{"id":54780,"date":"2020-01-04T00:00:00","date_gmt":"2020-01-03T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/nejroseti-kuda-eto-vse-dvizhetsya"},"modified":"2020-02-18T14:02:50","modified_gmt":"2020-02-18T11:02:50","slug":"nejroseti-kuda-eto-vse-dvizhetsya","status":"publish","type":"post","link":"https:\/\/prohoster.info\/bs\/blog\/news\/nejroseti-kuda-eto-vse-dvizhetsya","title":{"rendered":"Neuralne mre\u017ee. Kuda ovo sve ide","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<p>\u010clanak se sastoji od dva dela:<\/p>\n<p><\/p>\n<ol>\n<li>Kratak opis nekih arhitektura mre\u017ea za prepoznavanje objekata na slikama i segmentaciju slika sa najjasnijim resursima za mene. Poku\u0161ao sam da izaberem video obja\u0161njenja, po mogu\u0107stvu na ruskom jeziku.<\/li>\n<li>Drugi deo se sastoji u poku\u0161aju da se razume smer razvoja arhitektura neuronskih mre\u017ea i tehnologija zasnovanih na njima.<\/li>\n<\/ol>\n<p><\/p>\n<p><img decoding=\"async\" alt=\"Neuralne mre\u017ee. Kuda ovo sve ide\" src=\"\/wp-content\/uploads\/2020\/01\/3e0238547dc956dbb069c11241e4534f.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><\/p>\n<p>Slika 1 \u2013 Razumevanje arhitektura neuronskih mre\u017ea nije lako<\/p>\n<p><\/p>\n<p>Sve je po\u010delo kada sam napravio dva demonstraciona aplikacija za klasifikaciju i prepoznavanje objekata na Android telefonu:<\/p>\n<p><\/p>\n<ul>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/foobar167\/junkyard\/tree\/master\/object_classifier\">Back-end demo<\/a><\/noindex>, kada se podaci obra\u0111uju na serveru i \u0161alju na telefon. Klasifikacija slika (image classification) tri tipa medveda: sme\u0161nog, crnog i pli\u0161anog.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/foobar167\/android\/tree\/master\/object_detection_demo\">Front-end demo<\/a><\/noindex>, kada se podaci obra\u0111uju na samom telefonu. Prepoznavanje objekata (object detection) tri tipa: le\u0161nik, smokva i urme.<\/li>\n<\/ul>\n<p><noindex><a rel=\"nofollow\" name=\"habracut\"><\/a><\/noindex><\/p>\n<p>Postoji razlika izme\u0111u zadataka klasifikacije slika, prepoznavanja objekata na slici i <noindex><a rel=\"nofollow\" href=\"https:\/\/medium.com\/analytics-vidhya\/image-classification-vs-object-detection-vs-image-segmentation-f36db85fe81\">segmentacije slika<\/a><\/noindex>. Stoga se pojavila potreba da saznam koje arhitekture neuronskih mre\u017ea prepoznaju objekte na slikama i koje mogu segmenitirati. Prona\u0161ao sam slede\u0107e primere arhitektura sa najjasnijim resursima za mene:<\/p>\n<p><\/p>\n<ul>\n<li>Serija arhitektura zasnovanih na R-CNN (<strong>R<\/strong>egions with <strong>C<\/strong>onvolution <strong>N<\/strong>eural <strong>N<\/strong>etworks features): R-CNN, Fast R-CNN, <noindex><a rel=\"nofollow\" href=\"https:\/\/medium.com\/@smallfishbigsea\/faster-r-cnn-explained-864d4fb7e3f8\">Faster R-CNN<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/0vt05rQqk_I\">Mask R-CNN<\/a><\/noindex>. Za prepoznavanje objekta na slici koristi se mehanizam Region Proposal Network (RPN) za izdvajanje ograni\u010denih regiona (bounding boxes). Prvobitno je umesto RPN kori\u0161\u0107en sporiji mehanizam Selective Search. Zatim se izdvojeni ograni\u010deni regioni \u0161alju kao ulaz obi\u010dnoj neuronskoj mre\u017ei za klasifikaciju. U arhitekturi R-CNN postoje o\u010digledne \u201efor\u201c petlje za prolaz kroz ograni\u010dene regione, ukupno do 2000 prolaza kroz unutra\u0161nju mre\u017eu AlexNet. Zbog o\u010diglednih \u201efor\u201c petlji usporava se brzina obrade slika. Broj o\u010diglednih petlji i prolaza kroz unutra\u0161nju neuronsku mre\u017eu smanjuje se sa svakom novom verzijom arhitekture, a tako\u0111e se sprovode desetine drugih promena radi pove\u0107anja brzine i zamene zadatka prepoznavanja objekata sa segmentacijom objekata u Mask R-CNN.