Kodi kuwira kwa makina ophunzirira kwaphulika, kapena ndi chiyambi cha mbandakucha watsopano?

Zatulutsidwa posachedwa nkhani, zomwe zikuwonetsa njira yabwino yophunzirira makina m'zaka zaposachedwa. Mwachidule: chiwerengero cha makina ophunzirira makina chatsika kwambiri m'zaka ziwiri zapitazi.

Chabwino. Tiyeni tiwone "ngati kuwira kwaphulika", "momwe mungapitirizire kukhala ndi moyo" ndikulankhula za komwe squiggle iyi imachokera poyamba.

Choyamba, tiyeni tikambirane chomwe chinali cholimbikitsa pamapindikirawa. Kodi iye anachokera kuti? Iwo mwina adzakumbukira chirichonse chigonjetso Kuphunzira kwamakina mu 2012 pampikisano wa ImageNet. Kupatula apo, ichi ndi chochitika choyamba padziko lonse lapansi! Koma zoona zake n’zakuti sizili choncho. Ndipo kukula kwa mphukira kumayamba pang'onopang'ono. Ndikhoza kuzigawa mu mfundo zingapo.

1. 2008 idatulukira mawu akuti "deta yayikulu". Zogulitsa zenizeni zinayamba kuwonekera kuyambira 2010. Zambiri zimagwirizana mwachindunji ndi kuphunzira kwamakina. Popanda deta yaikulu, ntchito yokhazikika ya ma aligorivimu yomwe inalipo panthawiyo sizingatheke. Ndipo awa si ma neural network. Mpaka 2012, ma neural network anali osungidwa ochepa. Koma ma aligorivimu osiyana kwambiri anayamba kugwira ntchito, amene analipo kwa zaka, kapena zaka zambiri: SVM(1963,1993, XNUMX), Random Forest (1995), AdaBoost (2003)

   Chomwe chimachokera ku funde loyambali ndi magawo monga XGBoost, CatBoost, LightGBM, etc.

2. Mu 2011-2012 convolutional neural network adapambana mipikisano ingapo yozindikira zithunzi. Kugwiritsa ntchito kwawo kwenikweni kunachedwa. Ndinganene kuti zoyambira zazikulu komanso zothetsera zidayamba kuwonekera mu 2014. Zinatenga zaka ziwiri kuti ma neuroni agwirebe ntchito, kuti apange zomangira zosavuta zomwe zitha kukhazikitsidwa ndikukhazikitsidwa munthawi yoyenera, kuti apange njira zomwe zingakhazikitse ndikufulumizitsa nthawi yolumikizana.

   Ma Convolutional network adapangitsa kuti athe kuthana ndi vuto la masomphenya apakompyuta: kugawa zithunzi ndi zinthu zomwe zili pachithunzichi, kuzindikira zinthu, kuzindikira zinthu ndi anthu, kukonza zithunzi, ndi zina zambiri.

3. 2015-2017. Kukula kwa ma algorithms ndi mapulojekiti otengera maukonde obwereza kapena ma analogue awo (LSTM, GRU, TransformerNet, etc.). Ma algorithms oyenda bwino akulankhula ndi mawu komanso makina omasulira apezeka. Mwa zina zimatengera ma convolutional network kuti atenge zinthu zofunika. Pang'ono ndi pang'ono chifukwa chakuti tinaphunzira kusonkhanitsa ma dataset akuluakulu komanso abwino.

“Kodi kuwira kwaphulika? Kodi hype yatenthedwa? Kodi adamwalira ngati blockchain?"
Apo ayi! Mawa Siri adzasiya kugwira ntchito pa foni yanu, ndipo mawa Tesla sadzadziwa kusiyana pakati pa kutembenuka ndi kangaroo.

