kuiwala zambiri zanthawi zakale, zomwe ndizofunikira kwambiri pantchito monga kuzindikira mawu, kukonza zilankhulo zachilengedwe, komanso kulosera zanthawi.
Ma LSTM ndi opindulitsa kwambiri mulimonse momwe mulili ndi data yotsatizana yomwe imayenera kuwunikiridwa kapena kuneneratu. Nthawi zambiri amagwiritsidwa ntchito mu pulogalamu yozindikira mawu kuti asinthe mawu olankhulidwa kukhala mawu, kapena mkati malonda akhalire kusanthula kulosera mitengo yamtsogolo motengera zomwe zidachitika kale.
3. Mapu Odzipanga okha (SOMs)
Ma SOM ndi ochita kupanga neural network yomwe ingaphunzire ndikuyimira deta yovuta m'malo otsika kwambiri. Njirayi imagwira ntchito posintha zolowetsa zapamwamba kwambiri kukhala gulu la magawo awiri, ndipo unit iliyonse kapena neuroni ikuyimira gawo losiyana la malo olowetsamo.
Ma neurons amalumikizidwa palimodzi ndikupanga mawonekedwe a topological, kuwalola kuti aphunzire ndikusintha malinga ndi zomwe alowetsa. Chifukwa chake, SOM idakhazikitsidwa pamaphunziro osayang'aniridwa.
Algorithm sikufunika deta yolembedwa kuphunzira kuchokera. M'malo mwake, imagwiritsa ntchito ziwerengero za data yolowera kuti ipeze mawonekedwe ndi kulumikizana pakati pa zosinthazo.
Panthawi yophunzitsira, ma neuron amapikisana kuti akhale chisonyezero chabwino kwambiri cha zomwe zalowetsedwa. Ndipo, amadzipangira okha kukhala omveka bwino. Ma SOM ali ndi ntchito zosiyanasiyana, kuphatikizapo kuzindikira kwa zithunzi ndi kulankhula, migodi ya deta, ndi kuzindikira mawonekedwe.
Iwo ndi zothandiza kuwona deta yovuta, kusonkhanitsa malo okhudzana ndi deta, ndi kuzindikira zolakwika kapena zakunja.
4. Kulimbikitsa Kwambiri Kuphunzira
kwambiri Maphunziro Owonjezera ndi mtundu wa kuphunzira pamakina komwe wothandizira amaphunzitsidwa kupanga zisankho potengera mphotho. Imagwira ntchito polola wothandizira kuti azitha kulumikizana ndi zozungulira zake ndikuphunzira kudzera m'mayesero ndi zolakwika.
Wothandizira amalipidwa pa chilichonse chomwe amachita, ndipo cholinga chake ndikuphunzira momwe angakwaniritsire zabwino zake pakapita nthawi. Izi zitha kugwiritsidwa ntchito pophunzitsa othandizira kusewera masewera, kuyendetsa magalimoto, komanso kuyang'anira maloboti.
Q-Learning ndi njira yodziwika bwino ya Deep Reinforcement Learning. Imagwira ntchito poyesa kufunikira kochita chinthu china m'boma linalake ndikukonzanso kuyerekezera komwe wothandizila amagwirizana ndi chilengedwe.
Wothandizirayo amagwiritsa ntchito kuyerekezera uku kuti adziwe kuti ndi chiyani chomwe chingabweretse mphotho yayikulu kwambiri. Q-Learning yagwiritsidwa ntchito pophunzitsa othandizira kusewera masewera a Atari, komanso kupititsa patsogolo kugwiritsa ntchito mphamvu m'malo opangira deta.
Deep Q-Networks ndi njira ina yotchuka ya Deep Reinforcement Learning (DQN). Ma DQN ndi ofanana ndi Q-Learning chifukwa amayerekezera zochita pogwiritsa ntchito neural network yakuya osati tebulo.
Izi zimawathandiza kuthana ndi zosintha zazikulu, zovuta ndi zina zambiri. Ma DQN akhala akugwiritsidwa ntchito pophunzitsa othandizira kusewera masewera monga Go ndi Dota 2, komanso kupanga ma robot omwe angaphunzire kuyenda.
