{"id":284,"date":"2016-03-15T11:48:48","date_gmt":"2016-03-15T10:48:48","guid":{"rendered":"http:\/\/www.mlguru.cz\/?p=284"},"modified":"2017-08-12T15:55:57","modified_gmt":"2017-08-12T13:55:57","slug":"je-singularita-blizko","status":"publish","type":"post","link":"https:\/\/www.mlguru.com\/cs\/je-singularita-blizko\/","title":{"rendered":"Je singularita bl\u00edzko?"},"content":{"rendered":"

Dnes prob\u011bhlo posledn\u00ed kolo medi\u00e1ln\u011b ost\u0159e sledovan\u00e9ho pom\u011b\u0159ov\u00e1n\u00ed schopnost\u00ed \u010dlov\u011bka a stroje. Po\u010d\u00edta\u010dov\u00fd program AlphaGo firmy DeepMind ve h\u0159e Go definitivn\u011b potvrdil sv\u00e9 v\u00edt\u011bzstv\u00ed a porazil tak dosavadn\u00edho mistra sv\u011bta Lee Sedola s v\u00fdsledn\u00fdm sk\u00f3re 4:1<\/a>. Go se od \u0161achu nebo jin\u00fdch her, ve kter\u00fdch stroje porazily \u010dlov\u011bka u\u017e d\u00e1vno, li\u0161\u00ed v obrovsk\u00e9 velikosti prostoru hern\u00edch mo\u017enost\u00ed. Tento prostor nen\u00ed mo\u017en\u00e9 proch\u00e1zet cel\u00fd a hledat glob\u00e1ln\u011b optim\u00e1ln\u00ed tah, nam\u00edsto toho je pot\u0159eba vyu\u017e\u00edt jak\u00e9si strojov\u00e9 intuice, kter\u00e1 vych\u00e1z\u00ed pouze z lok\u00e1ln\u00edch znalost\u00ed. Intuice, kterou jsme doposud p\u0159isuzovali pouze lidem, se v souvislosti s AlphaGo v populariza\u010dn\u00edch textech zmi\u0148uje pom\u011brn\u011b \u010dasto. M\u016f\u017ee se tedy zd\u00e1t, \u017ee se p\u0159ibl\u00ed\u017eila doba, kdy n\u00e1s po\u010d\u00edta\u010de svoj\u00ed inteligenc\u00ed p\u0159ekonaj\u00ed.<\/p>\n

V\u00edt\u011bzstv\u00ed stroje nad \u0161ampi\u00f3nem v Go je bezpochyby v\u00fdznamn\u00fd \u00fasp\u011bch v\u00fdvoj\u00e1\u0159\u016f DeepMind, ale je singularita skute\u010dn\u011b bl\u00edzko? Nemysl\u00edm si.<\/p>\n

Singularita<\/h3>\n

Pojem singularita v kontextu um\u011bl\u00e9 inteligence zavedl matematik Stanislaw Ulam<\/a> v polovin\u011b dvac\u00e1t\u00e9ho stolet\u00ed. Jedn\u00e1 se o hypotetickou ud\u00e1lost, kdy univerz\u00e1ln\u00ed um\u011bl\u00e1 inteligence<\/a> dos\u00e1hne \u00farovn\u011b \u010dlov\u011bka a svoj\u00ed schopnost\u00ed vylep\u0161ovat sebe sama n\u00e1s za\u010dne rychle p\u0159ekon\u00e1vat. Nemal\u00e9 mno\u017estv\u00ed lid\u00ed v\u011b\u0159\u00ed, \u017ee se tato chv\u00edle velice rychle bl\u00ed\u017e\u00ed.<\/p>\n