Neural network ikugwira ntchito kale. Iwo ali mu zipangizo zambiri. Amakulolani kuti mupeze ndalama, kusintha msika ndi dziko lozungulira inu. Hype ikuwoneka mosiyana pang'ono:

Kungoti ma neural network salinso chatsopano. Inde, anthu ambiri amayembekezera zambiri. Koma makampani ambiri aphunzira kugwiritsa ntchito ma neuron ndikupanga zinthu potengera iwo. Ma Neuroni amapereka magwiridwe antchito atsopano, amakulolani kuti muchepetse ntchito, ndikuchepetsa mtengo wantchito:

• Makampani opanga zinthu akuphatikiza ma algorithms kuti awone zolakwika pamzere wopanga.
• Mafamu a ziweto amagula machitidwe owongolera ng'ombe.
• Zophatikiza zokha.
• Makina Oyimba Mafoni.
• Zosefera mu SnapChat. (Chabwino, china chake chothandiza!)

Koma chinthu chachikulu, osati chodziwikiratu: "Palibenso malingaliro atsopano, kapena sadzabweretsa likulu nthawi yomweyo." Neural network yathetsa mavuto ambiri. Ndipo adzasankha zambiri. Malingaliro onse odziwikiratu omwe analipo adayambitsa zoyambira zambiri. Koma zonse zimene zinali pamwamba zinali zitasonkhanitsidwa kale. Pazaka ziwiri zapitazi, sindinapeze lingaliro limodzi latsopano logwiritsa ntchito ma neural network. Palibe njira yatsopano (chabwino, pali zovuta zingapo ndi ma GAN).

Ndipo kuyambika kulikonse kotsatira kumakhala kovuta kwambiri. Sipafunikanso anyamata awiri omwe amaphunzitsa neuron pogwiritsa ntchito deta yotseguka. Zimafunikira opanga mapulogalamu, seva, gulu la zolembera, chithandizo chovuta, ndi zina.

Zotsatira zake, zoyambira zimachepa. Koma pali zambiri zopanga. Mukufuna kuwonjezera zizindikiritso za layisensi? Pali mazana a akatswiri omwe ali ndi chidziwitso chofunikira pamsika. Mutha kulemba ganyu wina ndipo m'miyezi ingapo wantchito wanu adzapanga dongosolo. Kapena gulani zopangidwa kale. Koma kupanga chiyambi chatsopano? .. Wopenga!

Muyenera kupanga njira yotsatirira alendo - chifukwa chiyani mumalipira zilolezo zambiri pomwe mutha kupanga zanu m'miyezi 3-4, kukulitsa bizinesi yanu.

Tsopano ma neural network akudutsa njira yomweyo yomwe matekinoloje ena ambiri adadutsamo.

Kodi mukukumbukira momwe lingaliro la "wopanga webusayiti" lasinthira kuyambira 1995? Msikawu sunakhudzebe akatswiri. Pali akatswiri ochepa. Koma nditha kubetcha kuti muzaka 5-10 sipadzakhala kusiyana kwakukulu pakati pa pulogalamu ya Java ndi wopanga ma neural network. Padzakhala okwanira onse akatswiri pa msika.

Padzangokhala gulu la zovuta zomwe zitha kuthetsedwa ndi ma neuron. Yabuka ntchito - ganyu katswiri.

"Chakudza ndi chiyani? Kodi nzeru zopangira zolonjezedwa zili kuti?"

Koma apa pali kusamvetsetsa kwakung'ono koma kosangalatsa :)

Kuchuluka kwaukadaulo komwe kulipo masiku ano, mwachiwonekere, sikudzatitsogolera ku luntha lochita kupanga. Malingaliro ndi zachilendo zawo zatopetsa okha. Tiye tikambirane zomwe zimagwira mulingo wamakono wachitukuko.

Zoletsa

Tiyeni tiyambe ndi magalimoto odziyendetsa okha. Zikuwoneka zomveka kuti ndizotheka kupanga magalimoto odziyimira pawokha ndiukadaulo wamakono. Koma zaka zingati izi zidzachitika sizidziwika. Tesla akukhulupirira kuti izi zichitika m'zaka zingapo -

Palinso ena ambiri akatswiri, amene amati ndi zaka 5-10.