5. Ma Neural Networks Okhazikika (RNNs)
Ma RNN ndi mtundu wa neural network womwe umatha kukonza zotsatizana ndikusunga zamkati. Lingalireni mofanana ndi munthu amene akuŵerenga bukhu, pamene liwu lililonse limagayidwa mogwirizana ndi limene linafikapo lisanafike.
Chifukwa chake ma RNN ndi abwino pantchito ngati kuzindikira mawu, kumasulira zilankhulo, komanso kulosera mawu otsatira m'mawu.
Ma RNN amagwira ntchito pogwiritsa ntchito malupu obwereza kuti alumikizane ndi zomwe zatuluka nthawi iliyonse ndikubwereranso ku sitepe yotsatira. Izi zimathandizira netiweki kugwiritsa ntchito zidziwitso zanthawi yayitali kuti zidziwitse zolosera zamtsogolo. Tsoka ilo, izi zikutanthawuzanso kuti ma RNN ali pachiwopsezo cha vuto lomwe likusokonekera, momwe ma gradients omwe amagwiritsidwa ntchito pophunzitsa amakhala ang'ono kwambiri ndipo netiweki imavutikira kuphunzira maubwenzi anthawi yayitali.
Ngakhale zili zovuta, ma RNN apeza ntchito m'mapulogalamu osiyanasiyana. Mapulogalamuwa akuphatikiza kukonza zilankhulo zachilengedwe, kuzindikira mawu, komanso kupanga nyimbo.
mtambasulira wa Google, mwachitsanzo, amagwiritsa ntchito makina a RNN kuti amasulire zilankhulo zonse, pamene Siri, wothandizira, amagwiritsa ntchito makina a RNN kuti azindikire mawu. Ma RNN adagwiritsidwanso ntchito kulosera mitengo yamasheya ndikupanga zolemba zenizeni ndi zithunzi.
6. Makapisozi Networks
Capsule Networks ndi mtundu watsopano wamapangidwe a neural network omwe amatha kuzindikira mawonekedwe ndi kulumikizana kwa data moyenera. Amapanga ma neuroni kukhala "makapisozi" omwe amasunga mbali zina za cholowa.
Mwanjira imeneyi amatha kulosera zolondola kwambiri. Ma Capsule Networks amachotsa zinthu zovuta pang'onopang'ono kuchokera pazolowera pogwiritsa ntchito zigawo zingapo za makapisozi.
Njira ya Capsule Networks imawapangitsa kuti aphunzire zoyimira zazomwe zaperekedwa. Amatha kuyika bwino kulumikizana kwa malo pakati pa zinthu zomwe zili mkati mwa chithunzi polumikizana pakati pa makapisozi.
Chizindikiritso cha chinthu, magawo azithunzi, ndikusintha zilankhulo zachilengedwe zonse ndikugwiritsa ntchito Capsule Networks.
Ma Capsule Networks ali ndi mwayi wogwiritsidwa ntchito kuyendetsa pawokha matekinoloje. Amathandizira dongosololi kuzindikira ndi kusiyanitsa pakati pa zinthu monga magalimoto, anthu, ndi zikwangwani zamagalimoto. Makinawa amatha kupewa kugundana polosera molondola za momwe zinthu zilili m'malo awo.
7. Variational Autoencoder (VAEs)
VAEs ndi chida chophunzirira mwakuya chomwe chimagwiritsidwa ntchito pophunzira mosayang'aniridwa. Poyika ma encode m'malo ocheperako ndikuyisinthanso kukhala momwe idayambira, angaphunzire kuwona mawonekedwe mu data.
Iwo ali ngati wamatsenga amene angasinthe kalulu kukhala chipewa kenako n’kukhala kalulu! Ma VAE ndi opindulitsa pakupanga zowonera kapena nyimbo zenizeni. Ndipo, angagwiritsidwe ntchito kupanga deta yatsopano yomwe ikufanana ndi deta yoyambirira.
VAEs ndi ofanana ndi codebreaker chinsinsi. Amatha kuzindikira maziko ake kapangidwe ka data pochiphwanya m’zidutswa zing’onozing’ono, mofanana ndi mmene chithunzithunzi chimagaŵira. Atha kugwiritsa ntchito chidziwitsocho kupanga zatsopano zomwe zimawoneka ngati zoyambirira atazikonza.
