{"id":35293,"date":"2019-10-31T22:03:28","date_gmt":"2019-10-31T19:03:28","guid":{"rendered":"https:\/\/prohoster.info\/blog\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari\/"},"modified":"2019-10-31T22:03:28","modified_gmt":"2019-10-31T19:03:28","slug":"lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari","status":"publish","type":"post","link":"https:\/\/prohoster.info\/et\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari","title":{"rendered":"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<p>Hiljuti ilmus <noindex><a rel=\"nofollow\" href=\"https:\/\/www.getrevue.co\/profile\/peterzhegin\/issues\/ai-investment-activity-trends-of-2018-issue-8-150825?fbclid=IwAR0tU3l4WSotv7pQpYm8PmyKgVbgUgfMLeue_IiV78lXXApH-cy9EcG2kDc\">artikkel<\/a><\/noindex>, mis annab head \u00fclevaadet viimaste aastate masin\u00f5ppe suundumustest. L\u00fchidalt: viimasel kahel aastal on masin\u00f5ppe alaste idufirmade arv j\u00e4rsult langenud.<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/68b58feab2da46b7bb6f412e088313c1.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\nNoh, r\u00e4\u00e4gime \"kas mull on l\u00f5hkenud\", \"kuidas edasi elada\" ja arutame, kust see \u00fcldse selline keerukus tuli.<br \/>\n<noindex><a rel=\"nofollow\" name=\"habracut\"><\/a><\/noindex><br \/>\nEsmalt r\u00e4\u00e4kige, mis oli selle k\u00f5vera t\u00f5ukej\u00f5ud. Kust see p\u00e4rines. K\u00f5ik ilmselt m\u00e4letavad <noindex><a rel=\"nofollow\" href=\"https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\">v\u00f5itu<\/a><\/noindex> masin\u00f5ppes 2012. aastal ImageNeti v\u00f5istlusel. See oli ju esimene globaalne \u00fcritus! Kuid tegelikult pole see nii. Ja k\u00f5vera t\u00f5us algab m\u00f5nev\u00f5rra varem. Jagaksin selle mitmeks hetkeks.<\/p>\n<ol>\n<li>Aasta 2008 oli suurandmete m\u00f5iste tekkimise aastaks. T\u00f5elised tooted hakkasid <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%91%D0%BE%D0%BB%D1%8C%D1%88%D0%B8%D0%B5_%D0%B4%D0%B0%D0%BD%D0%BD%D1%8B%D0%B5\">ilmuma<\/a><\/noindex> 2010. aastast. Suurandmed on tihedalt seotud masin\u00f5ppega. Ilma suurandmeteta ei ole olemasolevate algoritmide stabiilne t\u00f6\u00f6 v\u00f5imalik. Ja see ei ole n\u00e4rviv\u00f5rgud. Enne 2012. aastat olid n\u00e4rviv\u00f5rgud marginaalse v\u00e4hemuse p\u00e4rusmaa. Selle asemel hakkasid t\u00f6\u00f6le hoopis teised algoritmid, mis olid eksisteerinud juba aastaid, isegi aastak\u00fcmneid: <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%9C%D0%B5%D1%82%D0%BE%D0%B4_%D0%BE%D0%BF%D0%BE%D1%80%D0%BD%D1%8B%D1%85_%D0%B2%D0%B5%D0%BA%D1%82%D0%BE%D1%80%D0%BE%D0%B2\">SVM<\/a><\/noindex>(1963, 1993 aastad), <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/Random_forest\">Random Forest<\/a><\/noindex> (1995), <noindex>AdaBoost<\/noindex> (2003),... Selleaegsed idufirmad olid peamiselt seotud struktureeritud andmete automaatse t\u00f6\u00f6tlemisega: kassad, kasutajad, reklaam ja palju muud.\n<p>Selle esimesest lainest tulenev on hulk raamistikku, nagu XGBoost, CatBoost, LightGBM jne.\n<\/li>\n<li>Aastatel 2011-2012 <noindex><a rel=\"nofollow\" href=\"https:\/\/en.wikipedia.org\/wiki\/Convolutional_neural_network\">konvolutsioonilised n\u00e4rviv\u00f5rgud<\/a><\/noindex> on v\u00f5itnud mitmeid pildituvastuse konkurse. Nende tegelik kasutamine viibis m\u00f5nev\u00f5rra. \u00dctleksin, et massiliselt ettevaatlikud idufirmad ja lahendused hakkasid ilmuma alates 2014. aastast. Kahe aasta jooksul \u00f5piti, et n\u00e4rviv\u00f5rgud t\u00f5eliselt t\u00f6\u00f6tavad, loodi mugavad raamistikud, mida oli v\u00f5imalik paigaldada ja k\u00e4ivitada m\u00f5istliku ajaga ning t\u00f6\u00f6tati v\u00e4lja meetodeid, mis stabiliseeriksid ja kiirendaksid konvergentsi aega.\n<p>Konvolutsioonid v\u00f5imaldasid lahendada masin\u00f5ppe \u00fclesandeid: piltide ja objektide klassifitseerimine pildis, objektide tuvastamine, objektide ja inimestega tutvumine, piltide parandamine jne.<\/li>\n<li>2015\u20132017. Algoritmide ja projektides, mis on seotud korduvate v\u00f5rkude v\u00f5i nende analoogidega (LSTM, GRU, TransformerNet jne), toimus t\u00f5eline buum. Ilmusid h\u00e4sti t\u00f6\u00f6tavad \u201ek\u00f5ne-tekstiks\u201d algoritmid ja masint\u00f5lkes\u00fcsteemid. Osa neist p\u00f5hines konvolutsiooniv\u00f5rkudel, et tuvastada p\u00f5hiomadusi. Osaliselt on need loodud suurtest ja kvaliteetsetest andmestikest. <\/li>\n<\/ol>\n<p>\n<img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/d4b3a1dd2483ef1300862f5f61db645a.