{"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\/cs\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari","title":{"rendered":"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<p>Ned\u00e1vno vy\u0161lo <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\">neaktu\u00e1ln\u00ed statistika<\/a><\/noindex>, kter\u00e9 dob\u0159e ukazuje trendy ve strojov\u00e9m u\u010den\u00ed posledn\u00edch let. Stru\u010dn\u011b \u0159e\u010deno: po\u010det startup\u016f v oblasti strojov\u00e9ho u\u010den\u00ed v posledn\u00edch dvou letech rapidn\u011b klesl.<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/68b58feab2da46b7bb6f412e088313c1.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\nNo co. Pod\u00edv\u00e1me se na to, zda \u00abpraskl bubble\u00bb, \u00abjak d\u00e1l \u017e\u00edt\u00bb a probereme, odkud se to v\u0161echno vzalo.<br \/>\n<noindex><a rel=\"nofollow\" name=\"habracut\"><\/a><\/noindex><br \/>\nNejprve si pov\u00edme, co bylo t\u00edm impulsem tohoto trendu. Odkud p\u0159i\u0161el. Snad si v\u0161ichni vzpomenou <noindex><a rel=\"nofollow\" href=\"https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\">na v\u00edt\u011bzstv\u00ed<\/a><\/noindex> strojov\u00e9ho u\u010den\u00ed v roce 2012 v sout\u011b\u017ei ImageNet. P\u0159ece jen to bylo prvn\u00ed glob\u00e1ln\u00ed ud\u00e1lost! Ale ve skute\u010dnosti to tak nen\u00ed. A r\u016fst trendu za\u010d\u00edn\u00e1 o n\u011bco d\u0159\u00edve. Rozd\u011blil bych ho na n\u011bkolik moment\u016f.<\/p>\n<ol>\n<li>Rok 2008 znamenal vznik term\u00ednu \u201evelk\u00e1 data\u201c. Skute\u010dn\u00e9 produkty za\u010daly <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\">objevit<\/a><\/noindex> od roku 2010. Velk\u00e1 data jsou p\u0159\u00edmo spojena se strojov\u00fdm u\u010den\u00edm. Bez velk\u00fdch dat nen\u00ed stabiln\u00ed fungov\u00e1n\u00ed algoritm\u016f, kter\u00e9 v t\u00e9 dob\u011b existovaly. A to nejsou neuronov\u00e9 s\u00edt\u011b. Do roku 2012 byly neuronov\u00e9 s\u00edt\u011b z\u00e1le\u017eitost\u00ed margin\u00e1ln\u00ed men\u0161iny. Zato v t\u00e9 dob\u011b za\u010daly fungovat zcela jin\u00e9 algoritmy, kter\u00e9 existovaly u\u017e l\u00e9ta, nebo dokonce desetilet\u00ed: <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), <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/Random_forest\">Random Forest<\/a><\/noindex> (1995), <noindex>AdaBoost<\/noindex> (2003),\u2026 Startupy t\u00e9 doby byly p\u0159ev\u00e1\u017en\u011b spojeny s automatizovan\u00fdm zpracov\u00e1n\u00edm strukturovan\u00fdch dat: pokladny, u\u017eivatel\u00e9, reklama, a mnoh\u00e9 dal\u0161\u00ed.\n<p>Odvozenina t\u00e9to prvn\u00ed vlny \u2014 soubor framework\u016f, jako jsou XGBoost, CatBoost, LightGBM, atd.\n<\/li>\n<li>V letech 2011-2012 <noindex><a rel=\"nofollow\" href=\"https:\/\/en.wikipedia.org\/wiki\/Convolutional_neural_network\">konvolu\u010dn\u00ed neuronov\u00e9 s\u00edt\u011b<\/a><\/noindex> vyhr\u00e1ly s\u00e9rii sout\u011b\u017e\u00ed v rozpozn\u00e1v\u00e1n\u00ed obraz\u016f. Jejich skute\u010dn\u00e9 pou\u017eit\u00ed se trochu zpozdilo. \u0158ekl bych, \u017ee masov\u011b zam\u011b\u0159en\u00e9 startupy a \u0159e\u0161en\u00ed za\u010dala vznikat a\u017e od roku 2014. Dva roky trvalo, ne\u017e se str\u00e1vil fakt, \u017ee neuronky opravdu funguj\u00ed, vytvo\u0159it u\u017eivatelsky p\u0159\u00edv\u011btiv\u00e9 frameworky, kter\u00e9 mohly b\u00fdt nainstalov\u00e1ny a spu\u0161t\u011bny v rozumn\u00e9m \u010dase, vyvinout metody, kter\u00e9 by stabilizovaly a zrychlily \u010das konvergence.\n<p>Konvolu\u010dn\u00ed s\u00edt\u011b umo\u017enily \u0159e\u0161it \u00fakoly strojov\u00e9ho vid\u011bn\u00ed: klasifikace obraz\u016f a objekt\u016f na obr\u00e1zku, detekce objekt\u016f, rozpozn\u00e1v\u00e1n\u00ed objekt\u016f a lid\u00ed, zlep\u0161ov\u00e1n\u00ed obraz\u016f, atd., atp.