Ichangobva kuburitswa , iyo inoratidza maitiro akanaka mukudzidza kwemichina mumakore achangopfuura. Muchidimbu: huwandu hwemakina ekutanga ekudzidza akadzikira mumakore maviri apfuura.
Zvakanaka. Ngatitarisei "kana bubble rakaputika", "nzira yekuramba uchirarama" uye taura pamusoro pekuti iyi squiggle inobva kupi pakutanga.
Kutanga, ngatitaure nezve chii chaive kusimudzira kweiyi curve. Akabvepi? Zvichida vacharangarira zvose Kudzidza kwemichina muna 2012 pamakwikwi eImageNet. Mushure mezvose, ichi ndicho chiitiko chekutanga chepasi rose! Asi chaizvoizvo izvi hazvisizvo. Uye kukura kwecurve kunotanga zvishoma kare. Ndaizozvipatsanura kuita zvibodzwa zvakawanda.
- 2008 yakaona kubuda kweshoko rekuti "data hombe". Zvigadzirwa chaizvo zvakatanga kubva 2010. Data hombe yakanangana nekudzidza muchina. Pasina data hombe, kushanda kwakagadzikana kwealgorithms yaivepo panguva iyoyo hazvigoneke. Uye aya haasi neural network. Kusvika 2012, neural network yaive yakachengetedzwa yevashoma vashoma. Asi ipapo algorithms akasiyana zvachose akatanga kushanda, iyo yakanga iripo kwemakore, kana makumi emakore: (1963,1993), (1995), (2003)
Kunobva kune iyi yekutanga wave seti yezvimiro zvakaita seXGBoost, CatBoost, LightGBM, nezvimwe.
- Muna 2011-2012 akahwina makwikwi akawanda ekuzivikanwa kwemifananidzo. Kushandisa kwavo chaiko kwakati nonokei. Ndingati kutanga kwakakosha uye mhinduro dzakatanga kuoneka muna 2014. Zvakatora makore maviri kugaya kuti neuroni dzichiri kushanda, kugadzira masitadhi ari nyore anogona kuiswa uye kutangwa munguva inonzwisisika, kugadzira nzira dzinodzikamisa nekumhanyisa nguva yekusangana.
Convolutional network yakaita kuti zvikwanise kugadzirisa matambudziko ekuona komputa: kukamurwa kwemifananidzo uye zvinhu zviri mumufananidzo, kuona chinhu, kuzivikanwa kwezvinhu nevanhu, kuvandudzwa kwemifananidzo, nezvimwe, nezvimwe.
- 2015-2017. Iyo boom yealgorithms uye mapurojekiti akavakirwa pane anodzokororwa network kana analogues avo (LSTM, GRU, TransformerNet, nezvimwewo). Matauriro-kune-mavara ari kushanda zvakanaka uye masisitimu ekushandura muchina aonekwa. Iwo ari muchikamu akavakirwa pane convolutional network kubvisa zvakakosha maficha. Zvikamu nekuda kwekuti isu takadzidza kuunganidza akakura chaizvo uye akanaka datasets.
βFuro raputika here? Ko hype yanyanya kupisa here? Vakafa se blockchain here?"
Zvikasadaro! Mangwana Siri achamira kushanda pafoni yako, uye zuva rinotevera mangwana Tesla haazoziva mutsauko pakati pekutendeuka uye kangaroo.
Neural network dziri kutoshanda. Iwo ari mumichina yakawanda. Vanonyatso kukubvumira kuti uwane mari, shandura misika uye nyika yakakupoteredza. Hype inoratidzika zvakasiyana zvishoma:
Ingori kuti neural network haisisiri chinhu chitsva. Hungu, vanhu vakawanda vane kariro huru. Asi nhamba huru yemakambani akadzidza kushandisa neuroni uye kugadzira zvigadzirwa zvichibva pazviri. Neurons inopa mashandiro matsva, inobvumidza iwe kutema mabasa, uye kuderedza mutengo wemasevhisi:
- Makambani ekugadzira ari kubatanidza algorithms kuti aongorore kukanganisa pamutsara wekugadzira.