Jedn\u00edm z nejzn\u00e1m\u011bj\u0161\u00edch populariz\u00e1tor\u016f singularity je Ray Kurzweil, kter\u00fd se proslavil p\u0159edev\u0161\u00edm jako autor knihy The Singularity Is Near: When Humans Transcend Biology<\/a>. Hlavn\u00edm argumentem pro bl\u00ed\u017e\u00edc\u00ed se singularitu je dosavadn\u00ed exponenci\u00e1ln\u00ed r\u016fst r\u016fzn\u00fdch forem technologick\u00e9ho pokroku. Tato my\u0161lenka je roz\u0161\u00ed\u0159en\u00edm takzvan\u00e9ho Moorova z\u00e1kona<\/a>, podle kter\u00e9ho se po\u010det tranzistor\u016f um\u00edst\u011bn\u00fdch na po\u010d\u00edta\u010dov\u00fd \u010dip zdvojn\u00e1sobuje zhruba ka\u017ed\u00e9 dva roky. Jedn\u00edm z p\u0159\u00edklad\u016f roz\u0161\u00ed\u0159en\u00ed je exponenci\u00e1ln\u00ed r\u016fst v\u00fdpo\u010detn\u00ed rychlosti po\u010d\u00edta\u010d\u016f, ilustrovan\u00fd na obr\u00e1zku 1.<\/p>\n

\"Exponenci\u00e1ln\u00ed<\/a>

Obr\u00e1zek 1: Exponenci\u00e1ln\u00ed r\u016fst v\u00fdpo\u010detn\u00ed rychlosti. Zdroj: https:\/\/en.wikipedia.org\/wiki\/The_Singularity_Is_Near<\/p><\/div>\n

Je ale exponenci\u00e1ln\u00ed r\u016fst v\u00fdpo\u010detn\u00ed rychlosti a dal\u0161\u00edch technologick\u00fdch ukazatel\u016f dostate\u010dn\u00fdm p\u0159edpokladem k dosa\u017een\u00ed um\u011bl\u00e9 inteligence, p\u0159ekon\u00e1vaj\u00edc\u00ed ve v\u0161ech ohledech intelekt \u010dlov\u011bka?<\/p>\n

\u00dasp\u011bchy um\u011bl\u00fdch neuronov\u00fdch s\u00edt\u00ed<\/h3>\n

Nen\u00ed pochyb o tom, \u017ee existuje cel\u00e1 \u0159ada oblast\u00ed, ve kter\u00fdch stroje nad lidmi v\u00fdrazn\u011b vynikaj\u00ed. U\u017e od dob prvn\u00edch po\u010d\u00edtac\u00edch stroj\u016f to jsou \u00falohy, vy\u017eaduj\u00edc\u00ed mechanick\u00e9 opakov\u00e1n\u00ed rutinn\u00edch postup\u016f. V posledn\u00edch letech v\u0161ak slav\u00ed stroje \u00fasp\u011bchy i v oblastech, kter\u00e9 byly doposud pouze dom\u00e9nou \u010dlov\u011bka. Jmenujme nap\u0159\u00edklad rozpozn\u00e1n\u00ed obsahu obr\u00e1zku, kde sou\u010dasn\u00fd nejlep\u0161\u00ed algoritmus<\/a> dosahuje chybovosti 4,58 % (chybovost \u010dlov\u011bka se \u00fadajn\u011b pohybuje n\u011bkde kolem 5 %), algoritmy pro \u0159\u00edzen\u00ed aut a dron\u016f, projekt DeepDream<\/a> firmy Google, vytv\u00e1\u0159ej\u00edc\u00ed psychedelick\u00e9 obrazy, v \u00favodu zm\u00edn\u011bn\u00fd program AlphaGo, nebo t\u0159eba syst\u00e9m, kter\u00fd je schopn\u00fd ps\u00e1t b\u00e1sn\u011b<\/a>.<\/p>\n