Mwinamwake, m'malingaliro anga, m'zaka 15 zomangamanga za mizinda zidzasintha kotero kuti kutuluka kwa magalimoto odziyimira pawokha kudzakhala kosapeweka ndipo kudzakhala kupitiriza kwake. Koma izi sizingaganizidwe ngati nzeru. Tesla yamakono ndi payipi yovuta kwambiri pakusefa, kufufuza ndi kukonzanso. Awa ndi malamulo-malamulo-malamulo, kusonkhanitsa deta ndi zosefera pa iwo (apa apa Ndinalembanso pang'ono za izi, kapena penyani kuchokera izi zizindikiro).

Vuto loyamba

Ndipo apa ndi pamene tikuwona vuto loyamba lofunikira. Zambiri zambiri. Izi ndi zomwe zidabala mafunde apano a neural network ndi kuphunzira pamakina. Masiku ano, kuti muchite chinthu chovuta komanso chodziwikiratu, mumafunikira zambiri. Osati zambiri, koma kwambiri, kwambiri. Timafunikira ma aligorivimu odzipangira okha kuti azitolera, kuziyika, ndikugwiritsa ntchito. Tikufuna kuti galimotoyo iwonetse magalimoto omwe akuyang'ana dzuwa - choyamba tiyenera kusonkhanitsa chiwerengero chokwanira. Tikufuna kuti galimoto isachite misala ndi njinga yokhomeredwa ku thunthu - zitsanzo zambiri.

Komanso, chitsanzo chimodzi sichikwanira. Mazana? Zikwi?

Vuto lachiwiri

Vuto lachiwiri -Kuwonera zomwe neural network yathu yamvetsetsa. Iyi ndi ntchito yosachepera. Mpaka pano, anthu ochepa amamvetsa momwe angawonere izi. Nkhanizi ndi zaposachedwa kwambiri, izi ndi zitsanzo zochepa, ngakhale zitakhala kutali:
Kuwonetseratu kutengeka ndi ma textures. Imawonetsa bwino zomwe neuron imakonda kukonza + zomwe imawona ngati chidziwitso choyambira.

Kuwonetseratu Attention pa kumasulira. M'malo mwake, kukopa kumatha kugwiritsidwa ntchito ndendende kuwonetsa zomwe zidayambitsa ma network otere. Ndaziwonapo zinthu zotere pakuwongolera zolakwika komanso zothetsera. Pali nkhani zambiri pamutuwu. Koma zambiri zomwe zimakhala zovuta kwambiri, zimakhala zovuta kumvetsa momwe mungakwaniritsire maonekedwe amphamvu.

Chabwino, inde, gulu labwino lakale la "yang'anani zomwe zili mkati mwa mesh zosefera" Zithunzizi zinali zotchuka zaka 3-4 zapitazo, koma aliyense anazindikira mwamsanga kuti zithunzizo zinali zokongola, koma zinalibe tanthauzo lalikulu.

Sindinatchule zida zina zambiri, njira, ma hacks, kafukufuku wamomwe mungawonetsere zamkati mwamaneti. Kodi zida izi zimagwira ntchito? Kodi amakuthandizani kuti mumvetsetse msanga vuto ndikusintha maukonde?.. Pezani gawo lomaliza? Chabwino, ndizofanana:

Mutha kuwona mpikisano uliwonse pa Kaggle. Ndipo kufotokoza momwe anthu amapangira zisankho zomaliza. Tidasanjikiza mayunitsi 100-500-800 amitundu ndipo zidagwira ntchito!

Ndikukokomeza, ndithudi. Koma njirazi sizimapereka mayankho achangu komanso achindunji.

Kukhala ndi chidziwitso chokwanira, mutayang'ana zosankha zosiyanasiyana, mutha kupereka chigamulo cha chifukwa chomwe makina anu adapangira chisankho. Koma zidzakhala zovuta kukonza khalidwe la dongosolo. Ikani ndodo, sunthani poyambira, onjezani deta, tengani netiweki ina yakumbuyo.

Vuto lachitatu

Vuto Lachitatu Lofunika Kwambiri - ma grids amaphunzitsa ziwerengero, osati zomveka. Zowerengera izi лицо:

Zomveka, sizili zofanana kwambiri. Ma Neural network samaphunzira chilichonse chovuta pokhapokha atakakamizidwa. Nthawi zonse amaphunzitsa zizindikiro zosavuta. Kodi muli ndi maso, mphuno, mutu? Kotero iyi ndi nkhope! Kapena perekani chitsanzo pamene maso satanthauza nkhope. Ndipo kachiwiri - mamiliyoni a zitsanzo.

