id sid tid token lemma pos adeshpande3-github-io-4431 1 1 The the DT adeshpande3-github-io-4431 1 2 9 9 CD adeshpande3-github-io-4431 1 3 Deep Deep NNP adeshpande3-github-io-4431 1 4 Learning learning NN adeshpande3-github-io-4431 1 5 Papers paper NNS adeshpande3-github-io-4431 1 6 You -PRON- PRP adeshpande3-github-io-4431 1 7 Need need VBP adeshpande3-github-io-4431 1 8 To to TO adeshpande3-github-io-4431 1 9 Know know VB adeshpande3-github-io-4431 1 10 About about IN adeshpande3-github-io-4431 1 11 ( ( -LRB- adeshpande3-github-io-4431 1 12 Understanding understand VBG adeshpande3-github-io-4431 1 13 CNNs cnn NNS adeshpande3-github-io-4431 1 14 Part Part NNP adeshpande3-github-io-4431 1 15 3 3 CD adeshpande3-github-io-4431 1 16 ) ) -RRB- adeshpande3-github-io-4431 1 17 – – : adeshpande3-github-io-4431 1 18 Adit Adit NNP adeshpande3-github-io-4431 1 19 Deshpande Deshpande NNP adeshpande3-github-io-4431 1 20 – – : adeshpande3-github-io-4431 1 21 Engineering engineering NN adeshpande3-github-io-4431 1 22 at at IN adeshpande3-github-io-4431 1 23 Forward Forward NNP adeshpande3-github-io-4431 1 24 | | NNP adeshpande3-github-io-4431 1 25 UCLA UCLA NNP adeshpande3-github-io-4431 1 26 CS CS NNP adeshpande3-github-io-4431 1 27 ' ' POS adeshpande3-github-io-4431 1 28 19 19 CD adeshpande3-github-io-4431 1 29 Adit Adit NNP adeshpande3-github-io-4431 1 30 Deshpande Deshpande NNP adeshpande3-github-io-4431 1 31 Engineering Engineering NNP adeshpande3-github-io-4431 1 32 at at IN adeshpande3-github-io-4431 1 33 Forward Forward NNP adeshpande3-github-io-4431 1 34 | | NNP adeshpande3-github-io-4431 1 35 UCLA UCLA NNP adeshpande3-github-io-4431 1 36 CS CS NNP adeshpande3-github-io-4431 1 37 ' ' POS adeshpande3-github-io-4431 1 38 19 19 CD adeshpande3-github-io-4431 1 39 Blog blog NN adeshpande3-github-io-4431 1 40 About about IN adeshpande3-github-io-4431 1 41 GitHub GitHub NNP adeshpande3-github-io-4431 1 42 Projects Projects NNPS adeshpande3-github-io-4431 1 43 Resume resume VBP adeshpande3-github-io-4431 1 44 The the DT adeshpande3-github-io-4431 1 45 9 9 CD adeshpande3-github-io-4431 1 46 Deep Deep NNP adeshpande3-github-io-4431 1 47 Learning learning NN adeshpande3-github-io-4431 1 48 Papers paper NNS adeshpande3-github-io-4431 1 49 You -PRON- PRP adeshpande3-github-io-4431 1 50 Need need VBP adeshpande3-github-io-4431 1 51 To to TO adeshpande3-github-io-4431 1 52 Know know VB adeshpande3-github-io-4431 1 53 About about IN adeshpande3-github-io-4431 1 54 ( ( -LRB- adeshpande3-github-io-4431 1 55 Understanding understand VBG adeshpande3-github-io-4431 1 56 CNNs cnn NNS adeshpande3-github-io-4431 1 57 Part part NN adeshpande3-github-io-4431 1 58 3 3 CD adeshpande3-github-io-4431 1 59 ) ) -RRB- adeshpande3-github-io-4431 1 60 Introduction introduction NN adeshpande3-github-io-4431 1 61 Link Link NNP adeshpande3-github-io-4431 1 62 to to IN adeshpande3-github-io-4431 1 63 Part part NN adeshpande3-github-io-4431 1 64 1 1 CD adeshpande3-github-io-4431 1 65 Link link NN adeshpande3-github-io-4431 1 66 to to IN adeshpande3-github-io-4431 1 67 Part part NN adeshpande3-github-io-4431 1 68 2 2 CD adeshpande3-github-io-4431 1 69                                 _SP adeshpande3-github-io-4431 1 70 In in IN adeshpande3-github-io-4431 1 71 this this DT adeshpande3-github-io-4431 1 72 post post NN adeshpande3-github-io-4431 1 73 , , , adeshpande3-github-io-4431 1 74 we -PRON- PRP adeshpande3-github-io-4431 1 75 ’ll will MD adeshpande3-github-io-4431 1 76 go go VB adeshpande3-github-io-4431 1 77 into into IN adeshpande3-github-io-4431 1 78 summarizing summarize VBG adeshpande3-github-io-4431 1 79 a a DT adeshpande3-github-io-4431 1 80 lot lot NN adeshpande3-github-io-4431 1 81 of of IN adeshpande3-github-io-4431 1 82 the the DT adeshpande3-github-io-4431 1 83 new new JJ adeshpande3-github-io-4431 1 84 and and CC adeshpande3-github-io-4431 1 85 important important JJ adeshpande3-github-io-4431 1 86 developments development NNS adeshpande3-github-io-4431 1 87 in in IN adeshpande3-github-io-4431 1 88 the the DT adeshpande3-github-io-4431 1 89 field field NN adeshpande3-github-io-4431 1 90 of of IN adeshpande3-github-io-4431 1 91 computer computer NN adeshpande3-github-io-4431 1 92 vision vision NN adeshpande3-github-io-4431 1 93 and and CC adeshpande3-github-io-4431 1 94 convolutional convolutional JJ adeshpande3-github-io-4431 1 95 neural neural JJ adeshpande3-github-io-4431 1 96 networks network NNS adeshpande3-github-io-4431 1 97 . . . adeshpande3-github-io-4431 2 1 We -PRON- PRP adeshpande3-github-io-4431 2 2 ’ll will MD adeshpande3-github-io-4431 2 3 look look VB adeshpande3-github-io-4431 2 4 at at IN adeshpande3-github-io-4431 2 5 some some DT adeshpande3-github-io-4431 2 6 of of IN adeshpande3-github-io-4431 2 7 the the DT adeshpande3-github-io-4431 2 8 most most RBS adeshpande3-github-io-4431 2 9 important important JJ adeshpande3-github-io-4431 2 10 papers paper NNS adeshpande3-github-io-4431 2 11 that that WDT adeshpande3-github-io-4431 2 12 have have VBP adeshpande3-github-io-4431 2 13 been be VBN adeshpande3-github-io-4431 2 14 published publish VBN adeshpande3-github-io-4431 2 15 over over IN adeshpande3-github-io-4431 2 16 the the DT adeshpande3-github-io-4431 2 17 last last JJ adeshpande3-github-io-4431 2 18 5 5 CD adeshpande3-github-io-4431 2 19 years year NNS adeshpande3-github-io-4431 2 20 and and CC adeshpande3-github-io-4431 2 21 discuss discuss VB adeshpande3-github-io-4431 2 22 why why WRB adeshpande3-github-io-4431 2 23 they -PRON- PRP adeshpande3-github-io-4431 2 24 ’re be VBP adeshpande3-github-io-4431 2 25 so so RB adeshpande3-github-io-4431 2 26 important important JJ adeshpande3-github-io-4431 2 27 . . . adeshpande3-github-io-4431 3 1 The the DT adeshpande3-github-io-4431 3 2 first first JJ adeshpande3-github-io-4431 3 3 half half NN adeshpande3-github-io-4431 3 4 of of IN adeshpande3-github-io-4431 3 5 the the DT adeshpande3-github-io-4431 3 6 list list NN adeshpande3-github-io-4431 3 7 ( ( -LRB- adeshpande3-github-io-4431 3 8 AlexNet AlexNet NNP adeshpande3-github-io-4431 3 9 to to IN adeshpande3-github-io-4431 3 10 ResNet ResNet NNP adeshpande3-github-io-4431 3 11 ) ) -RRB- adeshpande3-github-io-4431 3 12 deals deal VBZ adeshpande3-github-io-4431 3 13 with with IN adeshpande3-github-io-4431 3 14 advancements advancement NNS adeshpande3-github-io-4431 3 15 in in IN adeshpande3-github-io-4431 3 16 general general JJ adeshpande3-github-io-4431 3 17 network network NN adeshpande3-github-io-4431 3 18 architecture architecture NN adeshpande3-github-io-4431 3 19 , , , adeshpande3-github-io-4431 3 20 while while IN adeshpande3-github-io-4431 3 21 the the DT adeshpande3-github-io-4431 3 22 second second JJ adeshpande3-github-io-4431 3 23 half half NN adeshpande3-github-io-4431 3 24 is be VBZ adeshpande3-github-io-4431 3 25 just just RB adeshpande3-github-io-4431 3 26 a a DT adeshpande3-github-io-4431 3 27 collection collection NN adeshpande3-github-io-4431 3 28 of of IN adeshpande3-github-io-4431 3 29 interesting interesting JJ adeshpande3-github-io-4431 3 30 papers paper NNS adeshpande3-github-io-4431 3 31 in in IN adeshpande3-github-io-4431 3 32 other other JJ adeshpande3-github-io-4431 3 33 subareas subarea NNS adeshpande3-github-io-4431 3 34 . . . adeshpande3-github-io-4431 4 1 AlexNet AlexNet NNP adeshpande3-github-io-4431 4 2   _SP adeshpande3-github-io-4431 4 3 ( ( -LRB- adeshpande3-github-io-4431 4 4 2012 2012 CD adeshpande3-github-io-4431 4 5 ) ) -RRB- adeshpande3-github-io-4431 4 6                                 _SP adeshpande3-github-io-4431 4 7 The the DT adeshpande3-github-io-4431 4 8 one one NN adeshpande3-github-io-4431 4 9 that that WDT adeshpande3-github-io-4431 4 10 started start VBD adeshpande3-github-io-4431 4 11 it -PRON- PRP adeshpande3-github-io-4431 4 12 all all DT adeshpande3-github-io-4431 4 13 ( ( -LRB- adeshpande3-github-io-4431 4 14 Though though IN adeshpande3-github-io-4431 4 15 some some DT adeshpande3-github-io-4431 4 16 may may MD adeshpande3-github-io-4431 4 17 say say VB adeshpande3-github-io-4431 4 18 that that IN adeshpande3-github-io-4431 4 19 Yann Yann NNP adeshpande3-github-io-4431 4 20 LeCun LeCun NNP adeshpande3-github-io-4431 4 21 ’s ’s POS adeshpande3-github-io-4431 4 22 paper paper NN adeshpande3-github-io-4431 4 23 in in IN adeshpande3-github-io-4431 4 24 1998 1998 CD adeshpande3-github-io-4431 4 25 was be VBD adeshpande3-github-io-4431 4 26 the the DT adeshpande3-github-io-4431 4 27 real real JJ adeshpande3-github-io-4431 4 28 pioneering pioneering JJ adeshpande3-github-io-4431 4 29 publication publication NN adeshpande3-github-io-4431 4 30 ) ) -RRB- adeshpande3-github-io-4431 4 31 . . . adeshpande3-github-io-4431 5 1 This this DT adeshpande3-github-io-4431 5 2 paper paper NN adeshpande3-github-io-4431 5 3 , , , adeshpande3-github-io-4431 5 4 titled title VBN adeshpande3-github-io-4431 5 5 “ " `` adeshpande3-github-io-4431 5 6 ImageNet ImageNet NNP adeshpande3-github-io-4431 5 7 Classification classification NN adeshpande3-github-io-4431 5 8 with with IN adeshpande3-github-io-4431 5 9 Deep Deep NNP adeshpande3-github-io-4431 5 10 Convolutional Convolutional NNP adeshpande3-github-io-4431 5 11 Networks Networks NNPS adeshpande3-github-io-4431 5 12 ” " '' adeshpande3-github-io-4431 5 13 , , , adeshpande3-github-io-4431 5 14 has have VBZ adeshpande3-github-io-4431 5 15 been be VBN adeshpande3-github-io-4431 5 16 cited cite VBN adeshpande3-github-io-4431 5 17 a a DT adeshpande3-github-io-4431 5 18 total total NN adeshpande3-github-io-4431 5 19 of of IN adeshpande3-github-io-4431 5 20 6,184 6,184 CD adeshpande3-github-io-4431 5 21 times time NNS adeshpande3-github-io-4431 5 22 and and CC adeshpande3-github-io-4431 5 23 is be VBZ adeshpande3-github-io-4431 5 24 widely widely RB adeshpande3-github-io-4431 5 25 regarded regard VBN adeshpande3-github-io-4431 5 26 as as IN adeshpande3-github-io-4431 5 27 one one CD adeshpande3-github-io-4431 5 28 of of IN adeshpande3-github-io-4431 5 29 the the DT adeshpande3-github-io-4431 5 30 most most RBS adeshpande3-github-io-4431 5 31 influential influential JJ adeshpande3-github-io-4431 5 32 publications publication NNS adeshpande3-github-io-4431 5 33 in in IN adeshpande3-github-io-4431 5 34 the the DT adeshpande3-github-io-4431 5 35 field field NN adeshpande3-github-io-4431 5 36 . . . adeshpande3-github-io-4431 6 1 Alex Alex NNP adeshpande3-github-io-4431 6 2 Krizhevsky Krizhevsky NNP adeshpande3-github-io-4431 6 3 , , , adeshpande3-github-io-4431 6 4 Ilya Ilya NNP adeshpande3-github-io-4431 6 5 Sutskever Sutskever NNP adeshpande3-github-io-4431 6 6 , , , adeshpande3-github-io-4431 6 7 and and CC adeshpande3-github-io-4431 6 8 Geoffrey Geoffrey NNP adeshpande3-github-io-4431 6 9 Hinton Hinton NNP adeshpande3-github-io-4431 6 10 created create VBD adeshpande3-github-io-4431 6 11 a a DT adeshpande3-github-io-4431 6 12 “ " `` adeshpande3-github-io-4431 6 13 large large JJ adeshpande3-github-io-4431 6 14 , , , adeshpande3-github-io-4431 6 15 deep deep JJ adeshpande3-github-io-4431 6 16 convolutional convolutional JJ adeshpande3-github-io-4431 6 17 neural neural JJ adeshpande3-github-io-4431 6 18 network network NN adeshpande3-github-io-4431 6 19 ” " '' adeshpande3-github-io-4431 6 20 that that WDT adeshpande3-github-io-4431 6 21 was be VBD adeshpande3-github-io-4431 6 22 used use VBN adeshpande3-github-io-4431 6 23 to to TO adeshpande3-github-io-4431 6 24 win win VB adeshpande3-github-io-4431 6 25 the the DT adeshpande3-github-io-4431 6 26 2012 2012 CD adeshpande3-github-io-4431 6 27 ILSVRC ILSVRC NNP adeshpande3-github-io-4431 6 28 ( ( -LRB- adeshpande3-github-io-4431 6 29 ImageNet ImageNet NNP adeshpande3-github-io-4431 6 30 Large large JJ adeshpande3-github-io-4431 6 31 - - HYPH adeshpande3-github-io-4431 6 32 Scale scale NN adeshpande3-github-io-4431 6 33 Visual visual JJ adeshpande3-github-io-4431 6 34 Recognition recognition NN adeshpande3-github-io-4431 6 35 Challenge Challenge NNP adeshpande3-github-io-4431 6 36 ) ) -RRB- adeshpande3-github-io-4431 6 37 . . . adeshpande3-github-io-4431 7 1 For for IN adeshpande3-github-io-4431 7 2 those those DT adeshpande3-github-io-4431 7 3 that that WDT adeshpande3-github-io-4431 7 4 are be VBP adeshpande3-github-io-4431 7 5 n’t not RB adeshpande3-github-io-4431 7 6 familiar familiar JJ adeshpande3-github-io-4431 7 7 , , , adeshpande3-github-io-4431 7 8 this this DT adeshpande3-github-io-4431 7 9 competition competition NN adeshpande3-github-io-4431 7 10 can can MD adeshpande3-github-io-4431 7 11 be be VB adeshpande3-github-io-4431 7 12 thought think VBN adeshpande3-github-io-4431 7 13 of of IN adeshpande3-github-io-4431 7 14 as as IN adeshpande3-github-io-4431 7 15 the the DT adeshpande3-github-io-4431 7 16 annual annual JJ adeshpande3-github-io-4431 7 17 Olympics Olympics NNPS adeshpande3-github-io-4431 7 18 of of IN adeshpande3-github-io-4431 7 19 computer computer NN adeshpande3-github-io-4431 7 20 vision vision NN adeshpande3-github-io-4431 7 21 , , , adeshpande3-github-io-4431 7 22 where where WRB adeshpande3-github-io-4431 7 23 teams team NNS adeshpande3-github-io-4431 7 24 from from IN adeshpande3-github-io-4431 7 25 across across IN adeshpande3-github-io-4431 7 26 the the DT adeshpande3-github-io-4431 7 27 world world NN adeshpande3-github-io-4431 7 28 compete compete VBP adeshpande3-github-io-4431 7 29 to to TO adeshpande3-github-io-4431 7 30 see see VB adeshpande3-github-io-4431 7 31 who who WP adeshpande3-github-io-4431 7 32 has have VBZ adeshpande3-github-io-4431 7 33 the the DT adeshpande3-github-io-4431 7 34 best good JJS adeshpande3-github-io-4431 7 35 computer computer NN adeshpande3-github-io-4431 7 36 vision vision NN adeshpande3-github-io-4431 7 37 model model NN adeshpande3-github-io-4431 7 38 for for IN adeshpande3-github-io-4431 7 39 tasks task NNS adeshpande3-github-io-4431 7 40 such such JJ adeshpande3-github-io-4431 7 41 as as IN adeshpande3-github-io-4431 7 42 classification classification NN adeshpande3-github-io-4431 7 43 , , , adeshpande3-github-io-4431 7 44 localization localization NN adeshpande3-github-io-4431 7 45 , , , adeshpande3-github-io-4431 7 46 detection detection NN adeshpande3-github-io-4431 7 47 , , , adeshpande3-github-io-4431 7 48 and and CC adeshpande3-github-io-4431 7 49 more more JJR adeshpande3-github-io-4431 7 50 . . . adeshpande3-github-io-4431 8 1 2012 2012 CD adeshpande3-github-io-4431 8 2 marked mark VBD adeshpande3-github-io-4431 8 3 the the DT adeshpande3-github-io-4431 8 4 first first JJ adeshpande3-github-io-4431 8 5 year year NN adeshpande3-github-io-4431 8 6 where where WRB adeshpande3-github-io-4431 8 7 a a DT adeshpande3-github-io-4431 8 8 CNN CNN NNP adeshpande3-github-io-4431 8 9 was be VBD adeshpande3-github-io-4431 8 10 used use VBN adeshpande3-github-io-4431 8 11 to to TO adeshpande3-github-io-4431 8 12 achieve achieve VB adeshpande3-github-io-4431 8 13 a a DT adeshpande3-github-io-4431 8 14 top top JJ adeshpande3-github-io-4431 8 15 5 5 CD adeshpande3-github-io-4431 8 16 test test NN adeshpande3-github-io-4431 8 17 error error NN adeshpande3-github-io-4431 8 18 rate rate NN adeshpande3-github-io-4431 8 19 of of IN adeshpande3-github-io-4431 8 20 15.4 15.4 CD adeshpande3-github-io-4431 8 21 % % NN adeshpande3-github-io-4431 8 22 ( ( -LRB- adeshpande3-github-io-4431 8 23 Top Top NNP adeshpande3-github-io-4431 8 24 5 5 CD adeshpande3-github-io-4431 8 25 error error NN adeshpande3-github-io-4431 8 26 is be VBZ adeshpande3-github-io-4431 8 27 the the DT adeshpande3-github-io-4431 8 28 rate rate NN adeshpande3-github-io-4431 8 29 at at IN adeshpande3-github-io-4431 8 30 which which WDT adeshpande3-github-io-4431 8 31 , , , adeshpande3-github-io-4431 8 32 given give VBN adeshpande3-github-io-4431 8 33 an an DT adeshpande3-github-io-4431 8 34 image image NN adeshpande3-github-io-4431 8 35 , , , adeshpande3-github-io-4431 8 36 the the DT adeshpande3-github-io-4431 8 37 model model NN adeshpande3-github-io-4431 8 38 does do VBZ adeshpande3-github-io-4431 8 39 not not RB adeshpande3-github-io-4431 8 40 output output VB adeshpande3-github-io-4431 8 41 the the DT adeshpande3-github-io-4431 8 42 correct correct JJ adeshpande3-github-io-4431 8 43 label label NN adeshpande3-github-io-4431 8 44 with with IN adeshpande3-github-io-4431 8 45 its -PRON- PRP$ adeshpande3-github-io-4431 8 46 top top JJ adeshpande3-github-io-4431 8 47 5 5 CD adeshpande3-github-io-4431 8 48 predictions prediction NNS adeshpande3-github-io-4431 8 49 ) ) -RRB- adeshpande3-github-io-4431 8 50 . . . adeshpande3-github-io-4431 9 1 The the DT adeshpande3-github-io-4431 9 2 next next JJ adeshpande3-github-io-4431 9 3 best good JJS adeshpande3-github-io-4431 9 4 entry entry NN adeshpande3-github-io-4431 9 5 achieved achieve VBD adeshpande3-github-io-4431 9 6 an an DT adeshpande3-github-io-4431 9 7 error error NN adeshpande3-github-io-4431 9 8 of of IN adeshpande3-github-io-4431 9 9 26.2 26.2 CD adeshpande3-github-io-4431 9 10 % % NN adeshpande3-github-io-4431 9 11 , , , adeshpande3-github-io-4431 9 12 which which WDT adeshpande3-github-io-4431 9 13 was be VBD adeshpande3-github-io-4431 9 14 an an DT adeshpande3-github-io-4431 9 15 astounding astounding JJ adeshpande3-github-io-4431 9 16 improvement improvement NN adeshpande3-github-io-4431 9 17 that that WDT adeshpande3-github-io-4431 9 18 pretty pretty RB adeshpande3-github-io-4431 9 19 much much RB adeshpande3-github-io-4431 9 20 shocked shock VBD adeshpande3-github-io-4431 9 21 the the DT adeshpande3-github-io-4431 9 22 computer computer NN adeshpande3-github-io-4431 9 23 vision vision NN adeshpande3-github-io-4431 9 24 community community NN adeshpande3-github-io-4431 9 25 . . . adeshpande3-github-io-4431 10 1 Safe safe JJ adeshpande3-github-io-4431 10 2 to to TO adeshpande3-github-io-4431 10 3 say say VB adeshpande3-github-io-4431 10 4 , , , adeshpande3-github-io-4431 10 5 CNNs cnn NNS adeshpande3-github-io-4431 10 6 became become VBD adeshpande3-github-io-4431 10 7 household household NN adeshpande3-github-io-4431 10 8 names name NNS adeshpande3-github-io-4431 10 9 in in IN adeshpande3-github-io-4431 10 10 the the DT adeshpande3-github-io-4431 10 11 competition competition NN adeshpande3-github-io-4431 10 12 from from IN adeshpande3-github-io-4431 10 13 then then RB adeshpande3-github-io-4431 10 14 on on RB adeshpande3-github-io-4431 10 15 out out RB adeshpande3-github-io-4431 10 16 . . . adeshpande3-github-io-4431 11 1 In in IN adeshpande3-github-io-4431 11 2 the the DT adeshpande3-github-io-4431 11 3 paper paper NN adeshpande3-github-io-4431 11 4 , , , adeshpande3-github-io-4431 11 5 the the DT adeshpande3-github-io-4431 11 6 group group NN adeshpande3-github-io-4431 11 7 discussed discuss VBD adeshpande3-github-io-4431 11 8 the the DT adeshpande3-github-io-4431 11 9 architecture architecture NN adeshpande3-github-io-4431 11 10 of of IN adeshpande3-github-io-4431 11 11 the the DT adeshpande3-github-io-4431 11 12 network network NN adeshpande3-github-io-4431 11 13 ( ( -LRB- adeshpande3-github-io-4431 11 14 which which WDT adeshpande3-github-io-4431 11 15 was be VBD adeshpande3-github-io-4431 11 16 called call VBN adeshpande3-github-io-4431 11 17 AlexNet AlexNet NNP adeshpande3-github-io-4431 11 18 ) ) -RRB- adeshpande3-github-io-4431 11 19 . . . adeshpande3-github-io-4431 12 1 They -PRON- PRP adeshpande3-github-io-4431 12 2 used use VBD adeshpande3-github-io-4431 12 3 a a DT adeshpande3-github-io-4431 12 4 relatively relatively RB adeshpande3-github-io-4431 12 5 simple simple JJ adeshpande3-github-io-4431 12 6 layout layout NN adeshpande3-github-io-4431 12 7 , , , adeshpande3-github-io-4431 12 8 compared compare VBN adeshpande3-github-io-4431 12 9 to to IN adeshpande3-github-io-4431 12 10 modern modern JJ adeshpande3-github-io-4431 12 11 architectures architecture NNS adeshpande3-github-io-4431 12 12 . . . adeshpande3-github-io-4431 13 1 The the DT adeshpande3-github-io-4431 13 2 network network NN adeshpande3-github-io-4431 13 3 was be VBD adeshpande3-github-io-4431 13 4 made make VBN adeshpande3-github-io-4431 13 5 up up RP adeshpande3-github-io-4431 13 6 of of IN adeshpande3-github-io-4431 13 7 5 5 CD adeshpande3-github-io-4431 13 8 conv conv NN adeshpande3-github-io-4431 13 9 layers layer NNS adeshpande3-github-io-4431 13 10 , , , adeshpande3-github-io-4431 13 11 max max NNP adeshpande3-github-io-4431 13 12 - - HYPH adeshpande3-github-io-4431 13 13 pooling pooling NN adeshpande3-github-io-4431 13 14 layers layer NNS adeshpande3-github-io-4431 13 15 , , , adeshpande3-github-io-4431 13 16 dropout dropout NN adeshpande3-github-io-4431 13 17 layers layer NNS adeshpande3-github-io-4431 13 18 , , , adeshpande3-github-io-4431 13 19 and and CC adeshpande3-github-io-4431 13 20 3 3 CD adeshpande3-github-io-4431 13 21 fully fully RB adeshpande3-github-io-4431 13 22 connected connect VBN adeshpande3-github-io-4431 13 23 layers layer NNS adeshpande3-github-io-4431 13 24 . . . adeshpande3-github-io-4431 14 1 The the DT adeshpande3-github-io-4431 14 2 network network NN adeshpande3-github-io-4431 14 3 they -PRON- PRP adeshpande3-github-io-4431 14 4 designed design VBD adeshpande3-github-io-4431 14 5 was be VBD adeshpande3-github-io-4431 14 6 used use VBN adeshpande3-github-io-4431 14 7 for for IN adeshpande3-github-io-4431 14 8 classification classification NN adeshpande3-github-io-4431 14 9 with with IN adeshpande3-github-io-4431 14 10 1000 1000 CD adeshpande3-github-io-4431 14 11 possible possible JJ adeshpande3-github-io-4431 14 12 categories category NNS adeshpande3-github-io-4431 14 13 . . . adeshpande3-github-io-4431 15 1 Main main JJ adeshpande3-github-io-4431 15 2 Points Points NNP adeshpande3-github-io-4431 15 3 Trained train VBD adeshpande3-github-io-4431 15 4 the the DT adeshpande3-github-io-4431 15 5 network network NN adeshpande3-github-io-4431 15 6 on on IN adeshpande3-github-io-4431 15 7 ImageNet ImageNet NNP adeshpande3-github-io-4431 15 8 data datum NNS adeshpande3-github-io-4431 15 9 , , , adeshpande3-github-io-4431 15 10 which which WDT adeshpande3-github-io-4431 15 11 contained contain VBD adeshpande3-github-io-4431 15 12 over over IN adeshpande3-github-io-4431 15 13 15 15 CD adeshpande3-github-io-4431 15 14 million million CD adeshpande3-github-io-4431 15 15 annotated annotate VBN adeshpande3-github-io-4431 15 16 images image NNS adeshpande3-github-io-4431 15 17 from from IN adeshpande3-github-io-4431 15 18 a a DT adeshpande3-github-io-4431 15 19 total total NN adeshpande3-github-io-4431 15 20 of of IN adeshpande3-github-io-4431 15 21 over over IN adeshpande3-github-io-4431 15 22 22,000 22,000 CD adeshpande3-github-io-4431 15 23 categories category NNS adeshpande3-github-io-4431 15 24 . . . adeshpande3-github-io-4431 16 1 Used use VBN adeshpande3-github-io-4431 16 2 ReLU relu NN adeshpande3-github-io-4431 16 3 for for IN adeshpande3-github-io-4431 16 4 the the DT adeshpande3-github-io-4431 16 5 nonlinearity nonlinearity NN adeshpande3-github-io-4431 16 6 functions function NNS adeshpande3-github-io-4431 16 7 ( ( -LRB- adeshpande3-github-io-4431 16 8 Found find VBN adeshpande3-github-io-4431 16 9 to to TO adeshpande3-github-io-4431 16 10 decrease decrease VB adeshpande3-github-io-4431 16 11 training training NN adeshpande3-github-io-4431 16 12 time time NN adeshpande3-github-io-4431 16 13 as as IN adeshpande3-github-io-4431 16 14 ReLUs relu NNS adeshpande3-github-io-4431 16 15 are be VBP adeshpande3-github-io-4431 16 16 several several JJ adeshpande3-github-io-4431 16 17 times time NNS adeshpande3-github-io-4431 16 18 faster fast RBR adeshpande3-github-io-4431 16 19 than than IN adeshpande3-github-io-4431 16 20 the the DT adeshpande3-github-io-4431 16 21 conventional conventional JJ adeshpande3-github-io-4431 16 22 tanh tanh NN adeshpande3-github-io-4431 16 23 function function NN adeshpande3-github-io-4431 16 24 ) ) -RRB- adeshpande3-github-io-4431 16 25 . . . adeshpande3-github-io-4431 17 1 Used use VBN adeshpande3-github-io-4431 17 2 data data NN adeshpande3-github-io-4431 17 3 augmentation augmentation NN adeshpande3-github-io-4431 17 4 techniques technique NNS adeshpande3-github-io-4431 17 5 that that WDT adeshpande3-github-io-4431 17 6 consisted consist VBD adeshpande3-github-io-4431 17 7 of of IN adeshpande3-github-io-4431 17 8 image image NN adeshpande3-github-io-4431 17 9 translations translation NNS adeshpande3-github-io-4431 17 10 , , , adeshpande3-github-io-4431 17 11 horizontal horizontal JJ adeshpande3-github-io-4431 17 12 reflections reflection NNS adeshpande3-github-io-4431 17 13 , , , adeshpande3-github-io-4431 17 14 and and CC adeshpande3-github-io-4431 17 15 patch patch VB adeshpande3-github-io-4431 17 16 extractions extraction NNS adeshpande3-github-io-4431 17 17 . . . adeshpande3-github-io-4431 18 1 Implemented implement VBN adeshpande3-github-io-4431 18 2 dropout dropout NN adeshpande3-github-io-4431 18 3 layers layer NNS adeshpande3-github-io-4431 18 4 in in IN adeshpande3-github-io-4431 18 5 order order NN adeshpande3-github-io-4431 18 6 to to TO adeshpande3-github-io-4431 18 7 combat combat VB adeshpande3-github-io-4431 18 8 the the DT adeshpande3-github-io-4431 18 9 problem problem NN adeshpande3-github-io-4431 18 10 of of IN adeshpande3-github-io-4431 18 11 overfitting overfitte VBG adeshpande3-github-io-4431 18 12 to to IN adeshpande3-github-io-4431 18 13 the the DT adeshpande3-github-io-4431 18 14 training training NN adeshpande3-github-io-4431 18 15 data datum NNS adeshpande3-github-io-4431 18 16 . . . adeshpande3-github-io-4431 19 1 Trained train VBN adeshpande3-github-io-4431 19 2 the the DT adeshpande3-github-io-4431 19 3 model model NN adeshpande3-github-io-4431 19 4 using use VBG adeshpande3-github-io-4431 19 5 batch batch NN adeshpande3-github-io-4431 19 6 stochastic stochastic JJ adeshpande3-github-io-4431 19 7 gradient gradient JJ adeshpande3-github-io-4431 19 8 descent descent NN adeshpande3-github-io-4431 19 9 , , , adeshpande3-github-io-4431 19 10 with with IN adeshpande3-github-io-4431 19 11 specific specific JJ adeshpande3-github-io-4431 19 12 values value NNS adeshpande3-github-io-4431 19 13 for for IN adeshpande3-github-io-4431 19 14 momentum momentum NN adeshpande3-github-io-4431 19 15 and and CC adeshpande3-github-io-4431 19 16 weight weight NN adeshpande3-github-io-4431 19 17 decay decay NN adeshpande3-github-io-4431 19 18 . . . adeshpande3-github-io-4431 20 1 Trained train VBN adeshpande3-github-io-4431 20 2 on on IN adeshpande3-github-io-4431 20 3 two two CD adeshpande3-github-io-4431 20 4 GTX GTX NNP adeshpande3-github-io-4431 20 5 580 580 CD adeshpande3-github-io-4431 20 6 GPUs gpu NNS adeshpande3-github-io-4431 20 7 for for IN adeshpande3-github-io-4431 20 8 five five CD adeshpande3-github-io-4431 20 9 to to TO adeshpande3-github-io-4431 20 10 six six CD adeshpande3-github-io-4431 20 11 days day NNS adeshpande3-github-io-4431 20 12 . . . adeshpande3-github-io-4431 21 1 Why why WRB adeshpande3-github-io-4431 21 2 It -PRON- PRP adeshpande3-github-io-4431 21 3 ’s ’ VBZ adeshpande3-github-io-4431 21 4 Important important JJ adeshpande3-github-io-4431 21 5                                 _SP adeshpande3-github-io-4431 21 6 The the DT adeshpande3-github-io-4431 21 7 neural neural JJ adeshpande3-github-io-4431 21 8 network network NN adeshpande3-github-io-4431 21 9 developed develop VBN adeshpande3-github-io-4431 21 10 by by IN adeshpande3-github-io-4431 21 11 Krizhevsky Krizhevsky NNP adeshpande3-github-io-4431 21 12 , , , adeshpande3-github-io-4431 21 13 Sutskever Sutskever NNP adeshpande3-github-io-4431 21 14 , , , adeshpande3-github-io-4431 21 15 and and CC adeshpande3-github-io-4431 21 16 Hinton Hinton NNP adeshpande3-github-io-4431 21 17 in in IN adeshpande3-github-io-4431 21 18 2012 2012 CD adeshpande3-github-io-4431 21 19 was be VBD adeshpande3-github-io-4431 21 20 the the DT adeshpande3-github-io-4431 21 21 coming come VBG adeshpande3-github-io-4431 21 22 out out RP adeshpande3-github-io-4431 21 23 party party NN adeshpande3-github-io-4431 21 24 for for IN adeshpande3-github-io-4431 21 25 CNNs cnn NNS adeshpande3-github-io-4431 21 26 in in IN adeshpande3-github-io-4431 21 27 the the DT adeshpande3-github-io-4431 21 28 computer computer NN adeshpande3-github-io-4431 21 29 vision vision NN adeshpande3-github-io-4431 21 30 community community NN adeshpande3-github-io-4431 21 31 . . . adeshpande3-github-io-4431 22 1 This this DT adeshpande3-github-io-4431 22 2 was be VBD adeshpande3-github-io-4431 22 3 the the DT adeshpande3-github-io-4431 22 4 first first JJ adeshpande3-github-io-4431 22 5 time time NN adeshpande3-github-io-4431 22 6 a a DT adeshpande3-github-io-4431 22 7 model model NN adeshpande3-github-io-4431 22 8 performed perform VBN adeshpande3-github-io-4431 22 9 so so RB adeshpande3-github-io-4431 22 10 well well RB adeshpande3-github-io-4431 22 11 on on IN adeshpande3-github-io-4431 22 12 a a DT adeshpande3-github-io-4431 22 13 historically historically RB adeshpande3-github-io-4431 22 14 difficult difficult JJ adeshpande3-github-io-4431 22 15 ImageNet ImageNet NNP adeshpande3-github-io-4431 22 16 dataset dataset NN adeshpande3-github-io-4431 22 17 . . . adeshpande3-github-io-4431 23 1 Utilizing utilize VBG adeshpande3-github-io-4431 23 2 techniques technique NNS adeshpande3-github-io-4431 23 3 that that WDT adeshpande3-github-io-4431 23 4 are be VBP adeshpande3-github-io-4431 23 5 still still RB adeshpande3-github-io-4431 23 6 used use VBN adeshpande3-github-io-4431 23 7 today today NN adeshpande3-github-io-4431 23 8 , , , adeshpande3-github-io-4431 23 9 such such JJ adeshpande3-github-io-4431 23 10 as as IN adeshpande3-github-io-4431 23 11 data datum NNS adeshpande3-github-io-4431 23 12 augmentation augmentation NN adeshpande3-github-io-4431 23 13 and and CC adeshpande3-github-io-4431 23 14 dropout dropout NN adeshpande3-github-io-4431 23 15 , , , adeshpande3-github-io-4431 23 16 this this DT adeshpande3-github-io-4431 23 17 paper paper NN adeshpande3-github-io-4431 23 18 really really RB adeshpande3-github-io-4431 23 19 illustrated illustrate VBD adeshpande3-github-io-4431 23 20 the the DT adeshpande3-github-io-4431 23 21 benefits benefit NNS adeshpande3-github-io-4431 23 22 of of IN adeshpande3-github-io-4431 23 23 CNNs cnn NNS adeshpande3-github-io-4431 23 24 and and CC adeshpande3-github-io-4431 23 25 backed back VBD adeshpande3-github-io-4431 23 26 them -PRON- PRP adeshpande3-github-io-4431 23 27 up up RP adeshpande3-github-io-4431 23 28 with with IN adeshpande3-github-io-4431 23 29 record record NN adeshpande3-github-io-4431 23 30 breaking break VBG adeshpande3-github-io-4431 23 31 performance performance NN adeshpande3-github-io-4431 23 32 in in IN adeshpande3-github-io-4431 23 33 the the DT adeshpande3-github-io-4431 23 34 competition competition NN adeshpande3-github-io-4431 23 35 . . . adeshpande3-github-io-4431 24 1 ZF ZF NNP adeshpande3-github-io-4431 24 2 Net net NN adeshpande3-github-io-4431 24 3 ( ( -LRB- adeshpande3-github-io-4431 24 4 2013 2013 CD adeshpande3-github-io-4431 24 5 ) ) -RRB- adeshpande3-github-io-4431 24 6                                 _SP adeshpande3-github-io-4431 24 7 With with IN adeshpande3-github-io-4431 24 8 AlexNet AlexNet NNP adeshpande3-github-io-4431 24 9 stealing steal VBG adeshpande3-github-io-4431 24 10 the the DT adeshpande3-github-io-4431 24 11 show show NN adeshpande3-github-io-4431 24 12 in in IN adeshpande3-github-io-4431 24 13 2012 2012 CD adeshpande3-github-io-4431 24 14 , , , adeshpande3-github-io-4431 24 15 there there EX adeshpande3-github-io-4431 24 16 was be VBD adeshpande3-github-io-4431 24 17 a a DT adeshpande3-github-io-4431 24 18 large large JJ adeshpande3-github-io-4431 24 19 increase increase NN adeshpande3-github-io-4431 24 20 in in IN adeshpande3-github-io-4431 24 21 the the DT adeshpande3-github-io-4431 24 22 number number NN adeshpande3-github-io-4431 24 23 of of IN adeshpande3-github-io-4431 24 24 CNN CNN NNP adeshpande3-github-io-4431 24 25 models model NNS adeshpande3-github-io-4431 24 26 submitted submit VBN adeshpande3-github-io-4431 24 27 to to IN adeshpande3-github-io-4431 24 28 ILSVRC ILSVRC NNP adeshpande3-github-io-4431 24 29 2013 2013 CD adeshpande3-github-io-4431 24 30 . . . adeshpande3-github-io-4431 25 1 The the DT