<\/li>\n<li>Roky 2015-2017. Boom algoritm\u016f a projekt\u016f zalo\u017een\u00fdch na rekurentn\u00edch s\u00edt\u00edch nebo jejich analog\u00edch (LSTM, GRU, TransformerNet atd.). Objevily se dob\u0159e funguj\u00edc\u00ed algoritmy pro \"\u0159e\u010d-na-text\" a syst\u00e9my strojov\u00e9ho p\u0159ekladu. \u010c\u00e1ste\u010dn\u011b jsou zalo\u017een\u00e9 na konvolu\u010dn\u00edch s\u00edt\u00edch pro vytv\u00e1\u0159en\u00ed z\u00e1kladn\u00edch prvk\u016f. \u010c\u00e1ste\u010dn\u011b na tom, \u017ee se n\u00e1m poda\u0159ilo shrom\u00e1\u017edit opravdu velk\u00e9 a kvalitn\u00ed datov\u00e9 sady. <\/li>\n<\/ol>\n<p>\n<img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/d4b3a1dd2483ef1300862f5f61db645a.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\n\"Bublina praskla? Hype je p\u0159eh\u0159\u00e1t\u00fd? Zem\u0159eli jako blockchain?\"<br \/>\nJasn\u011b! Z\u00edtra na va\u0161em telefonu p\u0159estane fungovat Siri a poz\u00edt\u0159\u00ed Tesla neodli\u0161\u00ed zat\u00e1\u010dku od klokana.<\/p>\n<p>Neuronov\u00e9 s\u00edt\u011b u\u017e funguj\u00ed. Jsou v des\u00edtk\u00e1ch za\u0159\u00edzen\u00ed. Opravdu umo\u017e\u0148uj\u00ed vyd\u011bl\u00e1vat, m\u011bn\u00ed trh a okoln\u00ed sv\u011bt. Hype vypad\u00e1 pon\u011bkud jinak:<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/a2b271c8eb1cf54fe59389d10cf8e17e.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nJednodu\u0161e \u0159e\u010deno, neuronov\u00e9 s\u00edt\u011b u\u017e p\u0159estaly b\u00fdt n\u011b\u010d\u00edm nov\u00fdm. Ano, mnoho lid\u00ed m\u00e1 p\u0159ehnan\u00e1 o\u010dek\u00e1v\u00e1n\u00ed. Ale velk\u00e9 mno\u017estv\u00ed firem se nau\u010dilo aplikovat neuronov\u00e9 s\u00edt\u011b a vytv\u00e1\u0159et produkty na jejich z\u00e1klad\u011b. Neuronov\u00e9 s\u00edt\u011b poskytuj\u00ed nov\u00e9 funkce, umo\u017e\u0148uj\u00ed zkr\u00e1tit pracovn\u00ed m\u00edsta a sn\u00ed\u017eit ceny slu\u017eeb:<\/p>\n<ul>\n<li>V\u00fdrobn\u00ed firmy integruj\u00ed algoritmy pro anal\u00fdzu chyb na link\u00e1ch. <\/li>\n<li>Hospod\u00e1\u0159stv\u00ed chovatele dobytka kupuj\u00ed syst\u00e9my pro kontrolu krav. <\/li>\n<li> Automatick\u00e9 kombajny. <\/li>\n<li>Automatizovan\u00e9 call centr\u00e1ly.<\/li>\n<li>Filtry v Snapchatu. (aspo\u0148 n\u011bco u\u017eite\u010dn\u00e9ho!)<\/li>\n<\/ul>\n<p>\nAle hlavn\u00ed, a ne tak z\u0159ejm\u00e9: \"Nov\u00e9 n\u00e1pady u\u017e nejsou, nebo nep\u0159inesou okam\u017eit\u00fd kapit\u00e1l.\" Neuronov\u00e9 s\u00edt\u011b vy\u0159e\u0161ily des\u00edtky probl\u00e9m\u016f. A vy\u0159e\u0161\u00ed je\u0161t\u011b v\u00edce. V\u0161echny z\u0159ejm\u00e9 n\u00e1pady, kter\u00e9 existovaly, daly vznik mnoha startup\u016fm. Ale v\u0161e, co bylo na povrchu, u\u017e bylo sebr\u00e1no. Za posledn\u00ed dva roky jsem nepotkal \u017e\u00e1dn\u00fd nov\u00fd n\u00e1pad pro aplikaci neuronov\u00fdch s\u00edt\u00ed. \u017d\u00e1dn\u00fd nov\u00fd p\u0159\u00edstup (no, ok, tam jsou u GAN\u016f n\u011bjak\u00e9 pot\u00ed\u017ee).<\/p>\n<p>A ka\u017ed\u00fd n\u00e1sleduj\u00edc\u00ed startup je st\u00e1le slo\u017eit\u011bj\u0161\u00ed. U\u017e to nevy\u017eaduje dva kluky, kte\u0159\u00ed tr\u00e9nuj\u00ed neuronovou s\u00ed\u0165 na otev\u0159en\u00fdch datech. Vy\u017eaduje program\u00e1tory, servery, t\u00fdm pro ozna\u010dov\u00e1n\u00ed, slo\u017eitou podporu atd.<\/p>\n<p>Jako v\u00fdsledek \u2014 startup\u016f ub\u00fdv\u00e1. Ale produkce roste. Pot\u0159ebujete implementovat rozpozn\u00e1v\u00e1n\u00ed SPZ? Na trhu je stovky odborn\u00edk\u016f s relevantn\u00edmi zku\u0161enostmi. M\u016f\u017eete najmout a za p\u00e1r m\u011bs\u00edc\u016f v\u00e1\u0161 zam\u011bstnanec ud\u011bl\u00e1 syst\u00e9m. Nebo koupit hotov\u00fd. Ale zakl\u00e1dat nov\u00fd startup?.. \u0160\u00edlenstv\u00ed!<\/p>\n<p>Je t\u0159eba vytvo\u0159it syst\u00e9m sledov\u00e1n\u00ed n\u00e1v\u0161t\u011bvn\u00edk\u016f - pro\u010d platit za spoustu licenc\u00ed, kdy\u017e si m\u016f\u017eete za 3-4 m\u011bs\u00edce vytvo\u0159it vlastn\u00ed, p\u0159izp\u016fsobenou pro sv\u00e9 podnik\u00e1n\u00ed. <\/p>\n<p>Nyn\u00ed proch\u00e1zej\u00ed neuronov\u00e9 s\u00edt\u011b stejnou cestou, jakou pro\u0161lo mnoho jin\u00fdch technologi\u00ed. <\/p>\n<p>Pamatujete, jak se od roku 1995 m\u011bnil pojem \"v\u00fdvoj\u00e1\u0159 webov\u00fdch str\u00e1nek\"? Dokud nen\u00ed trh nasycen specialisty. Profesion\u00e1l\u016f je velmi m\u00e1lo. Ale mohu se vsadit, \u017ee za 5-10 let nebude mezi program\u00e1torem v Jav\u011b a v\u00fdvoj\u00e1\u0159em neuronov\u00fdch s\u00edt\u00ed velk\u00fd rozd\u00edl. Ob\u011b profese budou na trhu dostate\u010dn\u011b zastoupeny.