Pali Malo Ambiri Pansi

Ndinganene kuti ndi mavuto atatu apadziko lonse lapansi omwe amachepetsa kukula kwa ma neural network ndi kuphunzira pamakina. Ndipo pamene mavutowa sanachepetse, amagwiritsidwa ntchito kale.

Awa ndi mathero? Kodi ma neural network ali pati?

Zosadziwika. Koma, ndithudi, aliyense sayembekezera.

Pali njira zambiri komanso njira zothetsera mavuto omwe ndawunikira pamwambapa. Koma mpaka pano, palibe imodzi mwa njira zimenezi imene yachititsa kuti munthu achite chinachake chatsopano, kuthetsa chinachake chimene sichinathetsedwe. Pakadali pano, ntchito zonse zofunika zikuchitika kutengera njira zokhazikika (Tesla), kapena kukhalabe ma projekiti oyesa a mabungwe kapena mabungwe (Google Brain, OpenAI).

Mwachidule, chitsogozo chachikulu ndikupanga chiwonetsero chapamwamba chazolowera. Mwanjira ina, "kukumbukira". Chitsanzo chosavuta cha kukumbukira ndi "Kuyika" kosiyanasiyana - zoyimira zithunzi. Chabwino, mwachitsanzo, machitidwe onse ozindikiritsa nkhope. Netiweki imaphunzira kupeza kuchokera kumaso mawonekedwe okhazikika omwe samatengera kuzungulira, kuyatsa, kapena kukonza. Kwenikweni, maukonde amachepetsa ma metric "nkhope zosiyanasiyana zili patali" komanso "nkhope zofananira zili pafupi."

Pa maphunziro oterowo, zitsanzo makumi ndi mazana a zikwi mazanamazana zikufunika. Koma zotsatira zake zimakhala ndi zoyambira za "Kuphunzira Kumodzi". Tsopano sitifunikira mazana a nkhope kuti tikumbukire munthu. Nkhope imodzi yokha ndipo ndizo zonse zomwe ife tiri tiyeni tifufuze!
Pali vuto limodzi lokha ... Gululi limatha kuphunzira zinthu zosavuta. Poyesera kusiyanitsa osati nkhope, koma, mwachitsanzo, "anthu ndi zovala" (ntchito Kuzindikiritsanso) - khalidwe limatsika ndi maulamuliro ambiri. Ndipo maukonde sangathenso kuphunzira mwachilungamo zoonekeratu kusintha ngodya.

Ndipo kuphunzira kuchokera ku mamiliyoni a zitsanzo kumakhalanso kosangalatsa.

Pali ntchito yochepetsa kwambiri zisankho. Mwachitsanzo, munthu akhoza kukumbukira nthawi yomweyo imodzi mwa ntchito zoyamba Kuphunzira kwa OneShot kuchokera ku Google:

Pali ntchito zambiri zotere, mwachitsanzo 1 kapena 2 kapena 3.

Pali kuchotsera kumodzi - nthawi zambiri maphunziro amagwira bwino pazitsanzo zosavuta, "MNIST". Ndipo mukamapita ku ntchito zovuta, mumafunika database yayikulu, chitsanzo cha zinthu, kapena matsenga amtundu wina.
Kawirikawiri, ntchito pa maphunziro a One-Shot ndi mutu wosangalatsa kwambiri. Mupeza malingaliro ambiri. Koma nthawi zambiri, mavuto awiri omwe ndidawalemba (kukonzekereratu pa dataset yayikulu / kusakhazikika pazovuta zovuta) amasokoneza kwambiri kuphunzira.

Kumbali ina, ma GAN - maukonde oyambitsa - amayandikira mutu wa Embedding. Mwina mwawerengapo zolemba zambiri za Habré pamutuwu. (1, 2,3)
Mbali ya GAN ndikupangidwa kwa malo ena amkati (makamaka Kuyika komweko), komwe kumakupatsani mwayi wojambulira chithunzi. Zitha kukhala nkhope, akhoza kukhala zochita.