adeshpande3-github-io-4431 25 2 winner winner NN adeshpande3-github-io-4431 25 3 of of IN adeshpande3-github-io-4431 25 4 the the DT adeshpande3-github-io-4431 25 5 competition competition NN adeshpande3-github-io-4431 25 6 that that DT adeshpande3-github-io-4431 25 7 year year NN adeshpande3-github-io-4431 25 8 was be VBD adeshpande3-github-io-4431 25 9 a a DT adeshpande3-github-io-4431 25 10 network network NN adeshpande3-github-io-4431 25 11 built build VBN adeshpande3-github-io-4431 25 12 by by IN adeshpande3-github-io-4431 25 13 Matthew Matthew NNP adeshpande3-github-io-4431 25 14 Zeiler Zeiler NNP adeshpande3-github-io-4431 25 15 and and CC adeshpande3-github-io-4431 25 16 Rob Rob NNP adeshpande3-github-io-4431 25 17 Fergus Fergus NNP adeshpande3-github-io-4431 25 18 from from IN adeshpande3-github-io-4431 25 19 NYU NYU NNP adeshpande3-github-io-4431 25 20 . . . adeshpande3-github-io-4431 26 1 Named name VBN adeshpande3-github-io-4431 26 2 ZF ZF NNP adeshpande3-github-io-4431 26 3 Net Net NNP adeshpande3-github-io-4431 26 4 , , , adeshpande3-github-io-4431 26 5 this this DT adeshpande3-github-io-4431 26 6 model model NN adeshpande3-github-io-4431 26 7 achieved achieve VBD adeshpande3-github-io-4431 26 8 an an DT adeshpande3-github-io-4431 26 9 11.2 11.2 CD adeshpande3-github-io-4431 26 10 % % NN adeshpande3-github-io-4431 26 11 error error NN adeshpande3-github-io-4431 26 12 rate rate NN adeshpande3-github-io-4431 26 13 . . . adeshpande3-github-io-4431 27 1 This this DT adeshpande3-github-io-4431 27 2 architecture architecture NN adeshpande3-github-io-4431 27 3 was be VBD adeshpande3-github-io-4431 27 4 more more JJR adeshpande3-github-io-4431 27 5 of of IN adeshpande3-github-io-4431 27 6 a a DT adeshpande3-github-io-4431 27 7 fine fine JJ adeshpande3-github-io-4431 27 8 tuning tuning NN adeshpande3-github-io-4431 27 9 to to IN adeshpande3-github-io-4431 27 10 the the DT adeshpande3-github-io-4431 27 11 previous previous JJ adeshpande3-github-io-4431 27 12 AlexNet AlexNet NNP adeshpande3-github-io-4431 27 13 structure structure NN adeshpande3-github-io-4431 27 14 , , , adeshpande3-github-io-4431 27 15 but but CC adeshpande3-github-io-4431 27 16 still still RB adeshpande3-github-io-4431 27 17 developed develop VBD adeshpande3-github-io-4431 27 18 some some DT adeshpande3-github-io-4431 27 19 very very JJ adeshpande3-github-io-4431 27 20 keys keys JJ adeshpande3-github-io-4431 27 21 ideas idea NNS adeshpande3-github-io-4431 27 22 about about IN adeshpande3-github-io-4431 27 23 improving improve VBG adeshpande3-github-io-4431 27 24 performance performance NN adeshpande3-github-io-4431 27 25 . . . adeshpande3-github-io-4431 28 1 Another another DT adeshpande3-github-io-4431 28 2 reason reason NN adeshpande3-github-io-4431 28 3 this this DT adeshpande3-github-io-4431 28 4 was be VBD adeshpande3-github-io-4431 28 5 such such PDT adeshpande3-github-io-4431 28 6 a a DT adeshpande3-github-io-4431 28 7 great great JJ adeshpande3-github-io-4431 28 8 paper paper NN adeshpande3-github-io-4431 28 9 is be VBZ adeshpande3-github-io-4431 28 10 that that IN adeshpande3-github-io-4431 28 11 the the DT adeshpande3-github-io-4431 28 12 authors author NNS adeshpande3-github-io-4431 28 13 spent spend VBD adeshpande3-github-io-4431 28 14 a a DT adeshpande3-github-io-4431 28 15 good good JJ adeshpande3-github-io-4431 28 16 amount amount NN adeshpande3-github-io-4431 28 17 of of IN adeshpande3-github-io-4431 28 18 time time NN adeshpande3-github-io-4431 28 19 explaining explain VBG adeshpande3-github-io-4431 28 20 a a DT adeshpande3-github-io-4431 28 21 lot lot NN adeshpande3-github-io-4431 28 22 of of IN adeshpande3-github-io-4431 28 23 the the DT adeshpande3-github-io-4431 28 24 intuition intuition NN adeshpande3-github-io-4431 28 25 behind behind IN adeshpande3-github-io-4431 28 26 ConvNets ConvNets NNP adeshpande3-github-io-4431 28 27 and and CC adeshpande3-github-io-4431 28 28 showing show VBG adeshpande3-github-io-4431 28 29 how how WRB adeshpande3-github-io-4431 28 30 to to TO adeshpande3-github-io-4431 28 31 visualize visualize VB adeshpande3-github-io-4431 28 32 the the DT adeshpande3-github-io-4431 28 33 filters filter NNS adeshpande3-github-io-4431 28 34 and and CC adeshpande3-github-io-4431 28 35 weights weight NNS adeshpande3-github-io-4431 28 36 correctly correctly RB adeshpande3-github-io-4431 28 37 . . . adeshpande3-github-io-4431 29 1 In in IN adeshpande3-github-io-4431 29 2 this this DT adeshpande3-github-io-4431 29 3 paper paper NN adeshpande3-github-io-4431 29 4 titled title VBN adeshpande3-github-io-4431 29 5 “ " `` adeshpande3-github-io-4431 29 6 Visualizing Visualizing NNP adeshpande3-github-io-4431 29 7 and and CC adeshpande3-github-io-4431 29 8 Understanding Understanding NNP adeshpande3-github-io-4431 29 9 Convolutional Convolutional NNP adeshpande3-github-io-4431 29 10 Neural Neural NNP adeshpande3-github-io-4431 29 11 Networks Networks NNPS adeshpande3-github-io-4431 29 12 ” " '' adeshpande3-github-io-4431 29 13 , , , adeshpande3-github-io-4431 29 14 Zeiler Zeiler NNP adeshpande3-github-io-4431 29 15 and and CC adeshpande3-github-io-4431 29 16 Fergus Fergus NNP adeshpande3-github-io-4431 29 17 begin begin VBP adeshpande3-github-io-4431 29 18 by by IN adeshpande3-github-io-4431 29 19 discussing discuss VBG adeshpande3-github-io-4431 29 20 the the DT adeshpande3-github-io-4431 29 21 idea idea NN adeshpande3-github-io-4431 29 22 that that IN adeshpande3-github-io-4431 29 23 this this DT adeshpande3-github-io-4431 29 24 renewed renew VBN adeshpande3-github-io-4431 29 25 interest interest NN adeshpande3-github-io-4431 29 26 in in IN adeshpande3-github-io-4431 29 27 CNNs cnn NNS adeshpande3-github-io-4431 29 28 is be VBZ adeshpande3-github-io-4431 29 29 due due JJ adeshpande3-github-io-4431 29 30 to to IN adeshpande3-github-io-4431 29 31 the the DT adeshpande3-github-io-4431 29 32 accessibility accessibility NN adeshpande3-github-io-4431 29 33 of of IN adeshpande3-github-io-4431 29 34 large large JJ adeshpande3-github-io-4431 29 35 training training NN adeshpande3-github-io-4431 29 36 sets set NNS adeshpande3-github-io-4431 29 37 and and CC adeshpande3-github-io-4431 29 38 increased increase VBN adeshpande3-github-io-4431 29 39 computational computational JJ adeshpande3-github-io-4431 29 40 power power NN adeshpande3-github-io-4431 29 41 with with IN adeshpande3-github-io-4431 29 42 the the DT adeshpande3-github-io-4431 29 43 usage usage NN adeshpande3-github-io-4431 29 44 of of IN adeshpande3-github-io-4431 29 45 GPUs gpu NNS adeshpande3-github-io-4431 29 46 . . . adeshpande3-github-io-4431 30 1 They -PRON- PRP adeshpande3-github-io-4431 30 2 also also RB adeshpande3-github-io-4431 30 3 talk talk VBP adeshpande3-github-io-4431 30 4 about about IN adeshpande3-github-io-4431 30 5 the the DT adeshpande3-github-io-4431 30 6 limited limited JJ adeshpande3-github-io-4431 30 7 knowledge knowledge NN adeshpande3-github-io-4431 30 8 that that WDT adeshpande3-github-io-4431 30 9 researchers researcher NNS adeshpande3-github-io-4431 30 10 had have VBD adeshpande3-github-io-4431 30 11 on on IN adeshpande3-github-io-4431 30 12 inner inner JJ adeshpande3-github-io-4431 30 13 mechanisms mechanism NNS adeshpande3-github-io-4431 30 14 of of IN adeshpande3-github-io-4431 30 15 these these DT adeshpande3-github-io-4431 30 16 models model NNS adeshpande3-github-io-4431 30 17 , , , adeshpande3-github-io-4431 30 18 saying say VBG adeshpande3-github-io-4431 30 19 that that IN adeshpande3-github-io-4431 30 20 without without IN adeshpande3-github-io-4431 30 21 this this DT adeshpande3-github-io-4431 30 22 insight insight NN adeshpande3-github-io-4431 30 23 , , , adeshpande3-github-io-4431 30 24 the the DT adeshpande3-github-io-4431 30 25 “ " `` adeshpande3-github-io-4431 30 26 development development NN adeshpande3-github-io-4431 30 27 of of IN adeshpande3-github-io-4431 30 28 better well JJR adeshpande3-github-io-4431 30 29 models model NNS adeshpande3-github-io-4431 30 30 is be VBZ adeshpande3-github-io-4431 30 31 reduced reduce VBN adeshpande3-github-io-4431 30 32 to to IN adeshpande3-github-io-4431 30 33 trial trial NN adeshpande3-github-io-4431 30 34 and and CC adeshpande3-github-io-4431 30 35 error error NN adeshpande3-github-io-4431 30 36 ” " '' adeshpande3-github-io-4431 30 37 . . . adeshpande3-github-io-4431 31 1 While while IN adeshpande3-github-io-4431 31 2 we -PRON- PRP adeshpande3-github-io-4431 31 3 do do VBP adeshpande3-github-io-4431 31 4 currently currently RB adeshpande3-github-io-4431 31 5 have have VB adeshpande3-github-io-4431 31 6 a a DT adeshpande3-github-io-4431 31 7 better well JJR adeshpande3-github-io-4431 31 8 understanding understanding NN adeshpande3-github-io-4431 31 9 than than IN adeshpande3-github-io-4431 31 10 3 3 CD adeshpande3-github-io-4431 31 11 years year NNS adeshpande3-github-io-4431 31 12 ago ago RB adeshpande3-github-io-4431 31 13 , , , adeshpande3-github-io-4431 31 14 this this DT adeshpande3-github-io-4431 31 15 still still RB adeshpande3-github-io-4431 31 16 remains remain VBZ adeshpande3-github-io-4431 31 17 an an DT adeshpande3-github-io-4431 31 18 issue issue NN adeshpande3-github-io-4431 31 19 for for IN adeshpande3-github-io-4431 31 20 a a DT adeshpande3-github-io-4431 31 21 lot lot NN adeshpande3-github-io-4431 31 22 of of IN adeshpande3-github-io-4431 31 23 researchers researcher NNS adeshpande3-github-io-4431 31 24 ! ! . adeshpande3-github-io-4431 32 1 The the DT adeshpande3-github-io-4431 32 2 main main JJ adeshpande3-github-io-4431 32 3 contributions contribution NNS adeshpande3-github-io-4431 32 4 of of IN adeshpande3-github-io-4431 32 5 this this DT adeshpande3-github-io-4431 32 6 paper paper NN adeshpande3-github-io-4431 32 7 are be VBP adeshpande3-github-io-4431 32 8 details detail NNS adeshpande3-github-io-4431 32 9 of of IN adeshpande3-github-io-4431 32 10 a a DT adeshpande3-github-io-4431 32 11 slightly slightly RB adeshpande3-github-io-4431 32 12 modified modify VBN adeshpande3-github-io-4431 32 13 AlexNet AlexNet NNP adeshpande3-github-io-4431 32 14 model model NN adeshpande3-github-io-4431 32 15 and and CC adeshpande3-github-io-4431 32 16 a a DT adeshpande3-github-io-4431 32 17 very very RB adeshpande3-github-io-4431 32 18 interesting interesting JJ adeshpande3-github-io-4431 32 19 way way NN adeshpande3-github-io-4431 32 20 of of IN adeshpande3-github-io-4431 32 21 visualizing visualize VBG adeshpande3-github-io-4431 32 22 feature feature NN adeshpande3-github-io-4431 32 23 maps map NNS adeshpande3-github-io-4431 32 24 . . . adeshpande3-github-io-4431 33 1 Main main JJ adeshpande3-github-io-4431 33 2 Points point NNS adeshpande3-github-io-4431 33 3 Very very RB adeshpande3-github-io-4431 33 4 similar similar JJ adeshpande3-github-io-4431 33 5 architecture architecture NN adeshpande3-github-io-4431 33 6 to to IN adeshpande3-github-io-4431 33 7 AlexNet AlexNet NNP adeshpande3-github-io-4431 33 8 , , , adeshpande3-github-io-4431 33 9 except except IN adeshpande3-github-io-4431 33 10 for for IN adeshpande3-github-io-4431 33 11 a a DT adeshpande3-github-io-4431 33 12 few few JJ adeshpande3-github-io-4431 33 13 minor minor JJ adeshpande3-github-io-4431 33 14 modifications modification NNS adeshpande3-github-io-4431 33 15 . . . adeshpande3-github-io-4431 34 1 AlexNet AlexNet NNP adeshpande3-github-io-4431 34 2 trained train VBD adeshpande3-github-io-4431 34 3 on on IN adeshpande3-github-io-4431 34 4 15 15 CD adeshpande3-github-io-4431 34 5 million million CD adeshpande3-github-io-4431 34 6 images image NNS adeshpande3-github-io-4431 34 7 , , , adeshpande3-github-io-4431 34 8 while while IN adeshpande3-github-io-4431 34 9 ZF ZF NNP adeshpande3-github-io-4431 34 10 Net net NN adeshpande3-github-io-4431 34 11 trained train VBN adeshpande3-github-io-4431 34 12 on on IN adeshpande3-github-io-4431 34 13 only only RB adeshpande3-github-io-4431 34 14 1.3 1.3 CD adeshpande3-github-io-4431 34 15 million million CD adeshpande3-github-io-4431 34 16 images image NNS adeshpande3-github-io-4431 34 17 . . . adeshpande3-github-io-4431 35 1 Instead instead RB adeshpande3-github-io-4431 35 2 of of IN adeshpande3-github-io-4431 35 3 using use VBG adeshpande3-github-io-4431 35 4 11x11 11x11 JJ adeshpande3-github-io-4431 35 5 sized sized JJ adeshpande3-github-io-4431 35 6 filters filter NNS adeshpande3-github-io-4431 35 7 in in IN adeshpande3-github-io-4431 35 8 the the DT adeshpande3-github-io-4431 35 9 first first JJ adeshpande3-github-io-4431 35 10 layer layer NN adeshpande3-github-io-4431 35 11 ( ( -LRB- adeshpande3-github-io-4431 35 12 which which WDT adeshpande3-github-io-4431 35 13 is be VBZ adeshpande3-github-io-4431 35 14 what what WP adeshpande3-github-io-4431 35 15 AlexNet AlexNet NNP adeshpande3-github-io-4431 35 16 implemented implement VBN adeshpande3-github-io-4431 35 17 ) ) -RRB- adeshpande3-github-io-4431 35 18 , , , adeshpande3-github-io-4431 35 19 ZF ZF NNP adeshpande3-github-io-4431 35 20 Net Net NNP adeshpande3-github-io-4431 35 21 used use VBD adeshpande3-github-io-4431 35 22 filters filter NNS adeshpande3-github-io-4431 35 23 of of IN adeshpande3-github-io-4431 35 24 size size NN adeshpande3-github-io-4431 35 25 7x7 7x7 NNP adeshpande3-github-io-4431 35 26 and and CC adeshpande3-github-io-4431 35 27 a a DT adeshpande3-github-io-4431 35 28 decreased decrease VBN adeshpande3-github-io-4431 35 29 stride stride NN adeshpande3-github-io-4431 35 30 value value NN adeshpande3-github-io-4431 35 31 . . . adeshpande3-github-io-4431 36 1 The the DT adeshpande3-github-io-4431 36 2 reasoning reasoning NN adeshpande3-github-io-4431 36 3 behind behind IN adeshpande3-github-io-4431 36 4 this this DT adeshpande3-github-io-4431 36 5 modification modification NN adeshpande3-github-io-4431 36 6 is be VBZ adeshpande3-github-io-4431 36 7 that that IN adeshpande3-github-io-4431 36 8 a a DT adeshpande3-github-io-4431 36 9 smaller small JJR adeshpande3-github-io-4431 36 10 filter filter NN adeshpande3-github-io-4431 36 11 size size NN adeshpande3-github-io-4431 36 12 in in IN adeshpande3-github-io-4431 36 13 the the DT adeshpande3-github-io-4431 36 14 first first JJ adeshpande3-github-io-4431 36 15 conv conv NN adeshpande3-github-io-4431 36 16 layer layer NN adeshpande3-github-io-4431 36 17 helps help VBZ adeshpande3-github-io-4431 36 18 retain retain VB adeshpande3-github-io-4431 36 19 a a DT adeshpande3-github-io-4431 36 20 lot lot NN adeshpande3-github-io-4431 36 21 of of IN adeshpande3-github-io-4431 36 22 original original JJ adeshpande3-github-io-4431 36 23 pixel pixel NN adeshpande3-github-io-4431 36 24 information information NN adeshpande3-github-io-4431 36 25 in in IN adeshpande3-github-io-4431 36 26 the the DT adeshpande3-github-io-4431 36 27 input input NN adeshpande3-github-io-4431 36 28 volume volume NN adeshpande3-github-io-4431 36 29 . . . adeshpande3-github-io-4431 37 1 A a DT adeshpande3-github-io-4431 37 2 filtering filtering NN adeshpande3-github-io-4431 37 3 of of IN adeshpande3-github-io-4431 37 4 size size NN adeshpande3-github-io-4431 37 5 11x11 11x11 CD adeshpande3-github-io-4431 37 6 proved prove VBD adeshpande3-github-io-4431 37 7 to to TO adeshpande3-github-io-4431 37 8 be be VB adeshpande3-github-io-4431 37 9 skipping skip VBG adeshpande3-github-io-4431 37 10 a a DT adeshpande3-github-io-4431 37 11 lot lot NN adeshpande3-github-io-4431 37 12 of of IN adeshpande3-github-io-4431 37 13 relevant relevant JJ adeshpande3-github-io-4431 37 14 information information NN adeshpande3-github-io-4431 37 15 , , , adeshpande3-github-io-4431 37 16 especially especially RB adeshpande3-github-io-4431 37 17 as as IN adeshpande3-github-io-4431 37 18 this this DT adeshpande3-github-io-4431 37 19 is be VBZ adeshpande3-github-io-4431 37 20 the the DT adeshpande3-github-io-4431 37 21 first first JJ adeshpande3-github-io-4431 37 22 conv conv NN adeshpande3-github-io-4431 37 23 layer layer NN adeshpande3-github-io-4431 37 24 . . . adeshpande3-github-io-4431 38 1 As as IN adeshpande3-github-io-4431 38 2 the the DT adeshpande3-github-io-4431 38 3 network network NN adeshpande3-github-io-4431 38 4 grows grow VBZ adeshpande3-github-io-4431 38 5 , , , adeshpande3-github-io-4431 38 6 we -PRON- PRP adeshpande3-github-io-4431 38 7 also also RB adeshpande3-github-io-4431 38 8 see see VBP adeshpande3-github-io-4431 38 9 a a DT adeshpande3-github-io-4431 38 10 rise rise NN adeshpande3-github-io-4431 38 11 in in IN adeshpande3-github-io-4431 38 12 the the DT adeshpande3-github-io-4431 38 13 number number NN adeshpande3-github-io-4431 38 14 of of IN adeshpande3-github-io-4431 38 15 filters filter NNS adeshpande3-github-io-4431 38 16 used use VBN adeshpande3-github-io-4431 38 17 . . . adeshpande3-github-io-4431 39 1 Used use VBN adeshpande3-github-io-4431 39 2 ReLUs relu NNS adeshpande3-github-io-4431 39 3 for for IN adeshpande3-github-io-4431 39 4 their -PRON- PRP$ adeshpande3-github-io-4431 39 5 activation activation NN adeshpande3-github-io-4431 39 6 functions function NNS adeshpande3-github-io-4431 39 7 , , , adeshpande3-github-io-4431 39 8 cross cross JJ adeshpande3-github-io-4431 39 9 - - JJ adeshpande3-github-io-4431 39 10 entropy entropy JJ adeshpande3-github-io-4431 39 11 loss loss NN adeshpande3-github-io-4431 39 12 for for IN adeshpande3-github-io-4431 39 13 the the DT adeshpande3-github-io-4431 39 14 error error NN adeshpande3-github-io-4431 39 15 function function NN adeshpande3-github-io-4431 39 16 , , , adeshpande3-github-io-4431 39 17 and and CC adeshpande3-github-io-4431 39 18 trained train VBD adeshpande3-github-io-4431 39 19 using use VBG adeshpande3-github-io-4431 39 20 batch batch NN adeshpande3-github-io-4431 39 21 stochastic stochastic JJ adeshpande3-github-io-4431 39 22 gradient gradient JJ adeshpande3-github-io-4431 39 23 descent descent NN adeshpande3-github-io-4431 39 24 . . . adeshpande3-github-io-4431 40 1 Trained train VBN adeshpande3-github-io-4431 40 2 on on IN adeshpande3-github-io-4431 40 3 a a DT adeshpande3-github-io-4431 40 4 GTX gtx NN adeshpande3-github-io-4431 40 5 580 580 CD adeshpande3-github-io-4431 40 6 GPU GPU NNP adeshpande3-github-io-4431 40 7 for for IN adeshpande3-github-io-4431 40 8 twelve twelve CD adeshpande3-github-io-4431 40 9 days day NNS adeshpande3-github-io-4431 40 10 . . . adeshpande3-github-io-4431 41 1 Developed develop VBN adeshpande3-github-io-4431 41 2 a a DT adeshpande3-github-io-4431 41 3 visualization visualization NN adeshpande3-github-io-4431 41 4 technique technique NN adeshpande3-github-io-4431 41 5 named name VBN adeshpande3-github-io-4431 41 6 Deconvolutional Deconvolutional NNP adeshpande3-github-io-4431 41 7 Network Network NNP adeshpande3-github-io-4431 41 8 , , , adeshpande3-github-io-4431 41 9 which which WDT adeshpande3-github-io-4431 41 10 helps help VBZ adeshpande3-github-io-4431 41 11 to to TO adeshpande3-github-io-4431 41 12 examine examine VB adeshpande3-github-io-4431 41 13 different different JJ adeshpande3-github-io-4431 41 14 feature feature NN adeshpande3-github-io-4431 41 15 activations activation NNS adeshpande3-github-io-4431 41 16 and and CC adeshpande3-github-io-4431 41 17 their -PRON- PRP$ adeshpande3-github-io-4431 41 18 relation relation NN adeshpande3-github-io-4431 41 19 to to IN adeshpande3-github-io-4431 41 20 the the DT adeshpande3-github-io-4431 41 21 input input NN adeshpande3-github-io-4431 41 22 space space NN adeshpande3-github-io-4431 41 23 . . . adeshpande3-github-io-4431 42 1 Called call VBN adeshpande3-github-io-4431 42 2 “ " `` adeshpande3-github-io-4431 42 3 deconvnet deconvnet NN adeshpande3-github-io-4431 42 4 ” " '' adeshpande3-github-io-4431 42 5 because because IN adeshpande3-github-io-4431 42 6 it -PRON- PRP adeshpande3-github-io-4431 42 7 maps map VBZ adeshpande3-github-io-4431 42 8 features feature VBZ adeshpande3-github-io-4431 42 9 to to IN adeshpande3-github-io-4431 42 10 pixels pixel NNS adeshpande3-github-io-4431 42 11 ( ( -LRB- adeshpande3-github-io-4431 42 12 the the DT adeshpande3-github-io-4431 42 13 opposite opposite NN adeshpande3-github-io-4431 42 14 of of IN adeshpande3-github-io-4431 42 15 what what WP adeshpande3-github-io-4431 42 16 a a DT adeshpande3-github-io-4431 42 17 convolutional convolutional JJ adeshpande3-github-io-4431 42 18 layer layer NN adeshpande3-github-io-4431 42 19 does do VBZ adeshpande3-github-io-4431 42 20 ) ) -RRB- adeshpande3-github-io-4431 42 21 . . . adeshpande3-github-io-4431 43 1 DeConvNet DeConvNet -RRB- adeshpande3-github-io-4431 43 2                                 _SP adeshpande3-github-io-4431 43 3 The the DT adeshpande3-github-io-4431 43 4 basic basic JJ adeshpande3-github-io-4431 43 5 idea idea NN adeshpande3-github-io-4431 43 6 behind behind IN adeshpande3-github-io-4431 43 7 how how WRB adeshpande3-github-io-4431 43 8 this this DT adeshpande3-github-io-4431 43 9 works work VBZ adeshpande3-github-io-4431 43 10 is be VBZ adeshpande3-github-io-4431 43 11 that that IN adeshpande3-github-io-4431 43 12 at at IN adeshpande3-github-io-4431 43 13 every every DT adeshpande3-github-io-4431 43 14 layer layer NN adeshpande3-github-io-4431 43 15 of of IN adeshpande3-github-io-4431 43 16 the the DT adeshpande3-github-io-4431 43 17 trained train VBN adeshpande3-github-io-4431 43 18 CNN CNN NNP adeshpande3-github-io-4431 43 19 , , , adeshpande3-github-io-4431 43 20 you -PRON- PRP adeshpande3-github-io-4431 43 21 attach attach VBP adeshpande3-github-io-4431 43 22 a a DT adeshpande3-github-io-4431 43 23 “ " `` adeshpande3-github-io-4431 43 24 deconvnet deconvnet NN adeshpande3-github-io-4431 43 25 ” " '' adeshpande3-github-io-4431 43 26 which which WDT adeshpande3-github-io-4431 43 27 has have VBZ adeshpande3-github-io-4431 43 28 a a DT adeshpande3-github-io-4431 43 29 path path NN adeshpande3-github-io-4431 43 30 back back RB adeshpande3-github-io-4431 43 31 to to IN adeshpande3-github-io-4431 43 32 the the DT adeshpande3-github-io-4431 43 33 image image NN adeshpande3-github-io-4431 43 34 pixels pixel NNS adeshpande3-github-io-4431 43 35 . . . adeshpande3-github-io-4431 44 1 An an DT adeshpande3-github-io-4431 44 2 input input NN adeshpande3-github-io-4431 44 3 image image NN adeshpande3-github-io-4431 44 4 is be VBZ adeshpande3-github-io-4431 44 5 fed feed VBN adeshpande3-github-io-4431 44 6 into into IN adeshpande3-github-io-4431 44 7 the the DT adeshpande3-github-io-4431 44 8 CNN CNN NNP adeshpande3-github-io-4431 44 9 and and CC adeshpande3-github-io-4431 44 10 activations activation NNS adeshpande3-github-io-4431 44 11 are be VBP adeshpande3-github-io-4431 44 12 computed compute VBN adeshpande3-github-io-4431 44 13 at at IN adeshpande3-github-io-4431 44 14 each each DT adeshpande3-github-io-4431 44 15 level level NN adeshpande3-github-io-4431 44 16 . . . adeshpande3-github-io-4431 45 1 This this DT adeshpande3-github-io-4431 45 2 is be VBZ adeshpande3-github-io-4431 45 3 the the DT adeshpande3-github-io-4431 45 4 forward forward JJ adeshpande3-github-io-4431 45 5 pass pass NN adeshpande3-github-io-4431 45 6 . . . adeshpande3-github-io-4431 46 1 Now now RB adeshpande3-github-io-4431 46 2 , , , adeshpande3-github-io-4431 46 3 let let VB adeshpande3-github-io-4431 46 4 ’s -PRON- PRP adeshpande3-github-io-4431 46 5 say say VB adeshpande3-github-io-4431 46 6 we -PRON- PRP adeshpande3-github-io-4431 46 7 want want VBP adeshpande3-github-io-4431 46 8 to to TO adeshpande3-github-io-4431 46 9 examine examine VB adeshpande3-github-io-4431 46 10 the the DT adeshpande3-github-io-4431 46 11 activations activation NNS adeshpande3-github-io-4431 46 12 of of IN adeshpande3-github-io-4431 46 13 a a DT adeshpande3-github-io-4431 46 14 certain certain JJ adeshpande3-github-io-4431 46 15 feature feature NN adeshpande3-github-io-4431 46 16 in in IN adeshpande3-github-io-4431 46 17 the the DT adeshpande3-github-io-4431 46 18 4th 4th JJ adeshpande3-github-io-4431 46 19 conv conv NN adeshpande3-github-io-4431 46 20 layer layer NN adeshpande3-github-io-4431 46 21 . . . adeshpande3-github-io-4431 47 1 We -PRON- PRP adeshpande3-github-io-4431 47 2 would would MD adeshpande3-github-io-4431 47 3 store store VB adeshpande3-github-io-4431 47 4 the the DT adeshpande3-github-io-4431 47 5 activations activation NNS adeshpande3-github-io-4431 47 6 of of IN adeshpande3-github-io-4431 47 7 this this DT adeshpande3-github-io-4431 47 8 one one CD adeshpande3-github-io-4431 47 9 feature feature NN adeshpande3-github-io-4431 47 10 map map NN adeshpande3-github-io-4431 47 11 , , , adeshpande3-github-io-4431 47 12 but but CC adeshpande3-github-io-4431 47 13 set set VBD adeshpande3-github-io-4431 47 14 all all DT adeshpande3-github-io-4431 47 15 of of IN adeshpande3-github-io-4431 47 16 the the DT adeshpande3-github-io-4431 47 17 other other JJ adeshpande3-github-io-4431 47 18 activations activation NNS adeshpande3-github-io-4431 47 19 in in IN adeshpande3-github-io-4431 47 20 the the DT adeshpande3-github-io-4431 47 21 layer layer NN adeshpande3-github-io-4431 47 22 to to IN adeshpande3-github-io-4431 47 23 0 0 CD adeshpande3-github-io-4431 47 24 , , , adeshpande3-github-io-4431 47 25 and and CC adeshpande3-github-io-4431 47 26 then then RB adeshpande3-github-io-4431 47 27 pass pass VB adeshpande3-github-io-4431 47 28 this this DT adeshpande3-github-io-4431 47 29 feature feature NN adeshpande3-github-io-4431 47 30 map map NN adeshpande3-github-io-4431 47 31 as as IN adeshpande3-github-io-4431 47 32 the the DT adeshpande3-github-io-4431 47 33 input input NN adeshpande3-github-io-4431 47 34 into into IN adeshpande3-github-io-4431 47 35 the the DT adeshpande3-github-io-4431 47 36 deconvnet deconvnet NN adeshpande3-github-io-4431 47 37 . . . adeshpande3-github-io-4431 48 1 This this DT adeshpande3-github-io-4431 48 2 deconvnet deconvnet NN adeshpande3-github-io-4431 48 3 has have VBZ adeshpande3-github-io-4431 48 4 the the DT adeshpande3-github-io-4431 48 5 same same JJ adeshpande3-github-io-4431 48 6 filters filter NNS adeshpande3-github-io-4431 48 7 as as IN adeshpande3-github-io-4431 48 8 the the DT adeshpande3-github-io-4431 48 9 original original JJ adeshpande3-github-io-4431 48 10 CNN CNN NNP adeshpande3-github-io-4431 48 11 . . . adeshpande3-github-io-4431 49 1 This this DT adeshpande3-github-io-4431 49 2 input input NN adeshpande3-github-io-4431 49 3 then then RB adeshpande3-github-io-4431 49 4 goes go VBZ adeshpande3-github-io-4431 49 5 through through IN adeshpande3-github-io-4431 49 6 a a DT adeshpande3-github-io-4431 49 7 series series NN adeshpande3-github-io-4431 49 8 of of IN adeshpande3-github-io-4431 49 9 unpool unpool NNP adeshpande3-github-io-4431 49 10 ( ( -LRB- adeshpande3-github-io-4431 49 11 reverse reverse JJ adeshpande3-github-io-4431 49 12 maxpooling maxpooling NN adeshpande3-github-io-4431 49 13 ) ) -RRB- adeshpande3-github-io-4431 49 14 , , , adeshpande3-github-io-4431 49 15 rectify rectify VB adeshpande3-github-io-4431 49 16 , , , adeshpande3-github-io-4431 49 17 and and CC adeshpande3-github-io-4431 49 18 filter filter NN adeshpande3-github-io-4431 49 19 operations operation NNS adeshpande3-github-io-4431 49 20 for for IN adeshpande3-github-io-4431 49 21 each each DT adeshpande3-github-io-4431 49 22 preceding precede VBG adeshpande3-github-io-4431 49 23 layer layer NN adeshpande3-github-io-4431 49 24 until until IN adeshpande3-github-io-4431 49 25 the the DT adeshpande3-github-io-4431 49 26 input input NN adeshpande3-github-io-4431 49 27 space space NN adeshpande3-github-io-4431 49 28 is be VBZ adeshpande3-github-io-4431 49 29 reached reach VBN adeshpande3-github-io-4431 49 30 . . . adeshpande3-github-io-4431 50 1 The the DT adeshpande3-github-io-4431 50 2 reasoning reasoning NN adeshpande3-github-io-4431 50 3 behind behind IN adeshpande3-github-io-4431 50 4 this this DT adeshpande3-github-io-4431 50 5 whole whole JJ adeshpande3-github-io-4431 50 6 process process NN adeshpande3-github-io-4431 50 7 is be VBZ adeshpande3-github-io-4431 50 8 that that IN adeshpande3-github-io-4431 50 9 we -PRON- PRP adeshpande3-github-io-4431 50 10 want want VBP adeshpande3-github-io-4431 50 11 to to TO adeshpande3-github-io-4431 50 12 examine examine VB adeshpande3-github-io-4431 50 13 what what WDT adeshpande3-github-io-4431 50 14 type type NN adeshpande3-github-io-4431 50 15 of of IN adeshpande3-github-io-4431 50 16 structures structure NNS adeshpande3-github-io-4431 50 17 excite excite VBP adeshpande3-github-io-4431 50 18 a a DT adeshpande3-github-io-4431 50 19 given give VBN adeshpande3-github-io-4431 50 20 feature feature NN adeshpande3-github-io-4431 50 21 map map NN adeshpande3-github-io-4431 50 22 . . . adeshpande3-github-io-4431 51 1 Let let VB adeshpande3-github-io-4431 51 2 ’s -PRON- PRP adeshpande3-github-io-4431 51 3 look look VB adeshpande3-github-io-4431 51 4 at at IN adeshpande3-github-io-4431 51 5 the the DT adeshpande3-github-io-4431 51 6 visualizations visualization NNS adeshpande3-github-io-4431 51 7 of of IN adeshpande3-github-io-4431 51 8 the the DT adeshpande3-github-io-4431 51 9 first first JJ adeshpande3-github-io-4431 51 10 and and CC adeshpande3-github-io-4431 51 11 second second JJ adeshpande3-github-io-4431 51 12 layers layer NNS adeshpande3-github-io-4431 51 13 . . . adeshpande3-github-io-4431 52 1 Like like IN adeshpande3-github-io-4431 52 2 we -PRON- PRP adeshpande3-github-io-4431 52 3 discussed discuss VBD adeshpande3-github-io-4431 52 4 in in IN adeshpande3-github-io-4431 52 5 Part Part NNP adeshpande3-github-io-4431 52 6 1 1 CD adeshpande3-github-io-4431 52 7 , , , adeshpande3-github-io-4431 52 8 the the DT adeshpande3-github-io-4431 52 9 first first JJ adeshpande3-github-io-4431 52 10 layer layer NN adeshpande3-github-io-4431 52 11 of of IN adeshpande3-github-io-4431 52 12 your -PRON- PRP$ adeshpande3-github-io-4431 52 13 ConvNet ConvNet NNP adeshpande3-github-io-4431 52 14 is be VBZ adeshpande3-github-io-4431 52 15 always always RB adeshpande3-github-io-4431 52 16 a a DT adeshpande3-github-io-4431 52 17 low low JJ adeshpande3-github-io-4431 52 18 level level NN adeshpande3-github-io-4431 52 19 feature feature NN adeshpande3-github-io-4431 52 20 detector detector NN adeshpande3-github-io-4431 52 21 that that WDT adeshpande3-github-io-4431 52 22 will will MD adeshpande3-github-io-4431 52 23 detect detect VB adeshpande3-github-io-4431 52 24 simple simple JJ adeshpande3-github-io-4431 52 25 edges edge NNS adeshpande3-github-io-4431 52 26 or or CC adeshpande3-github-io-4431 52 27 colors color NNS adeshpande3-github-io-4431 52 28 in in IN adeshpande3-github-io-4431 52 29 this this DT adeshpande3-github-io-4431 52 30 particular particular JJ adeshpande3-github-io-4431 52 31 case case NN adeshpande3-github-io-4431 52 32 . . . adeshpande3-github-io-4431 53 1 We -PRON- PRP adeshpande3-github-io-4431 53 2 can can MD adeshpande3-github-io-4431 53 3 see see VB adeshpande3-github-io-4431 53 4 that that DT adeshpande3-github-io-4431 53 5 with with IN adeshpande3-github-io-4431 53 6 the the DT adeshpande3-github-io-4431 53 7 second second JJ adeshpande3-github-io-4431 53 8 layer layer NN adeshpande3-github-io-4431 53 9 , , , adeshpande3-github-io-4431 53 10 we -PRON- PRP adeshpande3-github-io-4431 53 11 have have VBP adeshpande3-github-io-4431 53 12 more more RBR adeshpande3-github-io-4431 53 13 circular circular JJ adeshpande3-github-io-4431 53 14 features feature NNS adeshpande3-github-io-4431 53 15 that that WDT adeshpande3-github-io-4431 53 16 are be VBP adeshpande3-github-io-4431 53 17 being be VBG adeshpande3-github-io-4431 53 18 detected detect VBN adeshpande3-github-io-4431 53 19 . . . adeshpande3-github-io-4431 54 1 Let let VB adeshpande3-github-io-4431 54 2 ’s -PRON- PRP adeshpande3-github-io-4431 54 3 look look VB adeshpande3-github-io-4431 54 4 at at IN adeshpande3-github-io-4431 54 5 layers layer NNS adeshpande3-github-io-4431 54 6 3 3 CD adeshpande3-github-io-4431 54 7 , , , adeshpande3-github-io-4431 54 8 4 4 CD adeshpande3-github-io-4431 54 9 , , , adeshpande3-github-io-4431 54 10 and and CC adeshpande3-github-io-4431 54 11 5 5 CD adeshpande3-github-io-4431 54 12 . . . adeshpande3-github-io-4431 55 1 These these DT adeshpande3-github-io-4431 55 2 layers layer NNS adeshpande3-github-io-4431 55 3 show show VBP adeshpande3-github-io-4431 55 4 a a DT adeshpande3-github-io-4431 55 5 lot lot NN