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/L0tzmv--CGY\">YOLO<\/a><\/noindex> (<strong>Y<\/strong>ou <strong>O<\/strong>nly <strong>L<\/strong>ook <strong>O<\/strong>nce) \u2013 prva neuronska mre\u017ea koja prepoznaje objekte u realnom vremenu na mobilnim ure\u0111ajima. Odlikuje se sposobno\u0161\u0107u prepoznavanja objekata u jednom prolazu (samo jedan pogled je dovoljan). Tako da u arhitekturi YOLO nema o\u010diglednih \"for\" petlji, zbog \u010dega mre\u017ea radi brzo. Na primer, analogija: u NumPy-u pri operacijama sa matricama tako\u0111e nema o\u010diglednih \"for\" petlji, koje se u NumPy-u realizuju na ni\u017eim nivoima arhitekture kroz programski jezik C. YOLO koristi mre\u017eu unapred definiranih prozora. Da isti objekat ne bi bio vi\u0161e puta prepoznat, koristi se koeficijent preklapanja prozora (IoU, <strong>I<\/strong>ntersection <strong>o<\/strong>ver <strong>U<\/strong>nion). Ova arhitektura funkcioni\u0161e u \u0161irokom opsegu i karakteri\u0161e je visoka <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%A0%D0%BE%D0%B1%D0%B0%D1%81%D1%82%D0%BD%D0%BE%D1%81%D1%82%D1%8C\">robustnost<\/a><\/noindex>: model mo\u017ee biti obu\u010den na fotografijama, ali dobro funkcioni\u0161e i na crte\u017eima.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/P8e-G-Mhx4k\">SSD<\/a><\/noindex> (<strong>S<\/strong>ingle <strong>S<\/strong>hot MultiBox <strong>D<\/strong>etector) \u2013 koriste se najbolji \"hackovi\" arhitekture YOLO (na primer, non-maximum suppression) i dodaju se novi, kako bi neuronska mre\u017ea br\u017ee i preciznije radila. Odlikuje se sposobno\u0161\u0107u prepoznavanja objekata u jednom prolazu uz pomo\u0107 unapred definisane mre\u017ee prozora (default box) na piramidama slika. Piramida slika je kodirana u konvolucijskim tenzorima tokom sekvencijalnih operacija konvolucije i puliinga (prilikom operacije max-pooling prostorna dimenzija opada). Tako se prepoznaju i veliki i mali objekti u jednom prolazu mre\u017ee.<\/li>\n<li>MobileSSD (<strong>Mobilni<\/strong>NetV2 + <strong>SSD<\/strong>) \u2013 kombinacija dve arhitekture neuronskih mre\u017ea. Prva mre\u017ea <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/352804\/\">MobileNetV2<\/a><\/noindex> radi brzo i pove\u0107ava ta\u010dnost prepoznavanja. MobileNetV2 se koristi umesto VGG-16, koja je prvobitno kori\u0161\u0107ena u <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/abs\/1512.02325\">originalnom \u010dlanku<\/a><\/noindex>. Druga mre\u017ea SSD odre\u0111uje lokaciju objekata na slici.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/ge_RT5wvHvY\">SqueezeNet<\/a><\/noindex> \u2013 veoma mala, ali precizna neuronska mre\u017ea. Sama po sebi ne re\u0161ava zadatak prepoznavanja objekata. Ipak, mo\u017ee se koristiti u kombinaciji razli\u010ditih arhitektura. Mo\u017ee se koristiti na mobilnim ure\u0111ajima. Njena posebna karakteristika je to \u0161to se podaci najpre kompresuju na \u010detiri 1\u00d71 konvolucijska filtera, a zatim \u0161ire na \u010detiri 1\u00d71 i \u010detiri 3\u00d73 konvolucijska filtera. Jedna takva iteracija kompresije-\u0161irenja podataka naziva se \u201eFire Module\u201c.