V\u0161echny tyto \u00falohy maj\u00ed dv\u011b spole\u010dn\u00e9 vlastnosti. Prvn\u00ed z nich je vysok\u00e1 dimenzionalita probl\u00e9mu \u2013 u algoritm\u016f pro zpracov\u00e1n\u00ed obrazu je to mno\u017estv\u00ed vstupn\u00edch pixel\u016f, kter\u00e9 ned\u00e1vaj\u00ed velk\u00fd smysl jednotliv\u011b, ale je t\u0159eba se na n\u011b d\u00edvat jako na celek, u hry Go je to obrovsk\u00e9 mno\u017estv\u00ed mo\u017en\u00fdch hern\u00edch sc\u00e9n\u00e1\u0159\u016f, u generov\u00e1n\u00ed text\u016f souvis\u00ed vysok\u00e1 dimenzionalita s mno\u017estv\u00edm r\u016fzn\u00fdch kombinac\u00ed slov, kter\u00e9 mohou m\u00edt poka\u017ed\u00e9 \u00fapln\u011b jin\u00fd v\u00fdznam. Druh\u00fdm jednot\u00edc\u00edm prvkem je fakt, \u017ee sou\u010dasn\u00fdch fascinuj\u00edc\u00edch v\u00fdsledk\u016f bylo dosa\u017eeno pomoc\u00ed hlubok\u00fdch neuronov\u00fdch s\u00edt\u00ed<\/a>.<\/p>\n

Um\u011bl\u00e9 neuronov\u00e9 s\u00edt\u011b v\u0161ak zn\u00e1me ji\u017e n\u011bkolik des\u00edtek let, je tedy legitimn\u00ed pokl\u00e1dat ot\u00e1zku, co n\u00e1hl\u00fd posun umo\u017enilo. V prvn\u00ed \u0159ad\u011b je d\u016fle\u017eit\u00e9 si uv\u011bdomit, \u017ee se zm\u00edn\u011bn\u00e9 \u00falohy v principu neli\u0161\u00ed od jin\u00fdch \u00faloh strojov\u00e9ho u\u010den\u00ed, kter\u00e9 u\u017e dlouh\u00e9 roky pova\u017eujeme za uspokojiv\u011b vy\u0159e\u0161en\u00e9. Nikdo se dnes nepozastavuje nad t\u00edm, \u017ee e-mailov\u00fd server dok\u00e1\u017ee s velkou \u00fasp\u011b\u0161nost\u00ed spr\u00e1vn\u011b rozpoznat spam, nebo nad t\u00edm, \u017ee c\u00edlen\u00e1 reklama na internetu velice p\u0159esn\u011b reflektuje nikde explicitn\u011b nespecifikovan\u00e9 z\u00e1jmy u\u017eivatele. Rozpozn\u00e1va\u010d obr\u00e1zk\u016f nen\u00ed nic slo\u017eit\u011bj\u0161\u00edho ne\u017e klasifik\u00e1tor s komplexn\u00ed vnit\u0159n\u00ed strukturou. Strojov\u00fd gener\u00e1tor b\u00e1sn\u00ed je sv\u00fdm zp\u016fsobem tak\u00e9 klasifik\u00e1tor, jen\u017e se na z\u00e1klad\u011b p\u0159edchoz\u00edho textu sna\u017e\u00ed predikovat jeho nejpravd\u011bpodobn\u011bj\u0161\u00ed pokra\u010dov\u00e1n\u00ed.<\/p>\n

Syst\u00e9my \u0159\u00edzen\u00ed dron\u016f nebo algoritmus AlphaGo se od p\u0159edchoz\u00edch p\u0159\u00edklad\u016f li\u0161\u00ed v tom, \u017ee v sob\u011b zahrnuj\u00ed zp\u011btnou vazbu od okoln\u00edho prost\u0159ed\u00ed, na z\u00e1klad\u011b kter\u00e9 jsou schopny se d\u00e1le zlep\u0161ovat. Pokud dron provede n\u011bjakou akci, kv\u016fli kter\u00e9 za\u010dne neo\u010dek\u00e1van\u011b ztr\u00e1cet v\u00fd\u0161ku, je to pro n\u011bj jednozna\u010dn\u00e1 zpr\u00e1va o proveden\u00ed chybn\u00e9ho man\u00e9vru, ze kter\u00e9 se m\u016f\u017ee pou\u010dit. Podobn\u011b algoritmus AlphaGo sehr\u00e1l n\u011bkolik milion\u016f her s\u00e1m proti sob\u011b, d\u00edky \u010demu\u017e se dok\u00e1zal nau\u010dit l\u00e9pe odli\u0161ovat dobr\u00e9 strategie od hor\u0161\u00edch. Tomuto zp\u016fsobu u\u010den\u00ed se \u0159\u00edk\u00e1 reinforcement learning<\/a> a zdaleka nejde o \u017e\u00e1dn\u00fd nov\u00fd vyn\u00e1lez nato\u017e o n\u011bco principi\u00e1ln\u011b odli\u0161n\u00e9ho od neuronov\u00fdch s\u00edt\u00ed s jinou formou zp\u011btn\u00e9 vazby.<\/p>\n