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\n\u201eMull l\u00f5hkes? Hype on \u00fclekuumenenud? Kas nad on surnud nagu plokiahel?\u201d<br \/>\nJust nimelt! Homme lakkab teie telefonis Siri t\u00f6\u00f6tamast ja \u00fclehomme ei suuda Tesla eristada p\u00f6\u00f6rdeid kengurust.<\/p>\n<p>Neuraalv\u00f5rgud t\u00f6\u00f6tavad juba praegu. Need on k\u00fcmnetes seadmetes, nad v\u00f5imaldavad t\u00f5eliselt teenida raha, muutes turgu ja \u00fcmbritsevat maailma. Hype n\u00e4eb v\u00e4lja veidi teistsugune:<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/a2b271c8eb1cf54fe59389d10cf8e17e.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nLihtsalt neuraalv\u00f5rgud ei ole enam midagi uut. Jah, paljude inimeste ootused on liiga k\u00f5rged. Kuid suur hulk ettev\u00f5tteid on \u00f5ppinud neid rakendama ja looma nende alusel tooteid. Neuraalv\u00f5rgud pakuvad uut funktsionaalsust, v\u00f5imaldavad v\u00e4hendada t\u00f6\u00f6tajate arvu ja langetada teenuste hindu:<\/p>\n<ul>\n<li>Tootmisettev\u00f5tted integreerivad algoritme defektide anal\u00fc\u00fcsiks liinil. <\/li>\n<li>Loomakasvatustalud ostavad s\u00fcsteeme lehmade j\u00e4lgimiseks. <\/li>\n<li> Automaatne l\u00f5ikusmasin. <\/li>\n<li>Automatiseeritud k\u00f5nekeskused.<\/li>\n<li>Filtrid SnapChat'is. (noh, v\u00e4hemalt midagi m\u00f5istlikku!)<\/li>\n<\/ul>\n<p>\nAga peamine, ja mitte nii ilmne: \u201eUusi ideid pole enam, v\u00f5i need ei too kohest kapitali\u201d. Neuraalv\u00f5rgud on lahendanud k\u00fcmneid probleeme. Ja lahendavad veel rohkem. K\u00f5ik ilmsemad ideed, mis olemas olid, on tootnud hulgi alustavaid ettev\u00f5tteid. Kuid k\u00f5ik, mis oli pinnal \u2014 on juba kogutud. Viimase kahe aasta jooksul ei ole ma n\u00e4inud \u00fchtegi uut ideed neuraalv\u00f5rkude rakendamiseks. \u00dchtegi uut l\u00e4henemist (noh, ok, seal on GAN-ide osas veidi segadust).<\/p>\n<p>Ja iga j\u00e4rgmine startup on \u00fcha keerulisem. See n\u00f5uab enam mitte kahte kutti, kes koolitavad neuraalv\u00f5rku avatud andmetel. See n\u00f5uab programmeerijaid, serverit, m\u00e4rkijate meeskonda, keerukamat tuge jne.<\/p>\n<p>Tulemuseks on \u2014 alustavate ettev\u00f5tete arv v\u00e4heneb. Kuid tootmine suureneb. Kas on vaja auto numbrite tuvastamist? Turul on sadu spetsialiste, kellel on asjakohane kogemus. V\u00f5ite palgata ja paari kuuga teie t\u00f6\u00f6taja loob s\u00fcsteemi. V\u00f5i osta valmis lahendus. Kuid uue startupi loomine?.. Hullumeelsus!<\/p>\n<p>Vajame k\u00fclastajate j\u00e4lgimise s\u00fcsteemi - miks maksta hunniku litsentside eest, kui oma mugandatud s\u00fcsteemi saab luua 3-4 kuuga? <\/p>\n<p>Praegu l\u00e4bivad n\u00e4rviv\u00f5rgud sama teed, mida on l\u00e4binud k\u00fcmned teised tehnoloogiad. <\/p>\n<p>Kas m\u00e4letate, kuidas 1995. aastast alates on muutunud m\u00f5isted 'veebiarendaja'? Seni, kuni turg ei ole spetsialiste \u00fcle ujutatud, on professionaale v\u00e4ga v\u00e4he. Kuid ma ei saa vaidelda, et 5-10 aasta p\u00e4rast ei ole vahet Java-kodut\u00f6\u00f6 tegija ja n\u00e4rviv\u00f5rkude arendaja vahel. Nii \u00fcht kui teist spetsialisti on turul piisavalt.<\/p>\n<p>Lihtsalt on ilmne, et on tekkimas \u00fclesanne - palgatakse spetsialist.<\/p>\n<p><b>\"Ja mis edasi? Kus on lubatud tehisintellekt?\"<\/b><\/p>\n<p>Siin on v\u00e4ike, kuid huvitav segadus :)<\/p>\n<p>Praegune tehnoloogiate kogum, mis meil on, ei viida meid ilmselt tehisintellektini. Ideed ja nende uudsus on suuresti ammendunud. R\u00e4\u00e4gime, mis hoiab praegust arengutaset.