- Mapurazi ezvipfuyo anotenga masisitimu ekudzora mombe.
- Otomatiki inosanganisa.
- Automated Call Centers.
- Mafirita mu SnapChat. (zvakanaka, chimwe chinhu chinobatsira!)
Asi chinhu chikuru, uye kwete chinonyanya kuoneka: "Hapasisina pfungwa itsva, kana kuti havazounzi mari pakarepo." Neural network yakagadzirisa matambudziko mazhinji. Uye ivo vachasarudza zvakatowanda. Ese pfungwa dzakajeka dzaivepo dzakakonzera kutanga kwakawanda. Asi zvese zvaive pamusoro zvakange zvatounganidzwa. Kwemakore maviri apfuura, handina kuwana zano rimwechete idzva rekushandiswa kweneural network. Kwete imwe nzira nyowani (zvakanaka, ok, kune mashoma matambudziko neGANs).
Uye kutanga kwega kwega kunotevera kunowedzera kuoma. Hazvichadi vakomana vaviri vanodzidzisa neuron vachishandisa data rakavhurika. Zvinoda vanogadzira mapurogiramu, sevha, timu yemakaki, tsigiro yakaoma, nezvimwe.
Nekuda kweizvozvo, kune mashoma ekutanga. Asi pane zvakawanda kugadzirwa. Unoda kuwedzera rezinesi rekuziva here? Kune mazana enyanzvi dzine ruzivo rwakakodzera pamusika. Iwe unogona kuhaya mumwe munhu uye mumwedzi mishoma mushandi wako achagadzira iyo system. Kana kutenga yakagadzirira-yakagadzirwa. Asi kuita kutanga kutsva?.. Anopenga!
Iwe unofanirwa kugadzira yevaenzi yekutevera sisitimu - sei uchibhadharira boka remarezinesi iwe uchigona kugadzira yako mumwedzi 3-4, urodzere bhizinesi rako.
Iye zvino neural network iri kufamba nenzira imwechete iyo akati wandei emamwe matekinoroji apfuura.
Unoyeuka here kuti pfungwa ye "webhusaiti yekuvandudza" yakachinja sei kubva 1995? Musika hausati wazara nenyanzvi. Kune nyanzvi shoma shoma. Asi ini ndinogona kubheja kuti mumakore mashanu-5 hapazove nemusiyano wakawanda pakati peJava programmer uye neural network developer. Pachave nekukwana kwevaviri nyanzvi pamusika.
Pachave nekungove nekirasi yematambudziko anogona kugadziriswa nema neuron. Basa ramuka - kuhaya nyanzvi.
"Chii chinotevera? Hungwaru hwekunyepedzera hwakavimbiswa huripi?β
Asi pano pane kusanzwisisa kudiki asi kunonakidza :)
Iyo tekinoroji stack iripo nhasi, sezviri pachena, haizotitungamirire kuhungwaru hwekugadzira. Mazano uye hutsva hwawo hwakanyanya kuzvinetesa. Ngatitaurei pamusoro pezvakabata danho razvino rebudiriro.
Zvibvumirano
Ngatitange nemota dzinozvityaira. Zvinoita sezviri pachena kuti zvinokwanisika kugadzira mota dzakazvimiririra zvizere nehunyanzvi hwemazuva ano. Asi mumakore mangani izvi zvichaitika hazvina kujeka. Tesla anotenda kuti izvi zvichaitika mumakore mashoma -
Kune vamwe vakawanda , vanofungidzira kuti makore 5-10.
Zvingangodaro, mukuona kwangu, mumakore gumi nemashanu zvivakwa zvemaguta pachazvo zvichachinja nenzira yekuti kubuda kwemotokari dzinozvimiririra kuchava kusingadzivisiki uye kuchava kuenderera kwayo. Asi izvi hazvinganzi hungwaru. Tesla yemazuva ano ipombi yakaoma kwazvo yekusefa, kutsvaga uye kudzidzisazve. Iyi mitemo-mitemo-mitemo, kuunganidza data uye mafirita pamusoro pavo (pano Ndakanyora zvishoma nezve izvi, kana tarisa kubva marks).