adeshpande3-github-io-4431 55 6 more more JJR adeshpande3-github-io-4431 55 7 of of IN adeshpande3-github-io-4431 55 8 the the DT adeshpande3-github-io-4431 55 9 higher high JJR adeshpande3-github-io-4431 55 10 level level NN adeshpande3-github-io-4431 55 11 features feature VBZ adeshpande3-github-io-4431 55 12 such such JJ adeshpande3-github-io-4431 55 13 as as IN adeshpande3-github-io-4431 55 14 dogs dog NNS adeshpande3-github-io-4431 55 15 ’ ’ POS adeshpande3-github-io-4431 55 16 faces face NNS adeshpande3-github-io-4431 55 17 or or CC adeshpande3-github-io-4431 55 18 flowers flower NNS adeshpande3-github-io-4431 55 19 . . . adeshpande3-github-io-4431 56 1 One one CD adeshpande3-github-io-4431 56 2 thing thing NN adeshpande3-github-io-4431 56 3 to to TO adeshpande3-github-io-4431 56 4 note note VB adeshpande3-github-io-4431 56 5 is be VBZ adeshpande3-github-io-4431 56 6 that that IN adeshpande3-github-io-4431 56 7 as as IN adeshpande3-github-io-4431 56 8 you -PRON- PRP adeshpande3-github-io-4431 56 9 may may MD adeshpande3-github-io-4431 56 10 remember remember VB adeshpande3-github-io-4431 56 11 , , , adeshpande3-github-io-4431 56 12 after after IN adeshpande3-github-io-4431 56 13 the the DT adeshpande3-github-io-4431 56 14 first first JJ adeshpande3-github-io-4431 56 15 conv conv NN adeshpande3-github-io-4431 56 16 layer layer NN adeshpande3-github-io-4431 56 17 , , , adeshpande3-github-io-4431 56 18 we -PRON- PRP adeshpande3-github-io-4431 56 19 normally normally RB adeshpande3-github-io-4431 56 20 have have VBP adeshpande3-github-io-4431 56 21 a a DT adeshpande3-github-io-4431 56 22 pooling pool VBG adeshpande3-github-io-4431 56 23 layer layer NN adeshpande3-github-io-4431 56 24 that that WDT adeshpande3-github-io-4431 56 25 downsamples downsample VBZ adeshpande3-github-io-4431 56 26 the the DT adeshpande3-github-io-4431 56 27 image image NN adeshpande3-github-io-4431 56 28 ( ( -LRB- adeshpande3-github-io-4431 56 29 for for IN adeshpande3-github-io-4431 56 30 example example NN adeshpande3-github-io-4431 56 31 , , , adeshpande3-github-io-4431 56 32 turns turn VBZ adeshpande3-github-io-4431 56 33 a a DT adeshpande3-github-io-4431 56 34 32x32x3 32x32x3 CD adeshpande3-github-io-4431 56 35 volume volume NN adeshpande3-github-io-4431 56 36 into into IN adeshpande3-github-io-4431 56 37 a a DT adeshpande3-github-io-4431 56 38 16x16x3 16x16x3 CD adeshpande3-github-io-4431 56 39 volume volume NN adeshpande3-github-io-4431 56 40 ) ) -RRB- adeshpande3-github-io-4431 56 41 . . . adeshpande3-github-io-4431 57 1 The the DT adeshpande3-github-io-4431 57 2 effect effect NN adeshpande3-github-io-4431 57 3 this this DT adeshpande3-github-io-4431 57 4 has have VBZ adeshpande3-github-io-4431 57 5 is be VBZ adeshpande3-github-io-4431 57 6 that that IN adeshpande3-github-io-4431 57 7 the the DT adeshpande3-github-io-4431 57 8 2nd 2nd JJ adeshpande3-github-io-4431 57 9 layer layer NN adeshpande3-github-io-4431 57 10 has have VBZ adeshpande3-github-io-4431 57 11 a a DT adeshpande3-github-io-4431 57 12 broader broad JJR adeshpande3-github-io-4431 57 13 scope scope NN adeshpande3-github-io-4431 57 14 of of IN adeshpande3-github-io-4431 57 15 what what WP adeshpande3-github-io-4431 57 16 it -PRON- PRP adeshpande3-github-io-4431 57 17 can can MD adeshpande3-github-io-4431 57 18 see see VB adeshpande3-github-io-4431 57 19 in in IN adeshpande3-github-io-4431 57 20 the the DT adeshpande3-github-io-4431 57 21 original original JJ adeshpande3-github-io-4431 57 22 image image NN adeshpande3-github-io-4431 57 23 . . . adeshpande3-github-io-4431 58 1 For for IN adeshpande3-github-io-4431 58 2 more more JJR adeshpande3-github-io-4431 58 3 info info NN adeshpande3-github-io-4431 58 4 on on IN adeshpande3-github-io-4431 58 5 deconvnet deconvnet NN adeshpande3-github-io-4431 58 6 or or CC adeshpande3-github-io-4431 58 7 the the DT adeshpande3-github-io-4431 58 8 paper paper NN adeshpande3-github-io-4431 58 9 in in IN adeshpande3-github-io-4431 58 10 general general JJ adeshpande3-github-io-4431 58 11 , , , adeshpande3-github-io-4431 58 12 check check VB adeshpande3-github-io-4431 58 13 out out RP adeshpande3-github-io-4431 58 14 Zeiler Zeiler NNP adeshpande3-github-io-4431 58 15 himself -PRON- PRP adeshpande3-github-io-4431 58 16 presenting present VBG adeshpande3-github-io-4431 58 17 on on IN adeshpande3-github-io-4431 58 18 the the DT adeshpande3-github-io-4431 58 19 topic topic NN adeshpande3-github-io-4431 58 20 . . . adeshpande3-github-io-4431 59 1 Why why WRB adeshpande3-github-io-4431 59 2 It -PRON- PRP adeshpande3-github-io-4431 59 3 ’s ’ VBZ adeshpande3-github-io-4431 59 4 Important important JJ adeshpande3-github-io-4431 59 5                                 _SP adeshpande3-github-io-4431 59 6 ZF ZF NNP adeshpande3-github-io-4431 59 7 Net Net NNP adeshpande3-github-io-4431 59 8 was be VBD adeshpande3-github-io-4431 59 9 not not RB adeshpande3-github-io-4431 59 10 only only RB adeshpande3-github-io-4431 59 11 the the DT adeshpande3-github-io-4431 59 12 winner winner NN adeshpande3-github-io-4431 59 13 of of IN adeshpande3-github-io-4431 59 14 the the DT adeshpande3-github-io-4431 59 15 competition competition NN adeshpande3-github-io-4431 59 16 in in IN adeshpande3-github-io-4431 59 17 2013 2013 CD adeshpande3-github-io-4431 59 18 , , , adeshpande3-github-io-4431 59 19 but but CC adeshpande3-github-io-4431 59 20 also also RB adeshpande3-github-io-4431 59 21 provided provide VBN adeshpande3-github-io-4431 59 22 great great JJ adeshpande3-github-io-4431 59 23 intuition intuition NN adeshpande3-github-io-4431 59 24 as as IN adeshpande3-github-io-4431 59 25 to to IN adeshpande3-github-io-4431 59 26 the the DT adeshpande3-github-io-4431 59 27 workings working NNS adeshpande3-github-io-4431 59 28 on on IN adeshpande3-github-io-4431 59 29 CNNs cnn NNS adeshpande3-github-io-4431 59 30 and and CC adeshpande3-github-io-4431 59 31 illustrated illustrate VBN adeshpande3-github-io-4431 59 32 more more JJR adeshpande3-github-io-4431 59 33 ways way NNS adeshpande3-github-io-4431 59 34 to to TO adeshpande3-github-io-4431 59 35 improve improve VB adeshpande3-github-io-4431 59 36 performance performance NN adeshpande3-github-io-4431 59 37 . . . adeshpande3-github-io-4431 60 1 The the DT adeshpande3-github-io-4431 60 2 visualization visualization NN adeshpande3-github-io-4431 60 3 approach approach NN adeshpande3-github-io-4431 60 4 described describe VBN adeshpande3-github-io-4431 60 5 helps help VBZ adeshpande3-github-io-4431 60 6 not not RB adeshpande3-github-io-4431 60 7 only only RB adeshpande3-github-io-4431 60 8 to to TO adeshpande3-github-io-4431 60 9 explain explain VB adeshpande3-github-io-4431 60 10 the the DT adeshpande3-github-io-4431 60 11 inner inner JJ adeshpande3-github-io-4431 60 12 workings working NNS adeshpande3-github-io-4431 60 13 of of IN adeshpande3-github-io-4431 60 14 CNNs cnn NNS adeshpande3-github-io-4431 60 15 , , , adeshpande3-github-io-4431 60 16 but but CC adeshpande3-github-io-4431 60 17 also also RB adeshpande3-github-io-4431 60 18 provides provide VBZ adeshpande3-github-io-4431 60 19 insight insight NN adeshpande3-github-io-4431 60 20 for for IN adeshpande3-github-io-4431 60 21 improvements improvement NNS adeshpande3-github-io-4431 60 22 to to IN adeshpande3-github-io-4431 60 23 network network NN adeshpande3-github-io-4431 60 24 architectures architecture NNS adeshpande3-github-io-4431 60 25 . . . adeshpande3-github-io-4431 61 1 The the DT adeshpande3-github-io-4431 61 2 fascinating fascinating JJ adeshpande3-github-io-4431 61 3 deconv deconv NN adeshpande3-github-io-4431 61 4 visualization visualization NN adeshpande3-github-io-4431 61 5 approach approach NN adeshpande3-github-io-4431 61 6 and and CC adeshpande3-github-io-4431 61 7 occlusion occlusion NN adeshpande3-github-io-4431 61 8 experiments experiment NNS adeshpande3-github-io-4431 61 9 make make VBP adeshpande3-github-io-4431 61 10 this this DT adeshpande3-github-io-4431 61 11 one one CD adeshpande3-github-io-4431 61 12 of of IN adeshpande3-github-io-4431 61 13 my -PRON- PRP$ adeshpande3-github-io-4431 61 14 personal personal JJ adeshpande3-github-io-4431 61 15 favorite favorite JJ adeshpande3-github-io-4431 61 16 papers paper NNS adeshpande3-github-io-4431 61 17 . . . adeshpande3-github-io-4431 62 1 VGG VGG NNP adeshpande3-github-io-4431 62 2 Net Net NNP adeshpande3-github-io-4431 62 3 ( ( -LRB- adeshpande3-github-io-4431 62 4 2014 2014 CD adeshpande3-github-io-4431 62 5 ) ) -RRB- adeshpande3-github-io-4431 62 6                                 _SP adeshpande3-github-io-4431 62 7 Simplicity simplicity NN adeshpande3-github-io-4431 62 8 and and CC adeshpande3-github-io-4431 62 9 depth depth NN adeshpande3-github-io-4431 62 10 . . . adeshpande3-github-io-4431 63 1 That that DT adeshpande3-github-io-4431 63 2 ’s ’ VBZ adeshpande3-github-io-4431 63 3 what what WP adeshpande3-github-io-4431 63 4 a a DT adeshpande3-github-io-4431 63 5 model model NN adeshpande3-github-io-4431 63 6 created create VBN adeshpande3-github-io-4431 63 7 in in IN adeshpande3-github-io-4431 63 8 2014 2014 CD adeshpande3-github-io-4431 63 9 ( ( -LRB- adeshpande3-github-io-4431 63 10 were be VBD adeshpande3-github-io-4431 63 11 n’t not RB adeshpande3-github-io-4431 63 12 the the DT adeshpande3-github-io-4431 63 13 winners winner NNS adeshpande3-github-io-4431 63 14 of of IN adeshpande3-github-io-4431 63 15 ILSVRC ILSVRC NNP adeshpande3-github-io-4431 63 16 2014 2014 CD adeshpande3-github-io-4431 63 17 ) ) -RRB- adeshpande3-github-io-4431 63 18 best well RBS adeshpande3-github-io-4431 63 19 utilized utilize VBN adeshpande3-github-io-4431 63 20 with with IN adeshpande3-github-io-4431 63 21 its -PRON- PRP$ adeshpande3-github-io-4431 63 22 7.3 7.3 CD adeshpande3-github-io-4431 63 23 % % NN adeshpande3-github-io-4431 63 24 error error NN adeshpande3-github-io-4431 63 25 rate rate NN adeshpande3-github-io-4431 63 26 . . . adeshpande3-github-io-4431 64 1 Karen Karen NNP adeshpande3-github-io-4431 64 2 Simonyan Simonyan NNP adeshpande3-github-io-4431 64 3 and and CC adeshpande3-github-io-4431 64 4 Andrew Andrew NNP adeshpande3-github-io-4431 64 5 Zisserman Zisserman NNP adeshpande3-github-io-4431 64 6 of of IN adeshpande3-github-io-4431 64 7 the the DT adeshpande3-github-io-4431 64 8 University University NNP adeshpande3-github-io-4431 64 9 of of IN adeshpande3-github-io-4431 64 10 Oxford Oxford NNP adeshpande3-github-io-4431 64 11 created create VBD adeshpande3-github-io-4431 64 12 a a DT adeshpande3-github-io-4431 64 13 19 19 CD adeshpande3-github-io-4431 64 14 layer layer NN adeshpande3-github-io-4431 64 15 CNN CNN NNP adeshpande3-github-io-4431 64 16 that that WDT adeshpande3-github-io-4431 64 17 strictly strictly RB adeshpande3-github-io-4431 64 18 used use VBD adeshpande3-github-io-4431 64 19 3x3 3x3 CD adeshpande3-github-io-4431 64 20 filters filter NNS adeshpande3-github-io-4431 64 21 with with IN adeshpande3-github-io-4431 64 22 stride stride NN adeshpande3-github-io-4431 64 23 and and CC adeshpande3-github-io-4431 64 24 pad pad NN adeshpande3-github-io-4431 64 25 of of IN adeshpande3-github-io-4431 64 26 1 1 CD adeshpande3-github-io-4431 64 27 , , , adeshpande3-github-io-4431 64 28 along along IN adeshpande3-github-io-4431 64 29 with with IN adeshpande3-github-io-4431 64 30 2x2 2x2 CD adeshpande3-github-io-4431 64 31 maxpooling maxpooling NN adeshpande3-github-io-4431 64 32 layers layer NNS adeshpande3-github-io-4431 64 33 with with IN adeshpande3-github-io-4431 64 34 stride stride NN adeshpande3-github-io-4431 64 35 2 2 CD adeshpande3-github-io-4431 64 36 . . . adeshpande3-github-io-4431 65 1 Simple simple JJ adeshpande3-github-io-4431 65 2 enough enough RB adeshpande3-github-io-4431 65 3 right right NN adeshpande3-github-io-4431 65 4 ? ? . adeshpande3-github-io-4431 66 1 Main main JJ adeshpande3-github-io-4431 66 2 Points point VBZ adeshpande3-github-io-4431 66 3 The the DT adeshpande3-github-io-4431 66 4 use use NN adeshpande3-github-io-4431 66 5 of of IN adeshpande3-github-io-4431 66 6 only only JJ adeshpande3-github-io-4431 66 7 3x3 3x3 CD adeshpande3-github-io-4431 66 8 sized sized JJ adeshpande3-github-io-4431 66 9 filters filter NNS adeshpande3-github-io-4431 66 10 is be VBZ adeshpande3-github-io-4431 66 11 quite quite RB adeshpande3-github-io-4431 66 12 different different JJ adeshpande3-github-io-4431 66 13 from from IN adeshpande3-github-io-4431 66 14 AlexNet AlexNet NNP adeshpande3-github-io-4431 66 15 ’s ’s POS adeshpande3-github-io-4431 66 16 11x11 11x11 CD adeshpande3-github-io-4431 66 17 filters filter NNS adeshpande3-github-io-4431 66 18 in in IN adeshpande3-github-io-4431 66 19 the the DT adeshpande3-github-io-4431 66 20 first first JJ adeshpande3-github-io-4431 66 21 layer layer NN adeshpande3-github-io-4431 66 22 and and CC adeshpande3-github-io-4431 66 23 ZF ZF NNP adeshpande3-github-io-4431 66 24 Net net NN adeshpande3-github-io-4431 66 25 ’s ’s NN adeshpande3-github-io-4431 66 26 7x7 7x7 CD adeshpande3-github-io-4431 66 27 filters filter NNS adeshpande3-github-io-4431 66 28 . . . adeshpande3-github-io-4431 67 1 The the DT adeshpande3-github-io-4431 67 2 authors author NNS adeshpande3-github-io-4431 67 3 ’ ’ POS adeshpande3-github-io-4431 67 4 reasoning reasoning NN adeshpande3-github-io-4431 67 5 is be VBZ adeshpande3-github-io-4431 67 6 that that IN adeshpande3-github-io-4431 67 7 the the DT adeshpande3-github-io-4431 67 8 combination combination NN adeshpande3-github-io-4431 67 9 of of IN adeshpande3-github-io-4431 67 10 two two CD adeshpande3-github-io-4431 67 11 3x3 3x3 CD adeshpande3-github-io-4431 67 12 conv conv NN adeshpande3-github-io-4431 67 13 layers layer NNS adeshpande3-github-io-4431 67 14 has have VBZ adeshpande3-github-io-4431 67 15 an an DT adeshpande3-github-io-4431 67 16 effective effective JJ adeshpande3-github-io-4431 67 17 receptive receptive JJ adeshpande3-github-io-4431 67 18 field field NN adeshpande3-github-io-4431 67 19 of of IN adeshpande3-github-io-4431 67 20 5x5 5x5 CD adeshpande3-github-io-4431 67 21 . . . adeshpande3-github-io-4431 68 1 This this DT adeshpande3-github-io-4431 68 2 in in IN adeshpande3-github-io-4431 68 3 turn turn NN adeshpande3-github-io-4431 68 4 simulates simulate VBZ adeshpande3-github-io-4431 68 5 a a DT adeshpande3-github-io-4431 68 6 larger large JJR adeshpande3-github-io-4431 68 7 filter filter NN adeshpande3-github-io-4431 68 8 while while IN adeshpande3-github-io-4431 68 9 keeping keep VBG adeshpande3-github-io-4431 68 10 the the DT adeshpande3-github-io-4431 68 11 benefits benefit NNS adeshpande3-github-io-4431 68 12 of of IN adeshpande3-github-io-4431 68 13 smaller small JJR adeshpande3-github-io-4431 68 14 filter filter NN adeshpande3-github-io-4431 68 15 sizes size NNS adeshpande3-github-io-4431 68 16 . . . adeshpande3-github-io-4431 69 1 One one CD adeshpande3-github-io-4431 69 2 of of IN adeshpande3-github-io-4431 69 3 the the DT adeshpande3-github-io-4431 69 4 benefits benefit NNS adeshpande3-github-io-4431 69 5 is be VBZ adeshpande3-github-io-4431 69 6 a a DT adeshpande3-github-io-4431 69 7 decrease decrease NN adeshpande3-github-io-4431 69 8 in in IN adeshpande3-github-io-4431 69 9 the the DT adeshpande3-github-io-4431 69 10 number number NN adeshpande3-github-io-4431 69 11 of of IN adeshpande3-github-io-4431 69 12 parameters parameter NNS adeshpande3-github-io-4431 69 13 . . . adeshpande3-github-io-4431 70 1 Also also RB adeshpande3-github-io-4431 70 2 , , , adeshpande3-github-io-4431 70 3 with with IN adeshpande3-github-io-4431 70 4 two two CD adeshpande3-github-io-4431 70 5 conv conv NN adeshpande3-github-io-4431 70 6 layers layer NNS adeshpande3-github-io-4431 70 7 , , , adeshpande3-github-io-4431 70 8 we -PRON- PRP adeshpande3-github-io-4431 70 9 ’re be VBP adeshpande3-github-io-4431 70 10 able able JJ adeshpande3-github-io-4431 70 11 to to TO adeshpande3-github-io-4431 70 12 use use VB adeshpande3-github-io-4431 70 13 two two CD adeshpande3-github-io-4431 70 14 ReLU relu NN adeshpande3-github-io-4431 70 15 layers layer NNS adeshpande3-github-io-4431 70 16 instead instead RB adeshpande3-github-io-4431 70 17 of of IN adeshpande3-github-io-4431 70 18 one one CD adeshpande3-github-io-4431 70 19 . . . adeshpande3-github-io-4431 71 1 3 3 CD adeshpande3-github-io-4431 71 2 conv conv NN adeshpande3-github-io-4431 71 3 layers layer NNS adeshpande3-github-io-4431 71 4 back back RB adeshpande3-github-io-4431 71 5 to to IN adeshpande3-github-io-4431 71 6 back back RB adeshpande3-github-io-4431 71 7 have have VB adeshpande3-github-io-4431 71 8 an an DT adeshpande3-github-io-4431 71 9 effective effective JJ adeshpande3-github-io-4431 71 10 receptive receptive JJ adeshpande3-github-io-4431 71 11 field field NN adeshpande3-github-io-4431 71 12 of of IN adeshpande3-github-io-4431 71 13 7x7 7x7 CD adeshpande3-github-io-4431 71 14 . . . adeshpande3-github-io-4431 72 1 As as IN adeshpande3-github-io-4431 72 2 the the DT adeshpande3-github-io-4431 72 3 spatial spatial JJ adeshpande3-github-io-4431 72 4 size size NN adeshpande3-github-io-4431 72 5 of of IN adeshpande3-github-io-4431 72 6 the the DT adeshpande3-github-io-4431 72 7 input input NN adeshpande3-github-io-4431 72 8 volumes volume NNS adeshpande3-github-io-4431 72 9 at at IN adeshpande3-github-io-4431 72 10 each each DT adeshpande3-github-io-4431 72 11 layer layer NN adeshpande3-github-io-4431 72 12 decrease decrease NN adeshpande3-github-io-4431 72 13 ( ( -LRB- adeshpande3-github-io-4431 72 14 result result NN adeshpande3-github-io-4431 72 15 of of IN adeshpande3-github-io-4431 72 16 the the DT adeshpande3-github-io-4431 72 17 conv conv NN adeshpande3-github-io-4431 72 18 and and CC adeshpande3-github-io-4431 72 19 pool pool NN adeshpande3-github-io-4431 72 20 layers layer NNS adeshpande3-github-io-4431 72 21 ) ) -RRB- adeshpande3-github-io-4431 72 22 , , , adeshpande3-github-io-4431 72 23 the the DT adeshpande3-github-io-4431 72 24 depth depth NN adeshpande3-github-io-4431 72 25 of of IN adeshpande3-github-io-4431 72 26 the the DT adeshpande3-github-io-4431 72 27 volumes volume NNS adeshpande3-github-io-4431 72 28 increase increase VB adeshpande3-github-io-4431 72 29 due due IN adeshpande3-github-io-4431 72 30 to to IN adeshpande3-github-io-4431 72 31 the the DT adeshpande3-github-io-4431 72 32 increased increase VBN adeshpande3-github-io-4431 72 33 number number NN adeshpande3-github-io-4431 72 34 of of IN adeshpande3-github-io-4431 72 35 filters filter NNS adeshpande3-github-io-4431 72 36 as as IN adeshpande3-github-io-4431 72 37 you -PRON- PRP adeshpande3-github-io-4431 72 38 go go VBP adeshpande3-github-io-4431 72 39 down down IN adeshpande3-github-io-4431 72 40 the the DT adeshpande3-github-io-4431 72 41 network network NN adeshpande3-github-io-4431 72 42 . . . adeshpande3-github-io-4431 73 1 Interesting interesting JJ adeshpande3-github-io-4431 73 2 to to TO adeshpande3-github-io-4431 73 3 notice notice VB adeshpande3-github-io-4431 73 4 that that IN adeshpande3-github-io-4431 73 5 the the DT adeshpande3-github-io-4431 73 6 number number NN adeshpande3-github-io-4431 73 7 of of IN adeshpande3-github-io-4431 73 8 filters filter NNS adeshpande3-github-io-4431 73 9 doubles double NNS adeshpande3-github-io-4431 73 10 after after IN adeshpande3-github-io-4431 73 11 each each DT adeshpande3-github-io-4431 73 12 maxpool maxpool NN adeshpande3-github-io-4431 73 13 layer layer NN adeshpande3-github-io-4431 73 14 . . . adeshpande3-github-io-4431 74 1 This this DT adeshpande3-github-io-4431 74 2 reinforces reinforce VBZ adeshpande3-github-io-4431 74 3 the the DT adeshpande3-github-io-4431 74 4 idea idea NN adeshpande3-github-io-4431 74 5 of of IN adeshpande3-github-io-4431 74 6 shrinking shrink VBG adeshpande3-github-io-4431 74 7 spatial spatial JJ adeshpande3-github-io-4431 74 8 dimensions dimension NNS adeshpande3-github-io-4431 74 9 , , , adeshpande3-github-io-4431 74 10 but but CC adeshpande3-github-io-4431 74 11 growing grow VBG adeshpande3-github-io-4431 74 12 depth depth NN adeshpande3-github-io-4431 74 13 . . . adeshpande3-github-io-4431 75 1 Worked work VBN adeshpande3-github-io-4431 75 2 well well RB adeshpande3-github-io-4431 75 3 on on IN adeshpande3-github-io-4431 75 4 both both CC adeshpande3-github-io-4431 75 5 image image NN adeshpande3-github-io-4431 75 6 classification classification NN adeshpande3-github-io-4431 75 7 and and CC adeshpande3-github-io-4431 75 8 localization localization NN adeshpande3-github-io-4431 75 9 tasks task NNS adeshpande3-github-io-4431 75 10 . . . adeshpande3-github-io-4431 76 1 The the DT adeshpande3-github-io-4431 76 2 authors author NNS adeshpande3-github-io-4431 76 3 used use VBD adeshpande3-github-io-4431 76 4 a a DT adeshpande3-github-io-4431 76 5 form form NN adeshpande3-github-io-4431 76 6 of of IN adeshpande3-github-io-4431 76 7 localization localization NN adeshpande3-github-io-4431 76 8 as as IN adeshpande3-github-io-4431 76 9 regression regression NN adeshpande3-github-io-4431 76 10 ( ( -LRB- adeshpande3-github-io-4431 76 11 see see VB adeshpande3-github-io-4431 76 12 page page NN adeshpande3-github-io-4431 76 13 10 10 CD adeshpande3-github-io-4431 76 14 of of IN adeshpande3-github-io-4431 76 15 the the DT adeshpande3-github-io-4431 76 16 paper paper NN adeshpande3-github-io-4431 76 17 for for IN adeshpande3-github-io-4431 76 18 all all DT adeshpande3-github-io-4431 76 19 details detail NNS adeshpande3-github-io-4431 76 20 ) ) -RRB- adeshpande3-github-io-4431 76 21 . . . adeshpande3-github-io-4431 77 1 Built build VBN adeshpande3-github-io-4431 77 2 model model NN adeshpande3-github-io-4431 77 3 with with IN adeshpande3-github-io-4431 77 4 the the DT adeshpande3-github-io-4431 77 5 Caffe Caffe NNP adeshpande3-github-io-4431 77 6 toolbox toolbox NN adeshpande3-github-io-4431 77 7 . . . adeshpande3-github-io-4431 78 1 Used use VBN adeshpande3-github-io-4431 78 2 scale scale NN adeshpande3-github-io-4431 78 3 jittering jittering NN adeshpande3-github-io-4431 78 4 as as IN adeshpande3-github-io-4431 78 5 one one CD adeshpande3-github-io-4431 78 6 data data NN adeshpande3-github-io-4431 78 7 augmentation augmentation NN adeshpande3-github-io-4431 78 8 technique technique NN adeshpande3-github-io-4431 78 9 during during IN adeshpande3-github-io-4431 78 10 training training NN adeshpande3-github-io-4431 78 11 . . . adeshpande3-github-io-4431 79 1 Used use VBN adeshpande3-github-io-4431 79 2 ReLU relu NN adeshpande3-github-io-4431 79 3 layers layer NNS adeshpande3-github-io-4431 79 4 after after IN adeshpande3-github-io-4431 79 5 each each DT adeshpande3-github-io-4431 79 6 conv conv NN adeshpande3-github-io-4431 79 7 layer layer NN adeshpande3-github-io-4431 79 8 and and CC adeshpande3-github-io-4431 79 9 trained train VBN adeshpande3-github-io-4431 79 10 with with IN adeshpande3-github-io-4431 79 11 batch batch JJ adeshpande3-github-io-4431 79 12 gradient gradient NN adeshpande3-github-io-4431 79 13 descent descent NN adeshpande3-github-io-4431 79 14 . . . adeshpande3-github-io-4431 80 1 Trained train VBN adeshpande3-github-io-4431 80 2 on on IN adeshpande3-github-io-4431 80 3 4 4 CD adeshpande3-github-io-4431 80 4 Nvidia Nvidia NNP adeshpande3-github-io-4431 80 5 Titan Titan NNP adeshpande3-github-io-4431 80 6 Black black JJ adeshpande3-github-io-4431 80 7 GPUs gpu NNS adeshpande3-github-io-4431 80 8 for for IN adeshpande3-github-io-4431 80 9 two two CD adeshpande3-github-io-4431 80 10 to to TO adeshpande3-github-io-4431 80 11 three three CD adeshpande3-github-io-4431 80 12 weeks week NNS adeshpande3-github-io-4431 80 13 . . . adeshpande3-github-io-4431 81 1 Why why WRB adeshpande3-github-io-4431 81 2 It -PRON- PRP adeshpande3-github-io-4431 81 3 ’s ’ VBZ adeshpande3-github-io-4431 81 4 Important important JJ adeshpande3-github-io-4431 81 5                                 _SP adeshpande3-github-io-4431 81 6 VGG VGG NNP adeshpande3-github-io-4431 81 7 Net Net NNP adeshpande3-github-io-4431 81 8 is be VBZ adeshpande3-github-io-4431 81 9 one one CD adeshpande3-github-io-4431 81 10 of of IN adeshpande3-github-io-4431 81 11 the the DT adeshpande3-github-io-4431 81 12 most most RBS adeshpande3-github-io-4431 81 13 influential influential JJ adeshpande3-github-io-4431 81 14 papers paper NNS adeshpande3-github-io-4431 81 15 in in IN adeshpande3-github-io-4431 81 16 my -PRON- PRP$ adeshpande3-github-io-4431 81 17 mind mind NN adeshpande3-github-io-4431 81 18 because because IN adeshpande3-github-io-4431 81 19 it -PRON- PRP adeshpande3-github-io-4431 81 20 reinforced reinforce VBD adeshpande3-github-io-4431 81 21 the the DT adeshpande3-github-io-4431 81 22 notion notion NN adeshpande3-github-io-4431 81 23 that that IN adeshpande3-github-io-4431 81 24 convolutional convolutional JJ adeshpande3-github-io-4431 81 25 neural neural JJ adeshpande3-github-io-4431 81 26 networks network NNS adeshpande3-github-io-4431 81 27 have have VBP adeshpande3-github-io-4431 81 28 to to TO adeshpande3-github-io-4431 81 29 have have VB adeshpande3-github-io-4431 81 30 a a DT adeshpande3-github-io-4431 81 31 deep deep JJ adeshpande3-github-io-4431 81 32 network network NN adeshpande3-github-io-4431 81 33 of of IN adeshpande3-github-io-4431 81 34 layers layer NNS adeshpande3-github-io-4431 81 35 in in IN adeshpande3-github-io-4431 81 36 order order NN adeshpande3-github-io-4431 81 37 for for IN adeshpande3-github-io-4431 81 38 this this DT adeshpande3-github-io-4431 81 39 hierarchical hierarchical JJ adeshpande3-github-io-4431 81 40 representation representation NN adeshpande3-github-io-4431 81 41 of of IN adeshpande3-github-io-4431 81 42 visual visual JJ adeshpande3-github-io-4431 81 43 data datum NNS adeshpande3-github-io-4431 81 44 to to TO adeshpande3-github-io-4431 81 45 work work VB adeshpande3-github-io-4431 81 46 . . . adeshpande3-github-io-4431 82 1 Keep keep VB adeshpande3-github-io-4431 82 2 it -PRON- PRP adeshpande3-github-io-4431 82 3 deep deep JJ adeshpande3-github-io-4431 82 4 . . . adeshpande3-github-io-4431 83 1 Keep keep VB adeshpande3-github-io-4431 83 2 it -PRON- PRP adeshpande3-github-io-4431 83 3 simple simple JJ adeshpande3-github-io-4431 83 4 . . . adeshpande3-github-io-4431 84 1 GoogLeNet GoogLeNet NNP adeshpande3-github-io-4431 84 2 ( ( -LRB- adeshpande3-github-io-4431 84 3 2015 2015 CD adeshpande3-github-io-4431 84 4 ) ) -RRB- adeshpande3-github-io-4431 84 5                                 _SP adeshpande3-github-io-4431 84 6 You -PRON- PRP adeshpande3-github-io-4431 84 7 know know VBP adeshpande3-github-io-4431 84 8 that that DT adeshpande3-github-io-4431 84 9 idea idea NN adeshpande3-github-io-4431 84 10 of of IN adeshpande3-github-io-4431 84 11 simplicity simplicity NN adeshpande3-github-io-4431 84 12 in in IN adeshpande3-github-io-4431 84 13 network network NN adeshpande3-github-io-4431 84 14 architecture architecture NN adeshpande3-github-io-4431 84 15 that that IN adeshpande3-github-io-4431 84 16 we -PRON- PRP adeshpande3-github-io-4431 84 17 just just RB adeshpande3-github-io-4431 84 18 talked talk VBD adeshpande3-github-io-4431 84 19 about about IN adeshpande3-github-io-4431 84 20 ? ? . adeshpande3-github-io-4431 85 1 Well well UH adeshpande3-github-io-4431 85 2 , , , adeshpande3-github-io-4431 85 3 Google Google NNP adeshpande3-github-io-4431 85 4 kind kind RB adeshpande3-github-io-4431 85 5 of of RB adeshpande3-github-io-4431 85 6 threw throw VBD adeshpande3-github-io-4431 85 7 that that DT adeshpande3-github-io-4431 85 8 out out RP adeshpande3-github-io-4431 85 9 the the DT adeshpande3-github-io-4431 85 10 window window NN adeshpande3-github-io-4431 85 11 with with IN adeshpande3-github-io-4431 85 12 the the DT adeshpande3-github-io-4431 85 13 introduction introduction NN adeshpande3-github-io-4431 85 14 of of IN adeshpande3-github-io-4431 85 15 the the DT adeshpande3-github-io-4431 85 16 Inception Inception NNP adeshpande3-github-io-4431 85 17 module module NN adeshpande3-github-io-4431 85 18 . . . adeshpande3-github-io-4431 86 1 GoogLeNet GoogLeNet NNP adeshpande3-github-io-4431 86 2 is be VBZ adeshpande3-github-io-4431 86 3 a a DT adeshpande3-github-io-4431 86 4 22 22 CD adeshpande3-github-io-4431 86 5 layer layer NN adeshpande3-github-io-4431 86 6 CNN CNN NNP adeshpande3-github-io-4431 86 7 and and CC adeshpande3-github-io-4431 86 8 was be VBD adeshpande3-github-io-4431 86 9 the the DT adeshpande3-github-io-4431 86 10 winner winner NN adeshpande3-github-io-4431 86 11 of of IN adeshpande3-github-io-4431 86 12 ILSVRC ILSVRC NNP adeshpande3-github-io-4431 86 13 2014 2014 CD adeshpande3-github-io-4431 86 14 with with IN adeshpande3-github-io-4431 86 15 a a DT adeshpande3-github-io-4431 86 16 top top JJ adeshpande3-github-io-4431 86 17 5 5 CD adeshpande3-github-io-4431 86 18 error error NN adeshpande3-github-io-4431 86 19 rate rate NN adeshpande3-github-io-4431 86 20 of of IN adeshpande3-github-io-4431 86 21 6.7 6.7 CD adeshpande3-github-io-4431 86 22 % % NN adeshpande3-github-io-4431 86 23 . . . adeshpande3-github-io-4431 87 1 To to IN adeshpande3-github-io-4431 87 2 my -PRON- PRP$ adeshpande3-github-io-4431 87 3 knowledge knowledge NN adeshpande3-github-io-4431 87 4 , , , adeshpande3-github-io-4431 87 5 this this DT adeshpande3-github-io-4431 87 6 was be VBD adeshpande3-github-io-4431 87 7 one one CD adeshpande3-github-io-4431 87 8 of of IN adeshpande3-github-io-4431 87 9 the the DT adeshpande3-github-io-4431 87 10 first first JJ adeshpande3-github-io-4431 87 11 CNN CNN NNP adeshpande3-github-io-4431 87 12 architectures architecture NNS adeshpande3-github-io-4431 87 13 that that WDT adeshpande3-github-io-4431 87 14 really really RB adeshpande3-github-io-4431 87 15 strayed stray VBD adeshpande3-github-io-4431 87 16 from from IN adeshpande3-github-io-4431 87 17 the the DT adeshpande3-github-io-4431 87 18 general general JJ adeshpande3-github-io-4431 87 19 approach approach NN adeshpande3-github-io-4431 87 20 of of IN adeshpande3-github-io-4431 87 21 simply simply RB adeshpande3-github-io-4431 87 22 stacking stack VBG adeshpande3-github-io-4431 87 23 conv conv NN adeshpande3-github-io-4431 87 24 and and CC adeshpande3-github-io-4431 87 25 pooling pool VBG adeshpande3-github-io-4431 87 26 layers layer NNS adeshpande3-github-io-4431 87 27 on on IN adeshpande3-github-io-4431 87 28 top top NN adeshpande3-github-io-4431 87 29 of of IN adeshpande3-github-io-4431 87 30 each each DT adeshpande3-github-io-4431 87 31 other other JJ adeshpande3-github-io-4431 87 32 in in IN adeshpande3-github-io-4431 87 33 a a DT adeshpande3-github-io-4431 87 34 sequential sequential JJ adeshpande3-github-io-4431 87 35 structure structure NN adeshpande3-github-io-4431 87 36 . . . adeshpande3-github-io-4431 88 1 The the DT adeshpande3-github-io-4431 88 2 authors author NNS adeshpande3-github-io-4431 88 3 of of IN adeshpande3-github-io-4431 88 4 the the DT adeshpande3-github-io-4431 88 5 paper paper NN adeshpande3-github-io-4431 88 6 also also RB adeshpande3-github-io-4431 88 7 emphasized emphasize VBD adeshpande3-github-io-4431 88 8 that that IN adeshpande3-github-io-4431 88 9 this this DT adeshpande3-github-io-4431 88 10 new new JJ adeshpande3-github-io-4431 88 11 model model NN adeshpande3-github-io-4431 88 12 places place VBZ adeshpande3-github-io-4431 88 13 notable notable JJ adeshpande3-github-io-4431 88 14 consideration consideration NN adeshpande3-github-io-4431 88 15 on on IN adeshpande3-github-io-4431 88 16 memory memory NN adeshpande3-github-io-4431 88 17 and and CC adeshpande3-github-io-4431 88 18 power power NN adeshpande3-github-io-4431 88 19 usage usage NN adeshpande3-github-io-4431 88 20 ( ( -LRB- adeshpande3-github-io-4431 88 21 Important important JJ adeshpande3-github-io-4431 88 22 note note NN adeshpande3-github-io-4431 88 23 that that IN adeshpande3-github-io-4431 88 24 I -PRON- PRP adeshpande3-github-io-4431 88 25 sometimes sometimes RB adeshpande3-github-io-4431 88 26 forget forget VBP adeshpande3-github-io-4431 88 27 too too RB adeshpande3-github-io-4431 88 28 : : : adeshpande3-github-io-4431 88 29 Stacking stack VBG adeshpande3-github-io-4431 88 30 all all DT adeshpande3-github-io-4431 88 31 of of IN adeshpande3-github-io-4431 88 32 these these DT adeshpande3-github-io-4431 88 33 layers layer NNS adeshpande3-github-io-4431 88 34 and and CC adeshpande3-github-io-4431 