<\/p>\n<p>Prost\u011b budou existovat druhy \u00fakol\u016f, pro kter\u00e9 se neuronov\u00e9 s\u00edt\u011b vyu\u017e\u00edvaj\u00ed. Objevil se \u00fakol - naj\u00edm\u00e1te specialistu.<\/p>\n<p><b>\"A co d\u00e1l? Kde je slibovan\u00e1 um\u011bl\u00e1 inteligence?\"<\/b><\/p>\n<p>Tady je mal\u00e1, ale zaj\u00edmav\u00e1 nejasnost :)<\/p>\n<p>Ten technologick\u00fd stack, kter\u00fd existuje dnes, zjevn\u011b n\u00e1s k um\u011bl\u00e9 inteligenci nepovede. My\u0161lenky, jejich novost - do zna\u010dn\u00e9 m\u00edry vy\u010derpaly sv\u00e9 mo\u017enosti. Poj\u010fme si promluvit o tom, co dr\u017e\u00ed sou\u010dasnou \u00farove\u0148 v\u00fdvoje.<\/p>\n<h3>Omezen\u00ed<\/h3>\n<p>\nZa\u010dneme s autonomn\u00edmi vozidly. Zd\u00e1 se, \u017ee je jasn\u00e9, \u017ee vytvo\u0159it pln\u011b autonomn\u00ed automobily p\u0159i dne\u0161n\u00edch technologi\u00edch je mo\u017en\u00e9. Ale za jak dlouho se to stane - nen\u00ed jasn\u00e9. Tesla tvrd\u00ed, \u017ee se to stane za p\u00e1r let - <\/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=\"P\u0159ehr\u00e1t video\" loading=\"lazy\" width=\"480\" height=\"360\" style=\"width:100%;height:auto;\">\r\n        <div class=\"play-button\"><\/div>\r\n    <\/div><\/center><br \/>\nExistuje mnoho dal\u0161\u00edch <noindex><a rel=\"nofollow\" href=\"https:\/\/beth.technology\/truths-autonomous-vehicles\/\">specialist\u016f<\/a><\/noindex>, kte\u0159\u00ed to hodnot\u00ed na 5-10 let. <\/p>\n<p>Podle m\u00e9ho n\u00e1zoru, za 15 let se infrastruktura m\u011bst ji\u017e sama zm\u011bn\u00ed tak, \u017ee objev autonomn\u00edch automobil\u016f se stane nevyhnuteln\u00fdm, stane se jejich pokra\u010dov\u00e1n\u00edm. Ale to p\u0159ece nelze pova\u017eovat za inteligenci. Modern\u00ed Tesla je velmi slo\u017eit\u00fd stroj na filtrov\u00e1n\u00ed dat, jejich vyhled\u00e1v\u00e1n\u00ed a p\u0159eu\u010dov\u00e1n\u00ed. Jsou to pravidla-pravidla-pravidla, sb\u011br dat a filtry nad nimi (tohle <noindex><a rel=\"nofollow\" href=\"http:\/\/cv-blog.ru\/?p=279\">tady<\/a><\/noindex> jsem o tom napsal podrobn\u011bji, nebo se m\u016f\u017eete pod\u00edvat na <noindex><a rel=\"nofollow\" href=\"https:\/\/www.youtube.com\/watch?time_continue=7614&amp;v=Ucp0TTmvqOE\">tato<\/a><\/noindex> pozn\u00e1mky).<\/p>\n<h3>Prvn\u00ed probl\u00e9m<\/h3>\n<p>\nA pr\u00e1v\u011b zde vid\u00edme <b>prvn\u00ed z\u00e1sadn\u00ed probl\u00e9m<\/b>. Velk\u00e9 data. To je p\u0159esn\u011b to, co vyvolalo aktu\u00e1ln\u00ed vlnu neuronov\u00fdch s\u00edt\u00ed a strojov\u00e9ho u\u010den\u00ed. Nyn\u00ed, abychom vytvo\u0159ili n\u011bco slo\u017eit\u00e9ho a automatick\u00e9ho, pot\u0159ebujeme hodn\u011b dat. Ne jen hodn\u011b, ale opravdu velmi mnoho. Pot\u0159ebujeme automatizovan\u00e9 algoritmy pro jejich sb\u011br, ozna\u010dov\u00e1n\u00ed a pou\u017eit\u00ed. Chceme, aby stroj vid\u011bl n\u00e1kla\u010f\u00e1ky proti slunci - mus\u00edme nejprve shrom\u00e1\u017edit dostate\u010dn\u00fd jejich po\u010det. Chceme, aby stroj nezbl\u00e1znil z kola p\u0159ipevn\u011bn\u00e9ho k nosi\u010di - pot\u0159ebujeme v\u00edce vzork\u016f.<\/p>\n<p>A jedna uk\u00e1zka nesta\u010d\u00ed. Stovky? Tis\u00edce? <\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/923ef975804234f1b3dcbfee0143f2b4.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>Druh\u00fd probl\u00e9m<\/h3>\n<p>\n<b>Druh\u00fd probl\u00e9m <\/b> \u2014 vizualizace toho, co na\u0161e neuronov\u00e1 s\u00ed\u0165 pochopila. To je velmi netrivi\u00e1ln\u00ed \u00fakol. Dosud m\u00e1lo kdo rozum\u00ed, jak to vizualizovat. Tyto \u010dl\u00e1nky jsou velmi \u010derstv\u00e9, to jsou pouze n\u011bkter\u00e9 p\u0159\u00edklady, i kdy\u017e vzd\u00e1len\u00e9:<br \/>\n<noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/ods\/blog\/453788\/\">Vizualizace<\/a><\/noindex> cyklickosti na texturn\u00ed prvky. Dob\u0159e ukazuje, na \u010dem se neuronov\u00e1 s\u00ed\u0165 r\u00e1da cykl\u00ed + co vn\u00edm\u00e1 jako v\u00fdchoz\u00ed informace.