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/b6jhopSMit8\">DeepLab<\/a><\/noindex> (Semanti\u010dka Segmentacija Slika uz Duboke Konvolucijske Mre\u017ee) \u2013 segmentacija objekata na slici. Karakteristika arhitekture je razre\u0111ena (dilated convolution) konvolucija, koja zadr\u017eava prostornu rezoluciju. Potom slijedi faza postprocesiranja rezultata kori\u0161tenjem grafi\u010dkog probabilisti\u010dkog modela (conditional random field), \u0161to omogu\u0107ava uklanjanje malih \u0161umova u segmentaciji i pobolj\u0161anje kvaliteta segmentirane slike. Iza zastra\u0161uju\u0107eg naziva \u201egrafi\u010dki probabilisti\u010dki model\u201c krije se obi\u010dni Gaussov filtar, koji je aproksimiran na pet ta\u010daka.<\/li>\n<li>Poku\u0161avao sam da shvatim ure\u0111aj <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/abs\/1711.06897\">RefineDet<\/a><\/noindex> (Jednokratno <strong>Refine<\/strong>ment Neural Network for Object <strong>Det<\/strong>ekciju), ali nisam puno shvatio.<\/li>\n<li>Tako\u0111e sam pogledao kako funkcioni\u0161e tehnologija \u201epa\u017enja\u201c (attention): <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/W2rWgXJBZhU\">video1<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/iDulhoQ2pro\">video2<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/H6Qiegq_36c\">video3<\/a><\/noindex>. Karakteristika arhitekture \u201epa\u017enja\u201c je automatsko isticanje podru\u010dja pove\u0107ane pa\u017enje na slici (RoI, <strong>R<\/strong>regije <strong>o<\/strong>f <strong>I<\/strong>interesa) pomo\u0107u neuronske mre\u017ee nazvane Attention Unit. Podru\u010dja pove\u0107ane pa\u017enje sli\u010dna su ograni\u010denim regionima (bounding boxes), ali za razliku od njih nisu fiksna na slici i mogu imati nejasne granice. Iz podru\u010dja pove\u0107ane pa\u017enje izdvajaju se osobine (features) koje se \u201e\u0161alju\u201c rekurentnim neuronskim mre\u017eama s arhitekturama <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/5lUUrREboSk\">LSTM, GRU ili Vanilla RNN<\/a><\/noindex>. Rekurentne neuronske mre\u017ee mogu analizirati me\u0111uzavisnost osobina u sekvenci. Rekurentne neuronske mre\u017ee su prvobitno kori\u0161\u0107ene za prevo\u0111enje teksta na druge jezike, a sada i za prevo\u0111enje <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/e-WB4lfg30M\">slika u tekst<\/a><\/noindex> i <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/rAbhypxs1qQ\">teksta u sliku<\/a><\/noindex>.<\/li>\n<\/ul>\n<p><\/p>\n<p>Kako sam istra\u017eivao ove arhitekture <strong>shvatio sam da ni\u0161ta ne razumem.<\/strong>. I nije stvar u tome da moja neuronska mre\u017ea ima problema s mehanizmom pa\u017enje. Izrada svih ovih arhitektura li\u010di na neki ogroman hackathon, gde se autori takmi\u010de u hakovima. Hack je brzo re\u0161enje te\u0161ke programske zada\u0107e. To jest, izme\u0111u svih ovih arhitektura ne postoji vidljiva i jasna logi\u010dka povezanost. Sve \u0161to ih spaja je skup najuspe\u0161nijih hakova koje preuzimaju jedni od drugih, plus zajedni\u010dka za sve <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/Ilg3gGewQ5U\">operacija konvolucije s povratnom sprekom<\/a><\/noindex> (povratno razprostranjivanje gre\u0161ke, backpropagation). Nema <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/272473\/\">sistemskog razmi\u0161ljanja<\/a><\/noindex>! Nije jasno \u0161ta menjati i kako optimizovati postoje\u0107e dostignu\u0107e.