<\/p>\n<h3>Piirangud<\/h3>\n<p>\nAlustame isejuhtivatest autodest. Tundub, et t\u00e4ielikult autonoomsete autode valmistamine praeguste tehnoloogiatega on v\u00f5imalik. Kuid millal see juhtub - ei ole selge. Tesla arvab, et see juhtub paari aasta p\u00e4rast - <\/p>\n<p><center><div class=\"youtube-placeholder\" data-id=\"Ucp0TTmvqOE\" onclick=\"loadVideo(this)\">\r\n        <img decoding=\"async\" src=\"https:\/\/img.youtube.com\/vi\/Ucp0TTmvqOE\/hqdefault.jpg\" alt=\"M\u00e4ngi videot\" loading=\"lazy\" width=\"480\" height=\"360\" style=\"width:100%;height:auto;\">\r\n        <div class=\"play-button\"><\/div>\r\n    <\/div><\/center><br \/>\nOn palju teisi <noindex><a rel=\"nofollow\" href=\"https:\/\/beth.technology\/truths-autonomous-vehicles\/\">spetsialiste<\/a><\/noindex>, kes hindavad seda 5-10 aastaks. <\/p>\n<p>T\u00f5en\u00e4oliselt, minu arvates, 15 aasta p\u00e4rast muutuvad linnade infrastruktuurid nii, et autonoomsete autode ilmumine muutub v\u00e4ltimatuks, see saab nende j\u00e4tkuks. Kuid seda ei saa pidada intelligentsuseks. Kaasaegne Tesla on v\u00e4ga keeruline andmete filtreerimise, otsimise ja \u00fcmber\u00f5ppe teenus. See on reeglid-reeglid-reeglid, andmete kogumine ja filtrid nende \u00fcle (siin <noindex><a rel=\"nofollow\" href=\"http:\/\/cv-blog.ru\/?p=279\">siit<\/a><\/noindex> olen ma sellest veidi p\u00f5hjalikumalt kirjutanud, v\u00f5i vaata <noindex><a rel=\"nofollow\" href=\"https:\/\/www.youtube.com\/watch?time_continue=7614&amp;v=Ucp0TTmvqOE\">selle<\/a><\/noindex> marki).<\/p>\n<h3>Esimene probleem<\/h3>\n<p>\nJa just siin n\u00e4eme <b>esimest fundamentaalset probleemi<\/b>. Suured andmed. See on see, mis p\u00f5hjustas praeguse laine n\u00e4rviv\u00f5rkude ja masin\u00f5ppe vallas. Praegu, et midagi keerulist ja automaatset teha, on vaja palju andmeid. Mitte lihtsalt palju, vaid v\u00e4ga-v\u00e4ga palju. On vajalikud automatiseeritud algoritmid nende kogumiseks, t\u00e4histamiseks ja kasutamiseks. Kui tahame, et masin n\u00e4eks veokit p\u00e4ikese vastu \u2014 peame esmalt koguma piisavalt palju n\u00e4iteid. Kui tahame, et masin ei l\u00e4heks segadusse jalgrattast, mis on kinnitatud pagasiruumi \u2014 rohkem n\u00e4iteid.<\/p>\n<p>Ja \u00fchte n\u00e4idet ei piisa. Sadu? Tuhandeid? <\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/923ef975804234f1b3dcbfee0143f2b4.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>Teine probleem<\/h3>\n<p>\n<b>Teine probleem <\/b> \u2014 visualiseerimine, mida meie n\u00e4rviv\u00f5rk m\u00f5istis. See on v\u00e4ga mittetriviaalne \u00fclesanne. Siiani on v\u00e4he inimesi, kes m\u00f5istavad, kuidas seda visualiseerida. Need artiklid on \u00fcsna v\u00e4rsked, see on vaid m\u00f5ned n\u00e4ited, isegi kui need on kauged:<br \/>\n<noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/ods\/blog\/453788\/\">Visualiseerimine<\/a><\/noindex> teksteuride peatukesed. N\u00e4itab h\u00e4sti, millele n\u00e4rviv\u00f5rk kipub peatuma + mida ta tajub algteabe allikana.<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/0eab41c3aec71dc2b04794559dd4a651.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<noindex><a rel=\"nofollow\" href=\"http:\/\/jalammar.github.io\/visualizing-neural-machine-translation-mechanics-of-seq2seq-models-with-attention\/\">Visualiseerimine<\/a><\/noindex> t\u00e4helepanu juures <noindex><a rel=\"nofollow\" href=\"http:\/\/www.wildml.com\/2016\/01\/attention-and-memory-in-deep-learning-and-nlp\/\">t\u00f5lgetes<\/a><\/noindex>. T\u00f5eliselt saab t\u00e4helepanu sageli kasutada, et n\u00e4idata, mis p\u00f5hjustas sellist reageeringut v\u00f5rku. Olen selliseid asju kohanud nii silumise kui ka toote lahenduste jaoks. Selle teema kohta on v\u00e4ga palju artikleid. Aga mida keerulisemad on andmed, seda keerulisem on m\u00f5ista, kuidas saavutada stabiilne visualiseerimine.<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/e0c370724115f602e5bd35b20b56f6eb.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nJa jah, vana hea komplekt, \u00abvaata, mis on v\u00f5rgu sees <noindex><a rel=\"nofollow\" href=\"https:\/\/towardsdatascience.com\/how-to-visualize-convolutional-features-in-40-lines-of-code-70b7d87b0030\">filtrites<\/a><\/noindex>\u00bb. Need pildid olid populaarsed umbes 3-4 aastat tagasi, aga k\u00f5ik m\u00f5istsid kiiresti, et pildid on k\u00fcll ilusad, aga sisul on v\u00e4he.<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/87ace90924d5b900f9382ecc6ceef6d0.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nMa ei nimetanud k\u00fcmneid teisi nippe, viise, h\u00e4kke, uuringuid selle kohta, kuidas v\u00f5rgu sisemust kuvada. Kas need t\u00f6\u00f6riistad t\u00f6\u00f6tavad? Kas need aitavad kiiresti m\u00f5ista, mis on probleem ja v\u00f5rku siluda?.. Viimased protsendid v\u00e4lja t\u00f5mmata? No, umbes sama moodi:<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/6da71648c300ee3bc8673d08287b77e3.