Kdy\u017e opomeneme n\u011bkolik in\u017een\u00fdrsk\u00fdch trik\u016f urychluj\u00edc\u00edch tr\u00e9nov\u00e1n\u00ed, \u00fasp\u011bch hlubok\u00fdch neuronov\u00fdm s\u00edt\u00ed v posledn\u00ed dob\u011b je zp\u016fsoben pouze obrovsk\u00fdm n\u00e1r\u016fstem v\u00fdpo\u010detn\u00edho v\u00fdkonu, kter\u00fd s\u00edt\u00edm v rozumn\u00e9m \u010dase umo\u017e\u0148uje vytv\u00e1\u0159et velmi slo\u017eit\u00e9 vnit\u0159n\u00ed struktury a efektivn\u011b pracovat s daty vysok\u00e9 dimenzionality. To v\u0161ak samo o sob\u011b nen\u00ed v\u0161esp\u00e1sn\u00e9.<\/p>\n

Budoucnost um\u011bl\u00e9 inteligence<\/h3>\n

V tuto chv\u00edli um\u00edme strojov\u011b efektivn\u011b \u0159e\u0161it n\u011bkter\u00e9 jednotliv\u00e9 probl\u00e9my, kter\u00e9 se na prvn\u00ed pohled zdaj\u00ed b\u00fdt slo\u017eit\u00e9, ale princip\u00e1ln\u011b p\u0159\u00edli\u0161 obt\u00ed\u017en\u00e9 nejsou. Pokud bychom cht\u011bli, aby se stroje vyrovnaly lidem ve v\u0161em, pot\u0159ebovali bychom n\u011bco v\u00edc ne\u017e jen v\u00fdpo\u010detn\u00ed v\u00fdkon. Pro ilustraci uvedu dva p\u0159\u00edklady, kter\u00e9 p\u0159esahuj\u00ed r\u00e1mec schopnost\u00ed zn\u00e1m\u00fdch algoritm\u016f.<\/p>\n

Sou\u010dasn\u00e1 um\u011bl\u00e1 inteligence um\u00ed \u0159e\u0161it pouze probl\u00e9my, pro kter\u00e9 je vytvo\u0159en\u00e1. St\u00e1le je to jen algoritmus, vykon\u00e1vaj\u00edc\u00ed postupy naprogramovan\u00e9 \u010dlov\u011bkem, a p\u0159\u00edli\u0161 se tak neli\u0161\u00ed t\u0159eba od v\u00fdrobn\u00ed linky v tov\u00e1rn\u011b. Oproti tomu \u010dlov\u011bk je schopn\u00fd vid\u011bt sv\u011bt v souvislostech a b\u00fdt kreativn\u00ed. Pokud \u010dlov\u011bk vytvo\u0159\u00ed program, kter\u00fd umo\u017e\u0148uje generovat b\u00e1sn\u011b, p\u0159irozen\u011b ho napadne, \u017ee by se podobn\u00fd algoritmus dal pou\u017e\u00edt i pro generov\u00e1n\u00ed notov\u00e9ho z\u00e1pisu hudby. Stroj ale tuto souvislost nevid\u00ed. Kdy\u017e u\u010ditel na z\u00e1kladn\u00ed \u0161kole p\u0159ehraje d\u011btem \u00faryvek Smetanovy Vltavy a Blan\u00edku, v\u011b\u0159\u00edm, \u017ee v\u011bt\u0161ina \u017e\u00e1k\u016f dok\u00e1\u017ee rozpoznat, kter\u00fd \u00faryvek odkazuje k \u0159ece a kter\u00fd k ryt\u00ed\u0159\u016fm, ani\u017e by nad t\u00edm \u017e\u00e1ci kdykoliv p\u0159em\u00fd\u0161leli nebo jim n\u011bkdo \u0159ekl, podle jak\u00e9ho kl\u00ed\u010de se maj\u00ed rozhodnou.<\/p>\n