<\/p>\n<p><\/p>\n<p>Kao rezultat nedostatka logi\u010dke povezanosti izme\u0111u hakova, izuzetno je te\u0161ko zapamtiti ih i primeniti u praksi. To su fragmentisana znanja. U najboljem slu\u010daju, zapamti se nekoliko zanimljivih i neo\u010dekivanih momenata, ali ve\u0107ina shva\u0107enog i nesha\u0107enog nestaje iz pam\u0107enja ve\u0107 nakon nekoliko dana. Dobro bi bilo ako se nakon nedelje ipak seti barem naziva arhitekture. A zar nije potro\u0161eno nekoliko sati i \u010dak dana radnog vremena na \u010ditanje \u010dlanaka i gledanje preglednih video snimaka!<\/p>\n<p><\/p>\n<p><img decoding=\"async\" alt=\"Neuralne mre\u017ee. Kuda ovo sve ide\" src=\"\/wp-content\/uploads\/2020\/01\/86f0f24e3be1d0f1a210f3132897d981.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><\/p>\n<p>Slika 2 \u2013 <noindex><a rel=\"nofollow\" href=\"https:\/\/www.asimovinstitute.org\/neural-network-zoo\/\">Zoo neuronih mre\u017ea<\/a><\/noindex><\/p>\n<p><\/p>\n<p>Ve\u0107ina autora nau\u010dnih \u010dlanaka, po mom li\u010dnom mi\u0161ljenju, \u010dini sve \u0161to je mogu\u0107e da \u010dak i ova fragmentisana znanja ne budu shva\u0107ena od strane \u010ditaoca. No, dezinformacije u deset re\u010denica sa formulama koje su preuzete \u201es plafona\u201c \u2013 to je tema za poseban \u010dlanak (problem <noindex><a rel=\"nofollow\" href=\"https:\/\/en.wikipedia.org\/wiki\/Publish_or_perish\">publish or perish<\/a><\/noindex>).<\/p>\n<p><\/p>\n<p>Iz tog razloga pojavila se potreba za sistematizacijom informacija o neuronskim mre\u017eama i, tako, pove\u0107anjem kvaliteta razumevanja i pam\u0107enja. Stoga je glavna tema analize pojedinih tehnologija i arhitektura ve\u0161ta\u010dkih neuronskih mre\u017ea postala slede\u0107a: <strong>saznati kamo sve ovo ide<\/strong>, a ne struktura neke konkretne neuronske mre\u017ee posebno.<\/p>\n<p><\/p>\n<p>Kamo sve ovo ide. Glavni rezultati:<\/p>\n<p><\/p>\n<ul>\n<li>Broj startupa u oblasti ma\u0161inskog u\u010denja u poslednje dve godine <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/recognitor\/blog\/455676\/\">naglo je opao<\/a><\/noindex>. Mogu\u0107i razlog: \u201eneuronske mre\u017ee su prestale biti ne\u0161to novo\u201c.<\/li>\n<li>Svako \u0107e mo\u0107i da kreira radnu neuronsku mre\u017eu za re\u0161avanje jednostavnog zadatka. Za to \u0107e uzeti gotov model iz \u201ezoo modela\u201c (model zoo) i obu\u010diti poslednji sloj neuronske mre\u017ee (<noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/yofjFQddwHE\">transfer learning<\/a><\/noindex>) na gotovim podacima iz <noindex><a rel=\"nofollow\" href=\"https:\/\/toolbox.google.com\/datasetsearch\">Google Dataset Search<\/a><\/noindex> ili sa <noindex><a rel=\"nofollow\" href=\"https:\/\/www.kaggle.com\/datasets\">25.000 datasetova Kaggle<\/a><\/noindex> u besplatnom <noindex><a rel=\"nofollow\" href=\"https:\/\/www.dataschool.io\/cloud-services-for-jupyter-notebook\/\">Jupyter Notebook oblaku<\/a><\/noindex>.<\/li>\n<li>Veliki proizvo\u0111a\u010di neuronskih mre\u017ea po\u010deli su da kreiraju <strong>\u201ezoo modele\u201c<\/strong> (model zoo). Uz njih se mo\u017ee brzo napraviti komercijalna aplikacija: <noindex><a rel=\"nofollow\" href=\"https:\/\/tfhub.dev\/\">TF Hub<\/a><\/noindex> za TensorFlow, <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/open-mmlab\/mmdetection\">MMDetection<\/a><\/noindex> za PyTorch, <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/facebookresearch\/Detectron\">Detectron<\/a><\/noindex> za Caffe2, <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/wkentaro\/chainer-modelzoo\">chainer-modelzoo<\/a><\/noindex> za Chainer i <noindex><a rel=\"nofollow\" href=\"https:\/\/modelzoo.co\/\">druge<\/a><\/noindex>.