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nV\u00f5id vaadata \u00fcksk\u00f5ik millisest konkursist Kaggle'is. Ja kirjeldust sellest, kuidas inimesed oma l\u00f5plikke lahendusi teevad. Me ehitasime 100-500-800 miljonit mudelit ja see t\u00f6\u00f6tas!<\/p>\n<p>Ma muidugi liialdan. Aga need l\u00e4henemised ei anna kiireid ja otseseid vastuseid.<\/p>\n<p>Omades piisavalt kogemusi ja proovides erinevaid variante, v\u00f5ib anda otsuse, miks teie s\u00fcsteem nii otsustas. Aga s\u00fcsteemi k\u00e4itumise parandamine on keeruline. Kinnita h\u00e4daabi, t\u00f5sta l\u00e4ve, lisa andmestik, v\u00f5ta teine tagaosa v\u00f5rku.<\/p>\n<h3>Kolmas probleem<\/h3>\n<p>\n<b>Kolmas fundamentaalne probleem <\/b> \u2014 v\u00f5rgud \u00f5petavad mitte loogikat, vaid statistikat. Statistiliselt see <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/417405\/\">n\u00e4gu<\/a><\/noindex>:<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/def14bbc2f40e4e26656a2d8032b09c1.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nLoogiliselt \u2014 see ei tundu eriti sarnane. K\u00fcberv\u00f5rgud ei \u00f5pi midagi keerulist, kui neid ei sunnita. Nad \u00f5pivad alati maksimaalselt lihtsaid jooni. Kas on silmad, nina, pea? Siis on see n\u00e4gu! Too n\u00e4iteks, kus silmad ei t\u00e4henda n\u00e4gu. Ja j\u00e4lle \u2014 miljoneid n\u00e4iteid.<\/p>\n<h3>All on ruudu all<\/h3>\n<p>\nMa \u00fctleksin, et just need kolm globaalse probleemid t\u00e4na piiravad k\u00fcberv\u00f5rkude ja masin\u00f5ppe arengut. Ja see, kus need probleemid ei piiranud \u2014 seda kasutatakse juba aktiivselt.<\/p>\n<p><b>Kas see on l\u00f5pp? K\u00fcberv\u00f5rgud on seisma j\u00e4\u00e4nud?<\/b><\/p>\n<p>Ei ole teada. Aga loomulikult loodavad k\u00f5ik, et ei ole. <\/p>\n<p>On palju l\u00e4henemisviise ja suundi nende fundamentaalsete probleemide lahendamiseks, mida ma eespool k\u00e4sitlesin. Kuid seni ei ole \u00fckski neist l\u00e4henemistest v\u00f5imaldanud teha midagi fundamentaalselt uut, lahendada midagi, mis ei ole seni lahendatud. Seni k\u00f5ik fundamentaalsed projektid on tehtud stabiilsete l\u00e4henemiste p\u00f5hjal (Tesla), v\u00f5i j\u00e4\u00e4vad need katsetusprojektideks instituutide v\u00f5i korporatsioonide (Google Brain, OpenAI) jaoks.<\/p>\n<p>R\u00e4\u00e4kides k\u00f5vasti, siis peamine suund on teatud k\u00f5rgema tasandi esitus sisendandmetest. Teatud m\u00f5ttes \u201cm\u00e4lu\u201d. Lihtsaim n\u00e4ide m\u00e4lust on erinevad \u201cEmbedding\u201d \u2014 piltide esitused. N\u00e4iteks k\u00f5ik n\u00e4otuvastuss\u00fcsteemid. V\u00f5rk \u00f5pib saama n\u00e4ost m\u00f5ningast stabiilset esitatust, mis ei s\u00f5ltu p\u00f6\u00f6rdest, valgustusest, resolutsioonist. Sisuliselt minimiseerib v\u00f5rk m\u00f5\u00f5distust \u201cerinevad n\u00e4od \u2014 kaugel\u201d ja \u201c\u00fchesugused \u2014 l\u00e4hedal\u201d.<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/5e9d871fe096b7a77dbe11a6e315c5e4.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nSelleks \u00f5petamiseks on vaja k\u00fcmneid ja sadu tuhandeid n\u00e4iteid. Kuid tulemus kannab endas teatud algseid jooni \u201cOne-shot Learning\u201d. N\u00fc\u00fcd ei ole meil vaja sadu n\u00e4gusid, et inimest meeles pidada. Ainult \u00fcks n\u00e4gu ja k\u00f5ik \u2014 me <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/davidsandberg\/facenet\">tunneme \u00e4ra<\/a><\/noindex>!<br \/>\nAga probleem on\u2026 V\u00f5rk suudab \u00f5ppida ainult piisavalt lihtsaid objekte. Proovides eristada mitte n\u00e4gusid, vaid n\u00e4iteks \u201cinimest riiete j\u00e4rgi\u201d (\u00fclesanne <noindex><a rel=\"nofollow\" href=\"https:\/\/medium.com\/@alitech_2017\/reforming-person-re-identification-with-local-convolutional-neural-networks-17148f11f17b\">Re-identification<\/a><\/noindex>) \u2014 kvaliteet kukub mitme j\u00e4rgu v\u00f5rra. Ja v\u00f5rk ei suuda enam \u00f5ppida piisavalt ilmselgeid vaatenurga muutusi.<\/p>\n<p>Ja \u00f5ppida miljonitest n\u00e4idetest \u2014 on samuti kuidagi \u00fcsna vaevarikas. <\/p>\n<p>On teoseid m\u00e4rkimisv\u00e4\u00e4rse valimite v\u00e4hendamise kohta. N\u00e4iteks v\u00f5ib kohe meenuda \u00fcks esimesi t\u00f6id <b>OneShot Learning<\/b> <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/pdf\/1605.06065v1.pdf\">Google'ilt<\/a><\/noindex>:<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/da83c6f290248f2bcf963c3053a26688.