Vuto la GAN ndiloti chinthu chopangidwa ndi chovuta kwambiri, chimakhala chovuta kufotokoza momveka bwino "generator-discriminator". Zotsatira zake, ntchito zenizeni za GAN zomwe zimamveka ndi DeepFake, zomwe, zimagwiritsanso ntchito mawonekedwe amaso (omwe pali maziko akulu).

Ndawonapo zochepa zothandiza zina. Nthawi zambiri chinyengo chamtundu wina chomwe chimaphatikizapo kumaliza kujambula zithunzi.

Ndipo kachiwiri. Palibe amene akudziwa momwe izi zingatithandizire kupita ku tsogolo labwino. Kuyimira malingaliro / malo mu neural network ndizabwino. Koma timafunikira zitsanzo zambiri, sitikumvetsetsa momwe neuron imayimira izi palokha, sitikumvetsetsa momwe tingapangire neuron kukumbukira lingaliro lovuta kwambiri.

Zolimbikitsa kuphunzira - iyi ndi njira yochokera ku njira yosiyana kwambiri. Mosakayikira mukukumbukira momwe Google idagonjetsera aliyense mu Go. Kupambana kwaposachedwa mu Starcraft ndi Dota. Koma apa chirichonse chiri kutali kwambiri ndi kukongola ndi kulonjeza. Amalankhula bwino za RL ndi zovuta zake Nkhani iyi.

Kuti tifotokoze mwachidule zomwe wolemba analemba:

• Zitsanzo kunja kwa bokosi sizikwanira / zimagwira ntchito bwino nthawi zambiri
• Mavuto othandiza ndi osavuta kuthetsa m'njira zina. Boston Dynamics sagwiritsa ntchito RL chifukwa chazovuta zake/zosayembekezereka/zovuta
• Kuti RL igwire ntchito, mufunika ntchito yovuta. Nthawi zambiri zimakhala zovuta kupanga / kulemba
• Zovuta kuphunzitsa zitsanzo. Muyenera kuthera nthawi yochuluka kuti mupope ndikutuluka muzabwino zakomweko
• Zotsatira zake, zimakhala zovuta kubwereza chitsanzo, chitsanzocho sichikhazikika ndi kusintha pang'ono
• Nthawi zambiri imakwirira mitundu ina mwachisawawa, ngakhale jenereta ya manambala mwachisawawa

Mfundo yofunika ndi yakuti RL sikugwirabe ntchito pakupanga. Google ili ndi zoyeserera zina ( 1, 2 ). Koma sindinawone dongosolo limodzi lazinthu.

Memory. Choyipa cha zonse zomwe tafotokozazi ndi kusowa kwa kapangidwe. Imodzi mwa njira zoyesera kukonza zonsezi ndikupatsa neural network mwayi wofikira kukumbukira. Kuti athe kujambula ndikulembanso zotsatira zamayendedwe ake pamenepo. Ndiye neural network imatha kutsimikiziridwa ndi momwe kukumbukira komwe kulipo. Izi ndizofanana kwambiri ndi mapurosesa akale ndi makompyuta.

Odziwika kwambiri komanso otchuka nkhani - kuchokera ku DeepMind:

Zikuwoneka kuti iyi ndiye chinsinsi chomvetsetsa nzeru? Koma mwina ayi. Dongosolo limafunikirabe kuchuluka kwa data pakuphunzitsidwa. Ndipo imagwira ntchito makamaka ndi deta yokhazikika. Komanso, pamene Facebook anaganiza vuto lofananalo, kenako adatenga njirayo "kukumbukira kukumbukira, kupangitsa kuti neuron ikhale yovuta, ndikukhala ndi zitsanzo zambiri - ndipo iphunzira yokha."