<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" 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\/\">Vizualizace<\/a><\/noindex> pozornost p\u0159i <noindex><a rel=\"nofollow\" href=\"http:\/\/www.wildml.com\/2016\/01\/attention-and-memory-in-deep-learning-and-nlp\/\">p\u0159ekladech<\/a><\/noindex>. Opravdu lze pozornost \u010dasto pou\u017e\u00edt pr\u00e1v\u011b k tomu, aby uk\u00e1zala, co vyvolalo takovou reakci s\u00edt\u011b. Takov\u00e9 n\u00e1stroje jsem potk\u00e1val jak pro debugov\u00e1n\u00ed, tak pro produktov\u00e1 \u0159e\u0161en\u00ed. Na toto t\u00e9ma existuje opravdu hodn\u011b \u010dl\u00e1nk\u016f. Ale \u010d\u00edm slo\u017eit\u011bj\u0161\u00ed jsou data, t\u00edm obt\u00ed\u017en\u011bj\u0161\u00ed je pochopit, jak dos\u00e1hnout stabiln\u00ed vizualizace.<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/e0c370724115f602e5bd35b20b56f6eb.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nNu a ano, star\u00fd dobr\u00fd soubor z \u201epod\u00edvej se, co je uvnit\u0159 s\u00edt\u011b v <noindex><a rel=\"nofollow\" href=\"https:\/\/towardsdatascience.com\/how-to-visualize-convolutional-features-in-40-lines-of-code-70b7d87b0030\">filtrech<\/a><\/noindex>\u201c. Tyto obr\u00e1zky byly popul\u00e1rn\u00ed p\u0159ibli\u017en\u011b p\u0159ed 3-4 lety, ale v\u0161ichni rychle pochopili, \u017ee obr\u00e1zky jsou sice kr\u00e1sn\u00e9, ale nemaj\u00ed moc v\u00fdznam.<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/87ace90924d5b900f9382ecc6ceef6d0.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nNezm\u00ednil jsem des\u00edtky dal\u0161\u00edch zlep\u0161en\u00ed, zp\u016fsob\u016f, hack\u016f, v\u00fdzkum\u016f o tom, jak zobrazit vnit\u0159nosti s\u00edt\u011b. Funguj\u00ed tyto n\u00e1stroje? Pom\u00e1haj\u00ed rychle pochopit, v \u010dem je probl\u00e9m a debuggovat s\u00ed\u0165?.. Z\u00edskat posledn\u00ed procenta? No, p\u0159ibli\u017en\u011b takto:<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/6da71648c300ee3bc8673d08287b77e3.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nM\u016f\u017eete se pod\u00edvat na jakoukoli sout\u011b\u017e na Kaggle. A na popis toho, jak lid\u00e9 vytv\u00e1\u0159ej\u00ed kone\u010dn\u00e9 \u0159e\u0161en\u00ed. Nahromadili jsme 100-500-800 milion\u016f model\u016f a ono to fungovalo!<\/p>\n<p>Samoz\u0159ejm\u011b, p\u0159eh\u00e1n\u00edm. Ale rychl\u00e9 a p\u0159\u00edm\u00e9 odpov\u011bdi tyto p\u0159\u00edstupy ned\u00e1vaj\u00ed.<\/p>\n<p>M\u00edt dostatek zku\u0161enost\u00ed a prozkoumat r\u016fzn\u00e9 varianty m\u016f\u017eete vydat verdikt o tom, pro\u010d v\u00e1\u0161 syst\u00e9m p\u0159ijal takov\u00e9 rozhodnut\u00ed. Ale upravit chov\u00e1n\u00ed syst\u00e9mu bude slo\u017eit\u00e9. Nasadit z\u00e1platu, posunout pr\u00e1h, p\u0159idat datov\u00fd soubor, vz\u00edt jinou backendovou s\u00ed\u0165.<\/p>\n<h3>T\u0159et\u00ed probl\u00e9m<\/h3>\n<p>\n<b>T\u0159et\u00ed fundament\u00e1ln\u00ed probl\u00e9m <\/b> \u2014 s\u00edt\u011b se neu\u010d\u00ed logiku, ale statistiku. Statisticky to <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/417405\/\">tv\u00e1\u0159<\/a><\/noindex>:<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/def14bbc2f40e4e26656a2d8032b09c1.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nLogicky to p\u0159\u00edli\u0161 nevypad\u00e1. Neuronov\u00e9 s\u00edt\u011b se neu\u010d\u00ed nic slo\u017eit\u00e9ho, pokud je k tomu nenut\u00edme. V\u017edy se u\u010d\u00ed maxim\u00e1ln\u011b jednoduch\u00e9 rysy. Jsou tam o\u010di, nos, hlava? Pak je to tv\u00e1\u0159! Nebo p\u0159ive\u010f p\u0159\u00edklad, kde by o\u010di neznamenaly tv\u00e1\u0159. A op\u011bt \u2014 miliony p\u0159\u00edklad\u016f.<\/p>\n<h3>Na dn\u011b je spousta m\u00edsta<\/h3>\n<p>\n\u0158ekl bych, \u017ee pr\u00e1v\u011b tyto t\u0159i glob\u00e1ln\u00ed probl\u00e9my dnes omezuj\u00ed v\u00fdvoj neuronov\u00fdch s\u00edt\u00ed a strojov\u00e9ho u\u010den\u00ed. Tam, kde tyto probl\u00e9my neomezovaly, se ji\u017e aktivn\u011b vyu\u017e\u00edvaj\u00ed.