Dambudziko rekutanga
Uye apa ndipo patinoona dambudziko rekutanga. Data huru. Izvi ndizvo chaizvo zvakabereka iyo yazvino wave yeneural network uye kudzidza muchina. Mazuva ano, kuita chimwe chinhu chakaoma uye otomatiki, unoda data rakawanda. Kwete chete zvakawanda, asi zvikuru, zvikuru. Isu tinoda otomatiki algorithms ekuunganidza kwavo, kumaka, uye kushandiswa. Tinoda kuita kuti motokari ione marori akatarisana nezuva - tinofanira kutanga taunganidza nhamba yakakwana yawo. Tinoda kuti mota isapenge nebhasikoro rakasungirirwa pahunde - mamwe masampuli.
Uyezve, mumwe muenzaniso hauna kukwana. Mazana? Zviuru?
Dambudziko repiri
Dambudziko repiri -Kuona zvakanzwisiswa neural network yedu. Iri ibasa risiri diki. Kusvika ikozvino, vashoma vanhu vanonzwisisa maitiro ekuona izvi. Izvi zvinyorwa ndezvenguva pfupi yapfuura, iyi ingori mienzaniso mishoma, kunyangwe iri kure:
kufinhikana nemaumbirwo. Inoratidza zvakanaka izvo neuron inoda kugadzirisa pane + iyo yainoona sekutanga ruzivo.
Attention pa . Muchokwadi, kukwezva kunogona kushandiswa chaizvo kuratidza chakonzera kuita kwakadaro kwetiweki. Ndakaona zvinhu zvakadaro kune zvese debugging uye zvigadzirwa mhinduro. Pane zvakawanda zvinyorwa pamusoro penyaya iyi. Asi iyo yakanyanya kuomarara iyo data, zvakanyanya kuoma kuti unzwisise maitiro ekuita kuona kwakasimba.
Ehe, hongu, yakanaka yekare seti ye "tarisa zviri mukati me mesh mukati " Iyi mifananidzo yaizivikanwa makore 3-4 apfuura, asi munhu wose akakurumidza kuona kuti mifananidzo yakanga yakanaka, asi yakanga isina zvakawanda zvinoreva.
Ini handina kutaura akawanda emamwe magajeti, nzira, hacks, tsvagiridzo yekuti ungaratidza sei mukati metiweki. Zvishandiso izvi zvinoshanda here? Vanokubatsira nekukurumidza kunzwisisa kuti dambudziko nderei uye kugadzirisa network? .. Wana chikamu chekupedzisira? Zvakanaka, zvakangofanana:
Unogona kuona chero makwikwi paKaggle. Uye tsananguro yekuti vanhu vanoita sei sarudzo dzekupedzisira. Isu takarongedza 100-500-800 mayuniti emhando uye yakashanda!
Ndiri kuwedzeredza, hongu. Asi nzira idzi hadzipi mhinduro dzinokurumidza uye dzakananga.
Kuve neruzivo rwakakwana, uchitenderedza sarudzo dzakasiyana, unogona kupa mutongo wekuti sei system yako yakaita sarudzo yakadai. Asi zvichange zvakaoma kugadzirisa maitiro ehurongwa. Isa crutch, fambisa chikumbaridzo, wedzera dataset, tora imwe backend network.
Dambudziko rechitatu
Dambudziko Rechitatu Rakakosha - magridi anodzidzisa nhamba, kwete pfungwa. Statistically izvi :
Zvinonzwisisika, hazvina kunyanya kufanana. Neural network haidzidze chero chinhu chakaoma kunze kwekunge vamanikidzwa. Vanogara vachidzidzisa zviratidzo zvakapfava zvinobvira. Une maziso, mhino, musoro here? Saka ichi ndicho chiso! Kana kupa muenzaniso apo maziso asingarevi chiso. Uye zvakare - mamiriyoni emuenzaniso.