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nSelliseid teoseid on palju, n\u00e4iteks <noindex><a rel=\"nofollow\" href=\"https:\/\/pdfs.semanticscholar.org\/d1c4\/c4c7989102e85b5248cebfcb0cb000c3b837.pdf\">1<\/a><\/noindex> v\u00f5i <noindex><a rel=\"nofollow\" href=\"https:\/\/www.cs.cmu.edu\/~rsalakhu\/papers\/oneshot1.pdf\">2<\/a><\/noindex> v\u00f5i <noindex><a rel=\"nofollow\" href=\"http:\/\/www.robots.ox.ac.uk\/~tvg\/publications\/2018\/0431.pdf\">3<\/a><\/noindex>.<\/p>\n<p>Miinus \u00fcks \u2013 tavaliselt t\u00f6\u00f6tab \u00f5pe h\u00e4sti lihtsate, \"MNISTi\" n\u00e4idete korral. Kuid keerukamate \u00fclesannete korral on vaja suurt baasi, objektimudelit v\u00f5i mingit v\u00f5lu.<br \/>\n\u00dcldiselt on One-Shot \u00f5pe v\u00e4ga huvitav teema. Leiad palju ideid. Kuid suurem osa neist kahest probleemist, mida mainisin (enne\u00f5ppimine tohutu andmestiku peal \/ keerulistel andmetel ebastabiilsus) segab \u00f5ppet\u00f6\u00f6d t\u00f5siselt.<\/p>\n<p>Teisest k\u00fcljest sobivad Embedding teema jaoks GAN\u2019id \u2013 genereerivad vastasv\u00f5rgud. Olete kindlasti lugenud selle kohta palju artikleid Habr\u2019s.<noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/ods\/blog\/340154\/\">1<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/itsumma\/blog\/447896\/\">2<\/a><\/noindex>,<noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/ods\/blog\/322514\/\">3<\/a><\/noindex>)<br \/>\nGANide erip\u00e4ra on mingisuguse sisemise olekute ruumi (sisuliselt sama, mis Embedding) loomine, mis v\u00f5imaldab joonistada pilti. Need v\u00f5ivad olla <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/shaoanlu\/faceswap-GAN\">n\u00e4od<\/a><\/noindex>, need v\u00f5ivad olla <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/sergeytulyakov\/mocogan\">tegevused<\/a><\/noindex>. <\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/8c25c375dae3559b6895d6c4eb3f6cfd.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nGANide probleem on see, et mida keerulisem on genereeritav objekt, seda keerulisem on seda kirjeldada \"genereerija-diskrimineerija\" loogikas. Seet\u00f5ttu on ainus tuntud rakendus GAN-ide jaoks DeepFake, mis samuti manipuleerib n\u00e4o esitlusega (millest on tohutu andmebaas).<\/p>\n<p>Teisi kasulikke rakendusi olen kohanud \u00fcliharva. Tavaliselt on need lihtsalt mingid asjad piltide t\u00e4iendamiseks.<\/p>\n<p>Ja taas, kellelgi pole arusaamist, kuidas see v\u00f5iks meid viia valgusesse tulevikku. Loogika\/esitluse m\u00f5istmine neuroniv\u00f5rgus on hea. Kuid vajame tohutult n\u00e4iteid, me ei saa aru, kuidas neuroniv\u00f5rk seda endas esitab, me ei saa aru, kuidas panna neuroniv\u00f5rk meelde j\u00e4tma m\u00f5nda tegelikult keerulist esitust.<\/p>\n<p><b>Reinforcement learning<\/b> on t\u00e4iesti teise nurga alt l\u00e4henemine. Te kindlasti m\u00e4letate, kuidas Google v\u00f5itis k\u00f5iki Go m\u00e4ngus. Hiljutised v\u00f5idud Starcraftis ja Dota\u2019s. Kuid siin pole k\u00f5ik kaugeltki nii roosiline ja lootustandev. Parim selgitus RL ja selle raskuste kohta on <noindex><a rel=\"nofollow\" href=\"https:\/\/www.alexirpan.com\/2018\/02\/14\/rl-hard.html\">see artikkel<\/a><\/noindex>.<\/p>\n<p>Kui autorit kokku v\u00f5tta, siis<\/p>\n<ul>\n<li>Lahendused karbis ei sobi \/ t\u00f6\u00f6tavad enamasti halvasti<\/li>\n<li>Praktilisi \u00fclesandeid on lihtsam lahendada muude meetoditega. Boston Dynamics ei kasuta RL-i selle keerukuse \/ ettearvamatuse \/ arvutuste keerukuse t\u00f5ttu<\/li>\n<li>Kuna RL t\u00f6\u00f6le hakkab \u2013 on vaja keerulist funktsiooni. Tihti on selle loomine \/ kirjutamine keeruline.<\/li>\n<li>Mudelite \u00f5petamine on keeruline. Tuleb kulutada palju aega, et need loksutada ja v\u00e4lja tuua kohalikest optimaalseist.<\/li>\n<li>Seet\u00f5ttu on modelleerimine keeruline, mudeli ebastabiilsus v\u00e4iksemate muudatuste korral.<\/li>\n<li>Tihti satub see liiga tihedalt mingitesse vale seadusp\u00e4rasustesse, isegi juhuslike arvugeneraatoriteni v\u00e4lja.