<\/p>\n<p><b>Je to konec? Neuronov\u00e9 s\u00edt\u011b se zastavily?<\/b><\/p>\n<p>Nev\u00edm. Ale samoz\u0159ejm\u011b, v\u0161ichni doufaj\u00ed, \u017ee ne. <\/p>\n<p>Existuje mnoho p\u0159\u00edstup\u016f a sm\u011br\u016f k \u0159e\u0161en\u00ed t\u011bch fundament\u00e1ln\u00edch probl\u00e9m\u016f, kter\u00e9 jsem osv\u011btlil v\u00fd\u0161e. Ale zat\u00edm \u017e\u00e1dn\u00fd z t\u011bchto p\u0159\u00edstup\u016f nedovolil ud\u011blat n\u011bco fundament\u00e1ln\u011b nov\u00e9ho, vy\u0159e\u0161it n\u011bco, co dosud nebylo vy\u0159e\u0161eno. Zat\u00edm jsou v\u0161echny fundament\u00e1ln\u00ed projekty zalo\u017eeny na stabiln\u00edch p\u0159\u00edstupech (Tesla) nebo z\u016fst\u00e1vaj\u00ed testovac\u00edmi projekty instituc\u00ed nebo korporac\u00ed (Google Brain, OpenAI).<\/p>\n<p>Pokud bych m\u011bl hovo\u0159it hrub\u011b, hlavn\u00edm sm\u011brem je vytv\u00e1\u0159en\u00ed ur\u010dit\u00e9ho vysok\u00e9ho \u00farov\u0148ov\u00e9ho zn\u00e1zorn\u011bn\u00ed vstupn\u00edch dat. V ur\u010dit\u00e9m smyslu \u201epam\u011bti\u201c. Nejjednodu\u0161\u0161\u00ed p\u0159\u00edklad pam\u011bti jsou r\u016fzn\u00e9 \u201eEmbedding\u201c \u2014 reprezentace obr\u00e1zk\u016f. Nap\u0159\u00edklad v\u0161echny syst\u00e9my rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed. S\u00ed\u0165 se u\u010d\u00ed z\u00edsk\u00e1vat ze tv\u00e1\u0159e ur\u010dit\u00e9 stabiln\u00ed zn\u00e1zorn\u011bn\u00ed, kter\u00e9 nez\u00e1vis\u00ed na oto\u010den\u00ed, osv\u011btlen\u00ed, rozli\u0161en\u00ed. V podstat\u011b s\u00ed\u0165 minimalizuje metriku \u201eodli\u0161n\u00e9 tv\u00e1\u0159e \u2014 daleko\u201c a \u201estejn\u00e9 \u2014 bl\u00edzko\u201c.<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/5e9d871fe096b7a77dbe11a6e315c5e4.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nPro takov\u00e9 u\u010den\u00ed jsou pot\u0159eba des\u00edtky a stovky tis\u00edc p\u0159\u00edklad\u016f. Ale v\u00fdsledek nese ur\u010dit\u00e9 n\u00e1znaky \u201eOne-shot Learning\u201c. Nyn\u00ed nepot\u0159ebujeme stovky tv\u00e1\u0159\u00ed, abychom si zapamatovali \u010dlov\u011bka. Sta\u010d\u00ed pouze jedna tv\u00e1\u0159, a v\u0161e \u2014 my <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/davidsandberg\/facenet\">rozpozn\u00e1v\u00e1me<\/a><\/noindex>!<br \/>\nAle je tu probl\u00e9m\u2026 S\u00ed\u0165 se m\u016f\u017ee nau\u010dit pouze dost jednoduch\u00e9 objekty. P\u0159i pokusu rozli\u0161ovat ne tv\u00e1\u0159e, ale nap\u0159\u00edklad \u201elidi podle oble\u010den\u00ed\u201c (\u00fakol <noindex><a rel=\"nofollow\" href=\"https:\/\/medium.com\/@alitech_2017\/reforming-person-re-identification-with-local-convolutional-neural-networks-17148f11f17b\">Re-indentification<\/a><\/noindex>) \u2014 kvalita se propad\u00e1 o mnoho \u0159\u00e1d\u016f. A s\u00ed\u0165 ji\u017e nem\u016f\u017ee nau\u010dit dost z\u0159ejm\u00e9 zm\u011bny \u00fahl\u016f.<\/p>\n<p>A u\u010dit se na milionech p\u0159\u00edklad\u016f \u2014 tak\u00e9 takov\u00e9 si to u\u017e\u00edv\u00e1n\u00ed. <\/p>\n<p>Existuj\u00ed pr\u00e1ce na v\u00fdznamn\u00e9m zmen\u0161en\u00ed v\u00fdb\u011bru. Nap\u0159\u00edklad si hned vzpomenu na jednu z prvn\u00edch prac\u00ed na <b>OneShot Learning<\/b> <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/pdf\/1605.06065v1.pdf\">od Googlu<\/a><\/noindex>:<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/da83c6f290248f2bcf963c3053a26688.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nTakov\u00fdch prac\u00ed je mnoho, nap\u0159\u00edklad <noindex><a rel=\"nofollow\" href=\"https:\/\/pdfs.semanticscholar.org\/d1c4\/c4c7989102e85b5248cebfcb0cb000c3b837.pdf\">1<\/a><\/noindex> nebo <noindex><a rel=\"nofollow\" href=\"https:\/\/www.cs.cmu.edu\/~rsalakhu\/papers\/oneshot1.pdf\">2<\/a><\/noindex> nebo <noindex><a rel=\"nofollow\" href=\"http:\/\/www.robots.ox.ac.uk\/~tvg\/publications\/2018\/0431.pdf\">3<\/a><\/noindex>.