<\/li>\n<li>Neuronske mre\u017ee koje rade u <strong>realnom vremenu<\/strong> na mobilnim ure\u0111ajima. Od 10 do 50 frejmova u sekundi.<\/li>\n<li>Primena neuronskih mre\u017ea na telefonima (TF Lite), u pretra\u017eiva\u010dima (TF.js) i u <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/19ZNz2N79u4\">potro\u0161a\u010dkim ure\u0111ajima<\/a><\/noindex> (IoT, <strong>I<\/strong>internet <strong>o<\/strong>f <strong>T<\/strong>stvari). Posebno na telefonima, koji ve\u0107 podr\u017eavaju neuronske mre\u017ee na nivou \u201ehardvera\u201c (neuronski akceleratori).<\/li>\n<li>Svako ure\u0111aj, komadi odje\u0107e i, mo\u017eda, \u010dak i hrana \u0107e imati <strong>IP-v6 adresu<\/strong> i komunikovati jedni s drugima\" \u2013 <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/GG7H8Xa4m8I?t=85\">Sebastian Thrun<\/a><\/noindex>.<\/li>\n<li>Rast broja publikacija o ma\u0161inskom u\u010denju po\u010deo je <noindex><a rel=\"nofollow\" href=\"http:\/\/data-mining.philippe-fournier-viger.com\/too-many-machine-learning-papers\">prema\u0161ivati Mooreov zakon<\/a><\/noindex> (udvostru\u010denje svakih dvije godine) od 2015. godine. O\u010digledno su potrebne neuronske mre\u017ee za analizu \u010dlanaka.<\/li>\n<li>Slede\u0107e tehnologije sti\u010du popularnost:\n<ul>\n<li><strong>PyTorch<\/strong> \u2013 popularnost raste brzo i, izgleda, nadma\u0161uje TensorFlow.<\/li>\n<li>Automatsko pode\u0161avanje hiperparametara <strong>AutoML<\/strong> \u2013 popularnost raste postepeno.<\/li>\n<li>Postepeno smanjenje ta\u010dnosti i pove\u0107anje brzine prora\u010duna: <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/rln_kZbYaWc\">fuzzy logika<\/a><\/noindex>, algoritmi <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/MIPkK5ZAsms\">boostinga<\/a><\/noindex>, aproksimativni (pribli\u017eni) prora\u010duni, kvantizacija (kada se te\u017eine neuronske mre\u017ee prevode u cele brojeve i kvantizuju), neuroakseleratori.<\/li>\n<li>Prevod <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/e-WB4lfg30M\">slika u tekst<\/a><\/noindex> i <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/rAbhypxs1qQ\">teksta u sliku<\/a><\/noindex>.<\/li>\n<li>Kreiranje <noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/OrHLacCDZVQ\">trodimenzionalnih objekata u video<\/a><\/noindex>, sada ve\u0107 u realnom vremenu.<\/li>\n<li>Osnovno u DL \u2013 to su velike koli\u010dine podataka, ali prikupljanje i ozna\u010davanje nije lako. Zato se razvija automatizacija ozna\u010davanja (<noindex><a rel=\"nofollow\" href=\"https:\/\/youtu.be\/NcKTn4C91Yc\">automated annotation<\/a><\/noindex>) za neuronske mre\u017ee pomo\u0107u neuronskih mre\u017ea.<\/li>\n<\/ul>\n<\/li>\n<li>Sa neuronskim mre\u017eama, Ra\u010dunarske nauke su iznenada postale <strong>eksperimentalna nauka<\/strong> i do\u0161lo je do <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/480348\">krize reproduktivnosti<\/a><\/noindex>.<\/li>\n<li>IT-novac i popularnost neuronskih mre\u017ea su se pojavili istovremeno, kada su prora\u010duni postali tr\u017ei\u0161na vrednost. Ekonomija iz zlato-valutne postaje <strong>zlato-valutno-prora\u010dunska<\/strong>. Poglejte moj \u010dlanak o <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%AD%D0%BA%D0%BE%D0%BD%D0%BE%D1%84%D0%B8%D0%B7%D0%B8%D0%BA%D0%B0\">ekonofizici<\/a><\/noindex> i uzroku pojave IT-novca.