<\/li>\n<\/ul>\n<p>\nOluline punkt on see, et RL ei t\u00f6\u00f6ta hetkel tootmises. Google'il on mingid eksperimendid ( <noindex><a rel=\"nofollow\" href=\"https:\/\/ai.google\/research\/teams\/brain\/robotics\/\">1<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/ai.googleblog.com\/2018\/06\/scalable-deep-reinforcement-learning.html\">2<\/a><\/noindex> ). Kuid ma ei ole n\u00e4inud \u00fchtegi tootmisse suunatud s\u00fcsteemi.<\/p>\n<p><b>M\u00e4lu<\/b>. K\u00f5ik, mis eelpool on toodud, n\u00e4itab, et struktuuritus on probleem. \u00dcks l\u00e4henemine, kuidas seda p\u00fc\u00fcda korrastada, on anda n\u00e4rviv\u00f5rgule juurdep\u00e4\u00e4s eraldi m\u00e4lule. Nii saab ta tulemusi salvestada ja \u00fcle kirjutada. Siis saab n\u00e4rviv\u00f5rk m\u00e4\u00e4ratleda oma seisundi praeguse m\u00e4lu j\u00e4rgi. See on v\u00e4ga sarnane klassikaliste protsessorite ja arvutitega.<\/p>\n<p>K\u00f5ige kuulsam ja populaarsem <noindex>artikkel <\/noindex> \u2014 DeepMindist:<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/acc6bcd86c8c071fbd9776a91f990752.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nTundub, et see on m\u00f5istmise v\u00f5ti? Kuid pigem mitte. S\u00fcsteem vajab ikkagi tohutut andmemassi treenimiseks. Ja see t\u00f6\u00f6tab enamasti struktureeritud tabelandmetega. Samal ajal, kui Facebook <noindex><a rel=\"nofollow\" href=\"https:\/\/embodiedqa.org\/\">lahendas <\/a><\/noindex>samalaadset probleemi, l\u00e4ksid nad teed \"unustame m\u00e4lestuse, teeme lihtsalt n\u00e4rviv\u00f5rgu keerulisemaks ja toome rohkem n\u00e4iteid - ja ta \u00f5pib ise.\"<\/p>\n<p><b>Disentanglement<\/b>. Teine viis t\u00e4hendusrikka m\u00e4lu loomiseks on v\u00f5tta samad sisendandmed, kuid treenimisel sisestada t\u00e4iendavad kriteeriumid, mis v\u00f5imaldaksid nendes \"t\u00e4hendusi\" eristada. N\u00e4iteks kui soovime \u00f5petada n\u00e4rviv\u00f5rku eristama inimese k\u00e4itumist poes. Kui me j\u00e4rgiksime standaardset teed - peaksime tegema k\u00fcmneid v\u00f5rke. \u00dcks otsib inimest, teine m\u00e4\u00e4rab, mida ta teeb, kolmas tema vanuse, neljas - soo. Eraldi loogika j\u00e4lgib osa poest, kus ta teeb \\\/ \u00f5pib selle kallal. Kolmas m\u00e4\u00e4rab tema trajektoori jne.<\/p>\n<p>V\u00f5i kui oleks l\u00f5putult andmeid, v\u00f5iks \u00fche v\u00f5rgu \u00f5petada k\u00f5igi v\u00f5imalike tulemuste osas (ilmselgelt ei saa sellist andmemassi kokku korjata).<\/p>\n<p>Disentlement approach tells us \u2014 let\u2019s train the network so that it can distinguish concepts by itself. To have it create an embedding from video, where one area defines actions, one defines positions on the floor in time, one indicates a person's height, and another identifies their gender. Meanwhile, during training, we would like to provide minimal hints about these key concepts and let the network itself highlight and group areas. There are not many articles on such topics (some of them <noindex><a rel=\"nofollow\" href=\"https:\/\/ai.googleblog.com\/2019\/04\/evaluating-unsupervised-learning-of.html\">1<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"http:\/\/papers.nips.cc\/paper\/5851-deep-convolutional-inverse-graphics-network.pdf\">2<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/pdf\/1812.02230.pdf\">3<\/a><\/noindex>) and overall they are quite theoretical. <\/p>\n<p>However, this direction, at least theoretically, should address the problems mentioned at the start.<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/b36046dc3cbe3b849e1d3a60f56ec3a2.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nDecomposing the image by parameters \u201cwall color\/floor color\/object shape\/object color\/etc.\u201d<\/p>\n<p><img decoding=\"async\" alt=\"Kas masin\u00f5ppe mull l\u00f5hkes v\u00f5i algas uus koidik\" src=\"\/wp-content\/uploads\/4d5ac8dfc43f76b35f3490c26049c7db.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nDecomposing the face by parameters \u201csize, eyebrows, orientation, skin color, etc.\u201d<\/p>\n<h3>Muud<\/h3>\n<p>\nThere are many other less global directions that allow for some reduction of databases, working with more heterogeneous data, etc.