88 35 adding add VBG adeshpande3-github-io-4431 88 36 huge huge JJ adeshpande3-github-io-4431 88 37 numbers number NNS adeshpande3-github-io-4431 88 38 of of IN adeshpande3-github-io-4431 88 39 filters filter NNS adeshpande3-github-io-4431 88 40 has have VBZ adeshpande3-github-io-4431 88 41 a a DT adeshpande3-github-io-4431 88 42 computational computational JJ adeshpande3-github-io-4431 88 43 and and CC adeshpande3-github-io-4431 88 44 memory memory NN adeshpande3-github-io-4431 88 45 cost cost NN adeshpande3-github-io-4431 88 46 , , , adeshpande3-github-io-4431 88 47 as as RB adeshpande3-github-io-4431 88 48 well well RB adeshpande3-github-io-4431 88 49 as as IN adeshpande3-github-io-4431 88 50 an an DT adeshpande3-github-io-4431 88 51 increased increase VBN adeshpande3-github-io-4431 88 52 chance chance NN adeshpande3-github-io-4431 88 53 of of IN adeshpande3-github-io-4431 88 54 overfitting overfitting NN adeshpande3-github-io-4431 88 55 ) ) -RRB- adeshpande3-github-io-4431 88 56 . . . adeshpande3-github-io-4431 89 1 Inception Inception NNP adeshpande3-github-io-4431 89 2 Module Module NNP adeshpande3-github-io-4431 89 3                                 _SP adeshpande3-github-io-4431 89 4 When when WRB adeshpande3-github-io-4431 89 5 we -PRON- PRP adeshpande3-github-io-4431 89 6 first first RB adeshpande3-github-io-4431 89 7 take take VBP adeshpande3-github-io-4431 89 8 a a DT adeshpande3-github-io-4431 89 9 look look NN adeshpande3-github-io-4431 89 10 at at IN adeshpande3-github-io-4431 89 11 the the DT adeshpande3-github-io-4431 89 12 structure structure NN adeshpande3-github-io-4431 89 13 of of IN adeshpande3-github-io-4431 89 14 GoogLeNet GoogLeNet NNP adeshpande3-github-io-4431 89 15 , , , adeshpande3-github-io-4431 89 16 we -PRON- PRP adeshpande3-github-io-4431 89 17 notice notice VBP adeshpande3-github-io-4431 89 18 immediately immediately RB adeshpande3-github-io-4431 89 19 that that IN adeshpande3-github-io-4431 89 20 not not RB adeshpande3-github-io-4431 89 21 everything everything NN adeshpande3-github-io-4431 89 22 is be VBZ adeshpande3-github-io-4431 89 23 happening happen VBG adeshpande3-github-io-4431 89 24 sequentially sequentially RB adeshpande3-github-io-4431 89 25 , , , adeshpande3-github-io-4431 89 26 as as IN adeshpande3-github-io-4431 89 27 seen see VBN adeshpande3-github-io-4431 89 28 in in IN adeshpande3-github-io-4431 89 29 previous previous JJ adeshpande3-github-io-4431 89 30 architectures architecture NNS adeshpande3-github-io-4431 89 31 . . . adeshpande3-github-io-4431 90 1 We -PRON- PRP adeshpande3-github-io-4431 90 2 have have VBP adeshpande3-github-io-4431 90 3 pieces piece NNS adeshpande3-github-io-4431 90 4 of of IN adeshpande3-github-io-4431 90 5 the the DT adeshpande3-github-io-4431 90 6 network network NN adeshpande3-github-io-4431 90 7 that that WDT adeshpande3-github-io-4431 90 8 are be VBP adeshpande3-github-io-4431 90 9 happening happen VBG adeshpande3-github-io-4431 90 10 in in IN adeshpande3-github-io-4431 90 11 parallel parallel NN adeshpande3-github-io-4431 90 12 . . . adeshpande3-github-io-4431 91 1 This this DT adeshpande3-github-io-4431 91 2 box box NN adeshpande3-github-io-4431 91 3 is be VBZ adeshpande3-github-io-4431 91 4 called call VBN adeshpande3-github-io-4431 91 5 an an DT adeshpande3-github-io-4431 91 6 Inception Inception NNP adeshpande3-github-io-4431 91 7 module module NN adeshpande3-github-io-4431 91 8 . . . adeshpande3-github-io-4431 92 1 Let let VB adeshpande3-github-io-4431 92 2 ’s -PRON- PRP adeshpande3-github-io-4431 92 3 take take VB adeshpande3-github-io-4431 92 4 a a DT adeshpande3-github-io-4431 92 5 closer close JJR adeshpande3-github-io-4431 92 6 look look NN adeshpande3-github-io-4431 92 7 at at IN adeshpande3-github-io-4431 92 8 what what WP adeshpande3-github-io-4431 92 9 it -PRON- PRP adeshpande3-github-io-4431 92 10 ’s ’ VBZ adeshpande3-github-io-4431 92 11 made make VBN adeshpande3-github-io-4431 92 12 of of IN adeshpande3-github-io-4431 92 13 . . . adeshpande3-github-io-4431 93 1 The the DT adeshpande3-github-io-4431 93 2 bottom bottom JJ adeshpande3-github-io-4431 93 3 green green JJ adeshpande3-github-io-4431 93 4 box box NN adeshpande3-github-io-4431 93 5 is be VBZ adeshpande3-github-io-4431 93 6 our -PRON- PRP$ adeshpande3-github-io-4431 93 7 input input NN adeshpande3-github-io-4431 93 8 and and CC adeshpande3-github-io-4431 93 9 the the DT adeshpande3-github-io-4431 93 10 top top JJ adeshpande3-github-io-4431 93 11 one one CD adeshpande3-github-io-4431 93 12 is be VBZ adeshpande3-github-io-4431 93 13 the the DT adeshpande3-github-io-4431 93 14 output output NN adeshpande3-github-io-4431 93 15 of of IN adeshpande3-github-io-4431 93 16 the the DT adeshpande3-github-io-4431 93 17 model model NN adeshpande3-github-io-4431 93 18 ( ( -LRB- adeshpande3-github-io-4431 93 19 Turning turn VBG adeshpande3-github-io-4431 93 20 this this DT adeshpande3-github-io-4431 93 21 picture picture NN adeshpande3-github-io-4431 93 22 right right JJ adeshpande3-github-io-4431 93 23 90 90 CD adeshpande3-github-io-4431 93 24 degrees degree NNS adeshpande3-github-io-4431 93 25 would would MD adeshpande3-github-io-4431 93 26 let let VB adeshpande3-github-io-4431 93 27 you -PRON- PRP adeshpande3-github-io-4431 93 28 visualize visualize VB adeshpande3-github-io-4431 93 29 the the DT adeshpande3-github-io-4431 93 30 model model NN adeshpande3-github-io-4431 93 31 in in IN adeshpande3-github-io-4431 93 32 relation relation NN adeshpande3-github-io-4431 93 33 to to IN adeshpande3-github-io-4431 93 34 the the DT adeshpande3-github-io-4431 93 35 last last JJ adeshpande3-github-io-4431 93 36 picture picture NN adeshpande3-github-io-4431 93 37 which which WDT adeshpande3-github-io-4431 93 38 shows show VBZ adeshpande3-github-io-4431 93 39 the the DT adeshpande3-github-io-4431 93 40 full full JJ adeshpande3-github-io-4431 93 41 network network NN adeshpande3-github-io-4431 93 42 ) ) -RRB- adeshpande3-github-io-4431 93 43 . . . adeshpande3-github-io-4431 94 1 Basically basically RB adeshpande3-github-io-4431 94 2 , , , adeshpande3-github-io-4431 94 3 at at IN adeshpande3-github-io-4431 94 4 each each DT adeshpande3-github-io-4431 94 5 layer layer NN adeshpande3-github-io-4431 94 6 of of IN adeshpande3-github-io-4431 94 7 a a DT adeshpande3-github-io-4431 94 8 traditional traditional JJ adeshpande3-github-io-4431 94 9 ConvNet ConvNet NNP adeshpande3-github-io-4431 94 10 , , , adeshpande3-github-io-4431 94 11 you -PRON- PRP adeshpande3-github-io-4431 94 12 have have VBP adeshpande3-github-io-4431 94 13 to to TO adeshpande3-github-io-4431 94 14 make make VB adeshpande3-github-io-4431 94 15 a a DT adeshpande3-github-io-4431 94 16 choice choice NN adeshpande3-github-io-4431 94 17 of of IN adeshpande3-github-io-4431 94 18 whether whether IN adeshpande3-github-io-4431 94 19 to to TO adeshpande3-github-io-4431 94 20 have have VB adeshpande3-github-io-4431 94 21 a a DT adeshpande3-github-io-4431 94 22 pooling pooling NN adeshpande3-github-io-4431 94 23 operation operation NN adeshpande3-github-io-4431 94 24 or or CC adeshpande3-github-io-4431 94 25 a a DT adeshpande3-github-io-4431 94 26 conv conv NN adeshpande3-github-io-4431 94 27 operation operation NN adeshpande3-github-io-4431 94 28 ( ( -LRB- adeshpande3-github-io-4431 94 29 there there EX adeshpande3-github-io-4431 94 30 is be VBZ adeshpande3-github-io-4431 94 31 also also RB adeshpande3-github-io-4431 94 32 the the DT adeshpande3-github-io-4431 94 33 choice choice NN adeshpande3-github-io-4431 94 34 of of IN adeshpande3-github-io-4431 94 35 filter filter NN adeshpande3-github-io-4431 94 36 size size NN adeshpande3-github-io-4431 94 37 ) ) -RRB- adeshpande3-github-io-4431 94 38 . . . adeshpande3-github-io-4431 95 1 What what WDT adeshpande3-github-io-4431 95 2 an an DT adeshpande3-github-io-4431 95 3 Inception Inception NNP adeshpande3-github-io-4431 95 4 module module NN adeshpande3-github-io-4431 95 5 allows allow VBZ adeshpande3-github-io-4431 95 6 you -PRON- PRP adeshpande3-github-io-4431 95 7 to to TO adeshpande3-github-io-4431 95 8 do do VB adeshpande3-github-io-4431 95 9 is be VBZ adeshpande3-github-io-4431 95 10 perform perform VB adeshpande3-github-io-4431 95 11 all all DT adeshpande3-github-io-4431 95 12 of of IN adeshpande3-github-io-4431 95 13 these these DT adeshpande3-github-io-4431 95 14 operations operation NNS adeshpande3-github-io-4431 95 15 in in IN adeshpande3-github-io-4431 95 16 parallel parallel NN adeshpande3-github-io-4431 95 17 . . . adeshpande3-github-io-4431 96 1 In in IN adeshpande3-github-io-4431 96 2 fact fact NN adeshpande3-github-io-4431 96 3 , , , adeshpande3-github-io-4431 96 4 this this DT adeshpande3-github-io-4431 96 5 was be VBD adeshpande3-github-io-4431 96 6 exactly exactly RB adeshpande3-github-io-4431 96 7 the the DT adeshpande3-github-io-4431 96 8 “ " `` adeshpande3-github-io-4431 96 9 naïve naïve NN adeshpande3-github-io-4431 96 10 ” " '' adeshpande3-github-io-4431 96 11 idea idea NN adeshpande3-github-io-4431 96 12 that that IN adeshpande3-github-io-4431 96 13 the the DT adeshpande3-github-io-4431 96 14 authors author NNS adeshpande3-github-io-4431 96 15 came come VBD adeshpande3-github-io-4431 96 16 up up RP adeshpande3-github-io-4431 96 17 with with IN adeshpande3-github-io-4431 96 18 . . . adeshpande3-github-io-4431 97 1 Now now RB adeshpande3-github-io-4431 97 2 , , , adeshpande3-github-io-4431 97 3 why why WRB adeshpande3-github-io-4431 97 4 does do VBZ adeshpande3-github-io-4431 97 5 n’t not RB adeshpande3-github-io-4431 97 6 this this DT adeshpande3-github-io-4431 97 7 work work NN adeshpande3-github-io-4431 97 8 ? ? . adeshpande3-github-io-4431 98 1 It -PRON- PRP adeshpande3-github-io-4431 98 2 would would MD adeshpande3-github-io-4431 98 3 lead lead VB adeshpande3-github-io-4431 98 4 to to IN adeshpande3-github-io-4431 98 5 way way NN adeshpande3-github-io-4431 98 6 too too RB adeshpande3-github-io-4431 98 7 many many JJ adeshpande3-github-io-4431 98 8 outputs output NNS adeshpande3-github-io-4431 98 9 . . . adeshpande3-github-io-4431 99 1 We -PRON- PRP adeshpande3-github-io-4431 99 2 would would MD adeshpande3-github-io-4431 99 3 end end VB adeshpande3-github-io-4431 99 4 up up RP adeshpande3-github-io-4431 99 5 with with IN adeshpande3-github-io-4431 99 6 an an DT adeshpande3-github-io-4431 99 7 extremely extremely RB adeshpande3-github-io-4431 99 8 large large JJ adeshpande3-github-io-4431 99 9 depth depth NN adeshpande3-github-io-4431 99 10 channel channel NN adeshpande3-github-io-4431 99 11 for for IN adeshpande3-github-io-4431 99 12 the the DT adeshpande3-github-io-4431 99 13 output output NN adeshpande3-github-io-4431 99 14 volume volume NN adeshpande3-github-io-4431 99 15 . . . adeshpande3-github-io-4431 100 1 The the DT adeshpande3-github-io-4431 100 2 way way NN adeshpande3-github-io-4431 100 3 that that WDT adeshpande3-github-io-4431 100 4 the the DT adeshpande3-github-io-4431 100 5 authors author NNS adeshpande3-github-io-4431 100 6 address address NN adeshpande3-github-io-4431 100 7 this this DT adeshpande3-github-io-4431 100 8 is be VBZ adeshpande3-github-io-4431 100 9 by by IN adeshpande3-github-io-4431 100 10 adding add VBG adeshpande3-github-io-4431 100 11 1x1 1x1 CD adeshpande3-github-io-4431 100 12 conv conv NN adeshpande3-github-io-4431 100 13 operations operation NNS adeshpande3-github-io-4431 100 14 before before IN adeshpande3-github-io-4431 100 15 the the DT adeshpande3-github-io-4431 100 16 3x3 3x3 NNP adeshpande3-github-io-4431 100 17 and and CC adeshpande3-github-io-4431 100 18 5x5 5x5 CD adeshpande3-github-io-4431 100 19 layers layer NNS adeshpande3-github-io-4431 100 20 . . . adeshpande3-github-io-4431 101 1 The the DT adeshpande3-github-io-4431 101 2 1x1 1x1 CD adeshpande3-github-io-4431 101 3 convolutions convolution NNS adeshpande3-github-io-4431 101 4 ( ( -LRB- adeshpande3-github-io-4431 101 5 or or CC adeshpande3-github-io-4431 101 6 network network NN adeshpande3-github-io-4431 101 7 in in IN adeshpande3-github-io-4431 101 8 network network NN adeshpande3-github-io-4431 101 9 layer layer NN adeshpande3-github-io-4431 101 10 ) ) -RRB- adeshpande3-github-io-4431 101 11 provide provide VB adeshpande3-github-io-4431 101 12 a a DT adeshpande3-github-io-4431 101 13 method method NN adeshpande3-github-io-4431 101 14 of of IN adeshpande3-github-io-4431 101 15 dimensionality dimensionality NN adeshpande3-github-io-4431 101 16 reduction reduction NN adeshpande3-github-io-4431 101 17 . . . adeshpande3-github-io-4431 102 1 For for IN adeshpande3-github-io-4431 102 2 example example NN adeshpande3-github-io-4431 102 3 , , , adeshpande3-github-io-4431 102 4 let let VB adeshpande3-github-io-4431 102 5 ’s -PRON- PRP adeshpande3-github-io-4431 102 6 say say VB adeshpande3-github-io-4431 102 7 you -PRON- PRP adeshpande3-github-io-4431 102 8 had have VBD adeshpande3-github-io-4431 102 9 an an DT adeshpande3-github-io-4431 102 10 input input JJ adeshpande3-github-io-4431 102 11 volume volume NN adeshpande3-github-io-4431 102 12 of of IN adeshpande3-github-io-4431 102 13 100x100x60 100x100x60 NNS adeshpande3-github-io-4431 102 14 ( ( -LRB- adeshpande3-github-io-4431 102 15 This this DT adeshpande3-github-io-4431 102 16 is be VBZ adeshpande3-github-io-4431 102 17 n’t not RB adeshpande3-github-io-4431 102 18 necessarily necessarily RB adeshpande3-github-io-4431 102 19 the the DT adeshpande3-github-io-4431 102 20 dimensions dimension NNS adeshpande3-github-io-4431 102 21 of of IN adeshpande3-github-io-4431 102 22 the the DT adeshpande3-github-io-4431 102 23 image image NN adeshpande3-github-io-4431 102 24 , , , adeshpande3-github-io-4431 102 25 just just RB adeshpande3-github-io-4431 102 26 the the DT adeshpande3-github-io-4431 102 27 input input NN adeshpande3-github-io-4431 102 28 to to IN adeshpande3-github-io-4431 102 29 any any DT adeshpande3-github-io-4431 102 30 layer layer NN adeshpande3-github-io-4431 102 31 of of IN adeshpande3-github-io-4431 102 32 the the DT adeshpande3-github-io-4431 102 33 network network NN adeshpande3-github-io-4431 102 34 ) ) -RRB- adeshpande3-github-io-4431 102 35 . . . adeshpande3-github-io-4431 103 1 Applying apply VBG adeshpande3-github-io-4431 103 2 20 20 CD adeshpande3-github-io-4431 103 3 filters filter NNS adeshpande3-github-io-4431 103 4 of of IN adeshpande3-github-io-4431 103 5 1x1 1x1 CD adeshpande3-github-io-4431 103 6 convolution convolution NN adeshpande3-github-io-4431 103 7 would would MD adeshpande3-github-io-4431 103 8 allow allow VB adeshpande3-github-io-4431 103 9 you -PRON- PRP adeshpande3-github-io-4431 103 10 to to TO adeshpande3-github-io-4431 103 11 reduce reduce VB adeshpande3-github-io-4431 103 12 the the DT adeshpande3-github-io-4431 103 13 volume volume NN adeshpande3-github-io-4431 103 14 to to IN adeshpande3-github-io-4431 103 15 100x100x20 100x100x20 CD adeshpande3-github-io-4431 103 16 . . . adeshpande3-github-io-4431 104 1 This this DT adeshpande3-github-io-4431 104 2 means mean VBZ adeshpande3-github-io-4431 104 3 that that IN adeshpande3-github-io-4431 104 4 the the DT adeshpande3-github-io-4431 104 5 3x3 3x3 CD adeshpande3-github-io-4431 104 6 and and CC adeshpande3-github-io-4431 104 7 5x5 5x5 CD adeshpande3-github-io-4431 104 8 convolutions convolution NNS adeshpande3-github-io-4431 104 9 wo will MD adeshpande3-github-io-4431 104 10 n’t not RB adeshpande3-github-io-4431 104 11 have have VB adeshpande3-github-io-4431 104 12 as as RB adeshpande3-github-io-4431 104 13 large large JJ adeshpande3-github-io-4431 104 14 of of IN adeshpande3-github-io-4431 104 15 a a DT adeshpande3-github-io-4431 104 16 volume volume NN adeshpande3-github-io-4431 104 17 to to TO adeshpande3-github-io-4431 104 18 deal deal VB adeshpande3-github-io-4431 104 19 with with IN adeshpande3-github-io-4431 104 20 . . . adeshpande3-github-io-4431 105 1 This this DT adeshpande3-github-io-4431 105 2 can can MD adeshpande3-github-io-4431 105 3 be be VB adeshpande3-github-io-4431 105 4 thought think VBN adeshpande3-github-io-4431 105 5 of of IN adeshpande3-github-io-4431 105 6 as as IN adeshpande3-github-io-4431 105 7 a a DT adeshpande3-github-io-4431 105 8 “ " `` adeshpande3-github-io-4431 105 9 pooling pooling NN adeshpande3-github-io-4431 105 10 of of IN adeshpande3-github-io-4431 105 11 features feature NNS adeshpande3-github-io-4431 105 12 ” " '' adeshpande3-github-io-4431 105 13 because because IN adeshpande3-github-io-4431 105 14 we -PRON- PRP adeshpande3-github-io-4431 105 15 are be VBP adeshpande3-github-io-4431 105 16 reducing reduce VBG adeshpande3-github-io-4431 105 17 the the DT adeshpande3-github-io-4431 105 18 depth depth NN adeshpande3-github-io-4431 105 19 of of IN adeshpande3-github-io-4431 105 20 the the DT adeshpande3-github-io-4431 105 21 volume volume NN adeshpande3-github-io-4431 105 22 , , , adeshpande3-github-io-4431 105 23 similar similar JJ adeshpande3-github-io-4431 105 24 to to IN adeshpande3-github-io-4431 105 25 how how WRB adeshpande3-github-io-4431 105 26 we -PRON- PRP adeshpande3-github-io-4431 105 27 reduce reduce VBP adeshpande3-github-io-4431 105 28 the the DT adeshpande3-github-io-4431 105 29 dimensions dimension NNS adeshpande3-github-io-4431 105 30 of of IN adeshpande3-github-io-4431 105 31 height height NN adeshpande3-github-io-4431 105 32 and and CC adeshpande3-github-io-4431 105 33 width width NN adeshpande3-github-io-4431 105 34 with with IN adeshpande3-github-io-4431 105 35 normal normal JJ adeshpande3-github-io-4431 105 36 maxpooling maxpooling NN adeshpande3-github-io-4431 105 37 layers layer NNS adeshpande3-github-io-4431 105 38 . . . adeshpande3-github-io-4431 106 1 Another another DT adeshpande3-github-io-4431 106 2 note note NN adeshpande3-github-io-4431 106 3 is be VBZ adeshpande3-github-io-4431 106 4 that that IN adeshpande3-github-io-4431 106 5 these these DT adeshpande3-github-io-4431 106 6 1x1 1x1 CD adeshpande3-github-io-4431 106 7 conv conv NN adeshpande3-github-io-4431 106 8 layers layer NNS adeshpande3-github-io-4431 106 9 are be VBP adeshpande3-github-io-4431 106 10 followed follow VBN adeshpande3-github-io-4431 106 11 by by IN adeshpande3-github-io-4431 106 12 ReLU relu NN adeshpande3-github-io-4431 106 13 units unit NNS adeshpande3-github-io-4431 106 14 which which WDT adeshpande3-github-io-4431 106 15 definitely definitely RB adeshpande3-github-io-4431 106 16 ca can MD adeshpande3-github-io-4431 106 17 n’t not RB adeshpande3-github-io-4431 106 18 hurt hurt VB adeshpande3-github-io-4431 106 19 ( ( -LRB- adeshpande3-github-io-4431 106 20 See see VB adeshpande3-github-io-4431 106 21 Aaditya Aaditya NNP adeshpande3-github-io-4431 106 22 Prakash Prakash NNP adeshpande3-github-io-4431 106 23 ’s ’s POS adeshpande3-github-io-4431 106 24 great great JJ adeshpande3-github-io-4431 106 25 post post NN adeshpande3-github-io-4431 106 26 for for IN adeshpande3-github-io-4431 106 27 more more JJR adeshpande3-github-io-4431 106 28 info info NN adeshpande3-github-io-4431 106 29 on on IN adeshpande3-github-io-4431 106 30 the the DT adeshpande3-github-io-4431 106 31 effectiveness effectiveness NN adeshpande3-github-io-4431 106 32 of of IN adeshpande3-github-io-4431 106 33 1x1 1x1 CD adeshpande3-github-io-4431 106 34 convolutions convolution NNS adeshpande3-github-io-4431 106 35 ) ) -RRB- adeshpande3-github-io-4431 106 36 . . . adeshpande3-github-io-4431 107 1 Check check VB adeshpande3-github-io-4431 107 2 out out RP adeshpande3-github-io-4431 107 3 this this DT adeshpande3-github-io-4431 107 4 video video NN adeshpande3-github-io-4431 107 5 for for IN adeshpande3-github-io-4431 107 6 a a DT adeshpande3-github-io-4431 107 7 great great JJ adeshpande3-github-io-4431 107 8 visualization visualization NN adeshpande3-github-io-4431 107 9 of of IN adeshpande3-github-io-4431 107 10 the the DT adeshpande3-github-io-4431 107 11 filter filter NN adeshpande3-github-io-4431 107 12 concatenation concatenation NN adeshpande3-github-io-4431 107 13 at at IN adeshpande3-github-io-4431 107 14 the the DT adeshpande3-github-io-4431 107 15 end end NN adeshpande3-github-io-4431 107 16 . . . adeshpande3-github-io-4431 108 1 You -PRON- PRP adeshpande3-github-io-4431 108 2 may may MD adeshpande3-github-io-4431 108 3 be be VB adeshpande3-github-io-4431 108 4 asking ask VBG adeshpande3-github-io-4431 108 5 yourself -PRON- PRP adeshpande3-github-io-4431 108 6 “ " `` adeshpande3-github-io-4431 108 7 How how WRB adeshpande3-github-io-4431 108 8 does do VBZ adeshpande3-github-io-4431 108 9 this this DT adeshpande3-github-io-4431 108 10 architecture architecture NN adeshpande3-github-io-4431 108 11 help help NN adeshpande3-github-io-4431 108 12 ? ? . adeshpande3-github-io-4431 108 13 ” " '' adeshpande3-github-io-4431 108 14 . . . adeshpande3-github-io-4431 109 1 Well well UH adeshpande3-github-io-4431 109 2 , , , adeshpande3-github-io-4431 109 3 you -PRON- PRP adeshpande3-github-io-4431 109 4 have have VBP adeshpande3-github-io-4431 109 5 a a DT adeshpande3-github-io-4431 109 6 module module NN adeshpande3-github-io-4431 109 7 that that WDT adeshpande3-github-io-4431 109 8 consists consist VBZ adeshpande3-github-io-4431 109 9 of of IN adeshpande3-github-io-4431 109 10 a a DT adeshpande3-github-io-4431 109 11 network network NN adeshpande3-github-io-4431 109 12 in in IN adeshpande3-github-io-4431 109 13 network network NN adeshpande3-github-io-4431 109 14 layer layer NN adeshpande3-github-io-4431 109 15 , , , adeshpande3-github-io-4431 109 16 a a DT adeshpande3-github-io-4431 109 17 medium medium JJ adeshpande3-github-io-4431 109 18 sized sized JJ adeshpande3-github-io-4431 109 19 filter filter NN adeshpande3-github-io-4431 109 20 convolution convolution NN adeshpande3-github-io-4431 109 21 , , , adeshpande3-github-io-4431 109 22 a a DT adeshpande3-github-io-4431 109 23 large large JJ adeshpande3-github-io-4431 109 24 sized sized JJ adeshpande3-github-io-4431 109 25 filter filter NN adeshpande3-github-io-4431 109 26 convolution convolution NN adeshpande3-github-io-4431 109 27 , , , adeshpande3-github-io-4431 109 28 and and CC adeshpande3-github-io-4431 109 29 a a DT adeshpande3-github-io-4431 109 30 pooling pooling NN adeshpande3-github-io-4431 109 31 operation operation NN adeshpande3-github-io-4431 109 32 . . . adeshpande3-github-io-4431 110 1 The the DT adeshpande3-github-io-4431 110 2 network network NN adeshpande3-github-io-4431 110 3 in in IN adeshpande3-github-io-4431 110 4 network network NN adeshpande3-github-io-4431 110 5 conv conv NN adeshpande3-github-io-4431 110 6 is be VBZ adeshpande3-github-io-4431 110 7 able able JJ adeshpande3-github-io-4431 110 8 to to TO adeshpande3-github-io-4431 110 9 extract extract VB adeshpande3-github-io-4431 110 10 information information NN adeshpande3-github-io-4431 110 11 about about IN adeshpande3-github-io-4431 110 12 the the DT adeshpande3-github-io-4431 110 13 very very RB adeshpande3-github-io-4431 110 14 fine fine JJ adeshpande3-github-io-4431 110 15 grain grain NN adeshpande3-github-io-4431 110 16 details detail NNS adeshpande3-github-io-4431 110 17 in in IN adeshpande3-github-io-4431 110 18 the the DT adeshpande3-github-io-4431 110 19 volume volume NN adeshpande3-github-io-4431 110 20 , , , adeshpande3-github-io-4431 110 21 while while IN adeshpande3-github-io-4431 110 22 the the DT adeshpande3-github-io-4431 110 23 5x5 5x5 CD adeshpande3-github-io-4431 110 24 filter filter NN adeshpande3-github-io-4431 110 25 is be VBZ adeshpande3-github-io-4431 110 26 able able JJ adeshpande3-github-io-4431 110 27 to to TO adeshpande3-github-io-4431 110 28 cover cover VB adeshpande3-github-io-4431 110 29 a a DT adeshpande3-github-io-4431 110 30 large large JJ adeshpande3-github-io-4431 110 31 receptive receptive JJ adeshpande3-github-io-4431 110 32 field field NN adeshpande3-github-io-4431 110 33 of of IN adeshpande3-github-io-4431 110 34 the the DT adeshpande3-github-io-4431 110 35 input input NN adeshpande3-github-io-4431 110 36 , , , adeshpande3-github-io-4431 110 37 and and CC adeshpande3-github-io-4431 110 38 thus thus RB adeshpande3-github-io-4431 110 39 able able JJ adeshpande3-github-io-4431 110 40 to to TO adeshpande3-github-io-4431 110 41 extract extract VB adeshpande3-github-io-4431 110 42 its -PRON- PRP$ adeshpande3-github-io-4431 110 43 information information NN adeshpande3-github-io-4431 110 44 as as RB adeshpande3-github-io-4431 110 45 well well RB adeshpande3-github-io-4431 110 46 . . . adeshpande3-github-io-4431 111 1 You -PRON- PRP adeshpande3-github-io-4431 111 2 also also RB adeshpande3-github-io-4431 111 3 have have VBP adeshpande3-github-io-4431 111 4 a a DT adeshpande3-github-io-4431 111 5 pooling pooling NN adeshpande3-github-io-4431 111 6 operation operation NN adeshpande3-github-io-4431 111 7 that that WDT adeshpande3-github-io-4431 111 8 helps help VBZ adeshpande3-github-io-4431 111 9 to to TO adeshpande3-github-io-4431 111 10 reduce reduce VB adeshpande3-github-io-4431 111 11 spatial spatial JJ adeshpande3-github-io-4431 111 12 sizes size NNS adeshpande3-github-io-4431 111 13 and and CC adeshpande3-github-io-4431 111 14 combat combat NN adeshpande3-github-io-4431 111 15 overfitting overfitting NN adeshpande3-github-io-4431 111 16 . . . adeshpande3-github-io-4431 112 1 On on IN adeshpande3-github-io-4431 112 2 top top NN adeshpande3-github-io-4431 112 3 of of IN adeshpande3-github-io-4431 112 4 all all DT adeshpande3-github-io-4431 112 5 of of IN adeshpande3-github-io-4431 112 6 that that DT adeshpande3-github-io-4431 112 7 , , , adeshpande3-github-io-4431 112 8 you -PRON- PRP adeshpande3-github-io-4431 112 9 have have VBP adeshpande3-github-io-4431 112 10 ReLUs relu NNS adeshpande3-github-io-4431 112 11 after after IN adeshpande3-github-io-4431 112 12 each each DT adeshpande3-github-io-4431 112 13 conv conv NN adeshpande3-github-io-4431 112 14 layer layer NN adeshpande3-github-io-4431 112 15 , , , adeshpande3-github-io-4431 112 16 which which WDT adeshpande3-github-io-4431 112 17 help help VBP adeshpande3-github-io-4431 112 18 improve improve VB adeshpande3-github-io-4431 112 19 the the DT adeshpande3-github-io-4431 112 20 nonlinearity nonlinearity NN adeshpande3-github-io-4431 112 21 of of IN adeshpande3-github-io-4431 112 22 the the DT adeshpande3-github-io-4431 112 23 network network NN adeshpande3-github-io-4431 112 24 . . . adeshpande3-github-io-4431 113 1 Basically basically RB adeshpande3-github-io-4431 113 2 , , , adeshpande3-github-io-4431 113 3 the the DT adeshpande3-github-io-4431 113 4 network network NN adeshpande3-github-io-4431 113 5 is be VBZ adeshpande3-github-io-4431 113 6 able able JJ adeshpande3-github-io-4431 113 7 to to TO adeshpande3-github-io-4431 113 8 perform perform VB adeshpande3-github-io-4431 113 9 the the DT adeshpande3-github-io-4431 113 10 functions function NNS adeshpande3-github-io-4431 113 11 of of IN adeshpande3-github-io-4431 113 12 these these DT adeshpande3-github-io-4431 113 13 different different JJ adeshpande3-github-io-4431 113 14 operations operation NNS adeshpande3-github-io-4431 113 15 while while IN adeshpande3-github-io-4431 113 16 still still RB adeshpande3-github-io-4431 113 17 remaining remain VBG adeshpande3-github-io-4431 113 18 computationally computationally RB adeshpande3-github-io-4431 113 19 considerate considerate JJ adeshpande3-github-io-4431 113 20 . . . adeshpande3-github-io-4431 114 1 The the DT adeshpande3-github-io-4431 114 2 paper paper NN adeshpande3-github-io-4431 114 3 does do VBZ adeshpande3-github-io-4431 114 4 also also RB adeshpande3-github-io-4431 114 5 give give VB adeshpande3-github-io-4431 114 6 more more JJR adeshpande3-github-io-4431 114 7 of of IN adeshpande3-github-io-4431 114 8 a a DT adeshpande3-github-io-4431 114 9 high high JJ adeshpande3-github-io-4431 114 10 level level NN adeshpande3-github-io-4431 114 11 reasoning reasoning NN adeshpande3-github-io-4431 114 12 that that WDT adeshpande3-github-io-4431 114 13 involves involve VBZ adeshpande3-github-io-4431 114 14 topics topic NNS adeshpande3-github-io-4431 114 15 like like IN adeshpande3-github-io-4431 114 16 sparsity sparsity NN adeshpande3-github-io-4431 114 17 and and CC adeshpande3-github-io-4431 114 18 dense dense JJ adeshpande3-github-io-4431 114 19 connections connection NNS adeshpande3-github-io-4431 114 20 ( ( -LRB- adeshpande3-github-io-4431 114 21 read read VB adeshpande3-github-io-4431 114 22 Sections section NNS adeshpande3-github-io-4431 114 23 3 3 CD adeshpande3-github-io-4431 114 24 and and CC adeshpande3-github-io-4431 114 25 4 4 CD adeshpande3-github-io-4431 114 26 of of IN adeshpande3-github-io-4431 114 27 the the DT adeshpande3-github-io-4431 114 28 paper paper NN adeshpande3-github-io-4431 114 29 . . . adeshpande3-github-io-4431 115 1 Still still RB adeshpande3-github-io-4431 115 2 not not RB adeshpande3-github-io-4431 115 3 totally totally RB adeshpande3-github-io-4431 115 4 clear clear JJ adeshpande3-github-io-4431 115 5 to to IN adeshpande3-github-io-4431 115 6 me -PRON- PRP adeshpande3-github-io-4431 115 7 , , , adeshpande3-github-io-4431 115 8 but but CC adeshpande3-github-io-4431 115 9 if if IN adeshpande3-github-io-4431 115 10 anybody anybody NN adeshpande3-github-io-4431 115 11 has have VBZ adeshpande3-github-io-4431 115 12 any any DT adeshpande3-github-io-4431 115 13 insights insight NNS adeshpande3-github-io-4431 115 14 , , , adeshpande3-github-io-4431 115 15 I -PRON- PRP adeshpande3-github-io-4431 115 16 ’d ’d , adeshpande3-github-io-4431 115 17 love love VB adeshpande3-github-io-4431 115 18 to to TO adeshpande3-github-io-4431 115 19 hear hear VB adeshpande3-github-io-4431 115 20 them -PRON- PRP adeshpande3-github-io-4431 115 21 in in IN adeshpande3-github-io-4431 115 22 the the DT adeshpande3-github-io-4431 115 23 comments comment NNS adeshpande3-github-io-4431 115 24 ! ! . adeshpande3-github-io-4431 115 25 ) ) -RRB- adeshpande3-github-io-4431 115 26 . . . adeshpande3-github-io-4431 116 1 Main main JJ adeshpande3-github-io-4431 116 2 Points Points NNPS adeshpande3-github-io-4431 116 3 Used use VBD adeshpande3-github-io-4431 116 4 9 9 CD adeshpande3-github-io-4431 116 5 Inception inception NN adeshpande3-github-io-4431 116 6 modules module NNS adeshpande3-github-io-4431 116 7 in in IN adeshpande3-github-io-4431 116 8 the the DT adeshpande3-github-io-4431 116 9 whole whole JJ adeshpande3-github-io-4431 116 10 architecture architecture NN adeshpande3-github-io-4431 116 11 , , , adeshpande3-github-io-4431 116 12 with with IN adeshpande3-github-io-4431 116 13 over over IN adeshpande3-github-io-4431 116 14 100 100 CD adeshpande3-github-io-4431 116 15 layers layer NNS adeshpande3-github-io-4431 116 16 in in IN adeshpande3-github-io-4431 116 17 total total NN adeshpande3-github-io-4431 116 18 ! ! . adeshpande3-github-io-4431 117 1 Now now RB adeshpande3-github-io-4431 117 2 that that DT adeshpande3-github-io-4431 117 3 is be VBZ adeshpande3-github-io-4431 117 4 deep deep JJ adeshpande3-github-io-4431 117 5 … … NFP adeshpande3-github-io-4431 117 6 No no DT adeshpande3-github-io-4431 117 7 use use NN adeshpande3-github-io-4431 117 8 of of IN adeshpande3-github-io-4431 117 9 fully fully RB adeshpande3-github-io-4431 117 10 connected connect VBN adeshpande3-github-io-4431 117 11 layers layer NNS adeshpande3-github-io-4431 117 12 ! ! . adeshpande3-github-io-4431 118 1 They -PRON- PRP adeshpande3-github-io-4431 118 2 use use VBP adeshpande3-github-io-4431 118 3 an an DT adeshpande3-github-io-4431 118 4 average average JJ adeshpande3-github-io-4431 118 5 pool pool NN adeshpande3-github-io-4431 118 6 instead instead RB adeshpande3-github-io-4431 118 7 , , , adeshpande3-github-io-4431 118 8 to to TO adeshpande3-github-io-4431 118 9 go go VB adeshpande3-github-io-4431 118 10 from from IN adeshpande3-github-io-4431 118 11 a a DT adeshpande3-github-io-4431 118 12 7x7x1024 7x7x1024 NN adeshpande3-github-io-4431 118 13 volume volume NN adeshpande3-github-io-4431 118 14 to to IN adeshpande3-github-io-4431 118 15 a a DT adeshpande3-github-io-4431 118 16 1x1x1024 1x1x1024 CD adeshpande3-github-io-4431 118 17 volume volume NN adeshpande3-github-io-4431 118 18 . . . adeshpande3-github-io-4431 119 1 This this DT adeshpande3-github-io-4431 119 2 saves save VBZ adeshpande3-github-io-4431 119 3 a a DT adeshpande3-github-io-4431 119 4 huge