Pane Yakawanda Kamuri Kuzasi
Ndingati ndiwo matambudziko matatu epasi rose ari kudzikamisa kuvandudzwa kweneural network uye kudzidza kwemichina. Uye apo matambudziko aya haana kumisa, atove achishandiswa.
Uyu ndiwo magumo? Neural network yakakwira here?
Unknown. Asi, chokwadi, munhu wose haatarisiri.
Pane nzira dzakawanda uye nzira dzekugadzirisa matambudziko akakosha andakasimbisa pamusoro. Asi kusvika zvino, hapana kana imwe yenzira idzi yaita kuti zvikwanisike kuita chimwe chinhu chitsva, kugadzirisa chimwe chinhu chisati chagadziriswa. Parizvino, mapurojekiti ese akakosha ari kuitwa pahwaro hwenzira dzakagadzikana (Tesla), kana kuramba mapurojekiti ebvunzo emasangano kana makambani (Google Brain, OpenAI).
Zvichireva kutaura, iyo nhungamiro huru ndeyekugadzira imwe yepamusoro-yepamusoro inomiririra yedata rekuisa. Neimwe nzira, βndangariroβ. Muenzaniso wakapfava wendangariro wakasiyana "Embedding" - mifananidzo inomiririra. Zvakanaka, semuenzaniso, maitiro ese ekuzivikanwa kwechiso. Iyo network inodzidza kuwana kubva pachiso imwe inomiririra yakagadzikana isingaenderane nekutenderera, kuvheneka, kana kugadzirisa. Chaizvoizvo, network inoderedza metric "zviso zvakasiyana zviri kure" uye "zviso zvakafanana zviri pedyo."
Nokuda kwekudzidziswa kwakadaro, makumi nemazana ezviuru zvemienzaniso anodiwa. Asi mhedzisiro inobereka zvimwe zvekutanga zve "One-shot Kudzidza". Zvino hatidi mazana ezviso kuti tiyeuke munhu. Chiso chimwe chete uye ndizvo zvatiri !
Pane dambudziko rimwe chete ... Giridhi inogona kungodzidza zvinhu zviri nyore. Paunenge uchiedza kusiyanisa kwete zviso, asi, semuenzaniso, "vanhu nezvipfeko" (basa ) - mhando inodonha nemirairo yakawanda yehukuru. Uye network haichakwanisa kudzidza shanduko dziri pachena mumakona.
Uye kudzidza kubva kumamiriyoni emienzaniso imhando yekunakidzwa.
Pane basa rekuderedza zvakanyanya sarudzo. Semuenzaniso, munhu anogona kuyeuka pakarepo imwe yemabasa ekutanga pa OneShot Kudzidza :
Kune akawanda mabasa akadaro, semuenzaniso kana kana .
Pane minus imwe chete - kazhinji kudzidzisa kunoshanda zvakanaka pane mimwe mienzaniso yakapfava, "MNIST". Uye kana uchienda kumabasa akaoma, unoda dhatabhesi hombe, modhi yezvinhu, kana imwe mhando yemashiripiti.
Kazhinji, shanda pane One-Shot kudzidziswa inyaya inonakidza kwazvo. Unowana mazano akawanda. Asi kune chikamu chikuru, matambudziko maviri andakanyora (pretraining pane yakakura dataset / kusagadzikana pane yakaoma data) inokanganisa zvakanyanya kudzidza.
Kune rimwe divi, maGAN β adversarial networks β anoswedera kumusoro weEmbedding. Iwe unogona kunge wakaverenga boka rezvinyorwa paHabrΓ© pane iyi nyaya. (, ,)
Chinhu cheGAN kuumbwa kweimwe nzvimbo yemukati yenyika (chaizvoizvo zvakafanana Embedding), iyo inobvumidza iwe kudhirowa mufananidzo. Zvinogona kudaro , zvinogona kuva .