<\/p>\n<p><b>Attention<\/b>. It probably doesn't make sense to highlight this as a separate method. It\u2019s just an approach that enhances others. There are many articles dedicated to it (<noindex><a rel=\"nofollow\" href=\"http:\/\/www.wildml.com\/2016\/01\/attention-and-memory-in-deep-learning-and-nlp\/\">1<\/a><\/noindex>,<noindex><a rel=\"nofollow\" href=\"http:\/\/jalammar.github.io\/visualizing-neural-machine-translation-mechanics-of-seq2seq-models-with-attention\/\">2<\/a><\/noindex>,<noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/abs\/1706.03762\">3<\/a><\/noindex>). The essence of Attention is to enhance the network's response specifically to significant objects during training. Often this is done with some external target guidance or a small external network.<\/p>\n<p><b>3D simulation<\/b>. If you create a good 3D engine, it can cover about 90% of training data (I've even seen examples where almost 99% of data was covered by a good engine). There are many ideas and hacks to make a network trained on a 3D engine work with real data (fine-tuning, style transfer, etc.). But often creating a good engine is several orders of magnitude more complex than gathering data. Examples when engines were created:<br \/>\nRobot training (<noindex><a rel=\"nofollow\" href=\"https:\/\/ai.googleblog.com\/2018\/06\/teaching-uncalibrated-robots-to_22.html\">google<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/VZcmogKXC18\">braingarden<\/a><\/noindex>)<br \/>\nKoolitus <noindex><a rel=\"nofollow\" href=\"https:\/\/neuromation.io\/\">tuvastamiseks<\/a><\/noindex> goods in a store (but in the two projects we worked on \u2014 we managed without this).<br \/>\nTraining at Tesla (again the video mentioned earlier).<\/p>\n<h2>J\u00e4reldused<\/h2>\n<p>\nThe entire article is in some sense conclusions. Probably, the main message I wanted to convey is \u2014 'the free ride is over, neural networks no longer offer simple solutions.' Now, we need to work hard to build complex solutions. Or work hard conducting intricate scientific research.<\/p>\n<p>Overall, the topic is debatable. Perhaps the readers have more interesting examples?<br \/>\n<br \/>Allikas: <a content=\"nofollow\" rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/recognitor\/blog\/455676\/\">habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"<p>\u041d\u0435\u0434\u0430\u0432\u043d\u043e \u0432\u044b\u0448\u043b\u0430 \u0441\u0442\u0430\u0442\u044c\u044f, \u043a\u043e\u0442\u043e\u0440\u0430\u044f \u043d\u0435\u043f\u043b\u043e\u0445\u043e \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u0435\u0442 \u0442\u0435\u043d\u0434\u0435\u043d\u0446\u0438\u044e \u0432 \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u043c \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u0438 \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0438\u0445 \u043b\u0435\u0442. \u0415\u0441\u043b\u0438 \u043a\u043e\u0440\u043e\u0442\u043a\u043e: \u0447\u0438\u0441\u043b\u043e \u0441\u0442\u0430\u0440\u0442\u0430\u043f\u043e\u0432 \u0432 \u043e\u0431\u043b\u0430\u0441\u0442\u0438 \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u0433\u043e \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f \u0432 \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0438\u0435 \u0434\u0432\u0430 \u0433\u043e\u0434\u0430 \u0440\u0435\u0437\u043a\u043e \u0443\u043f\u0430\u043b\u043e. \u041d\u0443 \u0447\u0442\u043e. \u0420\u0430\u0437\u0431\u0435\u0440\u0451\u043c \u00ab\u043b\u043e\u043f\u043d\u0443\u043b \u043b\u0438 \u043f\u0443\u0437\u044b\u0440\u044c\u00bb, \u00ab\u043a\u0430\u043a \u0434\u0430\u043b\u044c\u0448\u0435 \u0436\u0438\u0442\u044c\u00bb \u0438 \u043f\u043e\u0433\u043e\u0432\u043e\u0440\u0438\u043c \u043e\u0442\u043a\u0443\u0434\u0430 \u0432\u043e\u043e\u0431\u0449\u0435 \u0442\u0430\u043a\u0430\u044f \u0437\u0430\u0433\u043e\u0433\u0443\u043b\u0438\u043d\u0430. \u0414\u043b\u044f \u043d\u0430\u0447\u0430\u043b\u0430 \u043f\u043e\u0433\u043e\u0432\u043e\u0440\u0438\u043c \u0447\u0442\u043e \u0431\u044b\u043b\u043e \u0431\u0443\u0441\u0442\u0435\u0440\u043e\u043c \u044d\u0442\u043e\u0439 \u043a\u0440\u0438\u0432\u043e\u0439. \u041e\u0442\u043a\u0443\u0434\u0430 \u043e\u043d\u0430 \u0432\u0437\u044f\u043b\u0430\u0441\u044c. \u041d\u0430\u0432\u0435\u0440\u043d\u043e\u0435 \u0432\u0441\u0451 \u0432\u0441\u043f\u043e\u043c\u043d\u044f\u0442 [&hellip;]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[702],"tags":[],"class_list":["post-35293","post","type-post","status-publish","format-standard","hentry","category-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 5.0.1.1 - aioseo.com -->\n\t<meta name=\"description\" content=\"\u041d\u0435\u0434\u0430\u0432\u043d\u043e \u0432\u044b\u0448\u043b\u0430\" \/>\n\t<meta name=\"robots\" content=\"max-image-preview:large\" \/>\n\t<meta name=\"author\" content=\"Yuri Gagarin\"\/>\n\t<link rel=\"canonical\" href=\"https:\/\/prohoster.info\/et\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari\" \/>\n\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.1.1\" \/>\n\t\t<meta property=\"og:locale\" content=\"et_EE\" \/>\n\t\t<meta property=\"og:site_name\" content=\"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b\" \/>\n\t\t<meta property=\"og:type\" content=\"article\" \/>\n\t\t<meta property=\"og:title\" content=\"\ud83e\udd47\u041b\u043e\u043f\u043d\u0443\u043b \u043b\u0438 \u043f\u0443\u0437\u044b\u0440\u044c \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u0433\u043e \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f, \u0438\u043b\u0438 \u043d\u0430\u0447\u0430\u043b\u043e \u043d\u043e\u0432\u043e\u0439 \u0437\u0430\u0440\u0438 | ProHoster\" \/>\n\t\t<meta property=\"og:description\" content=\"\u041d\u0435\u0434\u0430\u0432\u043d\u043e \u0432\u044b\u0448\u043b\u0430\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/prohoster.info\/et\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari\" \/>\n\t\t<meta property=\"og:image\" content=\"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg\" \/>\n\t\t<meta property=\"og:image:secure_url\" content=\"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg\" \/>\n\t\t<meta property=\"og:image:width\" content=\"350\" \/>\n\t\t<meta property=\"og:image:height\" content=\"350\" \/>\n\t\t<meta property=\"article:published_time\" content=\"2019-10-31T19:03:28+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2019-10-31T19:03:28+00:00\" \/>\n\t\t<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/prohoster\" \/>\n\t\t<meta property=\"article:author\" content=\"https:\/\/www.facebook.com\/prohoster\" \/>\n\t\t<!-- All in One SEO -->\n\n","aioseo_head_json":{"title":"\ud83e\udd47Has the bubble of machine learning burst, or is it the beginning of a new dawn | ProHoster","description":"Hiljuti ilmus","canonical_url":"https:\/\/prohoster.info\/et\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"et_EE","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47\u041b\u043e\u043f\u043d\u0443\u043b \u043b\u0438 \u043f\u0443\u0437\u044b\u0440\u044c \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u0433\u043e \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f, \u0438\u043b\u0438 \u043d\u0430\u0447\u0430\u043b\u043e \u043d\u043e\u0432\u043e\u0439 \u0437\u0430\u0440\u0438 | ProHoster","og:description":"\u041d\u0435\u0434\u0430\u0432\u043d\u043e \u0432\u044b\u0448\u043b\u0430","og:url":"https:\/\/prohoster.info\/et\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2019-10-31T19:03:28+00:00","article:modified_time":"2019-10-31T19:03:28+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"35293","title":null,"description":null,"keywords":null,"keyphrases":null,"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":null,"og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"","isEnabled":true},"graphs":[]},"schema_type":null,"schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":null,"robots_max_videopreview":null,"robots_max_imagepreview":"large","priority":null,"frequency":null,"local_seo":null,"seo_analyzer_scan_date":"2026-01-21 22:41:19","breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-03-01 02:06:24","updated":"2026-01-21 22:41:19","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/posts\/35293","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/comments?post=35293"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/posts\/35293\/revisions"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/media?parent=35293"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/categories?post=35293"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/et\/wp-json\/wp\/v2\/tags?post=35293"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}