Kusokoneza. Njira ina yopangira kukumbukira kwatanthauzo ndikutengera zoyika zomwezo, koma panthawi yophunzitsira, yambitsani zina zomwe zingakuthandizeni kuwunikira "matanthauzo" mwa iwo. Mwachitsanzo, tikufuna kuphunzitsa neural network kuti isiyanitse machitidwe amunthu m'sitolo. Ngati titsatira njira yokhazikika, tiyenera kupanga ma network khumi ndi awiri. Wina akufunafuna munthu, wachiwiri amazindikira zomwe akuchita, wachitatu ndi msinkhu wake, wachinayi ndi jenda. Malingaliro osiyana amayang'ana gawo la sitolo komwe amaphunzitsidwa kuchita izi. Chachitatu chimatsimikizira njira yake, etc.

Kapena, ngati panali chiwerengero chopanda malire, ndiye kuti zingatheke kuphunzitsa maukonde amodzi pazotsatira zonse zomwe zingatheke (mwachiwonekere, mndandanda woterewu sungathe kusonkhanitsidwa).

The disentanglement njira imatiuza - tiyeni kuphunzitsa maukonde kuti palokha akhoza kusiyanitsa mfundo. Kotero kuti ipange zopachika potengera vidiyoyo, pamene dera limodzi lidzatsimikizira zochita, wina adziwe malo pansi pa nthawi, wina adziwe kutalika kwa munthuyo, ndipo wina adziwe kuti munthuyo ndi ndani. Panthawi imodzimodziyo, pophunzitsa, ndikufuna kuti ndisamangoyambitsa maukonde ndi mfundo zazikuluzikuluzi, koma kuti ziwonetsetse ndi kugawa magawo. Pali zolemba zingapo zotere (zina mwa izo 1, 2, 3) ndipo ambiri amakhala ongoyerekeza.

Koma malangizowa, ongoyerekeza, akuyenera kufotokoza zovuta zomwe zalembedwa poyambira.

Kuwonongeka kwazithunzi molingana ndi magawo "mtundu wa khoma / mtundu wapansi / mawonekedwe a chinthu / mtundu wa chinthu / etc."

Kuwola kwa nkhope molingana ndi magawo "kukula, nsidze, mawonekedwe, khungu, ndi zina zambiri."

Прочее

Pali zina zambiri, osati zapadziko lonse lapansi, madera omwe amakulolani kuti muchepetse nkhokwe, kugwira ntchito ndi data yochulukirapo, ndi zina zambiri.

chisamaliro. Mwina sizomveka kulekanitsa izi ngati njira yosiyana. Njira yokhayo yomwe imawonjezera ena. Zolemba zambiri zidaperekedwa kwa iye (1,2,3). Cholinga cha chidwi ndikukulitsa kuyankha kwa netiweki makamaka kuzinthu zofunikira panthawi yamaphunziro. Nthawi zambiri ndi mtundu wina wa chandamale chakunja, kapena netiweki yaying'ono yakunja.

3D kayeseleledwe. Ngati mupanga injini yabwino ya 3D, nthawi zambiri mumatha kuphimba 90% ya data yophunzitsira nayo (ndinawonanso chitsanzo chomwe pafupifupi 99% ya deta idaphimbidwa ndi injini yabwino). Pali malingaliro ambiri ndi ma hacks a momwe mungapangire maukonde ophunzitsidwa pa injini ya 3D ntchito pogwiritsa ntchito deta yeniyeni (Kukonza bwino, kusamutsa kalembedwe, etc.). Koma nthawi zambiri kupanga injini yabwino kumakhala kovuta kwambiri kuposa kusonkhanitsa deta. Zitsanzo pamene injini anapangidwa:
Maphunziro a robot (sakani, bongo)
Zophunzitsa kuzindikira katundu m'sitolo (koma m'mapulojekiti awiri omwe tidachita, titha kuchita popanda izo).
Maphunziro ku Tesla (kachiwiri, kanema pamwambapa).

anapezazo

Nkhani yonseyo, m’lingaliro lina, ndi yomaliza. Mwina uthenga waukulu womwe ndimafuna kupanga unali "zaulere zatha, ma neuron saperekanso mayankho osavuta." Tsopano tifunika kuyesetsa kupanga zosankha zovuta. Kapena yesetsani kuchita kafukufuku wovuta wa sayansi.

Nthawi zambiri, mutuwu ndi wokambitsirana. Mwina owerenga ali ndi zitsanzo zosangalatsa?

Source: www.habr.com