Izi zitha kukhala zothandiza popanikiza mafayilo akulu kapena kupanga zithunzi zatsopano kapena nyimbo mwanjira inayake. Ma VAE amathanso kupanga zatsopano, monga nkhani kapena nyimbo.
8. Generative Adversarial Networks (GANs)
Ma GAN (Generative Adversarial Networks) ndi njira yophunzirira mwakuya yomwe imapanga deta yatsopano yofanana ndi yoyambirira. Amagwira ntchito pophunzitsa maukonde awiri: jenereta ndi netiweki ya tsankho.
Jenereta imapanga deta yatsopano yomwe ikufanana ndi yoyamba.
Ndipo, wosankhana amayesa kusiyanitsa pakati pa deta yoyambirira ndi yopangidwa. Maukonde awiriwa amaphunzitsidwa motsatira, jenereta ikuyesera kunyenga wosankhana komanso wosankhana kuyesa kuzindikira bwino deta yoyambirira.
Ganizirani ma GAN ngati mkangano pakati pa wabodza ndi wofufuza. Jeneretayo imagwira ntchito mofanana ndi yopeka, imapanga zojambula zatsopano zomwe zimafanana ndi zoyambirira.
Wosankhana amakhala ngati wapolisi wofufuza, kuyesa kusiyanitsa pakati pa zojambulajambula zenizeni ndi zabodza. Maukonde awiriwa amaphunzitsidwa motsatana, pomwe jenereta imachita bwino pakupanga zabodza zomveka komanso kusankhana bwino pakuzindikira.
Ma GAN ali ndi ntchito zingapo, kuyambira kupanga zithunzi zenizeni za anthu kapena nyama mpaka kupanga nyimbo zatsopano kapena zolemba. Atha kugwiritsidwanso ntchito powonjezera deta, zomwe zimaphatikizapo kuphatikiza deta yopangidwa ndi deta yeniyeni kuti apange gulu lalikulu la data yophunzitsira mitundu yophunzirira makina.
9. Deep Q-Networks (DQNs)
Deep Q-Networks (DQNs) ndi mtundu wa zisankho zolimbikitsira kuphunzira. Amagwira ntchito pophunzira ntchito ya Q yomwe imaneneratu mphotho yomwe ikuyembekezeka pochita chinthu china mumkhalidwe wina.
Q-function imaphunzitsidwa ndikuyesa ndi zolakwika, ndi algorithm kuyesa zochita zosiyanasiyana ndikuphunzira kuchokera pazotsatira zake.
Lingalirani ngati a masewera kanema woyeserera ndi zochita zosiyanasiyana ndikupeza zomwe zimatsogolera kuchita bwino! Ma DQN amaphunzitsa ma Q-function pogwiritsa ntchito neural network yakuya, kuwapanga kukhala zida zogwira ntchito popanga zisankho zovuta.
Iwo agonjetsa ngakhale akatswiri aumunthu pamasewera monga Go ndi chess, komanso mu robotics ndi magalimoto odziyendetsa okha. Chifukwa chake, zonse, ma DQN amagwira ntchito pophunzira kuchokera pazomwe adakumana nazo kuti apititse patsogolo luso lawo lopanga zisankho pakapita nthawi.
10. Ma Radial Basis Function Networks (RBFNs)
Ma Radial Basis Function Networks (RBFNs) ndi mtundu wa neural network womwe umagwiritsidwa ntchito kuyerekeza ntchito ndikuchita ntchito zamagulu. Amagwira ntchito posintha zolowetsazo kukhala malo apamwamba kwambiri pogwiritsa ntchito kusonkhanitsa kwa ma radial maziko.
Kutulutsa kwa netiweki ndikophatikizika kofananira kwa ntchito zoyambira, ndipo ntchito iliyonse yama radial maziko imayimira malo apakati pamalo olowera.
Ma RBFN ndi othandiza makamaka pazochitika zomwe zimakhala zovuta zolowera-zotulutsa, ndipo amatha kuphunzitsidwa pogwiritsa ntchito njira zosiyanasiyana, kuphatikizapo kuphunzira koyang'aniridwa ndi osayang'aniridwa. Zakhala zikugwiritsidwa ntchito pa chilichonse kuyambira zoneneratu zachuma kupita ku chithunzi ndi kuzindikira kwamawu mpaka kuwunika kwachipatala.