<\/p>\n<p>Jednou z nev\u00fdhod je, \u017ee u\u010den\u00ed obvykle funguje dob\u0159e na n\u011bjak\u00fdch jednoduch\u00fdch, \"MNISTovsk\u00fdch\" p\u0159\u00edkladech. P\u0159i p\u0159echodu k slo\u017eit\u011bj\u0161\u00edm \u00fakol\u016fm je pot\u0159eba v\u011bt\u0161\u00ed datab\u00e1ze, model objekt\u016f nebo n\u011bjak\u00e1 magie.<br \/>\nPr\u00e1ce na One-Shot u\u010den\u00ed je obecn\u011b velmi zaj\u00edmav\u00e9 t\u00e9ma. Najdete tu mnoho my\u0161lenek. Ale v\u011bt\u0161inou ty dv\u011b probl\u00e9my, kter\u00e9 jsem zm\u00ednil (p\u0159edtr\u00e9nov\u00e1n\u00ed na obrovsk\u00e9m datasetu \/ nestabilita na slo\u017eit\u00fdch datech), velmi br\u00e1n\u00ed u\u010den\u00ed.<\/p>\n<p>Na druhou stranu se k t\u00e9matu Embedding hod\u00ed GAN \u2014 generativn\u00ed adversari\u00e1ln\u00ed s\u00edt\u011b. Ur\u010dit\u011b jste \u010detli spoustu \u010dl\u00e1nk\u016f na tento t\u00e9ma na Habr.<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 \/>\nZvl\u00e1\u0161tnost\u00ed GAN je vytv\u00e1\u0159en\u00ed ur\u010dit\u00e9ho vnit\u0159n\u00edho prostoru stav\u016f (v podstat\u011b toho sam\u00e9ho Embeddingu), kter\u00fd umo\u017e\u0148uje generovat obr\u00e1zek. To mohou b\u00fdt <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/shaoanlu\/faceswap-GAN\">tv\u00e1\u0159e<\/a><\/noindex>, mohou to b\u00fdt <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/sergeytulyakov\/mocogan\">akce<\/a><\/noindex>. <\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/8c25c375dae3559b6895d6c4eb3f6cfd.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nProbl\u00e9m GAN je, \u017ee \u010d\u00edm slo\u017eit\u011bj\u0161\u00ed je generovan\u00fd objekt, t\u00edm slo\u017eit\u011bj\u0161\u00ed je jej popsat v logice \u201egener\u00e1tor-diskrimin\u00e1tor\u201c. V\u00fdsledkem je, \u017ee z re\u00e1ln\u00fdch aplikac\u00ed GAN, kter\u00e9 jsou na poslech - pouze DeepFake, kter\u00fd op\u011bt manipuluje s reprezentacemi tv\u00e1\u0159\u00ed (pro kter\u00e9 existuje obrovsk\u00e1 datab\u00e1ze).<\/p>\n<p>Dal\u0161\u00edch u\u017eite\u010dn\u00fdch aplikac\u00ed jsem vid\u011bl velmi m\u00e1lo. Obvykle to jsou n\u011bjak\u00e9 vymy\u0161len\u00e9 kousky s dopl\u0148ov\u00e1n\u00edm obr\u00e1zk\u016f.<\/p>\n<p>A op\u011bt. Nikdo nem\u00e1 p\u0159edstavu, jak n\u00e1m to pom\u016f\u017ee posunout se do sv\u011btl\u00e9 budoucnosti. P\u0159edstava logiky\/prostoru v neuronov\u00e9 s\u00edti je dobr\u00e1. Ale pot\u0159ebujeme obrovsk\u00e9 mno\u017estv\u00ed p\u0159\u00edklad\u016f, nen\u00ed n\u00e1m jasn\u00e9, jak si neuronov\u00e1 s\u00ed\u0165 d\u00e1v\u00e1 dohromady toto slo\u017eit\u00e9 zobrazen\u00ed.<\/p>\n<p><b>Reinforcement learning<\/b> \u2014 je p\u0159\u00edstup \u00fapln\u011b z jin\u00e9 strany. Ur\u010dit\u011b si pamatujete, jak Google porazil v\u0161echny v Go. Ned\u00e1vn\u00e9 v\u00edt\u011bzstv\u00ed ve Starcraftu a Dota. Ale tady to nen\u00ed tak r\u016f\u017eov\u00e9 a perspektivn\u00ed. Nejl\u00e9pe o RL a jeho komplikac\u00edch mluv\u00ed <noindex><a rel=\"nofollow\" href=\"https:\/\/www.alexirpan.com\/2018\/02\/14\/rl-hard.html\">this article<\/a><\/noindex>.<\/p>\n<p>Pokud bych m\u011bl stru\u010dn\u011b shrnout, co autor napsal:<\/p>\n<ul>\n<li>Modely z krabice nevyhovuj\u00ed \/ ve v\u011bt\u0161in\u011b p\u0159\u00edpad\u016f funguj\u00ed \u0161patn\u011b.<\/li>\n<li>Praktick\u00e9 \u00fakoly se daj\u00ed jednodu\u0161e vy\u0159e\u0161it jin\u00fdmi zp\u016fsoby. Boston Dynamics nepou\u017e\u00edv\u00e1 RL kv\u016fli jeho slo\u017eitosti \/ nep\u0159edv\u00eddatelnosti \/ v\u00fdpo\u010detn\u00edm slo\u017eitostem.<\/li>\n<li>Aby RL fungoval, pot\u0159ebujete slo\u017eitou funkci. \u010casto je obt\u00ed\u017en\u00e9 ji vytvo\u0159it \/ napsat.<\/li>\n<li>Je t\u011b\u017ek\u00e9 \u0161kolit modely. Mus\u00edte vynalo\u017eit spoustu \u010dasu na rozproud\u011bn\u00ed a p\u0159ekon\u00e1n\u00ed lok\u00e1ln\u00edch optim\u016f.<\/li>\n<li>V d\u016fsledku toho je obt\u00ed\u017en\u00e9 model opakovat, model je nestabiln\u00ed p\u0159i nejmen\u0161\u00edch zm\u011bn\u00e1ch.