huge JJ adeshpande3-github-io-4431 119 5 number number NN adeshpande3-github-io-4431 119 6 of of IN adeshpande3-github-io-4431 119 7 parameters parameter NNS adeshpande3-github-io-4431 119 8 . . . adeshpande3-github-io-4431 120 1 Uses use VBZ adeshpande3-github-io-4431 120 2 12x 12x CD adeshpande3-github-io-4431 120 3 fewer few JJR adeshpande3-github-io-4431 120 4 parameters parameter NNS adeshpande3-github-io-4431 120 5 than than IN adeshpande3-github-io-4431 120 6 AlexNet AlexNet NNP adeshpande3-github-io-4431 120 7 . . . adeshpande3-github-io-4431 121 1 During during IN adeshpande3-github-io-4431 121 2 testing testing NN adeshpande3-github-io-4431 121 3 , , , adeshpande3-github-io-4431 121 4 multiple multiple JJ adeshpande3-github-io-4431 121 5 crops crop NNS adeshpande3-github-io-4431 121 6 of of IN adeshpande3-github-io-4431 121 7 the the DT adeshpande3-github-io-4431 121 8 same same JJ adeshpande3-github-io-4431 121 9 image image NN adeshpande3-github-io-4431 121 10 were be VBD adeshpande3-github-io-4431 121 11 created create VBN adeshpande3-github-io-4431 121 12 , , , adeshpande3-github-io-4431 121 13 fed feed VBN adeshpande3-github-io-4431 121 14 into into IN adeshpande3-github-io-4431 121 15 the the DT adeshpande3-github-io-4431 121 16 network network NN adeshpande3-github-io-4431 121 17 , , , adeshpande3-github-io-4431 121 18 and and CC adeshpande3-github-io-4431 121 19 the the DT adeshpande3-github-io-4431 121 20 softmax softmax NNP adeshpande3-github-io-4431 121 21 probabilities probability NNS adeshpande3-github-io-4431 121 22 were be VBD adeshpande3-github-io-4431 121 23 averaged average VBN adeshpande3-github-io-4431 121 24 to to TO adeshpande3-github-io-4431 121 25 give give VB adeshpande3-github-io-4431 121 26 us -PRON- PRP adeshpande3-github-io-4431 121 27 the the DT adeshpande3-github-io-4431 121 28 final final JJ adeshpande3-github-io-4431 121 29 solution solution NN adeshpande3-github-io-4431 121 30 . . . adeshpande3-github-io-4431 122 1 Utilized utilize VBN adeshpande3-github-io-4431 122 2 concepts concept NNS adeshpande3-github-io-4431 122 3 from from IN adeshpande3-github-io-4431 122 4 R R NNP adeshpande3-github-io-4431 122 5 - - HYPH adeshpande3-github-io-4431 122 6 CNN CNN NNP adeshpande3-github-io-4431 122 7 ( ( -LRB- adeshpande3-github-io-4431 122 8 a a DT adeshpande3-github-io-4431 122 9 paper paper NN adeshpande3-github-io-4431 122 10 we -PRON- PRP adeshpande3-github-io-4431 122 11 ’ll will MD adeshpande3-github-io-4431 122 12 discuss discuss VB adeshpande3-github-io-4431 122 13 later later RB adeshpande3-github-io-4431 122 14 ) ) -RRB- adeshpande3-github-io-4431 122 15 for for IN adeshpande3-github-io-4431 122 16 their -PRON- PRP$ adeshpande3-github-io-4431 122 17 detection detection NN adeshpande3-github-io-4431 122 18 model model NN adeshpande3-github-io-4431 122 19 . . . adeshpande3-github-io-4431 123 1 There there EX adeshpande3-github-io-4431 123 2 are be VBP adeshpande3-github-io-4431 123 3 updated update VBN adeshpande3-github-io-4431 123 4 versions version NNS adeshpande3-github-io-4431 123 5 to to IN adeshpande3-github-io-4431 123 6 the the DT adeshpande3-github-io-4431 123 7 Inception inception NN adeshpande3-github-io-4431 123 8 module module NN adeshpande3-github-io-4431 123 9 ( ( -LRB- adeshpande3-github-io-4431 123 10 Versions version NNS adeshpande3-github-io-4431 123 11 6 6 CD adeshpande3-github-io-4431 123 12 and and CC adeshpande3-github-io-4431 123 13 7 7 CD adeshpande3-github-io-4431 123 14 ) ) -RRB- adeshpande3-github-io-4431 123 15 . . . adeshpande3-github-io-4431 124 1 Trained train VBN adeshpande3-github-io-4431 124 2 on on IN adeshpande3-github-io-4431 124 3 “ " `` adeshpande3-github-io-4431 124 4 a a DT adeshpande3-github-io-4431 124 5 few few JJ adeshpande3-github-io-4431 124 6 high high JJ adeshpande3-github-io-4431 124 7 - - HYPH adeshpande3-github-io-4431 124 8 end end NN adeshpande3-github-io-4431 124 9 GPUs gpu NNS adeshpande3-github-io-4431 124 10 within within IN adeshpande3-github-io-4431 124 11 a a DT adeshpande3-github-io-4431 124 12 week week NN adeshpande3-github-io-4431 124 13 ” " '' adeshpande3-github-io-4431 124 14 . . . adeshpande3-github-io-4431 125 1 Why why WRB adeshpande3-github-io-4431 125 2 It -PRON- PRP adeshpande3-github-io-4431 125 3 ’s ’ VBZ adeshpande3-github-io-4431 125 4 Important important JJ adeshpande3-github-io-4431 125 5                                 _SP adeshpande3-github-io-4431 125 6 GoogLeNet GoogLeNet NNP adeshpande3-github-io-4431 125 7 was be VBD adeshpande3-github-io-4431 125 8 one one CD adeshpande3-github-io-4431 125 9 of of IN adeshpande3-github-io-4431 125 10 the the DT adeshpande3-github-io-4431 125 11 first first JJ adeshpande3-github-io-4431 125 12 models model NNS adeshpande3-github-io-4431 125 13 that that WDT adeshpande3-github-io-4431 125 14 introduced introduce VBD adeshpande3-github-io-4431 125 15 the the DT adeshpande3-github-io-4431 125 16 idea idea NN adeshpande3-github-io-4431 125 17 that that IN adeshpande3-github-io-4431 125 18 CNN CNN NNP adeshpande3-github-io-4431 125 19 layers layer NNS adeshpande3-github-io-4431 125 20 did do VBD adeshpande3-github-io-4431 125 21 n’t not RB adeshpande3-github-io-4431 125 22 always always RB adeshpande3-github-io-4431 125 23 have have VB adeshpande3-github-io-4431 125 24 to to TO adeshpande3-github-io-4431 125 25 be be VB adeshpande3-github-io-4431 125 26 stacked stack VBN adeshpande3-github-io-4431 125 27 up up RP adeshpande3-github-io-4431 125 28 sequentially sequentially RB adeshpande3-github-io-4431 125 29 . . . adeshpande3-github-io-4431 126 1 Coming come VBG adeshpande3-github-io-4431 126 2 up up RP adeshpande3-github-io-4431 126 3 with with IN adeshpande3-github-io-4431 126 4 the the DT adeshpande3-github-io-4431 126 5 Inception Inception NNP adeshpande3-github-io-4431 126 6 module module NN adeshpande3-github-io-4431 126 7 , , , adeshpande3-github-io-4431 126 8 the the DT adeshpande3-github-io-4431 126 9 authors author NNS adeshpande3-github-io-4431 126 10 showed show VBD adeshpande3-github-io-4431 126 11 that that IN adeshpande3-github-io-4431 126 12 a a DT adeshpande3-github-io-4431 126 13 creative creative JJ adeshpande3-github-io-4431 126 14 structuring structuring NN adeshpande3-github-io-4431 126 15 of of IN adeshpande3-github-io-4431 126 16 layers layer NNS adeshpande3-github-io-4431 126 17 can can MD adeshpande3-github-io-4431 126 18 lead lead VB adeshpande3-github-io-4431 126 19 to to IN adeshpande3-github-io-4431 126 20 improved improve VBN adeshpande3-github-io-4431 126 21 performance performance NN adeshpande3-github-io-4431 126 22 and and CC adeshpande3-github-io-4431 126 23 computationally computationally RB adeshpande3-github-io-4431 126 24 efficiency efficiency NN adeshpande3-github-io-4431 126 25 . . . adeshpande3-github-io-4431 127 1 This this DT adeshpande3-github-io-4431 127 2 paper paper NN adeshpande3-github-io-4431 127 3 has have VBZ adeshpande3-github-io-4431 127 4 really really RB adeshpande3-github-io-4431 127 5 set set VBN adeshpande3-github-io-4431 127 6 the the DT adeshpande3-github-io-4431 127 7 stage stage NN adeshpande3-github-io-4431 127 8 for for IN adeshpande3-github-io-4431 127 9 some some DT adeshpande3-github-io-4431 127 10 amazing amazing JJ adeshpande3-github-io-4431 127 11 architectures architecture NNS adeshpande3-github-io-4431 127 12 that that WDT adeshpande3-github-io-4431 127 13 we -PRON- PRP adeshpande3-github-io-4431 127 14 could could MD adeshpande3-github-io-4431 127 15 see see VB adeshpande3-github-io-4431 127 16 in in IN adeshpande3-github-io-4431 127 17 the the DT adeshpande3-github-io-4431 127 18 coming come VBG adeshpande3-github-io-4431 127 19 years year NNS adeshpande3-github-io-4431 127 20 . . . adeshpande3-github-io-4431 128 1 Microsoft Microsoft NNP adeshpande3-github-io-4431 128 2 ResNet ResNet NNP adeshpande3-github-io-4431 128 3 ( ( -LRB- adeshpande3-github-io-4431 128 4 2015 2015 CD adeshpande3-github-io-4431 128 5 ) ) -RRB- adeshpande3-github-io-4431 128 6                                 _SP adeshpande3-github-io-4431 128 7 Imagine imagine VB adeshpande3-github-io-4431 128 8 a a DT adeshpande3-github-io-4431 128 9 deep deep JJ adeshpande3-github-io-4431 128 10 CNN CNN NNP adeshpande3-github-io-4431 128 11 architecture architecture NN adeshpande3-github-io-4431 128 12 . . . adeshpande3-github-io-4431 129 1 Take take VB adeshpande3-github-io-4431 129 2 that that DT adeshpande3-github-io-4431 129 3 , , , adeshpande3-github-io-4431 129 4 double double VB adeshpande3-github-io-4431 129 5 the the DT adeshpande3-github-io-4431 129 6 number number NN adeshpande3-github-io-4431 129 7 of of IN adeshpande3-github-io-4431 129 8 layers layer NNS adeshpande3-github-io-4431 129 9 , , , adeshpande3-github-io-4431 129 10 add add VB adeshpande3-github-io-4431 129 11 a a DT adeshpande3-github-io-4431 129 12 couple couple NN adeshpande3-github-io-4431 129 13 more more RBR adeshpande3-github-io-4431 129 14 , , , adeshpande3-github-io-4431 129 15 and and CC adeshpande3-github-io-4431 129 16 it -PRON- PRP adeshpande3-github-io-4431 129 17 still still RB adeshpande3-github-io-4431 129 18 probably probably RB adeshpande3-github-io-4431 129 19 is be VBZ adeshpande3-github-io-4431 129 20 n’t not RB adeshpande3-github-io-4431 129 21 as as RB adeshpande3-github-io-4431 129 22 deep deep JJ adeshpande3-github-io-4431 129 23 as as IN adeshpande3-github-io-4431 129 24 the the DT adeshpande3-github-io-4431 129 25 ResNet ResNet NNP adeshpande3-github-io-4431 129 26 architecture architecture NN adeshpande3-github-io-4431 129 27 that that IN adeshpande3-github-io-4431 129 28 Microsoft Microsoft NNP adeshpande3-github-io-4431 129 29 Research Research NNP adeshpande3-github-io-4431 129 30 Asia Asia NNP adeshpande3-github-io-4431 129 31 came come VBD adeshpande3-github-io-4431 129 32 up up RP adeshpande3-github-io-4431 129 33 with with IN adeshpande3-github-io-4431 129 34 in in IN adeshpande3-github-io-4431 129 35 late late JJ adeshpande3-github-io-4431 129 36 2015 2015 CD adeshpande3-github-io-4431 129 37 . . . adeshpande3-github-io-4431 130 1 ResNet resnet NN adeshpande3-github-io-4431 130 2 is be VBZ adeshpande3-github-io-4431 130 3 a a DT adeshpande3-github-io-4431 130 4 new new JJ adeshpande3-github-io-4431 130 5 152 152 CD adeshpande3-github-io-4431 130 6 layer layer NN adeshpande3-github-io-4431 130 7 network network NN adeshpande3-github-io-4431 130 8 architecture architecture NN adeshpande3-github-io-4431 130 9 that that WDT adeshpande3-github-io-4431 130 10 set set VBD adeshpande3-github-io-4431 130 11 new new JJ adeshpande3-github-io-4431 130 12 records record NNS adeshpande3-github-io-4431 130 13 in in IN adeshpande3-github-io-4431 130 14 classification classification NN adeshpande3-github-io-4431 130 15 , , , adeshpande3-github-io-4431 130 16 detection detection NN adeshpande3-github-io-4431 130 17 , , , adeshpande3-github-io-4431 130 18 and and CC adeshpande3-github-io-4431 130 19 localization localization NN adeshpande3-github-io-4431 130 20 through through IN adeshpande3-github-io-4431 130 21 one one CD adeshpande3-github-io-4431 130 22 incredible incredible JJ adeshpande3-github-io-4431 130 23 architecture architecture NN adeshpande3-github-io-4431 130 24 . . . adeshpande3-github-io-4431 131 1 Aside aside RB adeshpande3-github-io-4431 131 2 from from IN adeshpande3-github-io-4431 131 3 the the DT adeshpande3-github-io-4431 131 4 new new JJ adeshpande3-github-io-4431 131 5 record record NN adeshpande3-github-io-4431 131 6 in in IN adeshpande3-github-io-4431 131 7 terms term NNS adeshpande3-github-io-4431 131 8 of of IN adeshpande3-github-io-4431 131 9 number number NN adeshpande3-github-io-4431 131 10 of of IN adeshpande3-github-io-4431 131 11 layers layer NNS adeshpande3-github-io-4431 131 12 , , , adeshpande3-github-io-4431 131 13 ResNet ResNet NNP adeshpande3-github-io-4431 131 14 won win VBD adeshpande3-github-io-4431 131 15 ILSVRC ILSVRC NNP adeshpande3-github-io-4431 131 16 2015 2015 CD adeshpande3-github-io-4431 131 17 with with IN adeshpande3-github-io-4431 131 18 an an DT adeshpande3-github-io-4431 131 19 incredible incredible JJ adeshpande3-github-io-4431 131 20 error error NN adeshpande3-github-io-4431 131 21 rate rate NN adeshpande3-github-io-4431 131 22 of of IN adeshpande3-github-io-4431 131 23 3.6 3.6 CD adeshpande3-github-io-4431 131 24 % % NN adeshpande3-github-io-4431 131 25 ( ( -LRB- adeshpande3-github-io-4431 131 26 Depending depend VBG adeshpande3-github-io-4431 131 27 on on IN adeshpande3-github-io-4431 131 28 their -PRON- PRP$ adeshpande3-github-io-4431 131 29 skill skill NN adeshpande3-github-io-4431 131 30 and and CC adeshpande3-github-io-4431 131 31 expertise expertise NN adeshpande3-github-io-4431 131 32 , , , adeshpande3-github-io-4431 131 33 humans human NNS adeshpande3-github-io-4431 131 34 generally generally RB adeshpande3-github-io-4431 131 35 hover hover VBP adeshpande3-github-io-4431 131 36 around around RB adeshpande3-github-io-4431 131 37 a a DT adeshpande3-github-io-4431 131 38 5 5 CD adeshpande3-github-io-4431 131 39 - - SYM adeshpande3-github-io-4431 131 40 10 10 CD adeshpande3-github-io-4431 131 41 % % NN adeshpande3-github-io-4431 131 42 error error NN adeshpande3-github-io-4431 131 43 rate rate NN adeshpande3-github-io-4431 131 44 . . . adeshpande3-github-io-4431 132 1 See see VB adeshpande3-github-io-4431 132 2 Andrej Andrej NNP adeshpande3-github-io-4431 132 3 Karpathy Karpathy NNP adeshpande3-github-io-4431 132 4 ’s ’s POS adeshpande3-github-io-4431 132 5 great great JJ adeshpande3-github-io-4431 132 6 post post NN adeshpande3-github-io-4431 132 7 on on IN adeshpande3-github-io-4431 132 8 his -PRON- PRP$ adeshpande3-github-io-4431 132 9 experiences experience NNS adeshpande3-github-io-4431 132 10 with with IN adeshpande3-github-io-4431 132 11 competing compete VBG adeshpande3-github-io-4431 132 12 against against IN adeshpande3-github-io-4431 132 13 ConvNets ConvNets NNP adeshpande3-github-io-4431 132 14 on on IN adeshpande3-github-io-4431 132 15 the the DT adeshpande3-github-io-4431 132 16 ImageNet ImageNet NNP adeshpande3-github-io-4431 132 17 challenge challenge NN adeshpande3-github-io-4431 132 18 ) ) -RRB- adeshpande3-github-io-4431 132 19 . . . adeshpande3-github-io-4431 133 1 Residual Residual NNP adeshpande3-github-io-4431 133 2 Block Block NNP adeshpande3-github-io-4431 133 3                                 _SP adeshpande3-github-io-4431 133 4 The the DT adeshpande3-github-io-4431 133 5 idea idea NN adeshpande3-github-io-4431 133 6 behind behind IN adeshpande3-github-io-4431 133 7 a a DT adeshpande3-github-io-4431 133 8 residual residual JJ adeshpande3-github-io-4431 133 9 block block NN adeshpande3-github-io-4431 133 10 is be VBZ adeshpande3-github-io-4431 133 11 that that IN adeshpande3-github-io-4431 133 12 you -PRON- PRP adeshpande3-github-io-4431 133 13 have have VBP adeshpande3-github-io-4431 133 14 your -PRON- PRP$ adeshpande3-github-io-4431 133 15 input input NN adeshpande3-github-io-4431 133 16 x x TO adeshpande3-github-io-4431 133 17 go go VB adeshpande3-github-io-4431 133 18 through through IN adeshpande3-github-io-4431 133 19 conv conv NN adeshpande3-github-io-4431 133 20 - - HYPH adeshpande3-github-io-4431 133 21 relu relu JJ adeshpande3-github-io-4431 133 22 - - HYPH adeshpande3-github-io-4431 133 23 conv conv NN adeshpande3-github-io-4431 133 24 series series NN adeshpande3-github-io-4431 133 25 . . . adeshpande3-github-io-4431 134 1 This this DT adeshpande3-github-io-4431 134 2 will will MD adeshpande3-github-io-4431 134 3 give give VB adeshpande3-github-io-4431 134 4 you -PRON- PRP adeshpande3-github-io-4431 134 5 some some DT adeshpande3-github-io-4431 134 6 F(x f(x NN adeshpande3-github-io-4431 134 7 ) ) -RRB- adeshpande3-github-io-4431 134 8 . . . adeshpande3-github-io-4431 135 1 That that DT adeshpande3-github-io-4431 135 2 result result NN adeshpande3-github-io-4431 135 3 is be VBZ adeshpande3-github-io-4431 135 4 then then RB adeshpande3-github-io-4431 135 5 added add VBN adeshpande3-github-io-4431 135 6 to to IN adeshpande3-github-io-4431 135 7 the the DT adeshpande3-github-io-4431 135 8 original original JJ adeshpande3-github-io-4431 135 9 input input NN adeshpande3-github-io-4431 135 10 x. x. NNP adeshpande3-github-io-4431 135 11 Let let VB adeshpande3-github-io-4431 135 12 ’s -PRON- PRP adeshpande3-github-io-4431 135 13 call call VB adeshpande3-github-io-4431 135 14 that that DT adeshpande3-github-io-4431 135 15 H(x h(x NN adeshpande3-github-io-4431 135 16 ) ) -RRB- adeshpande3-github-io-4431 135 17 = = NFP adeshpande3-github-io-4431 135 18 F(x f(x NN adeshpande3-github-io-4431 135 19 ) ) -RRB- adeshpande3-github-io-4431 135 20 + + CC adeshpande3-github-io-4431 135 21 x. x. NN adeshpande3-github-io-4431 136 1 In in IN adeshpande3-github-io-4431 136 2 traditional traditional JJ adeshpande3-github-io-4431 136 3 CNNs cnn NNS adeshpande3-github-io-4431 136 4 , , , adeshpande3-github-io-4431 136 5 your -PRON- PRP$ adeshpande3-github-io-4431 136 6 H(x h(x NN adeshpande3-github-io-4431 136 7 ) ) -RRB- adeshpande3-github-io-4431 136 8 would would MD adeshpande3-github-io-4431 136 9 just just RB adeshpande3-github-io-4431 136 10 be be VB adeshpande3-github-io-4431 136 11 equal equal JJ adeshpande3-github-io-4431 136 12 to to IN adeshpande3-github-io-4431 136 13 F(x f(x CD adeshpande3-github-io-4431 136 14 ) ) -RRB- adeshpande3-github-io-4431 136 15 right right JJ adeshpande3-github-io-4431 136 16 ? ? . adeshpande3-github-io-4431 137 1 So so RB adeshpande3-github-io-4431 137 2 , , , adeshpande3-github-io-4431 137 3 instead instead RB adeshpande3-github-io-4431 137 4 of of IN adeshpande3-github-io-4431 137 5 just just RB adeshpande3-github-io-4431 137 6 computing compute VBG adeshpande3-github-io-4431 137 7 that that DT adeshpande3-github-io-4431 137 8 transformation transformation NN adeshpande3-github-io-4431 137 9 ( ( -LRB- adeshpande3-github-io-4431 137 10 straight straight RB adeshpande3-github-io-4431 137 11 from from IN adeshpande3-github-io-4431 137 12 x x NNS adeshpande3-github-io-4431 137 13 to to IN adeshpande3-github-io-4431 137 14 F(x f(x CD adeshpande3-github-io-4431 137 15 ) ) -RRB- adeshpande3-github-io-4431 137 16 ) ) -RRB- adeshpande3-github-io-4431 137 17 , , , adeshpande3-github-io-4431 137 18 we -PRON- PRP adeshpande3-github-io-4431 137 19 ’re be VBP adeshpande3-github-io-4431 137 20 computing compute VBG adeshpande3-github-io-4431 137 21 the the DT adeshpande3-github-io-4431 137 22 term term NN adeshpande3-github-io-4431 137 23 that that WDT adeshpande3-github-io-4431 137 24 you -PRON- PRP adeshpande3-github-io-4431 137 25 have have VBP adeshpande3-github-io-4431 137 26 to to TO adeshpande3-github-io-4431 137 27 add add VB adeshpande3-github-io-4431 137 28 , , , adeshpande3-github-io-4431 137 29 F(x f(x CD adeshpande3-github-io-4431 137 30 ) ) -RRB- adeshpande3-github-io-4431 137 31 , , , adeshpande3-github-io-4431 137 32 to to IN adeshpande3-github-io-4431 137 33 your -PRON- PRP$ adeshpande3-github-io-4431 137 34 input input NN adeshpande3-github-io-4431 137 35 , , , adeshpande3-github-io-4431 137 36 x. x. NNP adeshpande3-github-io-4431 138 1 Basically basically RB adeshpande3-github-io-4431 138 2 , , , adeshpande3-github-io-4431 138 3 the the DT adeshpande3-github-io-4431 138 4 mini mini JJ adeshpande3-github-io-4431 138 5 module module NN adeshpande3-github-io-4431 138 6 shown show VBN adeshpande3-github-io-4431 138 7 below below RB adeshpande3-github-io-4431 138 8 is be VBZ adeshpande3-github-io-4431 138 9 computing compute VBG adeshpande3-github-io-4431 138 10 a a DT adeshpande3-github-io-4431 138 11 “ " `` adeshpande3-github-io-4431 138 12 delta delta NN adeshpande3-github-io-4431 138 13 ” " '' adeshpande3-github-io-4431 138 14 or or CC adeshpande3-github-io-4431 138 15 a a DT adeshpande3-github-io-4431 138 16 slight slight JJ adeshpande3-github-io-4431 138 17 change change NN adeshpande3-github-io-4431 138 18 to to IN adeshpande3-github-io-4431 138 19 the the DT adeshpande3-github-io-4431 138 20 original original JJ adeshpande3-github-io-4431 138 21 input input NN adeshpande3-github-io-4431 138 22 x x NN adeshpande3-github-io-4431 138 23 to to TO adeshpande3-github-io-4431 138 24 get get VB adeshpande3-github-io-4431 138 25 a a DT adeshpande3-github-io-4431 138 26 slightly slightly RB adeshpande3-github-io-4431 138 27 altered alter VBN adeshpande3-github-io-4431 138 28 representation representation NN adeshpande3-github-io-4431 138 29 ( ( -LRB- adeshpande3-github-io-4431 138 30 When when WRB adeshpande3-github-io-4431 138 31 we -PRON- PRP adeshpande3-github-io-4431 138 32 think think VBP adeshpande3-github-io-4431 138 33 of of IN adeshpande3-github-io-4431 138 34 traditional traditional JJ adeshpande3-github-io-4431 138 35 CNNs cnn NNS adeshpande3-github-io-4431 138 36 , , , adeshpande3-github-io-4431 138 37 we -PRON- PRP adeshpande3-github-io-4431 138 38 go go VBP adeshpande3-github-io-4431 138 39 from from IN adeshpande3-github-io-4431 138 40 x x NNS adeshpande3-github-io-4431 138 41 to to IN adeshpande3-github-io-4431 138 42 F(x f(x CD adeshpande3-github-io-4431 138 43 ) ) -RRB- adeshpande3-github-io-4431 138 44 which which WDT adeshpande3-github-io-4431 138 45 is be VBZ adeshpande3-github-io-4431 138 46 a a DT adeshpande3-github-io-4431 138 47 completely completely RB adeshpande3-github-io-4431 138 48 new new JJ adeshpande3-github-io-4431 138 49 representation representation NN adeshpande3-github-io-4431 138 50 that that WDT adeshpande3-github-io-4431 138 51 does do VBZ adeshpande3-github-io-4431 138 52 n’t not RB adeshpande3-github-io-4431 138 53 keep keep VB adeshpande3-github-io-4431 138 54 any any DT adeshpande3-github-io-4431 138 55 information information NN adeshpande3-github-io-4431 138 56 about about IN adeshpande3-github-io-4431 138 57 the the DT adeshpande3-github-io-4431 138 58 original original JJ adeshpande3-github-io-4431 138 59 x x NN adeshpande3-github-io-4431 138 60 ) ) -RRB- adeshpande3-github-io-4431 138 61 . . . adeshpande3-github-io-4431 139 1 The the DT adeshpande3-github-io-4431 139 2 authors author NNS adeshpande3-github-io-4431 139 3 believe believe VBP adeshpande3-github-io-4431 139 4 that that IN adeshpande3-github-io-4431 139 5 “ " `` adeshpande3-github-io-4431 139 6 it -PRON- PRP adeshpande3-github-io-4431 139 7 is be VBZ adeshpande3-github-io-4431 139 8 easier easy JJR adeshpande3-github-io-4431 139 9 to to TO adeshpande3-github-io-4431 139 10 optimize optimize VB adeshpande3-github-io-4431 139 11 the the DT adeshpande3-github-io-4431 139 12 residual residual JJ adeshpande3-github-io-4431 139 13 mapping mapping NN adeshpande3-github-io-4431 139 14 than than IN adeshpande3-github-io-4431 139 15 to to TO adeshpande3-github-io-4431 139 16 optimize optimize VB adeshpande3-github-io-4431 139 17 the the DT adeshpande3-github-io-4431 139 18 original original JJ adeshpande3-github-io-4431 139 19 , , , adeshpande3-github-io-4431 139 20 unreferenced unreferenced JJ adeshpande3-github-io-4431 139 21 mapping mapping NN adeshpande3-github-io-4431 139 22 ” " '' adeshpande3-github-io-4431 139 23 . . . adeshpande3-github-io-4431 140 1 Another another DT adeshpande3-github-io-4431 140 2 reason reason NN adeshpande3-github-io-4431 140 3 for for IN adeshpande3-github-io-4431 140 4 why why WRB adeshpande3-github-io-4431 140 5 this this DT adeshpande3-github-io-4431 140 6 residual residual JJ adeshpande3-github-io-4431 140 7 block block NN adeshpande3-github-io-4431 140 8 might may MD adeshpande3-github-io-4431 140 9 be be VB adeshpande3-github-io-4431 140 10 effective effective JJ adeshpande3-github-io-4431 140 11 is be VBZ adeshpande3-github-io-4431 140 12 that that IN adeshpande3-github-io-4431 140 13 during during IN adeshpande3-github-io-4431 140 14 the the DT adeshpande3-github-io-4431 140 15 backward backward JJ adeshpande3-github-io-4431 140 16 pass pass NN adeshpande3-github-io-4431 140 17 of of IN adeshpande3-github-io-4431 140 18 backpropagation backpropagation NN adeshpande3-github-io-4431 140 19 , , , adeshpande3-github-io-4431 140 20 the the DT adeshpande3-github-io-4431 140 21 gradient gradient NN adeshpande3-github-io-4431 140 22 will will MD adeshpande3-github-io-4431 140 23 flow flow VB adeshpande3-github-io-4431 140 24 easily easily RB adeshpande3-github-io-4431 140 25 through through IN adeshpande3-github-io-4431 140 26 the the DT adeshpande3-github-io-4431 140 27 graph graph NN adeshpande3-github-io-4431 140 28 because because IN adeshpande3-github-io-4431 140 29 we -PRON- PRP adeshpande3-github-io-4431 140 30 have have VBP adeshpande3-github-io-4431 140 31 addition addition NN adeshpande3-github-io-4431 140 32 operations operation NNS adeshpande3-github-io-4431 140 33 , , , adeshpande3-github-io-4431 140 34 which which WDT adeshpande3-github-io-4431 140 35 distributes distribute VBZ adeshpande3-github-io-4431 140 36 the the DT adeshpande3-github-io-4431 140 37 gradient gradient NN adeshpande3-github-io-4431 140 38 . . . adeshpande3-github-io-4431 141 1 Main main JJ adeshpande3-github-io-4431 141 2 Points Points NNP adeshpande3-github-io-4431 141 3 “ " `` adeshpande3-github-io-4431 141 4 Ultra ultra JJ adeshpande3-github-io-4431 141 5 - - JJ adeshpande3-github-io-4431 141 6 deep deep JJ adeshpande3-github-io-4431 141 7 ” " '' adeshpande3-github-io-4431 141 8 – – : adeshpande3-github-io-4431 141 9 Yann Yann NNP adeshpande3-github-io-4431 141 10 LeCun LeCun NNP adeshpande3-github-io-4431 141 11 . . . adeshpande3-github-io-4431 142 1 152 152 CD adeshpande3-github-io-4431 142 2 layers layer NNS adeshpande3-github-io-4431 142 3 … … NFP adeshpande3-github-io-4431 142 4 Interesting interesting JJ adeshpande3-github-io-4431 142 5 note note NN adeshpande3-github-io-4431 142 6 that that IN adeshpande3-github-io-4431 142 7 after after IN adeshpande3-github-io-4431 142 8 only only RB adeshpande3-github-io-4431 142 9 the the DT adeshpande3-github-io-4431 142 10 first first JJ adeshpande3-github-io-4431 142 11 2 2 CD adeshpande3-github-io-4431 142 12 layers layer NNS adeshpande3-github-io-4431 142 13 , , , adeshpande3-github-io-4431 142 14 the the DT adeshpande3-github-io-4431 142 15 spatial spatial JJ adeshpande3-github-io-4431 142 16 size size NN adeshpande3-github-io-4431 142 17 gets get VBZ adeshpande3-github-io-4431 142 18 compressed compress VBN adeshpande3-github-io-4431 142 19 from from IN adeshpande3-github-io-4431 142 20 an an DT adeshpande3-github-io-4431 142 21 input input JJ adeshpande3-github-io-4431 142 22 volume volume NN adeshpande3-github-io-4431 142 23 of of IN adeshpande3-github-io-4431 142 24 224x224 224x224 CD adeshpande3-github-io-4431 142 25 to to IN adeshpande3-github-io-4431 142 26 a a DT adeshpande3-github-io-4431 142 27 56x56 56x56 NNP adeshpande3-github-io-4431 142 28 volume volume NN adeshpande3-github-io-4431 142 29 . . . adeshpande3-github-io-4431 143 1 Authors author NNS adeshpande3-github-io-4431 143 2 claim claim VBP adeshpande3-github-io-4431 143 3 that that IN adeshpande3-github-io-4431 143 4 a a DT adeshpande3-github-io-4431 143 5 naïve naïve JJ adeshpande3-github-io-4431 143 6 increase increase NN adeshpande3-github-io-4431 143 7 of of IN adeshpande3-github-io-4431 143 8 layers layer NNS adeshpande3-github-io-4431 143 9 in in IN adeshpande3-github-io-4431 143 10 plain plain JJ adeshpande3-github-io-4431 143 11 nets net NNS adeshpande3-github-io-4431 143 12 result result VBP adeshpande3-github-io-4431 143 13 in in IN adeshpande3-github-io-4431 143 14 higher high JJR adeshpande3-github-io-4431 143 15 training training NN adeshpande3-github-io-4431 143 16 and and CC adeshpande3-github-io-4431 143 17 test test NN adeshpande3-github-io-4431 143 18 error error NN adeshpande3-github-io-4431 143 19 ( ( -LRB- adeshpande3-github-io-4431 143 20 Figure figure NN adeshpande3-github-io-4431 143 21 1 1 CD adeshpande3-github-io-4431 143 22 in in IN adeshpande3-github-io-4431 143 23 the the DT adeshpande3-github-io-4431 143 24 paper paper NN adeshpande3-github-io-4431 143 25 ) ) -RRB- adeshpande3-github-io-4431 143 26 . . . adeshpande3-github-io-4431 144 1 The the DT adeshpande3-github-io-4431 144 2 group group NN adeshpande3-github-io-4431 144 3 tried try VBD adeshpande3-github-io-4431 144 4 a a DT adeshpande3-github-io-4431 144 5 1202-layer 1202-layer CD adeshpande3-github-io-4431 144 6 network network NN adeshpande3-github-io-4431 144 7 , , , adeshpande3-github-io-4431 144 8 but but CC adeshpande3-github-io-4431 144 9 got get VBD adeshpande3-github-io-4431 144 10 a a DT adeshpande3-github-io-4431 144 11 lower low JJR adeshpande3-github-io-4431 144 12 test test NN adeshpande3-github-io-4431 144 13 accuracy accuracy NN adeshpande3-github-io-4431 144 14 , , , adeshpande3-github-io-4431 144 15 presumably presumably RB adeshpande3-github-io-4431 144 16 due due IN adeshpande3-github-io-4431 144 17 to to IN adeshpande3-github-io-4431 144 18 overfitting overfitting NN adeshpande3-github-io-4431 144 19 . . . adeshpande3-github-io-4431 145 1 Trained train VBN adeshpande3-github-io-4431 145 2 on on IN adeshpande3-github-io-4431 145 3 an an DT adeshpande3-github-io-4431 145 4 8 8 CD adeshpande3-github-io-4431 145 5 GPU GPU NNP adeshpande3-github-io-4431 145 6 machine machine NN adeshpande3-github-io-4431 145 7 for for IN adeshpande3-github-io-4431 145 8 two two CD adeshpande3-github-io-4431 145 9 to to TO adeshpande3-github-io-4431 145 10 three three CD adeshpande3-github-io-4431 145 11 weeks week NNS adeshpande3-github-io-4431 145 12 . . . adeshpande3-github-io-4431 146 1 Why why WRB adeshpande3-github-io-4431 146 2 It -PRON- PRP adeshpande3-github-io-4431 146 3 ’s ’ VBZ adeshpande3-github-io-4431 146 4 Important important JJ adeshpande3-github-io-4431 146 5                                 _SP adeshpande3-github-io-4431 146 6 3.6 3.6 CD adeshpande3-github-io-4431 146 7 % % NN adeshpande3-github-io-4431 146 8 error error NN adeshpande3-github-io-4431 146 9 rate rate NN adeshpande3-github-io-4431 146 10 . . . adeshpande3-github-io-4431 147 1 That that IN adeshpande3-github-io-4431 147 2 itself -PRON- PRP adeshpande3-github-io-4431 147 3 should should MD adeshpande3-github-io-4431 147 4 be be VB adeshpande3-github-io-4431 147 5 enough enough JJ adeshpande3-github-io-4431 147 6 to to TO adeshpande3-github-io-4431 147 7 convince convince VB adeshpande3-github-io-4431 147 8 you -PRON- PRP adeshpande3-github-io-4431 147 9 . . . adeshpande3-github-io-4431 148 1 The the DT adeshpande3-github-io-4431 148 2 ResNet ResNet NNP adeshpande3-github-io-4431 148 3 model model NN adeshpande3-github-io-4431 148 4 is be VBZ adeshpande3-github-io-4431 148 5 the the DT adeshpande3-github-io-4431 148 6 best good JJS adeshpande3-github-io-4431 148 7 CNN CNN NNP adeshpande3-github-io-4431 148 8 architecture architecture NN adeshpande3-github-io-4431 148 9 that that IN adeshpande3-github-io-4431 148 10 we -PRON- PRP adeshpande3-github-io-4431 148 11 currently currently RB adeshpande3-github-io-4431 148 12 have have VBP adeshpande3-github-io-4431 148 13 and and CC adeshpande3-github-io-4431 148 14 is be VBZ adeshpande3-github-io-4431 148 15 a a DT adeshpande3-github-io-4431 148 16 great great JJ adeshpande3-github-io-4431 148 17 innovation innovation NN adeshpande3-github-io-4431 148 18 for for IN adeshpande3-github-io-4431 148 19 the the DT adeshpande3-github-io-4431 148 20 idea idea NN adeshpande3-github-io-4431 148 21 of of IN adeshpande3-github-io-4431 148 22 residual residual JJ adeshpande3-github-io-4431 148 23 learning learning NN adeshpande3-github-io-4431 148 24 . . . adeshpande3-github-io-4431 149 1 With with IN adeshpande3-github-io-4431 149 2 error error NN adeshpande3-github-io-4431 149 3 rates rate NNS adeshpande3-github-io-4431 149 4 dropping drop VBG adeshpande3-github-io-4431 149 5 every every DT adeshpande3-github-io-4431 149 6 year year NN adeshpande3-github-io-4431 149 7 since since IN adeshpande3-github-io-4431 149 8 2012 2012 CD adeshpande3-github-io-4431 149 9 , , , adeshpande3-github-io-4431 149 10 I -PRON- PRP adeshpande3-github-io-4431 149 11 ’m be VBP adeshpande3-github-io-4431 149 12 skeptical skeptical JJ adeshpande3-github-io-4431 149 13 about about IN adeshpande3-github-io-4431 149 14 whether whether IN adeshpande3-github-io-4431 