Dambudziko neGAN nderekuti iyo yakanyanya kuomarara chinhu chakagadzirwa, zvakanyanya kuoma kuitsanangura mu "jenareta-kusarura" logic. Nekuda kweizvozvo, iwo ega ega mashandisirwo eGAN anonzwika nezvayo ndeaya DeepFake, iyo, zvakare, inoshandura inomiririra yechiso (iyo ine hwaro hombe).
Ndakaona mamwe mashoma anobatsira anoshandiswa. Kazhinji imwe mhando yehunyengeri inosanganisira kupedzisa kudhirowa kwemifananidzo.
Uye zvakare. Hapana anoziva kuti izvi zvichatibvumira sei kuti tifambe mune ramangwana rakajeka. Kumiririra pfungwa / nzvimbo mune neural network yakanaka. Asi isu tinoda huwandu hukuru hwemienzaniso, isu hatinzwisise kuti neuron inomiririra sei izvi pachayo, hatinzwisise maitiro ekuita kuti neuron irangarire imwe pfungwa yakaoma.
Yekusimbisa kudzidza - iyi ndiyo nzira kubva kune imwe nzira yakasiyana zvachose. Chokwadi unorangarira kuti Google yakarova sei munhu wese muGo. Kukunda kwazvino mu Starcraft uye Dota. Asi pano zvinhu zvese zviri kure nekunaka uye kuvimbisa. Anotaura zvakanyanya nezveRL uye nekuoma kwayo .
Kupfupikisa zvakanyorwa nemunyori:
- Mamodheru kunze kwebhokisi haakwane / anoshanda zvisina kunaka muzviitiko zvakawanda
- Zvinetso zvinoshanda zviri nyore kugadzirisa nedzimwe nzira. Boston Dynamics haishandise RL nekuda kwekuoma kwayo / kusafungira / computational kuomarara.
- Kuti RL ishande, unoda basa rakaoma. Zvinowanzonetsa kugadzira/kunyora
- Zvakaoma kudzidzisa modhi. Iwe unofanirwa kupedza nguva yakawanda uchipomba uye kubuda kunze kwenzvimbo optima
- Somugumisiro, zvakaoma kudzokorora muenzaniso, muenzaniso hauna kugadzikana nekuchinja kuduku
- Kazhinji overfits mamwe mhando kuruboshwe-bapiro mapatani, kusvika random nhamba jenareta
Chinhu chakakosha ndechekuti RL haisati yashanda mukugadzira. Google ine zvimwe zviedzo ( , ) Asi ini handisati ndaona imwe chigadzirwa system.
ndangariro. Izvo zvakaderera zvezvose zvinotsanangurwa pamusoro apa kushaikwa kwechimiro. Imwe yenzira dzekuyedza kurongedza zvese izvi kupa iyo neural network mukana wekuparadzanisa ndangariro. Kuti akwanise kurekodha nekunyorazve mhedzisiro yematanho ake ipapo. Ipapo iyo neural network inogona kutariswa neyazvino ndangariro mamiriro. Izvi zvakafanana zvakanyanya neakare processors uye makomputa.
Iyo yakakurumbira uye yakakurumbira β kubva kuDeepMind:
Zvinoita sekuti iyi ndiyo kiyi yekunzwisisa njere? Asi zvichida kwete. Iyo sisitimu ichiri kuda huwandu hukuru hwe data yekudzidziswa. Uye inoshanda zvakanyanya neyakagadzirwa tabular data. Uyezve, kana Facebook dambudziko rakafanana, vakabva vatora nzira "screw memory, ingoita kuti neuron iwedzere kuoma, uye uve nemimwe mienzaniso - uye ichadzidza yega."
Disentanglement. Imwe nzira yekugadzira chiyeuchidzo chine musoro kutora zvakamisikidzwa zvakafanana, asi panguva yekudzidziswa, suma imwe nzira inobvumidza iwe kuburitsa "zvinoreva" mazviri. Semuenzaniso, tinoda kudzidzisa neural network kusiyanisa maitiro evanhu muchitoro. Kana isu tikatevedzera yakajairwa nzira, taizofanirwa kugadzira gumi nemaviri network. Mumwe ari kutsvaga munhu, wechipiri ari kuona zvaari kuita, wechitatu zera rake, wechina murume wake. Yakaparadzana pfungwa inotarisa chikamu chechitoro painoita / inodzidziswa kuita izvi. Yechitatu inosarudza nzira yayo, nezvimwe.