<\/li>\n<li>\u010casto se p\u0159izp\u016fsobuje n\u011bjak\u00fdm neobvykl\u00fdm vzorc\u016fm, a\u017e po gener\u00e1tor n\u00e1hodn\u00fdch \u010d\u00edsel.<\/li>\n<\/ul>\n<p>\nKl\u00ed\u010dov\u00fd moment \u2014 RL zat\u00edm nefunguje v produk\u010dn\u00edm prost\u0159ed\u00ed. Google m\u00e1 n\u011bjak\u00e9 experimenty ( <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> ). Ale nevid\u011bl jsem \u017e\u00e1dn\u00fd produktov\u00fd syst\u00e9m.<\/p>\n<p><b>Pam\u011b\u0165<\/b>. Nev\u00fdhodou v\u00fd\u0161e uveden\u00e9ho je neorganizovanost. Jedn\u00edm z p\u0159\u00edstup\u016f, jak to v\u0161echno zorganizovat, je poskytnout neuronov\u00e9 s\u00edti p\u0159\u00edstup k odd\u011blen\u00e9 pam\u011bti. Aby mohla ukl\u00e1dat a p\u0159episovat v\u00fdsledky sv\u00fdch krok\u016f. Pak se neuronov\u00e1 s\u00ed\u0165 m\u016f\u017ee definovat aktu\u00e1ln\u00edm stavem pam\u011bti. To je velmi podobn\u00e9 klasick\u00fdm procesor\u016fm a po\u010d\u00edta\u010d\u016fm.<\/p>\n<p>Nejzn\u00e1m\u011bj\u0161\u00ed a nejpopul\u00e1rn\u011bj\u0161\u00ed <noindex>neaktu\u00e1ln\u00ed statistika <\/noindex> \u2014 od DeepMind:<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/acc6bcd86c8c071fbd9776a91f990752.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nZd\u00e1 se, \u017ee to je kl\u00ed\u010d k pochopen\u00ed inteligence? Ale pravd\u011bpodobn\u011b ne. Syst\u00e9mu je i tak pot\u0159eba obrovsk\u00fd objem dat k tr\u00e9nov\u00e1n\u00ed. A funguje p\u0159ev\u00e1\u017en\u011b se strukturovan\u00fdmi tabulkov\u00fdmi daty. P\u0159itom, kdy\u017e Facebook <noindex><a rel=\"nofollow\" href=\"https:\/\/embodiedqa.org\/\">\u0159e\u0161il <\/a><\/noindex>podobn\u00fd probl\u00e9m, zvolili cestu \u201ena nic pam\u011bti, prost\u011b ud\u011bl\u00e1me neuronku slo\u017eit\u011bj\u0161\u00ed, a v\u00edce p\u0159\u00edklad\u016f \u2014 a ona se sama nau\u010d\u00ed\u201c.<\/p>\n<p><b>Disentanglement<\/b>. Dal\u0161\u00ed zp\u016fsob, jak vytvo\u0159it smysluplnou pam\u011b\u0165, je vz\u00edt stejn\u00e9 embeddingy, ale p\u0159i tr\u00e9nov\u00e1n\u00ed zav\u00e9st dal\u0161\u00ed krit\u00e9ria, kter\u00e1 by umo\u017enila v nich vy\u010dlenit \u201ev\u00fdznamy\u201c. Nap\u0159\u00edklad chceme nau\u010dit neuronovou s\u00ed\u0165 rozli\u0161ovat chov\u00e1n\u00ed \u010dlov\u011bka v obchod\u011b. Kdybychom \u0161li standardn\u00ed cestou \u2014 museli bychom vytvo\u0159it des\u00edtku s\u00edt\u00ed. Jedna hled\u00e1 \u010dlov\u011bka, druh\u00e1 ur\u010duje, co d\u011bl\u00e1, t\u0159et\u00ed jeho v\u011bk, \u010dtvrt\u00e1 \u2014 pohlav\u00ed. Dal\u0161\u00ed logika sleduje \u010d\u00e1st obchodu, kde se to d\u011bje\/nau\u010duje se to. T\u0159et\u00ed ur\u010duje jeho trajektorii atd.<\/p>\n<p>Nebo, pokud by bylo nekone\u010dn\u00e9 mno\u017estv\u00ed dat, mohlo by se nau\u010dit jednu s\u00ed\u0165 na v\u0161echny mo\u017en\u00e9 v\u00fdsledky (je z\u0159ejm\u00e9, \u017ee takov\u00fd objem dat nelze nasb\u00edrat).<\/p>\n<p>P\u0159\u00edstup k dysentelgimentu n\u00e1m \u0159\u00edk\u00e1, \u017ee bychom m\u011bli \u0161kolit s\u00ed\u0165 tak, aby sama byla schopna rozli\u0161ovat pojmy. Aby byla schopna z videa vytvo\u0159it embedding, kde jedna oblast by definovala akci, jedna - pozici na podlaze v \u010dase, jedna - v\u00fd\u0161ku \u010dlov\u011bka, a dal\u0161\u00ed - jeho pohlav\u00ed. P\u0159i \u0161kolen\u00ed bychom cht\u011bli co nejm\u00e9n\u011b nazna\u010dovat s\u00edti takov\u00e9 kl\u00ed\u010dov\u00e9 pojmy, ale aby si sama vyty\u010dila a seskupila oblasti. Takov\u00fdch \u010dl\u00e1nk\u016f je pom\u011brn\u011b m\u00e1lo (n\u011bkter\u00e9 z nich <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>) a celkov\u011b jsou dost teoretick\u00e9. <\/p>\n<p>Ale tento sm\u011br, alespo\u0148 teoreticky, by m\u011bl \u0159e\u0161it v\u00fd\u0161e uveden\u00e9 probl\u00e9my.<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/b36046dc3cbe3b849e1d3a60f56ec3a2.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nDecomposition obr\u00e1zku podle parametr\u016f \"barva st\u011bn \/ barva podlahy \/ tvar objektu \/ barva objektu \/ atd.