149 15 or or CC adeshpande3-github-io-4431 149 16 not not RB adeshpande3-github-io-4431 149 17 they -PRON- PRP adeshpande3-github-io-4431 149 18 will will MD adeshpande3-github-io-4431 149 19 go go VB adeshpande3-github-io-4431 149 20 down down RB adeshpande3-github-io-4431 149 21 for for IN adeshpande3-github-io-4431 149 22 ILSVRC ILSVRC NNP adeshpande3-github-io-4431 149 23 2016 2016 CD adeshpande3-github-io-4431 149 24 . . . adeshpande3-github-io-4431 150 1 I -PRON- PRP adeshpande3-github-io-4431 150 2 believe believe VBP adeshpande3-github-io-4431 150 3 we -PRON- PRP adeshpande3-github-io-4431 150 4 ’ve have VB adeshpande3-github-io-4431 150 5 gotten get VBN adeshpande3-github-io-4431 150 6 to to IN adeshpande3-github-io-4431 150 7 the the DT adeshpande3-github-io-4431 150 8 point point NN adeshpande3-github-io-4431 150 9 where where WRB adeshpande3-github-io-4431 150 10 stacking stack VBG adeshpande3-github-io-4431 150 11 more more JJR adeshpande3-github-io-4431 150 12 layers layer NNS adeshpande3-github-io-4431 150 13 on on IN adeshpande3-github-io-4431 150 14 top top NN adeshpande3-github-io-4431 150 15 of of IN adeshpande3-github-io-4431 150 16 each each DT adeshpande3-github-io-4431 150 17 other other JJ adeshpande3-github-io-4431 150 18 is be VBZ adeshpande3-github-io-4431 150 19 n’t not RB adeshpande3-github-io-4431 150 20 going go VBG adeshpande3-github-io-4431 150 21 to to TO adeshpande3-github-io-4431 150 22 result result VB adeshpande3-github-io-4431 150 23 in in IN adeshpande3-github-io-4431 150 24 a a DT adeshpande3-github-io-4431 150 25 substantial substantial JJ adeshpande3-github-io-4431 150 26 performance performance NN adeshpande3-github-io-4431 150 27 boost boost NN adeshpande3-github-io-4431 150 28 . . . adeshpande3-github-io-4431 151 1 There there EX adeshpande3-github-io-4431 151 2 would would MD adeshpande3-github-io-4431 151 3 definitely definitely RB adeshpande3-github-io-4431 151 4 have have VB adeshpande3-github-io-4431 151 5 to to TO adeshpande3-github-io-4431 151 6 be be VB adeshpande3-github-io-4431 151 7 creative creative JJ adeshpande3-github-io-4431 151 8 new new JJ adeshpande3-github-io-4431 151 9 architectures architecture NNS adeshpande3-github-io-4431 151 10 like like IN adeshpande3-github-io-4431 151 11 we -PRON- PRP adeshpande3-github-io-4431 151 12 ’ve have VB adeshpande3-github-io-4431 151 13 seen see VBN adeshpande3-github-io-4431 151 14 the the DT adeshpande3-github-io-4431 151 15 last last JJ adeshpande3-github-io-4431 151 16 2 2 CD adeshpande3-github-io-4431 151 17 years year NNS adeshpande3-github-io-4431 151 18 . . . adeshpande3-github-io-4431 152 1 On on IN adeshpande3-github-io-4431 152 2 September September NNP adeshpande3-github-io-4431 152 3 16th 16th NN adeshpande3-github-io-4431 152 4 , , , adeshpande3-github-io-4431 152 5 the the DT adeshpande3-github-io-4431 152 6 results result NNS adeshpande3-github-io-4431 152 7 for for IN adeshpande3-github-io-4431 152 8 this this DT adeshpande3-github-io-4431 152 9 year year NN adeshpande3-github-io-4431 152 10 ’s ’s POS adeshpande3-github-io-4431 152 11 competition competition NN adeshpande3-github-io-4431 152 12 will will MD adeshpande3-github-io-4431 152 13 be be VB adeshpande3-github-io-4431 152 14 released release VBN adeshpande3-github-io-4431 152 15 . . . adeshpande3-github-io-4431 153 1 Mark mark VB adeshpande3-github-io-4431 153 2 your -PRON- PRP$ adeshpande3-github-io-4431 153 3 calendar calendar NN adeshpande3-github-io-4431 153 4 . . . adeshpande3-github-io-4431 154 1 Bonus bonus NN adeshpande3-github-io-4431 154 2 : : : adeshpande3-github-io-4431 154 3 ResNets ResNets NNP adeshpande3-github-io-4431 154 4 inside inside RB adeshpande3-github-io-4431 154 5 of of IN adeshpande3-github-io-4431 154 6 ResNets ResNets NNP adeshpande3-github-io-4431 154 7 . . . adeshpande3-github-io-4431 155 1 Yeah yeah UH adeshpande3-github-io-4431 155 2 . . . adeshpande3-github-io-4431 156 1 I -PRON- PRP adeshpande3-github-io-4431 156 2 went go VBD adeshpande3-github-io-4431 156 3 there there RB adeshpande3-github-io-4431 156 4 . . . adeshpande3-github-io-4431 157 1 Region region NN adeshpande3-github-io-4431 157 2 Based base VBN adeshpande3-github-io-4431 157 3 CNNs cnn NNS adeshpande3-github-io-4431 157 4 ( ( -LRB- adeshpande3-github-io-4431 157 5 R R NNP adeshpande3-github-io-4431 157 6 - - HYPH adeshpande3-github-io-4431 157 7 CNN CNN NNP adeshpande3-github-io-4431 157 8 - - HYPH adeshpande3-github-io-4431 157 9 2013 2013 CD adeshpande3-github-io-4431 157 10 , , , adeshpande3-github-io-4431 157 11 Fast Fast NNP adeshpande3-github-io-4431 157 12 R R NNP adeshpande3-github-io-4431 157 13 - - HYPH adeshpande3-github-io-4431 157 14 CNN CNN NNP adeshpande3-github-io-4431 157 15 - - HYPH adeshpande3-github-io-4431 157 16 2015 2015 CD adeshpande3-github-io-4431 157 17 , , , adeshpande3-github-io-4431 157 18 Faster Faster NNP adeshpande3-github-io-4431 157 19 R R NNP adeshpande3-github-io-4431 157 20 - - HYPH adeshpande3-github-io-4431 157 21 CNN CNN NNP adeshpande3-github-io-4431 157 22 - - HYPH adeshpande3-github-io-4431 157 23 2015 2015 CD adeshpande3-github-io-4431 157 24 ) ) -RRB- adeshpande3-github-io-4431 157 25                                 _SP adeshpande3-github-io-4431 157 26 Some some DT adeshpande3-github-io-4431 157 27 may may MD adeshpande3-github-io-4431 157 28 argue argue VB adeshpande3-github-io-4431 157 29 that that IN adeshpande3-github-io-4431 157 30 the the DT adeshpande3-github-io-4431 157 31 advent advent NN adeshpande3-github-io-4431 157 32 of of IN adeshpande3-github-io-4431 157 33 R r NN adeshpande3-github-io-4431 157 34 - - HYPH adeshpande3-github-io-4431 157 35 CNNs cnn NNS adeshpande3-github-io-4431 157 36 has have VBZ adeshpande3-github-io-4431 157 37 been be VBN adeshpande3-github-io-4431 157 38 more more RBR adeshpande3-github-io-4431 157 39 impactful impactful JJ adeshpande3-github-io-4431 157 40 that that IN adeshpande3-github-io-4431 157 41 any any DT adeshpande3-github-io-4431 157 42 of of IN adeshpande3-github-io-4431 157 43 the the DT adeshpande3-github-io-4431 157 44 previous previous JJ adeshpande3-github-io-4431 157 45 papers paper NNS adeshpande3-github-io-4431 157 46 on on IN adeshpande3-github-io-4431 157 47 new new JJ adeshpande3-github-io-4431 157 48 network network NN adeshpande3-github-io-4431 157 49 architectures architecture NNS adeshpande3-github-io-4431 157 50 . . . adeshpande3-github-io-4431 158 1 With with IN adeshpande3-github-io-4431 158 2 the the DT adeshpande3-github-io-4431 158 3 first first JJ adeshpande3-github-io-4431 158 4 R R NNP adeshpande3-github-io-4431 158 5 - - HYPH adeshpande3-github-io-4431 158 6 CNN CNN NNP adeshpande3-github-io-4431 158 7 paper paper NN adeshpande3-github-io-4431 158 8 being being NN adeshpande3-github-io-4431 158 9 cited cite VBN adeshpande3-github-io-4431 158 10 over over IN adeshpande3-github-io-4431 158 11 1600 1600 CD adeshpande3-github-io-4431 158 12 times time NNS adeshpande3-github-io-4431 158 13 , , , adeshpande3-github-io-4431 158 14 Ross Ross NNP adeshpande3-github-io-4431 158 15 Girshick Girshick NNP adeshpande3-github-io-4431 158 16 and and CC adeshpande3-github-io-4431 158 17 his -PRON- PRP$ adeshpande3-github-io-4431 158 18 group group NN adeshpande3-github-io-4431 158 19 at at IN adeshpande3-github-io-4431 158 20 UC UC NNP adeshpande3-github-io-4431 158 21 Berkeley Berkeley NNP adeshpande3-github-io-4431 158 22 created create VBD adeshpande3-github-io-4431 158 23 one one CD adeshpande3-github-io-4431 158 24 of of IN adeshpande3-github-io-4431 158 25 the the DT adeshpande3-github-io-4431 158 26 most most RBS adeshpande3-github-io-4431 158 27 impactful impactful JJ adeshpande3-github-io-4431 158 28 advancements advancement NNS adeshpande3-github-io-4431 158 29 in in IN adeshpande3-github-io-4431 158 30 computer computer NN adeshpande3-github-io-4431 158 31 vision vision NN adeshpande3-github-io-4431 158 32 . . . adeshpande3-github-io-4431 159 1 As as IN adeshpande3-github-io-4431 159 2 evident evident JJ adeshpande3-github-io-4431 159 3 by by IN adeshpande3-github-io-4431 159 4 their -PRON- PRP$ adeshpande3-github-io-4431 159 5 titles title NNS adeshpande3-github-io-4431 159 6 , , , adeshpande3-github-io-4431 159 7 Fast Fast NNP adeshpande3-github-io-4431 159 8 R R NNP adeshpande3-github-io-4431 159 9 - - HYPH adeshpande3-github-io-4431 159 10 CNN CNN NNP adeshpande3-github-io-4431 159 11 and and CC adeshpande3-github-io-4431 159 12 Faster Faster NNP adeshpande3-github-io-4431 159 13 R R NNP adeshpande3-github-io-4431 159 14 - - HYPH adeshpande3-github-io-4431 159 15 CNN CNN NNP adeshpande3-github-io-4431 159 16 worked work VBD adeshpande3-github-io-4431 159 17 to to TO adeshpande3-github-io-4431 159 18 make make VB adeshpande3-github-io-4431 159 19 the the DT adeshpande3-github-io-4431 159 20 model model NN adeshpande3-github-io-4431 159 21 faster fast RBR adeshpande3-github-io-4431 159 22 and and CC adeshpande3-github-io-4431 159 23 better well RBR adeshpande3-github-io-4431 159 24 suited suit VBN adeshpande3-github-io-4431 159 25 for for IN adeshpande3-github-io-4431 159 26 modern modern JJ adeshpande3-github-io-4431 159 27 object object NN adeshpande3-github-io-4431 159 28 detection detection NN adeshpande3-github-io-4431 159 29 tasks task NNS adeshpande3-github-io-4431 159 30 . . . adeshpande3-github-io-4431 160 1 The the DT adeshpande3-github-io-4431 160 2 purpose purpose NN adeshpande3-github-io-4431 160 3 of of IN adeshpande3-github-io-4431 160 4 R r NN adeshpande3-github-io-4431 160 5 - - HYPH adeshpande3-github-io-4431 160 6 CNNs cnn NNS adeshpande3-github-io-4431 160 7 is be VBZ adeshpande3-github-io-4431 160 8 to to TO adeshpande3-github-io-4431 160 9 solve solve VB adeshpande3-github-io-4431 160 10 the the DT adeshpande3-github-io-4431 160 11 problem problem NN adeshpande3-github-io-4431 160 12 of of IN adeshpande3-github-io-4431 160 13 object object NN adeshpande3-github-io-4431 160 14 detection detection NN adeshpande3-github-io-4431 160 15 . . . adeshpande3-github-io-4431 161 1 Given give VBN adeshpande3-github-io-4431 161 2 a a DT adeshpande3-github-io-4431 161 3 certain certain JJ adeshpande3-github-io-4431 161 4 image image NN adeshpande3-github-io-4431 161 5 , , , adeshpande3-github-io-4431 161 6 we -PRON- PRP adeshpande3-github-io-4431 161 7 want want VBP adeshpande3-github-io-4431 161 8 to to TO adeshpande3-github-io-4431 161 9 be be VB adeshpande3-github-io-4431 161 10 able able JJ adeshpande3-github-io-4431 161 11 to to TO adeshpande3-github-io-4431 161 12 draw draw VB adeshpande3-github-io-4431 161 13 bounding bounding NN adeshpande3-github-io-4431 161 14 boxes box NNS adeshpande3-github-io-4431 161 15 over over IN adeshpande3-github-io-4431 161 16 all all DT adeshpande3-github-io-4431 161 17 of of IN adeshpande3-github-io-4431 161 18 the the DT adeshpande3-github-io-4431 161 19 objects object NNS adeshpande3-github-io-4431 161 20 . . . adeshpande3-github-io-4431 162 1 The the DT adeshpande3-github-io-4431 162 2 process process NN adeshpande3-github-io-4431 162 3 can can MD adeshpande3-github-io-4431 162 4 be be VB adeshpande3-github-io-4431 162 5 split split VBN adeshpande3-github-io-4431 162 6 into into IN adeshpande3-github-io-4431 162 7 two two CD adeshpande3-github-io-4431 162 8 general general JJ adeshpande3-github-io-4431 162 9 components component NNS adeshpande3-github-io-4431 162 10 , , , adeshpande3-github-io-4431 162 11 the the DT adeshpande3-github-io-4431 162 12 region region NN adeshpande3-github-io-4431 162 13 proposal proposal NN adeshpande3-github-io-4431 162 14 step step NN adeshpande3-github-io-4431 162 15 and and CC adeshpande3-github-io-4431 162 16 the the DT adeshpande3-github-io-4431 162 17 classification classification NN adeshpande3-github-io-4431 162 18 step step NN adeshpande3-github-io-4431 162 19 . . . adeshpande3-github-io-4431 163 1 The the DT adeshpande3-github-io-4431 163 2 authors author NNS adeshpande3-github-io-4431 163 3 note note VBP adeshpande3-github-io-4431 163 4 that that IN adeshpande3-github-io-4431 163 5 any any DT adeshpande3-github-io-4431 163 6 class class NN adeshpande3-github-io-4431 163 7 agnostic agnostic JJ adeshpande3-github-io-4431 163 8 region region NN adeshpande3-github-io-4431 163 9 proposal proposal NN adeshpande3-github-io-4431 163 10 method method NNP adeshpande3-github-io-4431 163 11 should should MD adeshpande3-github-io-4431 163 12 fit fit VB adeshpande3-github-io-4431 163 13 . . . adeshpande3-github-io-4431 164 1 Selective Selective NNP adeshpande3-github-io-4431 164 2 Search Search NNP adeshpande3-github-io-4431 164 3 is be VBZ adeshpande3-github-io-4431 164 4 used use VBN adeshpande3-github-io-4431 164 5 in in IN adeshpande3-github-io-4431 164 6 particular particular JJ adeshpande3-github-io-4431 164 7 for for IN adeshpande3-github-io-4431 164 8 RCNN RCNN NNP adeshpande3-github-io-4431 164 9 . . . adeshpande3-github-io-4431 165 1 Selective Selective NNP adeshpande3-github-io-4431 165 2 Search Search NNP adeshpande3-github-io-4431 165 3 performs perform VBZ adeshpande3-github-io-4431 165 4 the the DT adeshpande3-github-io-4431 165 5 function function NN adeshpande3-github-io-4431 165 6 of of IN adeshpande3-github-io-4431 165 7 generating generate VBG adeshpande3-github-io-4431 165 8 2000 2000 CD adeshpande3-github-io-4431 165 9 different different JJ adeshpande3-github-io-4431 165 10 regions region NNS adeshpande3-github-io-4431 165 11 that that WDT adeshpande3-github-io-4431 165 12 have have VBP adeshpande3-github-io-4431 165 13 the the DT adeshpande3-github-io-4431 165 14 highest high JJS adeshpande3-github-io-4431 165 15 probability probability NN adeshpande3-github-io-4431 165 16 of of IN adeshpande3-github-io-4431 165 17 containing contain VBG adeshpande3-github-io-4431 165 18 an an DT adeshpande3-github-io-4431 165 19 object object NN adeshpande3-github-io-4431 165 20 . . . adeshpande3-github-io-4431 166 1 After after IN adeshpande3-github-io-4431 166 2 we -PRON- PRP adeshpande3-github-io-4431 166 3 ’ve have VB adeshpande3-github-io-4431 166 4 come come VBN adeshpande3-github-io-4431 166 5 up up RP adeshpande3-github-io-4431 166 6 with with IN adeshpande3-github-io-4431 166 7 a a DT adeshpande3-github-io-4431 166 8 set set NN adeshpande3-github-io-4431 166 9 of of IN adeshpande3-github-io-4431 166 10 region region NN adeshpande3-github-io-4431 166 11 proposals proposal NNS adeshpande3-github-io-4431 166 12 , , , adeshpande3-github-io-4431 166 13 these these DT adeshpande3-github-io-4431 166 14 proposals proposal NNS adeshpande3-github-io-4431 166 15 are be VBP adeshpande3-github-io-4431 166 16 then then RB adeshpande3-github-io-4431 166 17 “ " `` adeshpande3-github-io-4431 166 18 warped warp VBN adeshpande3-github-io-4431 166 19 ” " '' adeshpande3-github-io-4431 166 20 into into IN adeshpande3-github-io-4431 166 21 an an DT adeshpande3-github-io-4431 166 22 image image NN adeshpande3-github-io-4431 166 23 size size NN adeshpande3-github-io-4431 166 24 that that WDT adeshpande3-github-io-4431 166 25 can can MD adeshpande3-github-io-4431 166 26 be be VB adeshpande3-github-io-4431 166 27 fed feed VBN adeshpande3-github-io-4431 166 28 into into IN adeshpande3-github-io-4431 166 29 a a DT adeshpande3-github-io-4431 166 30 trained train VBN adeshpande3-github-io-4431 166 31 CNN CNN NNP adeshpande3-github-io-4431 166 32 ( ( -LRB- adeshpande3-github-io-4431 166 33 AlexNet AlexNet NNP adeshpande3-github-io-4431 166 34 in in IN adeshpande3-github-io-4431 166 35 this this DT adeshpande3-github-io-4431 166 36 case case NN adeshpande3-github-io-4431 166 37 ) ) -RRB- adeshpande3-github-io-4431 166 38 that that WDT adeshpande3-github-io-4431 166 39 extracts extract VBZ adeshpande3-github-io-4431 166 40 a a DT adeshpande3-github-io-4431 166 41 feature feature NN adeshpande3-github-io-4431 166 42 vector vector NN adeshpande3-github-io-4431 166 43 for for IN adeshpande3-github-io-4431 166 44 each each DT adeshpande3-github-io-4431 166 45 region region NN adeshpande3-github-io-4431 166 46 . . . adeshpande3-github-io-4431 167 1 This this DT adeshpande3-github-io-4431 167 2 vector vector NN adeshpande3-github-io-4431 167 3 is be VBZ adeshpande3-github-io-4431 167 4 then then RB adeshpande3-github-io-4431 167 5 used use VBN adeshpande3-github-io-4431 167 6 as as IN adeshpande3-github-io-4431 167 7 the the DT adeshpande3-github-io-4431 167 8 input input NN adeshpande3-github-io-4431 167 9 to to IN adeshpande3-github-io-4431 167 10 a a DT adeshpande3-github-io-4431 167 11 set set NN adeshpande3-github-io-4431 167 12 of of IN adeshpande3-github-io-4431 167 13 linear linear JJ adeshpande3-github-io-4431 167 14 SVMs svm NNS adeshpande3-github-io-4431 167 15 that that WDT adeshpande3-github-io-4431 167 16 are be VBP adeshpande3-github-io-4431 167 17 trained train VBN adeshpande3-github-io-4431 167 18 for for IN adeshpande3-github-io-4431 167 19 each each DT adeshpande3-github-io-4431 167 20 class class NN adeshpande3-github-io-4431 167 21 and and CC adeshpande3-github-io-4431 167 22 output output NN adeshpande3-github-io-4431 167 23 a a DT adeshpande3-github-io-4431 167 24 classification classification NN adeshpande3-github-io-4431 167 25 . . . adeshpande3-github-io-4431 168 1 The the DT adeshpande3-github-io-4431 168 2 vector vector NN adeshpande3-github-io-4431 168 3 also also RB adeshpande3-github-io-4431 168 4 gets get VBZ adeshpande3-github-io-4431 168 5 fed feed VBN adeshpande3-github-io-4431 168 6 into into IN adeshpande3-github-io-4431 168 7 a a DT adeshpande3-github-io-4431 168 8 bounding bound VBG adeshpande3-github-io-4431 168 9 box box NN adeshpande3-github-io-4431 168 10 regressor regressor NN adeshpande3-github-io-4431 168 11 to to TO adeshpande3-github-io-4431 168 12 obtain obtain VB adeshpande3-github-io-4431 168 13 the the DT adeshpande3-github-io-4431 168 14 most most RBS adeshpande3-github-io-4431 168 15 accurate accurate JJ adeshpande3-github-io-4431 168 16 coordinates coordinate NNS adeshpande3-github-io-4431 168 17 . . . adeshpande3-github-io-4431 169 1 Non non JJ adeshpande3-github-io-4431 169 2 - - JJ adeshpande3-github-io-4431 169 3 maxima maxima JJ adeshpande3-github-io-4431 169 4 suppression suppression NN adeshpande3-github-io-4431 169 5 is be VBZ adeshpande3-github-io-4431 169 6 then then RB adeshpande3-github-io-4431 169 7 used use VBN adeshpande3-github-io-4431 169 8 to to TO adeshpande3-github-io-4431 169 9 suppress suppress VB adeshpande3-github-io-4431 169 10 bounding bounding NN adeshpande3-github-io-4431 169 11 boxes box NNS adeshpande3-github-io-4431 169 12 that that WDT adeshpande3-github-io-4431 169 13 have have VBP adeshpande3-github-io-4431 169 14 a a DT adeshpande3-github-io-4431 169 15 significant significant JJ adeshpande3-github-io-4431 169 16 overlap overlap NN adeshpande3-github-io-4431 169 17 with with IN adeshpande3-github-io-4431 169 18 each each DT adeshpande3-github-io-4431 169 19 other other JJ adeshpande3-github-io-4431 169 20 . . . adeshpande3-github-io-4431 170 1 Fast fast JJ adeshpande3-github-io-4431 170 2 R R NNP adeshpande3-github-io-4431 170 3 - - HYPH adeshpande3-github-io-4431 170 4 CNN CNN NNP adeshpande3-github-io-4431 170 5                                 _SP adeshpande3-github-io-4431 170 6 Improvements improvement NNS adeshpande3-github-io-4431 170 7 were be VBD adeshpande3-github-io-4431 170 8 made make VBN adeshpande3-github-io-4431 170 9 to to IN adeshpande3-github-io-4431 170 10 the the DT adeshpande3-github-io-4431 170 11 original original JJ adeshpande3-github-io-4431 170 12 model model NN adeshpande3-github-io-4431 170 13 because because IN adeshpande3-github-io-4431 170 14 of of IN adeshpande3-github-io-4431 170 15 3 3 CD adeshpande3-github-io-4431 170 16 main main JJ adeshpande3-github-io-4431 170 17 problems problem NNS adeshpande3-github-io-4431 170 18 . . . adeshpande3-github-io-4431 171 1 Training training NN adeshpande3-github-io-4431 171 2 took take VBD adeshpande3-github-io-4431 171 3 multiple multiple JJ adeshpande3-github-io-4431 171 4 stages stage NNS adeshpande3-github-io-4431 171 5 ( ( -LRB- adeshpande3-github-io-4431 171 6 ConvNets ConvNets NNP adeshpande3-github-io-4431 171 7 to to TO adeshpande3-github-io-4431 171 8 SVMs svm NNS adeshpande3-github-io-4431 171 9 to to IN adeshpande3-github-io-4431 171 10 bounding bound VBG adeshpande3-github-io-4431 171 11 box box NN adeshpande3-github-io-4431 171 12 regressors regressor NNS adeshpande3-github-io-4431 171 13 ) ) -RRB- adeshpande3-github-io-4431 171 14 , , , adeshpande3-github-io-4431 171 15 was be VBD adeshpande3-github-io-4431 171 16 computationally computationally RB adeshpande3-github-io-4431 171 17 expensive expensive JJ adeshpande3-github-io-4431 171 18 , , , adeshpande3-github-io-4431 171 19 and and CC adeshpande3-github-io-4431 171 20 was be VBD adeshpande3-github-io-4431 171 21 extremely extremely RB adeshpande3-github-io-4431 171 22 slow slow JJ adeshpande3-github-io-4431 171 23 ( ( -LRB- adeshpande3-github-io-4431 171 24 RCNN RCNN NNP adeshpande3-github-io-4431 171 25 took take VBD adeshpande3-github-io-4431 171 26 53 53 CD adeshpande3-github-io-4431 171 27 seconds second NNS adeshpande3-github-io-4431 171 28 per per IN adeshpande3-github-io-4431 171 29 image image NN adeshpande3-github-io-4431 171 30 ) ) -RRB- adeshpande3-github-io-4431 171 31 . . . adeshpande3-github-io-4431 172 1 Fast fast JJ adeshpande3-github-io-4431 172 2 R R NNP adeshpande3-github-io-4431 172 3 - - HYPH adeshpande3-github-io-4431 172 4 CNN CNN NNP adeshpande3-github-io-4431 172 5 was be VBD adeshpande3-github-io-4431 172 6 able able JJ adeshpande3-github-io-4431 172 7 to to TO adeshpande3-github-io-4431 172 8 solve solve VB adeshpande3-github-io-4431 172 9 the the DT adeshpande3-github-io-4431 172 10 problem problem NN adeshpande3-github-io-4431 172 11 of of IN adeshpande3-github-io-4431 172 12 speed speed NN adeshpande3-github-io-4431 172 13 by by IN adeshpande3-github-io-4431 172 14 basically basically RB adeshpande3-github-io-4431 172 15 sharing share VBG adeshpande3-github-io-4431 172 16 computation computation NN adeshpande3-github-io-4431 172 17 of of IN adeshpande3-github-io-4431 172 18 the the DT adeshpande3-github-io-4431 172 19 conv conv NN adeshpande3-github-io-4431 172 20 layers layer NNS adeshpande3-github-io-4431 172 21 between between IN adeshpande3-github-io-4431 172 22 different different JJ adeshpande3-github-io-4431 172 23 proposals proposal NNS adeshpande3-github-io-4431 172 24 and and CC adeshpande3-github-io-4431 172 25 swapping swap VBG adeshpande3-github-io-4431 172 26 the the DT adeshpande3-github-io-4431 172 27 order order NN adeshpande3-github-io-4431 172 28 of of IN adeshpande3-github-io-4431 172 29 generating generate VBG adeshpande3-github-io-4431 172 30 region region NN adeshpande3-github-io-4431 172 31 proposals proposal NNS adeshpande3-github-io-4431 172 32 and and CC adeshpande3-github-io-4431 172 33 running run VBG adeshpande3-github-io-4431 172 34 the the DT adeshpande3-github-io-4431 172 35 CNN CNN NNP adeshpande3-github-io-4431 172 36 . . . adeshpande3-github-io-4431 173 1 In in IN adeshpande3-github-io-4431 173 2 this this DT adeshpande3-github-io-4431 173 3 model model NN adeshpande3-github-io-4431 173 4 , , , adeshpande3-github-io-4431 173 5 the the DT adeshpande3-github-io-4431 173 6 image image NN adeshpande3-github-io-4431 173 7 is be VBZ adeshpande3-github-io-4431 173 8 first first RB adeshpande3-github-io-4431 173 9 fed feed VBN adeshpande3-github-io-4431 173 10 through through IN adeshpande3-github-io-4431 173 11 a a DT adeshpande3-github-io-4431 173 12 ConvNet ConvNet NNP adeshpande3-github-io-4431 173 13 , , , adeshpande3-github-io-4431 173 14 features feature NNS adeshpande3-github-io-4431 173 15 of of IN adeshpande3-github-io-4431 173 16 the the DT adeshpande3-github-io-4431 173 17 region region NN adeshpande3-github-io-4431 173 18 proposals proposal NNS adeshpande3-github-io-4431 173 19 are be VBP adeshpande3-github-io-4431 173 20 obtained obtain VBN adeshpande3-github-io-4431 173 21 from from IN adeshpande3-github-io-4431 173 22 the the DT adeshpande3-github-io-4431 173 23 last last JJ adeshpande3-github-io-4431 173 24 feature feature NN adeshpande3-github-io-4431 173 25 map map NN adeshpande3-github-io-4431 173 26 of of IN adeshpande3-github-io-4431 173 27 the the DT adeshpande3-github-io-4431 173 28 ConvNet ConvNet NNP adeshpande3-github-io-4431 173 29 ( ( -LRB- adeshpande3-github-io-4431 173 30 check check NN adeshpande3-github-io-4431 173 31 section section NN adeshpande3-github-io-4431 173 32 2.1 2.1 CD adeshpande3-github-io-4431 173 33 of of IN adeshpande3-github-io-4431 173 34 the the DT adeshpande3-github-io-4431 173 35 paper paper NN adeshpande3-github-io-4431 173 36 for for IN adeshpande3-github-io-4431 173 37 more more JJR adeshpande3-github-io-4431 173 38 details detail NNS adeshpande3-github-io-4431 173 39 ) ) -RRB- adeshpande3-github-io-4431 173 40 , , , adeshpande3-github-io-4431 173 41 and and CC adeshpande3-github-io-4431 173 42 lastly lastly RB adeshpande3-github-io-4431 173 43 we -PRON- PRP adeshpande3-github-io-4431 173 44 have have VBP adeshpande3-github-io-4431 173 45 our -PRON- PRP$ adeshpande3-github-io-4431 173 46 fully fully RB adeshpande3-github-io-4431 173 47 connected connect VBN adeshpande3-github-io-4431 173 48 layers layer NNS adeshpande3-github-io-4431 173 49 as as RB adeshpande3-github-io-4431 173 50 well well RB adeshpande3-github-io-4431 173 51 as as IN adeshpande3-github-io-4431 173 52 our -PRON- PRP$ adeshpande3-github-io-4431 173 53 regression regression NN adeshpande3-github-io-4431 173 54 and and CC adeshpande3-github-io-4431 173 55 classification classification NN adeshpande3-github-io-4431 173 56 heads head NNS adeshpande3-github-io-4431 173 57 . . . adeshpande3-github-io-4431 174 1 Faster fast JJR adeshpande3-github-io-4431 174 2 R R NNP adeshpande3-github-io-4431 174 3 - - HYPH adeshpande3-github-io-4431 174 4 CNN CNN NNP adeshpande3-github-io-4431 174 5                                 _SP adeshpande3-github-io-4431 174 6 Faster Faster NNP adeshpande3-github-io-4431 174 7 R R NNP adeshpande3-github-io-4431 174 8 - - HYPH adeshpande3-github-io-4431 174 9 CNN CNN NNP adeshpande3-github-io-4431 174 10 works work VBZ adeshpande3-github-io-4431 174 11 to to TO adeshpande3-github-io-4431 174 12 combat combat VB adeshpande3-github-io-4431 174 13 the the DT adeshpande3-github-io-4431 174 14 somewhat somewhat RB adeshpande3-github-io-4431 174 15 complex complex JJ adeshpande3-github-io-4431 174 16 training training NN adeshpande3-github-io-4431 174 17 pipeline pipeline NN adeshpande3-github-io-4431 174 18 that that IN adeshpande3-github-io-4431 174 19 both both CC adeshpande3-github-io-4431 174 20 R R NNP adeshpande3-github-io-4431 174 21 - - HYPH adeshpande3-github-io-4431 174 22 CNN CNN NNP adeshpande3-github-io-4431 174 23 and and CC adeshpande3-github-io-4431 174 24 Fast Fast NNP adeshpande3-github-io-4431 174 25 R R NNP adeshpande3-github-io-4431 174 26 - - HYPH adeshpande3-github-io-4431 174 27 CNN CNN NNP adeshpande3-github-io-4431 174 28 exhibited exhibit VBN adeshpande3-github-io-4431 174 29 . . . adeshpande3-github-io-4431 175 1 The the DT adeshpande3-github-io-4431 175 2 authors author NNS adeshpande3-github-io-4431 175 3 insert insert VBP adeshpande3-github-io-4431 175 4 a a DT adeshpande3-github-io-4431 175 5 region region NN adeshpande3-github-io-4431 175 6 proposal proposal NN adeshpande3-github-io-4431 175 7 network network NN adeshpande3-github-io-4431 175 8 ( ( -LRB- adeshpande3-github-io-4431 175 9 RPN RPN NNP adeshpande3-github-io-4431 175 10 ) ) -RRB- adeshpande3-github-io-4431 175 11 after after IN adeshpande3-github-io-4431 175 12 the the DT adeshpande3-github-io-4431 175 13 last last JJ adeshpande3-github-io-4431 175 14 convolutional convolutional JJ adeshpande3-github-io-4431 175 15 layer layer NN adeshpande3-github-io-4431 175 16 . . . adeshpande3-github-io-4431 176 1 This this DT adeshpande3-github-io-4431 176 2 network network NN adeshpande3-github-io-4431 176 3 is be VBZ adeshpande3-github-io-4431 176 4 able able JJ adeshpande3-github-io-4431 176 5 to to TO adeshpande3-github-io-4431 176 6 just just RB adeshpande3-github-io-4431 176 7 look look VB adeshpande3-github-io-4431 176 8 at at IN adeshpande3-github-io-4431 176 9 the the DT adeshpande3-github-io-4431 176 10 last last JJ adeshpande3-github-io-4431 176 11 convolutional convolutional JJ adeshpande3-github-io-4431 176 12 feature feature NN adeshpande3-github-io-4431 176 13 map map NN adeshpande3-github-io-4431 176 14 and and CC adeshpande3-github-io-4431 176 15 produce produce VB adeshpande3-github-io-4431 176 16 region region NN adeshpande3-github-io-4431 176 17 proposals proposal NNS adeshpande3-github-io-4431 176 18 from from IN adeshpande3-github-io-4431 176 19 that that DT adeshpande3-github-io-4431 176 20 . . . adeshpande3-github-io-4431 177 1 From from IN adeshpande3-github-io-4431 177 2 that that DT adeshpande3-github-io-4431 177 3 stage stage NN adeshpande3-github-io-4431 177 4 , , , adeshpande3-github-io-4431 177 5 the the DT adeshpande3-github-io-4431 177 6 same same JJ adeshpande3-github-io-4431 177 7 pipeline pipeline NN adeshpande3-github-io-4431 177 8 as as IN adeshpande3-github-io-4431 177 9 R R NNP adeshpande3-github-io-4431 177 10 - - HYPH adeshpande3-github-io-4431 177 11 CNN CNN NNP adeshpande3-github-io-4431 177 12 is be VBZ adeshpande3-github-io-4431 177 13 used use VBN adeshpande3-github-io-4431 177 14 ( ( -LRB- adeshpande3-github-io-4431 177 15 ROI ROI NNP adeshpande3-github-io-4431 177 16 pooling pooling NN adeshpande3-github-io-4431 177 17 , , , adeshpande3-github-io-4431 177 18 FC fc NN adeshpande3-github-io-4431 177 19 , , , adeshpande3-github-io-4431 177 20 and and CC adeshpande3-github-io-4431 177 21 then then RB adeshpande3-github-io-4431 177 22 classification classification NN adeshpande3-github-io-4431 177 23 and and CC adeshpande3-github-io-4431 177 24 regression regression NN adeshpande3-github-io-4431 177 25 heads head NNS adeshpande3-github-io-4431 177 26 ) ) -RRB- adeshpande3-github-io-4431 177 27 . . . adeshpande3-github-io-4431 178 1 Why why WRB adeshpande3-github-io-4431 178 2 It -PRON- PRP adeshpande3-github-io-4431 178 3 ’s ’ VBZ adeshpande3-github-io-4431 178 4 Important important JJ adeshpande3-github-io-4431 178 5                                 _SP adeshpande3-github-io-4431 178 6 Being be VBG adeshpande3-github-io-4431 178 7 able able JJ adeshpande3-github-io-4431 178 8 to to TO adeshpande3-github-io-4431 178 9 determine determine VB adeshpande3-github-io-4431 178 10 that that IN adeshpande3-github-io-4431 178 11 a a DT adeshpande3-github-io-4431 178 12 specific specific JJ adeshpande3-github-io-4431 178 13 object object NN adeshpande3-github-io-4431 178 14 is be VBZ adeshpande3-github-io-4431 178 15 in in IN adeshpande3-github-io-4431 178 16 an an DT adeshpande3-github-io-4431 178 17 image image NN adeshpande3-github-io-4431 178 18 is be VBZ adeshpande3-github-io-4431 178 19 one one CD adeshpande3-github-io-4431 178 20 thing thing NN adeshpande3-github-io-4431 178 21 , , , adeshpande3-github-io-4431 178 22 but but CC adeshpande3-github-io-4431 178 23 being be VBG adeshpande3-github-io-4431 178 24 able able JJ adeshpande3-github-io-4431 178 25 to to TO adeshpande3-github-io-4431 178 26 determine determine VB adeshpande3-github-io-4431 178 27 that that DT adeshpande3-github-io-4431 178 28 object object NN adeshpande3-github-io-4431 178 29 ’s ’s POS adeshpande3-github-io-4431 178 30 exact exact JJ adeshpande3-github-io-4431 178 31 location location NN adeshpande3-github-io-4431 178 32 is be VBZ adeshpande3-github-io-4431 178 33 a a DT adeshpande3-github-io-4431 178 34 huge huge JJ adeshpande3-github-io-4431 178 35 jump jump NN adeshpande3-github-io-4431 178 36 in in IN adeshpande3-github-io-4431 178 37 knowledge knowledge NN adeshpande3-github-io-4431 178 38 for for IN adeshpande3-github-io-4431 178 39 the the DT adeshpande3-github-io-4431 178 40 computer computer NN adeshpande3-github-io-4431 178 41 . . . adeshpande3-github-io-4431 179 1 Faster fast JJR adeshpande3-github-io-4431 179 2 R R NNP adeshpande3-github-io-4431 179 3 - - HYPH adeshpande3-github-io-4431 