Kana, kana paive nehuwandu husingagumi hwe data, zvino zvingave zvichikwanisika kudzidzisa imwe network kune zvese zvinogoneka mhedzisiro (zviri pachena, dhata rakadaro harigone kuunganidzwa).
Iyo disentanglement nzira inotitaurira - ngatidzidzisei network kuitira kuti iyo pachayo inogona kusiyanisa pakati pepfungwa. Kuitira kuti iite kunyudzwa zvichibva pavhidhiyo, apo imwe nzvimbo yaizosarudza chiito, imwe yaizosarudza nzvimbo iri pasi nenguva, imwe yaizosarudza urefu hwomunhu, uye imwe yaizosarudza kuti munhurume here. Panguva imwecheteyo, pakudzidziswa, ndinoda kupotsa ndisingakurumidzise network ine pfungwa dzakakosha, asi kuti iratidze uye nzvimbo dzeboka. Kune akati wandei zvinyorwa zvakadaro (zvimwe zvacho , , ) uye kazhinji ivo vanonyanya kufungidzira.
Asi nhungamiro iyi, zvirinani nedzidziso, inofanira kuvhara matambudziko akanyorwa pakutanga.
Kuora kwechifananidzo zvinoenderana nemaparamita "ruvara rwemadziro / ruvara rwepasi / chimiro chechinhu / ruvara rwechinhu / nezvimwe."
Kuora kwechiso zvinoenderana nema parameter "saizi, tsiye, kutaridzika, ruvara rweganda, nezvimwe."
ΠΡΠΎΡΠ΅Π΅
Kune dzimwe dzakawanda, dzisiri dzepasirese, nzvimbo dzinokutendera kuti uderedze dhatabhesi, shanda neakawanda heterogeneous data, nezvimwe.
Attention. Zvichida hazvina musoro kuparadzanisa izvi senzira yakaparadzana. Ingori nzira inosimudzira vamwe. Zvinyorwa zvakawanda zvakatsaurirwa kwaari (,,) Pfungwa yekutarisisa ndeyekusimudzira mhinduro yetiweki zvakananga kune zvakakosha zvinhu panguva yekudzidziswa. Kazhinji nerumwe rudzi rwekutariswa kwekunze, kana netiweki yekunze.
3D simulation. Kana iwe ukaita yakanaka 3D injini, iwe unogona kazhinji kuvhara 90% yedata rekudzidzira nayo (Ndakatoona muenzaniso apo inenge 99% yedata yakavharwa neinjini yakanaka). Kune akawanda mazano uye hacks ekuita sei network yakadzidziswa pane 3D injini kushanda uchishandisa chaiyo data (Fine tuning, kutamisa maitiro, nezvimwewo). Asi kazhinji kugadzira injini yakanaka kune akati wandei mirairo yehukuru yakaoma pane kuunganidza data. Mienzaniso apo injini dzakagadzirwa:
Robhoti kudzidzisa (, )
Kudzidzisa zvidzidzo zvinhu muchitoro (asi mumapurojekiti maviri atakaita, taigona kuita pasina iwo).
Kudzidzira kuTesla (zvakare, vhidhiyo iri pamusoro).
zvakawanikwa
Nyaya yacho yose, neimwe nzira, imhedziso. Zvichida meseji huru yandaida kuita yaive "yemahara yapera, neurons haachape mhinduro dzakareruka." Iye zvino tinofanira kushanda nesimba kuti tiite zvisarudzo zvakaoma. Kana kushanda nesimba kuita tsvakiridzo yesainzi yakaoma.
Kazhinji, musoro wacho unokakavadzana. Zvichida vaverengi vane mimwe mienzaniso inofadza?
Source: www.habr.com