Ganizirani ma RBFN ngati njira ya GPS yomwe imagwiritsa ntchito ma nangula angapo kuti ipeze njira yodutsa malo ovuta. Kutulutsa kwa maukonde ndi kuphatikiza kwa nangula, zomwe zimayimira ntchito za radial maziko.
Titha kuyang'ana pazidziwitso zovuta ndikupanga kulosera molondola za momwe zinthu zidzakhalire pogwiritsa ntchito ma RBFN.
11. Multilayer Perceptrons (MLPs)
Mtundu wodziwika bwino wa neural network wotchedwa multilayer perceptron (MLP) umagwiritsidwa ntchito poyang'anira ntchito zophunzirira monga kugawa ndi kuyambiranso. Amagwira ntchito posanjikiza zigawo zingapo za ma node olumikizidwa, kapena ma neurons, ndi gawo lililonse losintha mosagwirizana ndi zomwe zikubwera.
Mu MLP, neuron iliyonse imalowetsa kuchokera ku ma neuroni omwe ali pansipa ndikutumiza chizindikiro kumanyuroni omwe ali pamwamba. Kutulutsa kulikonse kwa neuron kumatsimikiziridwa pogwiritsa ntchito ntchito yoyambitsa, yomwe imapangitsa kuti netiweki ikhale yosagwirizana.
Amatha kuphunzira zoyimira zapamwamba zazomwe alowetsa chifukwa amatha kukhala ndi zigawo zingapo zobisika.
Ma MLP agwiritsidwa ntchito pazinthu zosiyanasiyana, monga kusanthula malingaliro, kuzindikira zachinyengo, kuzindikira mawu ndi zithunzi. Ma MLP angayerekezedwe ndi gulu la ofufuza omwe akugwira ntchito limodzi kuti awononge mlandu wovuta.
Pamodzi, atha kuphatikiza zowona ndikuthetsa upandu ngakhale kuti aliyense ali ndi gawo linalake lapadera.
12. Convolutional Neural Networks (CNNs)
Zithunzi ndi makanema amakonzedwa pogwiritsa ntchito ma convolutional neural network (CNNs), mawonekedwe a neural network. Amagwira ntchito pogwiritsa ntchito zosefera zomwe tingaphunzire, kapena maso, kuti atenge mawonekedwe ofunikira kuchokera muzolowera.
Zosefera zimadutsa pachithunzichi, ndikuchita ma convolutions kuti apange mapu omwe amajambula zofunikira pachithunzichi.
Popeza ma CNN amatha kuphunzira momwe amayimira mawonekedwe azithunzi, ndizothandiza makamaka pazochitika zomwe zikukhudza kuchuluka kwazithunzi. Mapulogalamu angapo azigwiritsa ntchito, monga kuzindikira zinthu, kugawa zithunzi, komanso kuzindikira nkhope.
Ganizirani za CNN ngati wojambula yemwe amagwiritsa ntchito maburashi angapo kuti apange mwaluso. Burashi iliyonse ndi kernel, ndipo wojambula akhoza kupanga chithunzi chovuta, chowona mwa kusakaniza maso ambiri. Titha kuchotsa mawonekedwe ofunikira pazithunzi ndikugwiritsa ntchito kulosera molondola zomwe zili pachithunzichi pogwiritsa ntchito ma CNN.
13. Deep Belief Networks (DBNs)
Ma DBN ndi mtundu wa neural network womwe umagwiritsidwa ntchito pazinthu zophunzirira zosayang'aniridwa monga kuchepetsa kukula ndi kuphunzira kwazinthu. Amagwira ntchito posanjikiza magawo angapo a Restricted Boltzmann Machines (RBMs), omwe ndi ma neural network osanjikiza awiri omwe amatha kuphunzira kukonzanso zomwe zalowa.
Ma DBN ndi opindulitsa kwambiri pazinthu zapamwamba za data chifukwa amatha kuphunzira kuyimira kophatikizana komanso koyenera kwa zomwe alowetsa. Amagwiritsidwa ntchito pachilichonse kuyambira kuzindikira mawu mpaka kugawa zithunzi mpaka kupeza mankhwala osokoneza bongo.