179 4 CNN CNN NNP adeshpande3-github-io-4431 179 5 has have VBZ adeshpande3-github-io-4431 179 6 become become VBN adeshpande3-github-io-4431 179 7 the the DT adeshpande3-github-io-4431 179 8 standard standard NN adeshpande3-github-io-4431 179 9 for for IN adeshpande3-github-io-4431 179 10 object object NN adeshpande3-github-io-4431 179 11 detection detection NN adeshpande3-github-io-4431 179 12 programs program NNS adeshpande3-github-io-4431 179 13 today today NN adeshpande3-github-io-4431 179 14 . . . adeshpande3-github-io-4431 180 1 Generative Generative NNP adeshpande3-github-io-4431 180 2 Adversarial Adversarial NNP adeshpande3-github-io-4431 180 3 Networks Networks NNPS adeshpande3-github-io-4431 180 4 ( ( -LRB- adeshpande3-github-io-4431 180 5 2014 2014 CD adeshpande3-github-io-4431 180 6 ) ) -RRB- adeshpande3-github-io-4431 180 7                                 _SP adeshpande3-github-io-4431 180 8 According accord VBG adeshpande3-github-io-4431 180 9 to to IN adeshpande3-github-io-4431 180 10 Yann Yann NNP adeshpande3-github-io-4431 180 11 LeCun LeCun NNP adeshpande3-github-io-4431 180 12 , , , adeshpande3-github-io-4431 180 13 these these DT adeshpande3-github-io-4431 180 14 networks network NNS adeshpande3-github-io-4431 180 15 could could MD adeshpande3-github-io-4431 180 16 be be VB adeshpande3-github-io-4431 180 17 the the DT adeshpande3-github-io-4431 180 18 next next JJ adeshpande3-github-io-4431 180 19 big big JJ adeshpande3-github-io-4431 180 20 development development NN adeshpande3-github-io-4431 180 21 . . . adeshpande3-github-io-4431 181 1 Before before IN adeshpande3-github-io-4431 181 2 talking talk VBG adeshpande3-github-io-4431 181 3 about about IN adeshpande3-github-io-4431 181 4 this this DT adeshpande3-github-io-4431 181 5 paper paper NN adeshpande3-github-io-4431 181 6 , , , adeshpande3-github-io-4431 181 7 let let VB adeshpande3-github-io-4431 181 8 ’s -PRON- PRP adeshpande3-github-io-4431 181 9 talk talk VB adeshpande3-github-io-4431 181 10 a a DT adeshpande3-github-io-4431 181 11 little little JJ adeshpande3-github-io-4431 181 12 about about IN adeshpande3-github-io-4431 181 13 adversarial adversarial JJ adeshpande3-github-io-4431 181 14 examples example NNS adeshpande3-github-io-4431 181 15 . . . adeshpande3-github-io-4431 182 1 For for IN adeshpande3-github-io-4431 182 2 example example NN adeshpande3-github-io-4431 182 3 , , , adeshpande3-github-io-4431 182 4 let let VB adeshpande3-github-io-4431 182 5 ’s -PRON- PRP adeshpande3-github-io-4431 182 6 consider consider VB adeshpande3-github-io-4431 182 7 a a DT adeshpande3-github-io-4431 182 8 trained train VBN adeshpande3-github-io-4431 182 9 CNN CNN NNP adeshpande3-github-io-4431 182 10 that that WDT adeshpande3-github-io-4431 182 11 works work VBZ adeshpande3-github-io-4431 182 12 well well RB adeshpande3-github-io-4431 182 13 on on IN adeshpande3-github-io-4431 182 14 ImageNet ImageNet NNP adeshpande3-github-io-4431 182 15 data datum NNS adeshpande3-github-io-4431 182 16 . . . adeshpande3-github-io-4431 183 1 Let let VB adeshpande3-github-io-4431 183 2 ’s -PRON- PRP adeshpande3-github-io-4431 183 3 take take VB adeshpande3-github-io-4431 183 4 an an DT adeshpande3-github-io-4431 183 5 example example NN adeshpande3-github-io-4431 183 6 image image NN adeshpande3-github-io-4431 183 7 and and CC adeshpande3-github-io-4431 183 8 apply apply VB adeshpande3-github-io-4431 183 9 a a DT adeshpande3-github-io-4431 183 10 perturbation perturbation NN adeshpande3-github-io-4431 183 11 , , , adeshpande3-github-io-4431 183 12 or or CC adeshpande3-github-io-4431 183 13 a a DT adeshpande3-github-io-4431 183 14 slight slight JJ adeshpande3-github-io-4431 183 15 modification modification NN adeshpande3-github-io-4431 183 16 , , , adeshpande3-github-io-4431 183 17 so so IN adeshpande3-github-io-4431 183 18 that that IN adeshpande3-github-io-4431 183 19 the the DT adeshpande3-github-io-4431 183 20 prediction prediction NN adeshpande3-github-io-4431 183 21 error error NN adeshpande3-github-io-4431 183 22 is be VBZ adeshpande3-github-io-4431 183 23 maximized maximize VBN adeshpande3-github-io-4431 183 24 . . . adeshpande3-github-io-4431 184 1 Thus thus RB adeshpande3-github-io-4431 184 2 , , , adeshpande3-github-io-4431 184 3 the the DT adeshpande3-github-io-4431 184 4 object object NN adeshpande3-github-io-4431 184 5 category category NN adeshpande3-github-io-4431 184 6 of of IN adeshpande3-github-io-4431 184 7 the the DT adeshpande3-github-io-4431 184 8 prediction prediction NN adeshpande3-github-io-4431 184 9 changes change NNS adeshpande3-github-io-4431 184 10 , , , adeshpande3-github-io-4431 184 11 while while IN adeshpande3-github-io-4431 184 12 the the DT adeshpande3-github-io-4431 184 13 image image NN adeshpande3-github-io-4431 184 14 itself -PRON- PRP adeshpande3-github-io-4431 184 15 looks look VBZ adeshpande3-github-io-4431 184 16 the the DT adeshpande3-github-io-4431 184 17 same same JJ adeshpande3-github-io-4431 184 18 when when WRB adeshpande3-github-io-4431 184 19 compared compare VBN adeshpande3-github-io-4431 184 20 to to IN adeshpande3-github-io-4431 184 21 the the DT adeshpande3-github-io-4431 184 22 image image NN adeshpande3-github-io-4431 184 23 without without IN adeshpande3-github-io-4431 184 24 the the DT adeshpande3-github-io-4431 184 25 perturbation perturbation NN adeshpande3-github-io-4431 184 26 . . . adeshpande3-github-io-4431 185 1 From from IN adeshpande3-github-io-4431 185 2 the the DT adeshpande3-github-io-4431 185 3 highest high JJS adeshpande3-github-io-4431 185 4 level level NN adeshpande3-github-io-4431 185 5 , , , adeshpande3-github-io-4431 185 6 adversarial adversarial JJ adeshpande3-github-io-4431 185 7 examples example NNS adeshpande3-github-io-4431 185 8 are be VBP adeshpande3-github-io-4431 185 9 basically basically RB adeshpande3-github-io-4431 185 10 the the DT adeshpande3-github-io-4431 185 11 images image NNS adeshpande3-github-io-4431 185 12 that that WDT adeshpande3-github-io-4431 185 13 fool fool VBP adeshpande3-github-io-4431 185 14 ConvNets ConvNets NNP adeshpande3-github-io-4431 185 15 . . . adeshpande3-github-io-4431 186 1 Adversarial adversarial JJ adeshpande3-github-io-4431 186 2 examples example NNS adeshpande3-github-io-4431 186 3 ( ( -LRB- adeshpande3-github-io-4431 186 4 paper paper NN adeshpande3-github-io-4431 186 5 ) ) -RRB- adeshpande3-github-io-4431 186 6 definitely definitely RB adeshpande3-github-io-4431 186 7 surprised surprise VBD adeshpande3-github-io-4431 186 8 a a DT adeshpande3-github-io-4431 186 9 lot lot NN adeshpande3-github-io-4431 186 10 of of IN adeshpande3-github-io-4431 186 11 researchers researcher NNS adeshpande3-github-io-4431 186 12 and and CC adeshpande3-github-io-4431 186 13 quickly quickly RB adeshpande3-github-io-4431 186 14 became become VBD adeshpande3-github-io-4431 186 15 a a DT adeshpande3-github-io-4431 186 16 topic topic NN adeshpande3-github-io-4431 186 17 of of IN adeshpande3-github-io-4431 186 18 interest interest NN adeshpande3-github-io-4431 186 19 . . . adeshpande3-github-io-4431 187 1 Now now RB adeshpande3-github-io-4431 187 2 let let VB adeshpande3-github-io-4431 187 3 ’s -PRON- PRP adeshpande3-github-io-4431 187 4 talk talk VB adeshpande3-github-io-4431 187 5 about about IN adeshpande3-github-io-4431 187 6 the the DT adeshpande3-github-io-4431 187 7 generative generative JJ adeshpande3-github-io-4431 187 8 adversarial adversarial JJ adeshpande3-github-io-4431 187 9 networks network NNS adeshpande3-github-io-4431 187 10 . . . adeshpande3-github-io-4431 188 1 Let let VB adeshpande3-github-io-4431 188 2 ’s -PRON- PRP adeshpande3-github-io-4431 188 3 think think VB adeshpande3-github-io-4431 188 4 of of IN adeshpande3-github-io-4431 188 5 two two CD adeshpande3-github-io-4431 188 6 models model NNS adeshpande3-github-io-4431 188 7 , , , adeshpande3-github-io-4431 188 8 a a DT adeshpande3-github-io-4431 188 9 generative generative JJ adeshpande3-github-io-4431 188 10 model model NN adeshpande3-github-io-4431 188 11 and and CC adeshpande3-github-io-4431 188 12 a a DT adeshpande3-github-io-4431 188 13 discriminative discriminative JJ adeshpande3-github-io-4431 188 14 model model NN adeshpande3-github-io-4431 188 15 . . . adeshpande3-github-io-4431 189 1 The the DT adeshpande3-github-io-4431 189 2 discriminative discriminative JJ adeshpande3-github-io-4431 189 3 model model NN adeshpande3-github-io-4431 189 4 has have VBZ adeshpande3-github-io-4431 189 5 the the DT adeshpande3-github-io-4431 189 6 task task NN adeshpande3-github-io-4431 189 7 of of IN adeshpande3-github-io-4431 189 8 determining determine VBG adeshpande3-github-io-4431 189 9 whether whether IN adeshpande3-github-io-4431 189 10 a a DT adeshpande3-github-io-4431 189 11 given give VBN adeshpande3-github-io-4431 189 12 image image NN adeshpande3-github-io-4431 189 13 looks look VBZ adeshpande3-github-io-4431 189 14 natural natural JJ adeshpande3-github-io-4431 189 15 ( ( -LRB- adeshpande3-github-io-4431 189 16 an an DT adeshpande3-github-io-4431 189 17 image image NN adeshpande3-github-io-4431 189 18 from from IN adeshpande3-github-io-4431 189 19 the the DT adeshpande3-github-io-4431 189 20 dataset dataset NN adeshpande3-github-io-4431 189 21 ) ) -RRB- adeshpande3-github-io-4431 189 22 or or CC adeshpande3-github-io-4431 189 23 looks look VBZ adeshpande3-github-io-4431 189 24 like like IN adeshpande3-github-io-4431 189 25 it -PRON- PRP adeshpande3-github-io-4431 189 26 has have VBZ adeshpande3-github-io-4431 189 27 been be VBN adeshpande3-github-io-4431 189 28 artificially artificially RB adeshpande3-github-io-4431 189 29 created create VBN adeshpande3-github-io-4431 189 30 . . . adeshpande3-github-io-4431 190 1 The the DT adeshpande3-github-io-4431 190 2 task task NN adeshpande3-github-io-4431 190 3 of of IN adeshpande3-github-io-4431 190 4 the the DT adeshpande3-github-io-4431 190 5 generator generator NN adeshpande3-github-io-4431 190 6 is be VBZ adeshpande3-github-io-4431 190 7 to to TO adeshpande3-github-io-4431 190 8 create create VB adeshpande3-github-io-4431 190 9 images image NNS adeshpande3-github-io-4431 190 10 so so IN adeshpande3-github-io-4431 190 11 that that IN adeshpande3-github-io-4431 190 12 the the DT adeshpande3-github-io-4431 190 13 discriminator discriminator NN adeshpande3-github-io-4431 190 14 gets get VBZ adeshpande3-github-io-4431 190 15 trained train VBN adeshpande3-github-io-4431 190 16 to to TO adeshpande3-github-io-4431 190 17 produce produce VB adeshpande3-github-io-4431 190 18 the the DT adeshpande3-github-io-4431 190 19 correct correct JJ adeshpande3-github-io-4431 190 20 outputs output NNS adeshpande3-github-io-4431 190 21 . . . adeshpande3-github-io-4431 191 1 This this DT adeshpande3-github-io-4431 191 2 can can MD adeshpande3-github-io-4431 191 3 be be VB adeshpande3-github-io-4431 191 4 thought think VBN adeshpande3-github-io-4431 191 5 of of IN adeshpande3-github-io-4431 191 6 as as IN adeshpande3-github-io-4431 191 7 a a DT adeshpande3-github-io-4431 191 8 zero zero CD adeshpande3-github-io-4431 191 9 - - HYPH adeshpande3-github-io-4431 191 10 sum sum NN adeshpande3-github-io-4431 191 11 or or CC adeshpande3-github-io-4431 191 12 minimax minimax VB adeshpande3-github-io-4431 191 13 two two CD adeshpande3-github-io-4431 191 14 player player NN adeshpande3-github-io-4431 191 15 game game NN adeshpande3-github-io-4431 191 16 . . . adeshpande3-github-io-4431 192 1 The the DT adeshpande3-github-io-4431 192 2 analogy analogy NN adeshpande3-github-io-4431 192 3 used use VBN adeshpande3-github-io-4431 192 4 in in IN adeshpande3-github-io-4431 192 5 the the DT adeshpande3-github-io-4431 192 6 paper paper NN adeshpande3-github-io-4431 192 7 is be VBZ adeshpande3-github-io-4431 192 8 that that IN adeshpande3-github-io-4431 192 9 the the DT adeshpande3-github-io-4431 192 10 generative generative JJ adeshpande3-github-io-4431 192 11 model model NN adeshpande3-github-io-4431 192 12 is be VBZ adeshpande3-github-io-4431 192 13 like like IN adeshpande3-github-io-4431 192 14 “ " `` adeshpande3-github-io-4431 192 15 a a DT adeshpande3-github-io-4431 192 16 team team NN adeshpande3-github-io-4431 192 17 of of IN adeshpande3-github-io-4431 192 18 counterfeiters counterfeiter NNS adeshpande3-github-io-4431 192 19 , , , adeshpande3-github-io-4431 192 20 trying try VBG adeshpande3-github-io-4431 192 21 to to TO adeshpande3-github-io-4431 192 22 produce produce VB adeshpande3-github-io-4431 192 23 and and CC adeshpande3-github-io-4431 192 24 use use VB adeshpande3-github-io-4431 192 25 fake fake JJ adeshpande3-github-io-4431 192 26 currency currency NN adeshpande3-github-io-4431 192 27 ” " '' adeshpande3-github-io-4431 192 28 while while IN adeshpande3-github-io-4431 192 29 the the DT adeshpande3-github-io-4431 192 30 discriminative discriminative JJ adeshpande3-github-io-4431 192 31 model model NN adeshpande3-github-io-4431 192 32 is be VBZ adeshpande3-github-io-4431 192 33 like like IN adeshpande3-github-io-4431 192 34 “ " `` adeshpande3-github-io-4431 192 35 the the DT adeshpande3-github-io-4431 192 36 police police NN adeshpande3-github-io-4431 192 37 , , , adeshpande3-github-io-4431 192 38 trying try VBG adeshpande3-github-io-4431 192 39 to to TO adeshpande3-github-io-4431 192 40 detect detect VB adeshpande3-github-io-4431 192 41 the the DT adeshpande3-github-io-4431 192 42 counterfeit counterfeit JJ adeshpande3-github-io-4431 192 43 currency currency NN adeshpande3-github-io-4431 192 44 ” " '' adeshpande3-github-io-4431 192 45 . . . adeshpande3-github-io-4431 193 1 The the DT adeshpande3-github-io-4431 193 2 generator generator NN adeshpande3-github-io-4431 193 3 is be VBZ adeshpande3-github-io-4431 193 4 trying try VBG adeshpande3-github-io-4431 193 5 to to TO adeshpande3-github-io-4431 193 6 fool fool VB adeshpande3-github-io-4431 193 7 the the DT adeshpande3-github-io-4431 193 8 discriminator discriminator NN adeshpande3-github-io-4431 193 9 while while IN adeshpande3-github-io-4431 193 10 the the DT adeshpande3-github-io-4431 193 11 discriminator discriminator NN adeshpande3-github-io-4431 193 12 is be VBZ adeshpande3-github-io-4431 193 13 trying try VBG adeshpande3-github-io-4431 193 14 to to TO adeshpande3-github-io-4431 193 15 not not RB adeshpande3-github-io-4431 193 16 get get VB adeshpande3-github-io-4431 193 17 fooled fool VBN adeshpande3-github-io-4431 193 18 by by IN adeshpande3-github-io-4431 193 19 the the DT adeshpande3-github-io-4431 193 20 generator generator NN adeshpande3-github-io-4431 193 21 . . . adeshpande3-github-io-4431 194 1 As as IN adeshpande3-github-io-4431 194 2 the the DT adeshpande3-github-io-4431 194 3 models model NNS adeshpande3-github-io-4431 194 4 train train VBP adeshpande3-github-io-4431 194 5 , , , adeshpande3-github-io-4431 194 6 both both DT adeshpande3-github-io-4431 194 7 methods method NNS adeshpande3-github-io-4431 194 8 are be VBP adeshpande3-github-io-4431 194 9 improved improve VBN adeshpande3-github-io-4431 194 10 until until IN adeshpande3-github-io-4431 194 11 a a DT adeshpande3-github-io-4431 194 12 point point NN adeshpande3-github-io-4431 194 13 where where WRB adeshpande3-github-io-4431 194 14 the the DT adeshpande3-github-io-4431 194 15 “ " `` adeshpande3-github-io-4431 194 16 counterfeits counterfeit NNS adeshpande3-github-io-4431 194 17 are be VBP adeshpande3-github-io-4431 194 18 indistinguishable indistinguishable JJ adeshpande3-github-io-4431 194 19 from from IN adeshpande3-github-io-4431 194 20 the the DT adeshpande3-github-io-4431 194 21 genuine genuine JJ adeshpande3-github-io-4431 194 22 articles article NNS adeshpande3-github-io-4431 194 23 ” " '' adeshpande3-github-io-4431 194 24 . . . adeshpande3-github-io-4431 195 1 Why why WRB adeshpande3-github-io-4431 195 2 It -PRON- PRP adeshpande3-github-io-4431 195 3 ’s ’ VBZ adeshpande3-github-io-4431 195 4 Important important JJ adeshpande3-github-io-4431 195 5                                 _SP adeshpande3-github-io-4431 195 6 Sounds sound VBZ adeshpande3-github-io-4431 195 7 simple simple JJ adeshpande3-github-io-4431 195 8 enough enough RB adeshpande3-github-io-4431 195 9 , , , adeshpande3-github-io-4431 195 10 but but CC adeshpande3-github-io-4431 195 11 why why WRB adeshpande3-github-io-4431 195 12 do do VBP adeshpande3-github-io-4431 195 13 we -PRON- PRP adeshpande3-github-io-4431 195 14 care care VB adeshpande3-github-io-4431 195 15 about about IN adeshpande3-github-io-4431 195 16 these these DT adeshpande3-github-io-4431 195 17 networks network NNS adeshpande3-github-io-4431 195 18 ? ? . adeshpande3-github-io-4431 196 1 As as IN adeshpande3-github-io-4431 196 2 Yann Yann NNP adeshpande3-github-io-4431 196 3 LeCun LeCun NNP adeshpande3-github-io-4431 196 4 stated state VBD adeshpande3-github-io-4431 196 5 in in IN adeshpande3-github-io-4431 196 6 his -PRON- PRP$ adeshpande3-github-io-4431 196 7 Quora Quora NNP adeshpande3-github-io-4431 196 8 post post NN adeshpande3-github-io-4431 196 9 , , , adeshpande3-github-io-4431 196 10 the the DT adeshpande3-github-io-4431 196 11 discriminator discriminator NN adeshpande3-github-io-4431 196 12 now now RB adeshpande3-github-io-4431 196 13 is be VBZ adeshpande3-github-io-4431 196 14 aware aware JJ adeshpande3-github-io-4431 196 15 of of IN adeshpande3-github-io-4431 196 16 the the DT adeshpande3-github-io-4431 196 17 “ " `` adeshpande3-github-io-4431 196 18 internal internal JJ adeshpande3-github-io-4431 196 19 representation representation NN adeshpande3-github-io-4431 196 20 of of IN adeshpande3-github-io-4431 196 21 the the DT adeshpande3-github-io-4431 196 22 data datum NNS adeshpande3-github-io-4431 196 23 ” " '' adeshpande3-github-io-4431 196 24 because because IN adeshpande3-github-io-4431 196 25 it -PRON- PRP adeshpande3-github-io-4431 196 26 has have VBZ adeshpande3-github-io-4431 196 27 been be VBN adeshpande3-github-io-4431 196 28 trained train VBN adeshpande3-github-io-4431 196 29 to to TO adeshpande3-github-io-4431 196 30 understand understand VB adeshpande3-github-io-4431 196 31 the the DT adeshpande3-github-io-4431 196 32 differences difference NNS adeshpande3-github-io-4431 196 33 between between IN adeshpande3-github-io-4431 196 34 real real JJ adeshpande3-github-io-4431 196 35 images image NNS adeshpande3-github-io-4431 196 36 from from IN adeshpande3-github-io-4431 196 37 the the DT adeshpande3-github-io-4431 196 38 dataset dataset NN adeshpande3-github-io-4431 196 39 and and CC adeshpande3-github-io-4431 196 40 artificially artificially RB adeshpande3-github-io-4431 196 41 created create VBN adeshpande3-github-io-4431 196 42 ones one NNS adeshpande3-github-io-4431 196 43 . . . adeshpande3-github-io-4431 197 1 Thus thus RB adeshpande3-github-io-4431 197 2 , , , adeshpande3-github-io-4431 197 3 it -PRON- PRP adeshpande3-github-io-4431 197 4 can can MD adeshpande3-github-io-4431 197 5 be be VB adeshpande3-github-io-4431 197 6 used use VBN adeshpande3-github-io-4431 197 7 as as IN adeshpande3-github-io-4431 197 8 a a DT adeshpande3-github-io-4431 197 9 feature feature NN adeshpande3-github-io-4431 197 10 extractor extractor NN adeshpande3-github-io-4431 197 11 that that WDT adeshpande3-github-io-4431 197 12 you -PRON- PRP adeshpande3-github-io-4431 197 13 can can MD adeshpande3-github-io-4431 197 14 use use VB adeshpande3-github-io-4431 197 15 in in IN adeshpande3-github-io-4431 197 16 a a DT adeshpande3-github-io-4431 197 17 CNN CNN NNP adeshpande3-github-io-4431 197 18 . . . adeshpande3-github-io-4431 198 1 Plus plus CC adeshpande3-github-io-4431 198 2 , , , adeshpande3-github-io-4431 198 3 you -PRON- PRP adeshpande3-github-io-4431 198 4 can can MD adeshpande3-github-io-4431 198 5 just just RB adeshpande3-github-io-4431 198 6 create create VB adeshpande3-github-io-4431 198 7 really really RB adeshpande3-github-io-4431 198 8 cool cool JJ adeshpande3-github-io-4431 198 9 artificial artificial JJ adeshpande3-github-io-4431 198 10 images image NNS adeshpande3-github-io-4431 198 11 that that WDT adeshpande3-github-io-4431 198 12 look look VBP adeshpande3-github-io-4431 198 13 pretty pretty RB adeshpande3-github-io-4431 198 14 natural natural JJ adeshpande3-github-io-4431 198 15 to to IN adeshpande3-github-io-4431 198 16 me -PRON- PRP adeshpande3-github-io-4431 198 17 ( ( -LRB- adeshpande3-github-io-4431 198 18 link link NN adeshpande3-github-io-4431 198 19 ) ) -RRB- adeshpande3-github-io-4431 198 20 . . . adeshpande3-github-io-4431 199 1 Generating generate VBG adeshpande3-github-io-4431 199 2 Image image NN adeshpande3-github-io-4431 199 3 Descriptions description NNS adeshpande3-github-io-4431 199 4 ( ( -LRB- adeshpande3-github-io-4431 199 5 2014 2014 CD adeshpande3-github-io-4431 199 6 ) ) -RRB- adeshpande3-github-io-4431 199 7                                 _SP adeshpande3-github-io-4431 199 8 What what WP adeshpande3-github-io-4431 199 9 happens happen VBZ adeshpande3-github-io-4431 199 10 when when WRB adeshpande3-github-io-4431 199 11 you -PRON- PRP adeshpande3-github-io-4431 199 12 combine combine VBP adeshpande3-github-io-4431 199 13 CNNs cnn NNS adeshpande3-github-io-4431 199 14 with with IN adeshpande3-github-io-4431 199 15 RNNs rnn NNS adeshpande3-github-io-4431 199 16 ( ( -LRB- adeshpande3-github-io-4431 199 17 No no UH adeshpande3-github-io-4431 199 18 , , , adeshpande3-github-io-4431 199 19 you -PRON- PRP adeshpande3-github-io-4431 199 20 do do VBP adeshpande3-github-io-4431 199 21 n’t not RB adeshpande3-github-io-4431 199 22 get get VB adeshpande3-github-io-4431 199 23 R r NN adeshpande3-github-io-4431 199 24 - - HYPH adeshpande3-github-io-4431 199 25 CNNs cnn NNS adeshpande3-github-io-4431 199 26 , , , adeshpande3-github-io-4431 199 27 sorry sorry JJ adeshpande3-github-io-4431 199 28 ) ) -RRB- adeshpande3-github-io-4431 199 29 ? ? . adeshpande3-github-io-4431 199 30 But but CC adeshpande3-github-io-4431 199 31 you -PRON- PRP adeshpande3-github-io-4431 199 32 do do VBP adeshpande3-github-io-4431 199 33 get get VB adeshpande3-github-io-4431 199 34 one one NN adeshpande3-github-io-4431 199 35 really really RB adeshpande3-github-io-4431 199 36 amazing amazing JJ adeshpande3-github-io-4431 199 37 application application NN adeshpande3-github-io-4431 199 38 . . . adeshpande3-github-io-4431 200 1 Written write VBN adeshpande3-github-io-4431 200 2 by by IN adeshpande3-github-io-4431 200 3 Andrej Andrej NNP adeshpande3-github-io-4431 200 4 Karpathy Karpathy NNP adeshpande3-github-io-4431 200 5 ( ( -LRB- adeshpande3-github-io-4431 200 6 one one CD adeshpande3-github-io-4431 200 7 of of IN adeshpande3-github-io-4431 200 8 my -PRON- PRP$ adeshpande3-github-io-4431 200 9 personal personal JJ adeshpande3-github-io-4431 200 10 favorite favorite JJ adeshpande3-github-io-4431 200 11 authors author NNS adeshpande3-github-io-4431 200 12 ) ) -RRB- adeshpande3-github-io-4431 200 13 and and CC adeshpande3-github-io-4431 200 14 Fei Fei NNP adeshpande3-github-io-4431 200 15 - - HYPH adeshpande3-github-io-4431 200 16 Fei Fei NNP adeshpande3-github-io-4431 200 17 Li Li NNP adeshpande3-github-io-4431 200 18 , , , adeshpande3-github-io-4431 200 19 this this DT adeshpande3-github-io-4431 200 20 paper paper NN adeshpande3-github-io-4431 200 21 looks look VBZ adeshpande3-github-io-4431 200 22 into into IN adeshpande3-github-io-4431 200 23 a a DT adeshpande3-github-io-4431 200 24 combination combination NN adeshpande3-github-io-4431 200 25 of of IN adeshpande3-github-io-4431 200 26 CNNs cnn NNS adeshpande3-github-io-4431 200 27 and and CC adeshpande3-github-io-4431 200 28 bidirectional bidirectional JJ adeshpande3-github-io-4431 200 29 RNNs rnn NNS adeshpande3-github-io-4431 200 30 ( ( -LRB- adeshpande3-github-io-4431 200 31 Recurrent recurrent JJ adeshpande3-github-io-4431 200 32 Neural Neural NNP adeshpande3-github-io-4431 200 33 Networks Networks NNPS adeshpande3-github-io-4431 200 34 ) ) -RRB- adeshpande3-github-io-4431 200 35 to to TO adeshpande3-github-io-4431 200 36 generate generate VB adeshpande3-github-io-4431 200 37 natural natural JJ adeshpande3-github-io-4431 200 38 language language NN adeshpande3-github-io-4431 200 39 descriptions description NNS adeshpande3-github-io-4431 200 40 of of IN adeshpande3-github-io-4431 200 41 different different JJ adeshpande3-github-io-4431 200 42 image image NN adeshpande3-github-io-4431 200 43 regions region NNS adeshpande3-github-io-4431 200 44 . . . adeshpande3-github-io-4431 201 1 Basically basically RB adeshpande3-github-io-4431 201 2 , , , adeshpande3-github-io-4431 201 3 the the DT adeshpande3-github-io-4431 201 4 model model NN adeshpande3-github-io-4431 201 5 is be VBZ adeshpande3-github-io-4431 201 6 able able JJ adeshpande3-github-io-4431 201 7 to to TO adeshpande3-github-io-4431 201 8 take take VB adeshpande3-github-io-4431 201 9 in in RP adeshpande3-github-io-4431 201 10 an an DT adeshpande3-github-io-4431 201 11 image image NN adeshpande3-github-io-4431 201 12 , , , adeshpande3-github-io-4431 201 13 and and CC adeshpande3-github-io-4431 201 14 output output NN adeshpande3-github-io-4431 201 15 this this DT adeshpande3-github-io-4431 201 16 : : : adeshpande3-github-io-4431 201 17 That that DT adeshpande3-github-io-4431 201 18 ’s ’ VBZ adeshpande3-github-io-4431 201 19 pretty pretty RB adeshpande3-github-io-4431 201 20 incredible incredible JJ adeshpande3-github-io-4431 201 21 . . . adeshpande3-github-io-4431 202 1 Let let VB adeshpande3-github-io-4431 202 2 ’s -PRON- PRP adeshpande3-github-io-4431 202 3 look look VB adeshpande3-github-io-4431 202 4 at at IN adeshpande3-github-io-4431 202 5 how how WRB adeshpande3-github-io-4431 202 6 this this DT adeshpande3-github-io-4431 202 7 compares compare VBZ adeshpande3-github-io-4431 202 8 to to IN adeshpande3-github-io-4431 202 9 normal normal JJ adeshpande3-github-io-4431 202 10 CNNs cnn NNS adeshpande3-github-io-4431 202 11 . . . adeshpande3-github-io-4431 203 1 With with IN adeshpande3-github-io-4431 203 2 traditional traditional JJ adeshpande3-github-io-4431 203 3 CNNs cnn NNS adeshpande3-github-io-4431 203 4 , , , adeshpande3-github-io-4431 203 5 there there EX adeshpande3-github-io-4431 203 6 is be VBZ adeshpande3-github-io-4431 203 7 a a DT adeshpande3-github-io-4431 203 8 single single JJ adeshpande3-github-io-4431 203 9 clear clear JJ adeshpande3-github-io-4431 203 10 label label NN adeshpande3-github-io-4431 203 11 associated associate VBN adeshpande3-github-io-4431 203 12 with with IN adeshpande3-github-io-4431 203 13 each each DT adeshpande3-github-io-4431 203 14 image image NN adeshpande3-github-io-4431 203 15 in in IN adeshpande3-github-io-4431 203 16 the the DT adeshpande3-github-io-4431 203 17 training training NN adeshpande3-github-io-4431 203 18 data datum NNS adeshpande3-github-io-4431 203 19 . . . adeshpande3-github-io-4431 204 1 The the DT adeshpande3-github-io-4431 204 2 model model NN adeshpande3-github-io-4431 204 3 described describe VBN adeshpande3-github-io-4431 204 4 in in IN adeshpande3-github-io-4431 204 5 the the DT adeshpande3-github-io-4431 204 6 paper paper NN adeshpande3-github-io-4431 204 7 has have VBZ adeshpande3-github-io-4431 204 8 training training NN adeshpande3-github-io-4431 204 9 examples example NNS adeshpande3-github-io-4431 204 10 that that WDT adeshpande3-github-io-4431 204 11 have have VBP adeshpande3-github-io-4431 204 12 a a DT adeshpande3-github-io-4431 204 13 sentence sentence NN adeshpande3-github-io-4431 204 14 ( ( -LRB- adeshpande3-github-io-4431 204 15 or or CC adeshpande3-github-io-4431 204 16 caption caption NN adeshpande3-github-io-4431 204 17 ) ) -RRB- adeshpande3-github-io-4431 204 18 associated associate VBN adeshpande3-github-io-4431 204 19 with with IN adeshpande3-github-io-4431 204 20 each each DT adeshpande3-github-io-4431 204 21 image image NN adeshpande3-github-io-4431 204 22 . . . adeshpande3-github-io-4431 205 1 This this DT adeshpande3-github-io-4431 205 2 type type NN adeshpande3-github-io-4431 205 3 of of IN adeshpande3-github-io-4431 205 4 label label NN adeshpande3-github-io-4431 205 5 is be VBZ adeshpande3-github-io-4431 205 6 called call VBN adeshpande3-github-io-4431 205 7 a a DT adeshpande3-github-io-4431 205 8 weak weak JJ adeshpande3-github-io-4431 205 9 label label NN adeshpande3-github-io-4431 205 10 , , , adeshpande3-github-io-4431 205 11 where where WRB adeshpande3-github-io-4431 205 12 segments segment NNS adeshpande3-github-io-4431 205 13 of of IN adeshpande3-github-io-4431 205 14 the the DT adeshpande3-github-io-4431 205 15 sentence sentence NN adeshpande3-github-io-4431 205 16 refer refer VBP adeshpande3-github-io-4431 205 17 to to IN adeshpande3-github-io-4431 205 18 ( ( -LRB- adeshpande3-github-io-4431 205 19 unknown unknown JJ adeshpande3-github-io-4431 205 20 ) ) -RRB- adeshpande3-github-io-4431 205 21 parts part NNS adeshpande3-github-io-4431 205 22 of of IN adeshpande3-github-io-4431 205 23 the the DT adeshpande3-github-io-4431 205 24 image image NN adeshpande3-github-io-4431 205 25 . . . adeshpande3-github-io-4431 206 1 Using use VBG adeshpande3-github-io-4431 206 2 this this DT adeshpande3-github-io-4431 206 3 training training NN adeshpande3-github-io-4431 206 4 data datum NNS adeshpande3-github-io-4431 206 5 , , , adeshpande3-github-io-4431 206 6 a a DT adeshpande3-github-io-4431 206 7 deep deep JJ adeshpande3-github-io-4431 206 8 neural neural JJ adeshpande3-github-io-4431 206 9 network network NN adeshpande3-github-io-4431 206 10 “ " `` adeshpande3-github-io-4431 206 11 infers infer VBZ adeshpande3-github-io-4431 206 12 the the DT adeshpande3-github-io-4431 206 13 latent latent NN adeshpande3-github-io-4431 206 14 alignment alignment NN adeshpande3-github-io-4431 206 15 between between IN adeshpande3-github-io-4431 206 16 segments segment NNS adeshpande3-github-io-4431 206 17 of of IN adeshpande3-github-io-4431 206 18 the the DT adeshpande3-github-io-4431 206 19 sentences sentence NNS adeshpande3-github-io-4431 206 20 and and CC adeshpande3-github-io-4431 206 21 the the DT adeshpande3-github-io-4431 206 22 region region NN adeshpande3-github-io-4431 206 23 that that IN adeshpande3-github-io-4431 206 24 they -PRON- PRP adeshpande3-github-io-4431 206 25 describe describe VBP adeshpande3-github-io-4431 206 26 ” " '' adeshpande3-github-io-4431 206 27 ( ( -LRB- adeshpande3-github-io-4431 206 28 quote quote UH adeshpande3-github-io-4431 206 29 from from IN adeshpande3-github-io-4431 206 30 the the DT adeshpande3-github-io-4431 206 31 paper paper NN adeshpande3-github-io-4431 206 32 ) ) -RRB- adeshpande3-github-io-4431 206 33 . . . adeshpande3-github-io-4431 207 1 Another another DT adeshpande3-github-io-4431 207 2 neural neural JJ adeshpande3-github-io-4431 207 3 net net NN adeshpande3-github-io-4431 207 4 takes take VBZ adeshpande3-github-io-4431 207 5 in in RP adeshpande3-github-io-4431 207 6 the the DT adeshpande3-github-io-4431 207 7 image image NN adeshpande3-github-io-4431 207 8 as as IN adeshpande3-github-io-4431 207 9 input input NN adeshpande3-github-io-4431 207 10 and and CC adeshpande3-github-io-4431 207 11 generates generate VBZ adeshpande3-github-io-4431 207 12 a a DT adeshpande3-github-io-4431 207 13 description description NN adeshpande3-github-io-4431 207 14 in in IN adeshpande3-github-io-4431 207 15 text text NN adeshpande3-github-io-4431 207 16 . . . adeshpande3-github-io-4431 208 1 Let let VB adeshpande3-github-io-4431 208 2 ’s -PRON- PRP adeshpande3-github-io-4431 208 3 take take VB adeshpande3-github-io-4431 208 4 a a DT adeshpande3-github-io-4431 208 5 separate separate JJ adeshpande3-github-io-4431 208 6 look look NN adeshpande3-github-io-4431 208 7 at at IN adeshpande3-github-io-4431 208 8 the the DT adeshpande3-github-io-4431 208 9 two two CD adeshpande3-github-io-4431 208 10 components component NNS adeshpande3-github-io-4431 208 11 , , , adeshpande3-github-io-4431 208 12 alignment alignment NN adeshpande3-github-io-4431 208 13 and and CC adeshpande3-github-io-4431 208 14 generation generation NN adeshpande3-github-io-4431 208 15 . . . adeshpande3-github-io-4431 209 1 Alignment Alignment NNP adeshpande3-github-io-4431 209 2 Model Model NNP adeshpande3-github-io-4431 209 3                                 _SP adeshpande3-github-io-4431 209 4 The the DT adeshpande3-github-io-4431 209 5 goal goal NN adeshpande3-github-io-4431 