\"<\/p>\n<p><img decoding=\"async\" alt=\"Praskl bubble strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9 \u00e9ry\" src=\"\/wp-content\/uploads\/4d5ac8dfc43f76b35f3490c26049c7db.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nDecomposition obli\u010deje podle parametr\u016f \"velikost, obo\u010d\u00ed, orientace, barva pleti, atd.\"<\/p>\n<h3>Ostatn\u00ed<\/h3>\n<p>\nExistuje mnoho dal\u0161\u00edch, ne tolik glob\u00e1ln\u00edch sm\u011br\u016f, kter\u00e9 umo\u017e\u0148uj\u00ed n\u011bjak zmen\u0161it datab\u00e1ze, pracovat s r\u016fznorod\u011bj\u0161\u00edmi daty, atd.<\/p>\n<p><b>Pozornost<\/b>. Asi nem\u00e1 smysl vy\u010dle\u0148ovat to jako samostatnou metodu. Jen p\u0159\u00edstup, kter\u00fd posiluje ostatn\u00ed. Je mu v\u011bnov\u00e1no mnoho \u010dl\u00e1nk\u016f (<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>). Smyslem pozornosti je zv\u00fd\u0161it reakci s\u00edt\u011b pr\u00e1v\u011b na v\u00fdznamn\u00e9 objekty b\u011bhem \u0161kolen\u00ed. \u010casto n\u011bjak\u00fdm extern\u00edm c\u00edlov\u00fdm ukazatelem, nebo malou extern\u00ed s\u00edt\u00ed.<\/p>\n<p><b>3D simulace<\/b>. Pokud ud\u011bl\u00e1te dobr\u00fd 3D engine, m\u016f\u017ee j\u00edm \u010dasto pokr\u00fdt 90 % tr\u00e9ninkov\u00fdch dat (vid\u011bl jsem p\u0159\u00edklad, kdy skoro 99 % dat bylo pokryto dobr\u00fdm enginem). Existuje mnoho n\u00e1pad\u016f a hack\u016f, jak donutit s\u00ed\u0165, kter\u00e1 byla \u0161kolena na 3D enginu, pracovat s re\u00e1ln\u00fdmi daty (Fine tuning, style transfer, atd.). Ale \u010dasto ud\u011blat dobr\u00fd engine je o n\u011bkolik \u0159\u00e1d\u016f slo\u017eit\u011bj\u0161\u00ed ne\u017e shrom\u00e1\u017edit data. P\u0159\u00edklady, kdy byly vytvo\u0159eny enginy:<br \/>\nV\u00fduka robot\u016f (<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 \/>\nStudium <noindex><a rel=\"nofollow\" href=\"https:\/\/neuromation.io\/\">rozpozn\u00e1v\u00e1n\u00ed<\/a><\/noindex> produkty v obchod\u011b (ale ve dvou projektech, kter\u00e9 jsme d\u011blali - jsme se bez toho bez probl\u00e9m\u016f obe\u0161li).<br \/>\n\u0160kolen\u00ed ve spole\u010dnosti Tesla (op\u011bt to video, kter\u00e9 bylo v\u00fd\u0161e).<\/p>\n<h2>Z\u00e1v\u011bry<\/h2>\n<p>\nCel\u00fd \u010dl\u00e1nek je v jist\u00e9m smyslu shrnut\u00edm. Asi hlavn\u00ed poselstv\u00ed, kter\u00e9 jsem cht\u011bl sd\u011blit - \"zdarma skon\u010dilo, neuronov\u00e9 s\u00edt\u011b ji\u017e nenab\u00edzej\u00ed jednoduch\u00e1 \u0159e\u0161en\u00ed\". Nyn\u00ed je t\u0159eba pracovat na slo\u017eit\u00fdch \u0159e\u0161en\u00edch. Nebo pracovat na slo\u017eit\u00fdch v\u011bdeck\u00fdch v\u00fdzkumech.<\/p>\n<p>A v\u016fbec, t\u00e9ma je diskutabiln\u00ed. Mo\u017en\u00e1 maj\u00ed \u010dten\u00e1\u0159i zaj\u00edmav\u011bj\u0161\u00ed p\u0159\u00edklady?<br \/>\n<br \/>Zdroj: <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 5.0.3 - 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\/cs\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari\" \/>\n\t\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.3\" \/>\n\t\t<meta property=\"og:locale\" content=\"cs_CZ\" \/>\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\/cs\/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\udd47Praskla bublina strojov\u00e9ho u\u010den\u00ed, nebo za\u010d\u00e1tek nov\u00e9ho \u00fasvitu | ProHoster","description":"Ned\u00e1vno vy\u0161lo","canonical_url":"https:\/\/prohoster.info\/cs\/blog\/news\/lopnul-li-puzyr-mashinnogo-obucheniya-ili-nachalo-novoj-zari","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"cs_CZ","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\/cs\/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\/cs\/wp-json\/wp\/v2\/posts\/35293","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/comments?post=35293"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/35293\/revisions"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media?parent=35293"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/categories?post=35293"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/tags?post=35293"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}