<\/li>\n<\/ul>\n<p><\/p>\n<p>Postepeno se pojavljuje nova <noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/481844\">metodologija programiranja ML\/DL<\/a><\/noindex> (Ma\u0161insko U\u010denje &amp; Duboko U\u010denje), koja se temelji na predstavi programa kao skupa obu\u010denih modela neuronskih mre\u017ea.<\/p>\n<p><\/p>\n<p><img decoding=\"async\" alt=\"Neuralne mre\u017ee. Kuda ovo sve ide\" src=\"\/wp-content\/uploads\/2020\/01\/17df76a613191522145abeba177c2685.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><\/p>\n<p>Slika 3 \u2013 ML\/DL kao nova metodologija programiranja<\/p>\n<p><\/p>\n<p>Me\u0111utim, jo\u0161 uvek nije postojala <strong>\"teorija neuronskih mre\u017ea\"<\/strong>, u okviru koje se mo\u017ee razmi\u0161ljati i raditi sistematski. Ono \u0161to se sada naziva \"teorijom\" u stvari su eksperimentalni, heuristi\u010dki algoritmi.<\/p>\n<p><\/p>\n<p>Reference na moje i druge resurse:<\/p>\n<p><\/p>\n<ul>\n<li>Newsletter o Data Science. Uglavnom, o obradi slika. Ko \u017eeli da prima, neka po\u0161alje e-mail (foobar167&lt;gaff-gaff&gt;gmail&lt;dot&gt;com). Linkove na \u010dlanke i video \u0161aljem kako se materijal skuplja.<\/li>\n<li>Op\u0161ti <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/foobar167\/articles\/blob\/master\/Machine_Learning\/courses_on_machine_learning.md\">spisak kurseva i \u010dlanaka<\/a><\/noindex>, koje sam pro\u0161ao i koje bih \u017eeleo da pro\u0111em.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/foobar167\/articles\/blob\/master\/Ubuntu\/13_Keras_and_TensorFlow_how-tos.md#exercises\">Kurseti i video za po\u010detnike<\/a><\/noindex>, sa kojima treba po\u010deti u\u010diti o neuronskim mre\u017eama. Plus bro\u0161ura <noindex><a rel=\"nofollow\" href=\"https:\/\/foobar167.github.io\/page\/vvedeniye-v-mashinnoye-obucheniye-i-iskusstvennyye-neyronnyye-seti.html\">\"Uvod u ma\u0161insko u\u010denje i ve\u0161ta\u010dke neuronske mre\u017ee\"<\/a><\/noindex>.<\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/foobar167\/articles\/blob\/master\/Ubuntu\/13_Keras_and_TensorFlow_how-tos.md#tools\">Koristan alati<\/a><\/noindex>, gdje svako mo\u017ee prona\u0107i ne\u0161to zanimljivo za sebe.<\/li>\n<li>Iznimno korisni su bili <strong>video kanali koji razla\u017eu nau\u010dne \u010dlanke<\/strong> o Data Science. Prona\u0111ite ih, pretplatite se i podijelite linkove s kolegama i sa mnom tako\u0111er. Primjeri:\n<ul>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/www.youtube.com\/user\/keeroyz\">Two Minute Papers<\/a><\/noindex><\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/www.youtube.com\/channel\/UCHB9VepY6kYvZjj0Bgxnpbw\">Henry AI Labs<\/a><\/noindex><\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/www.youtube.com\/channel\/UCZHmQk67mSJgfCCTn7xBfew\">Yannic Kilcher<\/a><\/noindex><\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/www.youtube.com\/channel\/UC5_6ZD6s8klmMu9TXEB_1IA\">CodeEmporium<\/a><\/noindex><\/li>\n<li><noindex><a rel=\"nofollow\" href=\"https:\/\/www.dlology.com\">Blog Chengwei Zhang<\/a><\/noindex> poznat kao <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/Tony607\">Tony607<\/a><\/noindex> s korak-po-korak uputama i otvorenim kodom.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><\/p>\n<p>Hvala na pa\u017enji!