Jin\u00fdm p\u0159\u00edkladem omezenosti stroje je absence jeho vlastn\u00edho v\u011bdom\u00ed. Stroje se nerozhoduj\u00ed podle sv\u00e9ho vlastn\u00edho \u00fasudku, ale bu\u010f podle pravidel, kter\u00e1 jsou z\u00edskan\u00e1 z dat (p\u0159\u00edpadn\u011b jsou p\u0159\u00edmo sou\u010d\u00e1st\u00ed jejich algoritmu), nebo statisticky podle napo\u010d\u00edtan\u00fdch pravd\u011bpodobnost\u00ed. Algoritmus pak sice dok\u00e1\u017ee prov\u00e9st spr\u00e1vn\u00e9 rozhodnut\u00ed, ale je velice obt\u00ed\u017en\u00e9 zjistit, pro\u010d se tak rozhodl. Nato\u017e, aby sv\u00e9 rozhodnut\u00ed dok\u00e1zal s\u00e1m zd\u016fvodnit. P\u0159i generov\u00e1n\u00ed b\u00e1sn\u00ed nem\u00e1 po\u010d\u00edta\u010d pon\u011bt\u00ed o tom, co vytv\u00e1\u0159\u00ed, pouze aplikuje obecn\u00e9 vzory z\u00edskan\u00e9 z dat a vytv\u00e1\u0159\u00ed statisticky o\u010dek\u00e1van\u00e9 posloupnosti p\u00edsmen.<\/p>\n

Na z\u00e1klad\u011b sv\u00fdch znalost\u00ed a zku\u0161enost\u00ed si netrouf\u00e1m predikovat, zda se jednou poda\u0159\u00ed vytvo\u0159it univerz\u00e1ln\u00ed um\u011blou inteligenci, kter\u00e1 se ve v\u0161em vyrovn\u00e1 \u010dlov\u011bku a postupn\u011b ho p\u0159ekon\u00e1. Jsem v\u0161ak p\u0159esv\u011bd\u010den\u00fd, \u017ee k tomu te\u010f nejsme o mnoho bl\u00ed\u017ee ne\u017e v polovin\u011b minul\u00e9ho stolet\u00ed, kdy Alan Turing<\/a> vytvo\u0159il prvn\u00ed teoretick\u00fd model v\u00fdpo\u010detn\u00edho stroje.<\/p>\n","protected":false},"excerpt":{"rendered":"

Dnes prob\u011bhlo posledn\u00ed kolo medi\u00e1ln\u011b ost\u0159e sledovan\u00e9ho pom\u011b\u0159ov\u00e1n\u00ed schopnost\u00ed \u010dlov\u011bka a stroje. Po\u010d\u00edta\u010dov\u00fd program AlphaGo firmy DeepMind ve h\u0159e Go definitivn\u011b potvrdil sv\u00e9 v\u00edt\u011bzstv\u00ed a porazil tak dosavadn\u00edho mistra sv\u011bta Lee Sedola s v\u00fdsledn\u00fdm sk\u00f3re 4:1. Go se od \u0161achu nebo jin\u00fdch her, ve kter\u00fdch stroje porazily \u010dlov\u011bka u\u017e d\u00e1vno, li\u0161\u00ed v obrovsk\u00e9 velikosti prostoru […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_exactmetrics_skip_tracking":false,"_exactmetrics_sitenote_active":false,"_exactmetrics_sitenote_note":"","_exactmetrics_sitenote_category":0,"footnotes":""},"categories":[12],"tags":[],"class_list":["post-284","post","type-post","status-publish","format-standard","hentry","category-umela-inteligence"],"_links":{"self":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts\/284","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/comments?post=284"}],"version-history":[{"count":11,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts\/284\/revisions"}],"predecessor-version":[{"id":435,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts\/284\/revisions\/435"}],"wp:attachment":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/media?parent=284"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/categories?post=284"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/tags?post=284"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}