Mwachitsanzo, ofufuza adagwiritsa ntchito DBN kuti ayese mgwirizano womangirira wa ofuna mankhwala ku estrogen receptor. DBN idaphunzitsidwa kuphatikizika kwazinthu zamakina ndi ma affinities omangika, ndipo idatha kuneneratu molondola za mgwirizano womwe umakhalapo wa omwe akufuna mankhwala osokoneza bongo.
Izi zikuwonetsa kugwiritsa ntchito ma DBN popanga mankhwala osokoneza bongo komanso ma data apamwamba kwambiri.
14. Ma Autoencoder
Ma Autoencoder ndi ma neural network omwe amagwiritsidwa ntchito pamaphunziro osayang'aniridwa. Amapangidwa kuti apangenso zomwe zalowa, zomwe zikutanthauza kuti aphunzira kuyika chidziwitsocho kuti chikhale choyimira chophatikizika ndikuchiyikanso m'mawu oyamba.
Ma Autoencoder ndi othandiza kwambiri pakupondereza kwa data, kuchotsa phokoso, komanso kuzindikira molakwika. Atha kugwiritsidwanso ntchito pophunzirira mbali, pomwe mawonekedwe a autoencoder amaphatikizidwa mu ntchito yophunzirira yoyang'aniridwa.
Ganizirani ma autoencoder kukhala ophunzira omwe amalemba manotsi mkalasi. Wophunzira amamvetsera nkhaniyo ndi kulemba mfundo zofunika kwambiri m’njira yachidule komanso yothandiza.
Pambuyo pake, wophunzirayo angaphunzire ndi kukumbukira phunzirolo pogwiritsa ntchito manotsi ake. Komano, makina ojambulira otopa okha, amaika data yolowera m'chimake chophatikizika chomwe pambuyo pake chingagwiritsidwe ntchito pazifukwa zosiyanasiyana monga kuzindikira zolakwika kapena kuphatikizika kwa data.
15. Makina a Boltzmann Oletsedwa (RBMs)
Ma RBM (Makina Oletsedwa a Boltzmann) ndi mtundu wa neural network yomwe imagwiritsidwa ntchito pophunzira mosayang'aniridwa. Amapangidwa ndi chigawo chowoneka ndi chobisika, chokhala ndi ma neuroni mumtundu uliwonse, wolumikizidwa koma osati mkati mwawo.
Ma RBM amaphunzitsidwa pogwiritsa ntchito njira yomwe imadziwika kuti kusiyana kosiyana, komwe kumaphatikizapo kusintha zolemera pakati pa zigawo zooneka ndi zobisika kuti athe kupititsa patsogolo mwayi wa maphunziro. Ma RBM atha kupanga deta yatsopano ataphunzitsidwa ndi sampuli kuchokera pakugawa komwe adaphunzira.
Kuzindikira zithunzi ndi malankhulidwe, kusefa kogwirizana, ndi kuzindikira mosadziwika bwino ndi mapulogalamu omwe agwiritsa ntchito ma RBM. Amagwiritsidwanso ntchito m'machitidwe opangira kupanga malingaliro ogwirizana pophunzira machitidwe a ogwiritsa ntchito.
Ma RBM akhala akugwiritsidwanso ntchito pophunzira mbali kuti apange chithunzi chophatikizika komanso choyenera cha data yapamwamba kwambiri.
Kumaliza ndi Kulonjeza Zachitukuko Patsogolo
Njira zophunzirira mozama, monga Convolutional Neural Networks (CNNs) ndi Recurrent Neural Networks (RNNs), ndi zina mwa njira zapamwamba kwambiri zanzeru zopangira. Ma CNN asintha kuzindikira kwa zithunzi ndi mawu, pomwe ma RNN apita patsogolo kwambiri pakukonza zilankhulo zachilengedwe komanso kusanthula kwa data motsatizana.
Gawo lotsatira pakusinthidwa kwa njirazi kuyenera kuyang'ana kwambiri pakuwongolera magwiridwe antchito awo komanso kusanja, kuwalola kusanthula ma dataseti akulu ndi ovuta, komanso kukulitsa kutanthauzira kwawo komanso kuthekera kophunzira kuchokera kuzinthu zosalembedwa zochepa.
Kuphunzira mozama kumakhala ndi mwayi wololeza kupita patsogolo m'magawo monga chisamaliro chaumoyo, zachuma, ndi machitidwe odziyimira pawokha pomwe ikupita patsogolo.
Siyani Mumakonda