<\/p>\n<p>Izvor: <a content=\"nofollow\" rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/482794\/\">habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"<p>\u0421\u0442\u0430\u0442\u044c\u044f \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u0447\u0430\u0441\u0442\u0435\u0439: \u041a\u0440\u0430\u0442\u043a\u043e\u0435 \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440 \u0441\u0435\u0442\u0435\u0439 \u043f\u043e \u043e\u0431\u043d\u0430\u0440\u0443\u0436\u0435\u043d\u0438\u044e \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432 \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u0438 \u0441\u0435\u0433\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439 \u0441 \u0441\u0430\u043c\u044b\u043c\u0438 \u043f\u043e\u043d\u044f\u0442\u043d\u044b\u043c\u0438 \u0434\u043b\u044f \u043c\u0435\u043d\u044f \u0441\u0441\u044b\u043b\u043a\u0430\u043c\u0438 \u043d\u0430 \u0440\u0435\u0441\u0443\u0440\u0441\u044b. \u0421\u0442\u0430\u0440\u0430\u043b\u0441\u044f \u0432\u044b\u0431\u0438\u0440\u0430\u0442\u044c \u0432\u0438\u0434\u0435\u043e \u043f\u043e\u044f\u0441\u043d\u0435\u043d\u0438\u044f \u0438 \u0436\u0435\u043b\u0430\u0442\u0435\u043b\u044c\u043d\u043e \u043d\u0430 \u0440\u0443\u0441\u0441\u043a\u043e\u043c \u044f\u0437\u044b\u043a\u0435. \u0412\u0442\u043e\u0440\u0430\u044f \u0447\u0430\u0441\u0442\u044c \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0432 \u043f\u043e\u043f\u044b\u0442\u043a\u0435 \u043e\u0441\u043e\u0437\u043d\u0430\u0442\u044c \u043d\u0430\u043f\u0440\u0430\u0432\u043b\u0435\u043d\u0438\u0435 \u0440\u0430\u0437\u0432\u0438\u0442\u0438\u044f \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u0435\u0439. \u0418 \u0442\u0435\u0445\u043d\u043e\u043b\u043e\u0433\u0438\u0439 \u043d\u0430 \u0438\u0445 \u043e\u0441\u043d\u043e\u0432\u0435. \u0420\u0438\u0441\u0443\u043d\u043e\u043a 1 \u2013 \u041f\u043e\u043d\u0438\u043c\u0430\u0442\u044c [&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-54780","post","type-post","status-publish","format-standard","hentry","category-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 5.0.5.1 - aioseo.com -->\n\t<meta name=\"description\" content=\"\u0421\u0442\u0430\u0442\u044c\u044f \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u0447\u0430\u0441\u0442\u0435\u0439: \u041a\u0440\u0430\u0442\u043a\u043e\u0435 \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440 \u0441\u0435\u0442\u0435\u0439 \u043f\u043e \u043e\u0431\u043d\u0430\u0440\u0443\u0436\u0435\u043d\u0438\u044e \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432 \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u0438 \u0441\u0435\u0433\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439 \u0441 \u0441\u0430\u043c\u044b\u043c\u0438 \u043f\u043e\u043d\u044f\u0442\u043d\u044b\u043c\u0438 \u0434\u043b\u044f \u043c\u0435\u043d\u044f \u0441\u0441\u044b\u043b\u043a\u0430\u043c\u0438 \u043d\u0430 \u0440\u0435\u0441\u0443\u0440\u0441\u044b.\" \/>\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\/bs\/blog\/news\/nejroseti-kuda-eto-vse-dvizhetsya\" \/>\n\t\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.5.1\" \/>\n\t\t<meta property=\"og:locale\" content=\"bs_BA\" \/>\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\u041d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438. \u041a\u0443\u0434\u0430 \u044d\u0442\u043e \u0432\u0441\u0435 \u0434\u0432\u0438\u0436\u0435\u0442\u0441\u044f | ProHoster\" \/>\n\t\t<meta property=\"og:description\" content=\"\u0421\u0442\u0430\u0442\u044c\u044f \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u0447\u0430\u0441\u0442\u0435\u0439: \u041a\u0440\u0430\u0442\u043a\u043e\u0435 \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440 \u0441\u0435\u0442\u0435\u0439 \u043f\u043e \u043e\u0431\u043d\u0430\u0440\u0443\u0436\u0435\u043d\u0438\u044e \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432 \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u0438 \u0441\u0435\u0433\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439 \u0441 \u0441\u0430\u043c\u044b\u043c\u0438 \u043f\u043e\u043d\u044f\u0442\u043d\u044b\u043c\u0438 \u0434\u043b\u044f \u043c\u0435\u043d\u044f \u0441\u0441\u044b\u043b\u043a\u0430\u043c\u0438 \u043d\u0430 \u0440\u0435\u0441\u0443\u0440\u0441\u044b.\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/prohoster.info\/bs\/blog\/news\/nejroseti-kuda-eto-vse-dvizhetsya\" \/>\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=\"2020-01-03T21:00:00+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2020-02-18T11:02:50+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\udd47Neuro mre\u017ee. Kuda sve ovo vodi | ProHoster","description":"\u010clanak se sastoji od dva dijela: Kratak prikaz nekih arhitektura mre\u017ea za prepoznavanje objekata na slici i segmentaciju slika s najjasnijim linkovima na resurse za mene.","canonical_url":"https:\/\/prohoster.info\/bs\/blog\/news\/nejroseti-kuda-eto-vse-dvizhetsya","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"bs_BA","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\u041d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438. \u041a\u0443\u0434\u0430 \u044d\u0442\u043e \u0432\u0441\u0435 \u0434\u0432\u0438\u0436\u0435\u0442\u0441\u044f | ProHoster","og:description":"\u0421\u0442\u0430\u0442\u044c\u044f \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u0447\u0430\u0441\u0442\u0435\u0439: \u041a\u0440\u0430\u0442\u043a\u043e\u0435 \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440 \u0441\u0435\u0442\u0435\u0439 \u043f\u043e \u043e\u0431\u043d\u0430\u0440\u0443\u0436\u0435\u043d\u0438\u044e \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432 \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u0438 \u0441\u0435\u0433\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439 \u0441 \u0441\u0430\u043c\u044b\u043c\u0438 \u043f\u043e\u043d\u044f\u0442\u043d\u044b\u043c\u0438 \u0434\u043b\u044f \u043c\u0435\u043d\u044f \u0441\u0441\u044b\u043b\u043a\u0430\u043c\u0438 \u043d\u0430 \u0440\u0435\u0441\u0443\u0440\u0441\u044b.","og:url":"https:\/\/prohoster.info\/bs\/blog\/news\/nejroseti-kuda-eto-vse-dvizhetsya","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":"2020-01-03T21:00:00+00:00","article:modified_time":"2020-02-18T11:02:50+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"54780","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":"2026-01-24 12:43:20","breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-02-28 19:57:38","updated":"2026-01-24 12:43:20","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/posts\/54780","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/comments?post=54780"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/posts\/54780\/revisions"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/media?parent=54780"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/categories?post=54780"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/bs\/wp-json\/wp\/v2\/tags?post=54780"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}