209 6 of of IN adeshpande3-github-io-4431 209 7 this this DT adeshpande3-github-io-4431 209 8 part part NN adeshpande3-github-io-4431 209 9 of of IN adeshpande3-github-io-4431 209 10 the the DT adeshpande3-github-io-4431 209 11 model model NN adeshpande3-github-io-4431 209 12 is be VBZ adeshpande3-github-io-4431 209 13 to to TO adeshpande3-github-io-4431 209 14 be be VB adeshpande3-github-io-4431 209 15 able able JJ adeshpande3-github-io-4431 209 16 to to TO adeshpande3-github-io-4431 209 17 align align VB adeshpande3-github-io-4431 209 18 the the DT adeshpande3-github-io-4431 209 19 visual visual JJ adeshpande3-github-io-4431 209 20 and and CC adeshpande3-github-io-4431 209 21 textual textual JJ adeshpande3-github-io-4431 209 22 data datum NNS adeshpande3-github-io-4431 209 23 ( ( -LRB- adeshpande3-github-io-4431 209 24 the the DT adeshpande3-github-io-4431 209 25 image image NN adeshpande3-github-io-4431 209 26 and and CC adeshpande3-github-io-4431 209 27 its -PRON- PRP$ adeshpande3-github-io-4431 209 28 sentence sentence NN adeshpande3-github-io-4431 209 29 description description NN adeshpande3-github-io-4431 209 30 ) ) -RRB- adeshpande3-github-io-4431 209 31 . . . adeshpande3-github-io-4431 210 1 The the DT adeshpande3-github-io-4431 210 2 model model NN adeshpande3-github-io-4431 210 3 works work VBZ adeshpande3-github-io-4431 210 4 by by IN adeshpande3-github-io-4431 210 5 accepting accept VBG adeshpande3-github-io-4431 210 6 an an DT adeshpande3-github-io-4431 210 7 image image NN adeshpande3-github-io-4431 210 8 and and CC adeshpande3-github-io-4431 210 9 a a DT adeshpande3-github-io-4431 210 10 sentence sentence NN adeshpande3-github-io-4431 210 11 as as IN adeshpande3-github-io-4431 210 12 input input NN adeshpande3-github-io-4431 210 13 , , , adeshpande3-github-io-4431 210 14 where where WRB adeshpande3-github-io-4431 210 15 the the DT adeshpande3-github-io-4431 210 16 output output NN adeshpande3-github-io-4431 210 17 is be VBZ adeshpande3-github-io-4431 210 18 a a DT adeshpande3-github-io-4431 210 19 score score NN adeshpande3-github-io-4431 210 20 for for IN adeshpande3-github-io-4431 210 21 how how WRB adeshpande3-github-io-4431 210 22 well well RB adeshpande3-github-io-4431 210 23 they -PRON- PRP adeshpande3-github-io-4431 210 24 match match VBP adeshpande3-github-io-4431 210 25 ( ( -LRB- adeshpande3-github-io-4431 210 26 Now now RB adeshpande3-github-io-4431 210 27 , , , adeshpande3-github-io-4431 210 28 Karpathy Karpathy NNP adeshpande3-github-io-4431 210 29 refers refer VBZ adeshpande3-github-io-4431 210 30 a a DT adeshpande3-github-io-4431 210 31 different different JJ adeshpande3-github-io-4431 210 32 paper paper NN adeshpande3-github-io-4431 210 33 which which WDT adeshpande3-github-io-4431 210 34 goes go VBZ adeshpande3-github-io-4431 210 35 into into IN adeshpande3-github-io-4431 210 36 the the DT adeshpande3-github-io-4431 210 37 specifics specific NNS adeshpande3-github-io-4431 210 38 of of IN adeshpande3-github-io-4431 210 39 how how WRB adeshpande3-github-io-4431 210 40 this this DT adeshpande3-github-io-4431 210 41 works work VBZ adeshpande3-github-io-4431 210 42 . . . adeshpande3-github-io-4431 211 1 This this DT adeshpande3-github-io-4431 211 2 model model NN adeshpande3-github-io-4431 211 3 is be VBZ adeshpande3-github-io-4431 211 4 trained train VBN adeshpande3-github-io-4431 211 5 on on IN adeshpande3-github-io-4431 211 6 compatible compatible JJ adeshpande3-github-io-4431 211 7 and and CC adeshpande3-github-io-4431 211 8 incompatible incompatible JJ adeshpande3-github-io-4431 211 9 image image NN adeshpande3-github-io-4431 211 10 - - HYPH adeshpande3-github-io-4431 211 11 sentence sentence NN adeshpande3-github-io-4431 211 12 pairs pair NNS adeshpande3-github-io-4431 211 13 ) ) -RRB- adeshpande3-github-io-4431 211 14 . . . adeshpande3-github-io-4431 212 1 Now now RB adeshpande3-github-io-4431 212 2 let let VB adeshpande3-github-io-4431 212 3 ’s -PRON- PRP adeshpande3-github-io-4431 212 4 think think VB adeshpande3-github-io-4431 212 5 about about IN adeshpande3-github-io-4431 212 6 representing represent VBG adeshpande3-github-io-4431 212 7 the the DT adeshpande3-github-io-4431 212 8 images image NNS adeshpande3-github-io-4431 212 9 . . . adeshpande3-github-io-4431 213 1 The the DT adeshpande3-github-io-4431 213 2 first first JJ adeshpande3-github-io-4431 213 3 step step NN adeshpande3-github-io-4431 213 4 is be VBZ adeshpande3-github-io-4431 213 5 feeding feed VBG adeshpande3-github-io-4431 213 6 the the DT adeshpande3-github-io-4431 213 7 image image NN adeshpande3-github-io-4431 213 8 into into IN adeshpande3-github-io-4431 213 9 an an DT adeshpande3-github-io-4431 213 10 R R NNP adeshpande3-github-io-4431 213 11 - - HYPH adeshpande3-github-io-4431 213 12 CNN CNN NNP adeshpande3-github-io-4431 213 13 in in IN adeshpande3-github-io-4431 213 14 order order NN adeshpande3-github-io-4431 213 15 to to TO adeshpande3-github-io-4431 213 16 detect detect VB adeshpande3-github-io-4431 213 17 the the DT adeshpande3-github-io-4431 213 18 individual individual JJ adeshpande3-github-io-4431 213 19 objects object NNS adeshpande3-github-io-4431 213 20 . . . adeshpande3-github-io-4431 214 1 This this DT adeshpande3-github-io-4431 214 2 R R NNP adeshpande3-github-io-4431 214 3 - - HYPH adeshpande3-github-io-4431 214 4 CNN CNN NNP adeshpande3-github-io-4431 214 5 was be VBD adeshpande3-github-io-4431 214 6 trained train VBN adeshpande3-github-io-4431 214 7 on on IN adeshpande3-github-io-4431 214 8 ImageNet ImageNet NNP adeshpande3-github-io-4431 214 9 data datum NNS adeshpande3-github-io-4431 214 10 . . . adeshpande3-github-io-4431 215 1 The the DT adeshpande3-github-io-4431 215 2 top top JJ adeshpande3-github-io-4431 215 3 19 19 CD adeshpande3-github-io-4431 215 4 ( ( -LRB- adeshpande3-github-io-4431 215 5 plus plus CC adeshpande3-github-io-4431 215 6 the the DT adeshpande3-github-io-4431 215 7 original original JJ adeshpande3-github-io-4431 215 8 image image NN adeshpande3-github-io-4431 215 9 ) ) -RRB- adeshpande3-github-io-4431 215 10 object object NN adeshpande3-github-io-4431 215 11 regions region NNS adeshpande3-github-io-4431 215 12 are be VBP adeshpande3-github-io-4431 215 13 embedded embed VBN adeshpande3-github-io-4431 215 14 to to IN adeshpande3-github-io-4431 215 15 a a DT adeshpande3-github-io-4431 215 16 500 500 CD adeshpande3-github-io-4431 215 17 dimensional dimensional JJ adeshpande3-github-io-4431 215 18 space space NN adeshpande3-github-io-4431 215 19 . . . adeshpande3-github-io-4431 216 1 Now now RB adeshpande3-github-io-4431 216 2 we -PRON- PRP adeshpande3-github-io-4431 216 3 have have VBP adeshpande3-github-io-4431 216 4 20 20 CD adeshpande3-github-io-4431 216 5 different different JJ adeshpande3-github-io-4431 216 6 500 500 CD adeshpande3-github-io-4431 216 7 dimensional dimensional JJ adeshpande3-github-io-4431 216 8 vectors vector NNS adeshpande3-github-io-4431 216 9 ( ( -LRB- adeshpande3-github-io-4431 216 10 represented represent VBN adeshpande3-github-io-4431 216 11 by by IN adeshpande3-github-io-4431 216 12 v v NN adeshpande3-github-io-4431 216 13 in in IN adeshpande3-github-io-4431 216 14 the the DT adeshpande3-github-io-4431 216 15 paper paper NN adeshpande3-github-io-4431 216 16 ) ) -RRB- adeshpande3-github-io-4431 216 17 for for IN adeshpande3-github-io-4431 216 18 each each DT adeshpande3-github-io-4431 216 19 image image NN adeshpande3-github-io-4431 216 20 . . . adeshpande3-github-io-4431 217 1 We -PRON- PRP adeshpande3-github-io-4431 217 2 have have VBP adeshpande3-github-io-4431 217 3 information information NN adeshpande3-github-io-4431 217 4 about about IN adeshpande3-github-io-4431 217 5 the the DT adeshpande3-github-io-4431 217 6 image image NN adeshpande3-github-io-4431 217 7 . . . adeshpande3-github-io-4431 218 1 Now now RB adeshpande3-github-io-4431 218 2 , , , adeshpande3-github-io-4431 218 3 we -PRON- PRP adeshpande3-github-io-4431 218 4 want want VBP adeshpande3-github-io-4431 218 5 information information NN adeshpande3-github-io-4431 218 6 about about IN adeshpande3-github-io-4431 218 7 the the DT adeshpande3-github-io-4431 218 8 sentence sentence NN adeshpande3-github-io-4431 218 9 . . . adeshpande3-github-io-4431 219 1 We -PRON- PRP adeshpande3-github-io-4431 219 2 ’re be VBP adeshpande3-github-io-4431 219 3 going go VBG adeshpande3-github-io-4431 219 4 to to IN adeshpande3-github-io-4431 219 5 embed embe VBN adeshpande3-github-io-4431 219 6 words word NNS adeshpande3-github-io-4431 219 7 into into IN adeshpande3-github-io-4431 219 8 this this DT adeshpande3-github-io-4431 219 9 same same JJ adeshpande3-github-io-4431 219 10 multimodal multimodal NN adeshpande3-github-io-4431 219 11 space space NN adeshpande3-github-io-4431 219 12 . . . adeshpande3-github-io-4431 220 1 This this DT adeshpande3-github-io-4431 220 2 is be VBZ adeshpande3-github-io-4431 220 3 done do VBN adeshpande3-github-io-4431 220 4 by by IN adeshpande3-github-io-4431 220 5 using use VBG adeshpande3-github-io-4431 220 6 a a DT adeshpande3-github-io-4431 220 7 bidirectional bidirectional JJ adeshpande3-github-io-4431 220 8 recurrent recurrent NN adeshpande3-github-io-4431 220 9 neural neural JJ adeshpande3-github-io-4431 220 10 network network NN adeshpande3-github-io-4431 220 11 . . . adeshpande3-github-io-4431 221 1 From from IN adeshpande3-github-io-4431 221 2 the the DT adeshpande3-github-io-4431 221 3 highest high JJS adeshpande3-github-io-4431 221 4 level level NN adeshpande3-github-io-4431 221 5 , , , adeshpande3-github-io-4431 221 6 this this DT adeshpande3-github-io-4431 221 7 serves serve VBZ adeshpande3-github-io-4431 221 8 to to TO adeshpande3-github-io-4431 221 9 illustrate illustrate VB adeshpande3-github-io-4431 221 10 information information NN adeshpande3-github-io-4431 221 11 about about IN adeshpande3-github-io-4431 221 12 the the DT adeshpande3-github-io-4431 221 13 context context NN adeshpande3-github-io-4431 221 14 of of IN adeshpande3-github-io-4431 221 15 words word NNS adeshpande3-github-io-4431 221 16 in in IN adeshpande3-github-io-4431 221 17 a a DT adeshpande3-github-io-4431 221 18 given give VBN adeshpande3-github-io-4431 221 19 sentence sentence NN adeshpande3-github-io-4431 221 20 . . . adeshpande3-github-io-4431 222 1 Since since IN adeshpande3-github-io-4431 222 2 this this DT adeshpande3-github-io-4431 222 3 information information NN adeshpande3-github-io-4431 222 4 about about IN adeshpande3-github-io-4431 222 5 the the DT adeshpande3-github-io-4431 222 6 picture picture NN adeshpande3-github-io-4431 222 7 and and CC adeshpande3-github-io-4431 222 8 the the DT adeshpande3-github-io-4431 222 9 sentence sentence NN adeshpande3-github-io-4431 222 10 are be VBP adeshpande3-github-io-4431 222 11 both both DT adeshpande3-github-io-4431 222 12 in in IN adeshpande3-github-io-4431 222 13 the the DT adeshpande3-github-io-4431 222 14 same same JJ adeshpande3-github-io-4431 222 15 space space NN adeshpande3-github-io-4431 222 16 , , , adeshpande3-github-io-4431 222 17 we -PRON- PRP adeshpande3-github-io-4431 222 18 can can MD adeshpande3-github-io-4431 222 19 compute compute VB adeshpande3-github-io-4431 222 20 inner inner JJ adeshpande3-github-io-4431 222 21 products product NNS adeshpande3-github-io-4431 222 22 to to TO adeshpande3-github-io-4431 222 23 show show VB adeshpande3-github-io-4431 222 24 a a DT adeshpande3-github-io-4431 222 25 measure measure NN adeshpande3-github-io-4431 222 26 of of IN adeshpande3-github-io-4431 222 27 similarity similarity NN adeshpande3-github-io-4431 222 28 . . . adeshpande3-github-io-4431 223 1 Generation Generation NNP adeshpande3-github-io-4431 223 2 Model Model NNP adeshpande3-github-io-4431 223 3                                 _SP adeshpande3-github-io-4431 223 4 The the DT adeshpande3-github-io-4431 223 5 alignment alignment NN adeshpande3-github-io-4431 223 6 model model NN adeshpande3-github-io-4431 223 7 has have VBZ adeshpande3-github-io-4431 223 8 the the DT adeshpande3-github-io-4431 223 9 main main JJ adeshpande3-github-io-4431 223 10 purpose purpose NN adeshpande3-github-io-4431 223 11 of of IN adeshpande3-github-io-4431 223 12 creating create VBG adeshpande3-github-io-4431 223 13 a a DT adeshpande3-github-io-4431 223 14 dataset dataset NN adeshpande3-github-io-4431 223 15 where where WRB adeshpande3-github-io-4431 223 16 you -PRON- PRP adeshpande3-github-io-4431 223 17 have have VBP adeshpande3-github-io-4431 223 18 a a DT adeshpande3-github-io-4431 223 19 set set NN adeshpande3-github-io-4431 223 20 of of IN adeshpande3-github-io-4431 223 21 image image NN adeshpande3-github-io-4431 223 22 regions region NNS adeshpande3-github-io-4431 223 23 ( ( -LRB- adeshpande3-github-io-4431 223 24 found find VBN adeshpande3-github-io-4431 223 25 by by IN adeshpande3-github-io-4431 223 26 the the DT adeshpande3-github-io-4431 223 27 RCNN RCNN NNP adeshpande3-github-io-4431 223 28 ) ) -RRB- adeshpande3-github-io-4431 223 29 and and CC adeshpande3-github-io-4431 223 30 corresponding correspond VBG adeshpande3-github-io-4431 223 31 text text NN adeshpande3-github-io-4431 223 32 ( ( -LRB- adeshpande3-github-io-4431 223 33 thanks thank NNS adeshpande3-github-io-4431 223 34 to to IN adeshpande3-github-io-4431 223 35 the the DT adeshpande3-github-io-4431 223 36 BRNN BRNN NNP adeshpande3-github-io-4431 223 37 ) ) -RRB- adeshpande3-github-io-4431 223 38 . . . adeshpande3-github-io-4431 224 1 Now now RB adeshpande3-github-io-4431 224 2 , , , adeshpande3-github-io-4431 224 3 the the DT adeshpande3-github-io-4431 224 4 generation generation NN adeshpande3-github-io-4431 224 5 model model NN adeshpande3-github-io-4431 224 6 is be VBZ adeshpande3-github-io-4431 224 7 going go VBG adeshpande3-github-io-4431 224 8 to to TO adeshpande3-github-io-4431 224 9 learn learn VB adeshpande3-github-io-4431 224 10 from from IN adeshpande3-github-io-4431 224 11 that that DT adeshpande3-github-io-4431 224 12 dataset dataset NN adeshpande3-github-io-4431 224 13 in in IN adeshpande3-github-io-4431 224 14 order order NN adeshpande3-github-io-4431 224 15 to to TO adeshpande3-github-io-4431 224 16 generate generate VB adeshpande3-github-io-4431 224 17 descriptions description NNS adeshpande3-github-io-4431 224 18 given give VBN adeshpande3-github-io-4431 224 19 an an DT adeshpande3-github-io-4431 224 20 image image NN adeshpande3-github-io-4431 224 21 . . . adeshpande3-github-io-4431 225 1 The the DT adeshpande3-github-io-4431 225 2 model model NN adeshpande3-github-io-4431 225 3 takes take VBZ adeshpande3-github-io-4431 225 4 in in RP adeshpande3-github-io-4431 225 5 an an DT adeshpande3-github-io-4431 225 6 image image NN adeshpande3-github-io-4431 225 7 and and CC adeshpande3-github-io-4431 225 8 feeds feed VBZ adeshpande3-github-io-4431 225 9 it -PRON- PRP adeshpande3-github-io-4431 225 10 through through IN adeshpande3-github-io-4431 225 11 a a DT adeshpande3-github-io-4431 225 12 CNN CNN NNP adeshpande3-github-io-4431 225 13 . . . adeshpande3-github-io-4431 226 1 The the DT adeshpande3-github-io-4431 226 2 softmax softmax NNP adeshpande3-github-io-4431 226 3 layer layer NN adeshpande3-github-io-4431 226 4 is be VBZ adeshpande3-github-io-4431 226 5 disregarded disregard VBN adeshpande3-github-io-4431 226 6 as as IN adeshpande3-github-io-4431 226 7 the the DT adeshpande3-github-io-4431 226 8 outputs output NNS adeshpande3-github-io-4431 226 9 of of IN adeshpande3-github-io-4431 226 10 the the DT adeshpande3-github-io-4431 226 11 fully fully RB adeshpande3-github-io-4431 226 12 connected connected JJ adeshpande3-github-io-4431 226 13 layer layer NN adeshpande3-github-io-4431 226 14 become become VB adeshpande3-github-io-4431 226 15 the the DT adeshpande3-github-io-4431 226 16 inputs input NNS adeshpande3-github-io-4431 226 17 to to IN adeshpande3-github-io-4431 226 18 another another DT adeshpande3-github-io-4431 226 19 RNN RNN NNP adeshpande3-github-io-4431 226 20 . . . adeshpande3-github-io-4431 227 1 For for IN adeshpande3-github-io-4431 227 2 those those DT adeshpande3-github-io-4431 227 3 that that WDT adeshpande3-github-io-4431 227 4 are be VBP adeshpande3-github-io-4431 227 5 n’t not RB adeshpande3-github-io-4431 227 6 as as RB adeshpande3-github-io-4431 227 7 familiar familiar JJ adeshpande3-github-io-4431 227 8 with with IN adeshpande3-github-io-4431 227 9 RNNs rnn NNS adeshpande3-github-io-4431 227 10 , , , adeshpande3-github-io-4431 227 11 their -PRON- PRP$ adeshpande3-github-io-4431 227 12 function function NN adeshpande3-github-io-4431 227 13 is be VBZ adeshpande3-github-io-4431 227 14 to to TO adeshpande3-github-io-4431 227 15 basically basically RB adeshpande3-github-io-4431 227 16 form form VB adeshpande3-github-io-4431 227 17 probability probability NN adeshpande3-github-io-4431 227 18 distributions distribution NNS adeshpande3-github-io-4431 227 19 on on IN adeshpande3-github-io-4431 227 20 the the DT adeshpande3-github-io-4431 227 21 different different JJ adeshpande3-github-io-4431 227 22 words word NNS adeshpande3-github-io-4431 227 23 in in IN adeshpande3-github-io-4431 227 24 a a DT adeshpande3-github-io-4431 227 25 sentence sentence NN adeshpande3-github-io-4431 227 26 ( ( -LRB- adeshpande3-github-io-4431 227 27 RNNs rnn NNS adeshpande3-github-io-4431 227 28 also also RB adeshpande3-github-io-4431 227 29 need need VBP adeshpande3-github-io-4431 227 30 to to TO adeshpande3-github-io-4431 227 31 be be VB adeshpande3-github-io-4431 227 32 trained train VBN adeshpande3-github-io-4431 227 33 just just RB adeshpande3-github-io-4431 227 34 like like IN adeshpande3-github-io-4431 227 35 CNNs cnn NNS adeshpande3-github-io-4431 227 36 do do VBP adeshpande3-github-io-4431 227 37 ) ) -RRB- adeshpande3-github-io-4431 227 38 . . . adeshpande3-github-io-4431 228 1 Disclaimer disclaimer NN adeshpande3-github-io-4431 228 2 : : : adeshpande3-github-io-4431 228 3 This this DT adeshpande3-github-io-4431 228 4 was be VBD adeshpande3-github-io-4431 228 5 definitely definitely RB adeshpande3-github-io-4431 228 6 one one CD adeshpande3-github-io-4431 228 7 of of IN adeshpande3-github-io-4431 228 8 the the DT adeshpande3-github-io-4431 228 9 more more RBR adeshpande3-github-io-4431 228 10 dense dense JJ adeshpande3-github-io-4431 228 11 papers paper NNS adeshpande3-github-io-4431 228 12 in in IN adeshpande3-github-io-4431 228 13 this this DT adeshpande3-github-io-4431 228 14 section section NN adeshpande3-github-io-4431 228 15 , , , adeshpande3-github-io-4431 228 16 so so CC adeshpande3-github-io-4431 228 17 if if IN adeshpande3-github-io-4431 228 18 anyone anyone NN adeshpande3-github-io-4431 228 19 has have VBZ adeshpande3-github-io-4431 228 20 any any DT adeshpande3-github-io-4431 228 21 corrections correction NNS adeshpande3-github-io-4431 228 22 or or CC adeshpande3-github-io-4431 228 23 other other JJ adeshpande3-github-io-4431 228 24 explanations explanation NNS adeshpande3-github-io-4431 228 25 , , , adeshpande3-github-io-4431 228 26 I -PRON- PRP adeshpande3-github-io-4431 228 27 ’d ’d , adeshpande3-github-io-4431 228 28 love love VB adeshpande3-github-io-4431 228 29 to to TO adeshpande3-github-io-4431 228 30 hear hear VB adeshpande3-github-io-4431 228 31 them -PRON- PRP adeshpande3-github-io-4431 228 32 in in IN adeshpande3-github-io-4431 228 33 the the DT adeshpande3-github-io-4431 228 34 comments comment NNS adeshpande3-github-io-4431 228 35 ! ! . adeshpande3-github-io-4431 229 1 Why why WRB adeshpande3-github-io-4431 229 2 It -PRON- PRP adeshpande3-github-io-4431 229 3 ’s ’ VBZ adeshpande3-github-io-4431 229 4 Important important JJ adeshpande3-github-io-4431 229 5                                 _SP adeshpande3-github-io-4431 229 6 The the DT adeshpande3-github-io-4431 229 7 interesting interesting JJ adeshpande3-github-io-4431 229 8 idea idea NN adeshpande3-github-io-4431 229 9 for for IN adeshpande3-github-io-4431 229 10 me -PRON- PRP adeshpande3-github-io-4431 229 11 was be VBD adeshpande3-github-io-4431 229 12 that that DT adeshpande3-github-io-4431 229 13 of of IN adeshpande3-github-io-4431 229 14 using use VBG adeshpande3-github-io-4431 229 15 these these DT adeshpande3-github-io-4431 229 16 seemingly seemingly RB adeshpande3-github-io-4431 229 17 different different JJ adeshpande3-github-io-4431 229 18 RNN RNN NNP adeshpande3-github-io-4431 229 19 and and CC adeshpande3-github-io-4431 229 20 CNN CNN NNP adeshpande3-github-io-4431 229 21 models model NNS adeshpande3-github-io-4431 229 22 to to TO adeshpande3-github-io-4431 229 23 create create VB adeshpande3-github-io-4431 229 24 a a DT adeshpande3-github-io-4431 229 25 very very RB adeshpande3-github-io-4431 229 26 useful useful JJ adeshpande3-github-io-4431 229 27 application application NN adeshpande3-github-io-4431 229 28 that that IN adeshpande3-github-io-4431 229 29 in in IN adeshpande3-github-io-4431 229 30 a a DT adeshpande3-github-io-4431 229 31 way way NN adeshpande3-github-io-4431 229 32 combines combine VBZ adeshpande3-github-io-4431 229 33 the the DT adeshpande3-github-io-4431 229 34 fields field NNS adeshpande3-github-io-4431 229 35 of of IN adeshpande3-github-io-4431 229 36 Computer Computer NNP adeshpande3-github-io-4431 229 37 Vision Vision NNP adeshpande3-github-io-4431 229 38 and and CC adeshpande3-github-io-4431 229 39 Natural Natural NNP adeshpande3-github-io-4431 229 40 Language Language NNP adeshpande3-github-io-4431 229 41 Processing Processing NNP adeshpande3-github-io-4431 229 42 . . . adeshpande3-github-io-4431 230 1 It -PRON- PRP adeshpande3-github-io-4431 230 2 opens open VBZ adeshpande3-github-io-4431 230 3 the the DT adeshpande3-github-io-4431 230 4 door door NN adeshpande3-github-io-4431 230 5 for for IN adeshpande3-github-io-4431 230 6 new new JJ adeshpande3-github-io-4431 230 7 ideas idea NNS adeshpande3-github-io-4431 230 8 in in IN adeshpande3-github-io-4431 230 9 terms term NNS adeshpande3-github-io-4431 230 10 of of IN adeshpande3-github-io-4431 230 11 how how WRB adeshpande3-github-io-4431 230 12 to to TO adeshpande3-github-io-4431 230 13 make make VB adeshpande3-github-io-4431 230 14 computers computer NNS adeshpande3-github-io-4431 230 15 and and CC adeshpande3-github-io-4431 230 16 models model NNS adeshpande3-github-io-4431 230 17 smarter smarter RBR adeshpande3-github-io-4431 230 18 when when WRB adeshpande3-github-io-4431 230 19 dealing deal VBG adeshpande3-github-io-4431 230 20 with with IN adeshpande3-github-io-4431 230 21 tasks task NNS adeshpande3-github-io-4431 230 22 that that WDT adeshpande3-github-io-4431 230 23 cross cross VBP adeshpande3-github-io-4431 230 24 different different JJ adeshpande3-github-io-4431 230 25 fields field NNS adeshpande3-github-io-4431 230 26 . . . adeshpande3-github-io-4431 231 1 Spatial spatial JJ adeshpande3-github-io-4431 231 2 Transformer Transformer NNP adeshpande3-github-io-4431 231 3 Networks Networks NNP adeshpande3-github-io-4431 231 4 ( ( -LRB- adeshpande3-github-io-4431 231 5 2015 2015 CD adeshpande3-github-io-4431 231 6 ) ) -RRB- adeshpande3-github-io-4431 231 7                                 _SP adeshpande3-github-io-4431 231 8 Last last RB adeshpande3-github-io-4431 231 9 , , , adeshpande3-github-io-4431 231 10 but but CC adeshpande3-github-io-4431 231 11 not not RB adeshpande3-github-io-4431 231 12 least least JJS adeshpande3-github-io-4431 231 13 , , , adeshpande3-github-io-4431 231 14 let let VB adeshpande3-github-io-4431 231 15 ’s -PRON- PRP adeshpande3-github-io-4431 231 16 get get VB adeshpande3-github-io-4431 231 17 into into IN adeshpande3-github-io-4431 231 18 one one CD adeshpande3-github-io-4431 231 19 of of IN adeshpande3-github-io-4431 231 20 the the DT adeshpande3-github-io-4431 231 21 more more RBR adeshpande3-github-io-4431 231 22 recent recent JJ adeshpande3-github-io-4431 231 23 papers paper NNS adeshpande3-github-io-4431 231 24 in in IN adeshpande3-github-io-4431 231 25 the the DT adeshpande3-github-io-4431 231 26 field field NN adeshpande3-github-io-4431 231 27 . . . adeshpande3-github-io-4431 232 1 This this DT adeshpande3-github-io-4431 232 2 paper paper NN adeshpande3-github-io-4431 232 3 was be VBD adeshpande3-github-io-4431 232 4 written write VBN adeshpande3-github-io-4431 232 5 by by IN adeshpande3-github-io-4431 232 6 a a DT adeshpande3-github-io-4431 232 7 group group NN adeshpande3-github-io-4431 232 8 at at IN adeshpande3-github-io-4431 232 9 Google Google NNP adeshpande3-github-io-4431 232 10 Deepmind Deepmind NNP adeshpande3-github-io-4431 232 11 a a DT adeshpande3-github-io-4431 232 12 little little JJ adeshpande3-github-io-4431 232 13 over over IN adeshpande3-github-io-4431 232 14 a a DT adeshpande3-github-io-4431 232 15 year year NN adeshpande3-github-io-4431 232 16 ago ago RB adeshpande3-github-io-4431 232 17 . . . adeshpande3-github-io-4431 233 1 The the DT adeshpande3-github-io-4431 233 2 main main JJ adeshpande3-github-io-4431 233 3 contribution contribution NN adeshpande3-github-io-4431 233 4 is be VBZ adeshpande3-github-io-4431 233 5 the the DT adeshpande3-github-io-4431 233 6 introduction introduction NN adeshpande3-github-io-4431 233 7 of of IN adeshpande3-github-io-4431 233 8 a a DT adeshpande3-github-io-4431 233 9 Spatial Spatial NNP adeshpande3-github-io-4431 233 10 Transformer Transformer NNP adeshpande3-github-io-4431 233 11 module module NN adeshpande3-github-io-4431 233 12 . . . adeshpande3-github-io-4431 234 1 The the DT adeshpande3-github-io-4431 234 2 basic basic JJ adeshpande3-github-io-4431 234 3 idea idea NN adeshpande3-github-io-4431 234 4 is be VBZ adeshpande3-github-io-4431 234 5 that that IN adeshpande3-github-io-4431 234 6 this this DT adeshpande3-github-io-4431 234 7 module module NN adeshpande3-github-io-4431 234 8 transforms transform VBZ adeshpande3-github-io-4431 234 9 the the DT adeshpande3-github-io-4431 234 10 input input NN adeshpande3-github-io-4431 234 11 image image NN adeshpande3-github-io-4431 234 12 in in IN adeshpande3-github-io-4431 234 13 a a DT adeshpande3-github-io-4431 234 14 way way NN adeshpande3-github-io-4431 234 15 so so IN adeshpande3-github-io-4431 234 16 that that IN adeshpande3-github-io-4431 234 17 the the DT adeshpande3-github-io-4431 234 18 subsequent subsequent JJ adeshpande3-github-io-4431 234 19 layers layer NNS adeshpande3-github-io-4431 234 20 have have VBP adeshpande3-github-io-4431 234 21 an an DT adeshpande3-github-io-4431 234 22 easier easy JJR adeshpande3-github-io-4431 234 23 time time NN adeshpande3-github-io-4431 234 24 making make VBG adeshpande3-github-io-4431 234 25 a a DT adeshpande3-github-io-4431 234 26 classification classification NN adeshpande3-github-io-4431 234 27 . . . adeshpande3-github-io-4431 235 1 Instead instead RB adeshpande3-github-io-4431 235 2 of of IN adeshpande3-github-io-4431 235 3 making make VBG adeshpande3-github-io-4431 235 4 changes change NNS adeshpande3-github-io-4431 235 5 to to IN adeshpande3-github-io-4431 235 6 the the DT adeshpande3-github-io-4431 235 7 main main JJ adeshpande3-github-io-4431 235 8 CNN CNN NNP adeshpande3-github-io-4431 235 9 architecture architecture NN adeshpande3-github-io-4431 235 10 itself -PRON- PRP adeshpande3-github-io-4431 235 11 , , , adeshpande3-github-io-4431 235 12 the the DT adeshpande3-github-io-4431 235 13 authors author NNS adeshpande3-github-io-4431 235 14 worry worry VBP adeshpande3-github-io-4431 235 15 about about IN adeshpande3-github-io-4431 235 16 making make VBG adeshpande3-github-io-4431 235 17 changes change NNS adeshpande3-github-io-4431 235 18 to to IN adeshpande3-github-io-4431 235 19 the the DT adeshpande3-github-io-4431 235 20 image image NN adeshpande3-github-io-4431 235 21 before before IN adeshpande3-github-io-4431 235 22 it -PRON- PRP adeshpande3-github-io-4431 235 23 is be VBZ adeshpande3-github-io-4431 235 24 fed feed VBN adeshpande3-github-io-4431 235 25 into into IN adeshpande3-github-io-4431 235 26 the the DT adeshpande3-github-io-4431 235 27 specific specific JJ adeshpande3-github-io-4431 235 28 conv conv NN adeshpande3-github-io-4431 235 29 layer layer NN adeshpande3-github-io-4431 235 30 . . . adeshpande3-github-io-4431 236 1 The the DT adeshpande3-github-io-4431 236 2 2 2 CD adeshpande3-github-io-4431 236 3 things thing NNS adeshpande3-github-io-4431 236 4 that that WDT adeshpande3-github-io-4431 236 5 this this DT adeshpande3-github-io-4431 236 6 module module NN adeshpande3-github-io-4431 236 7 hopes hope VBZ adeshpande3-github-io-4431 236 8 to to TO adeshpande3-github-io-4431 236 9 correct correct VB adeshpande3-github-io-4431 236 10 are be VBP adeshpande3-github-io-4431 236 11 pose pose NN adeshpande3-github-io-4431 236 12 normalization normalization NN adeshpande3-github-io-4431 236 13 ( ( -LRB- adeshpande3-github-io-4431 236 14 scenarios scenario NNS adeshpande3-github-io-4431 236 15 where where WRB adeshpande3-github-io-4431 236 16 the the DT adeshpande3-github-io-4431 236 17 object object NN adeshpande3-github-io-4431 236 18 is be VBZ adeshpande3-github-io-4431 236 19 tilted tilt VBN adeshpande3-github-io-4431 236 20 or or CC adeshpande3-github-io-4431 236 21 scaled scale VBN adeshpande3-github-io-4431 236 22 ) ) -RRB- adeshpande3-github-io-4431 236 23 and and CC adeshpande3-github-io-4431 236 24 spatial spatial JJ adeshpande3-github-io-4431 236 25 attention attention NN adeshpande3-github-io-4431 236 26 ( ( -LRB- adeshpande3-github-io-4431 236 27 bringing bring VBG adeshpande3-github-io-4431 236 28 attention attention NN adeshpande3-github-io-4431 236 29 to to IN adeshpande3-github-io-4431 236 30 the the DT adeshpande3-github-io-4431 236 31 correct correct JJ adeshpande3-github-io-4431 236 32 object object NN adeshpande3-github-io-4431 236 33 in in IN adeshpande3-github-io-4431 236 34 a a DT adeshpande3-github-io-4431 236 35 crowded crowded JJ adeshpande3-github-io-4431 236 36 image image NN adeshpande3-github-io-4431 236 37 ) ) -RRB- adeshpande3-github-io-4431 236 38 . . . adeshpande3-github-io-4431 237 1 For for IN adeshpande3-github-io-4431 237 2 traditional traditional JJ adeshpande3-github-io-4431 237 3 CNNs cnn NNS adeshpande3-github-io-4431 237 4 , , , adeshpande3-github-io-4431 237 5 if if IN adeshpande3-github-io-4431 237 6 you -PRON- PRP adeshpande3-github-io-4431 237 7 wanted want VBD adeshpande3-github-io-4431 237 8 to to TO adeshpande3-github-io-4431 237 9 make make VB adeshpande3-github-io-4431 237 10 your -PRON- PRP$ adeshpande3-github-io-4431 237 11 model model NN adeshpande3-github-io-4431 237 12 invariant invariant JJ adeshpande3-github-io-4431 237 13 to to IN adeshpande3-github-io-4431 237 14 images image NNS adeshpande3-github-io-4431 237 15 with with IN adeshpande3-github-io-4431 237 16 different different JJ adeshpande3-github-io-4431 237 17 scales scale NNS adeshpande3-github-io-4431 237 18 and and CC adeshpande3-github-io-4431 237 19 rotations rotation NNS adeshpande3-github-io-4431 237 20 , , , adeshpande3-github-io-4431 237 21 you -PRON- PRP adeshpande3-github-io-4431 237 22 ’d ’d VBP adeshpande3-github-io-4431 237 23 need need VB adeshpande3-github-io-4431 237 24 a a DT adeshpande3-github-io-4431 237 25 lot lot NN adeshpande3-github-io-4431 237 26 of of IN adeshpande3-github-io-4431 237 27 training training NN adeshpande3-github-io-4431 237 28 examples example NNS adeshpande3-github-io-4431 237 29 for for IN adeshpande3-github-io-4431 237 30 the the DT adeshpande3-github-io-4431 237 31 model model NN adeshpande3-github-io-4431 237 32 to to TO adeshpande3-github-io-4431 237 33 learn learn VB adeshpande3-github-io-4431 237 34 properly properly RB adeshpande3-github-io-4431 237 35 . . . adeshpande3-github-io-4431 238 1 Let let VB adeshpande3-github-io-4431 238 2 ’s -PRON- PRP adeshpande3-github-io-4431 238 3 get get VB adeshpande3-github-io-4431 238 4 into into IN adeshpande3-github-io-4431 238 5 the the DT adeshpande3-github-io-4431 238 6 specifics specific NNS adeshpande3-github-io-4431 238 7 of of IN adeshpande3-github-io-4431 238 8 how how WRB adeshpande3-github-io-4431 238 9 this this DT adeshpande3-github-io-4431 238 10 transformer transformer NN adeshpande3-github-io-4431 238 11 module module NN adeshpande3-github-io-4431 238 12 helps help VBZ adeshpande3-github-io-4431 238 13 combat combat VB adeshpande3-github-io-4431 238 14 that that DT adeshpande3-github-io-4431 238 15 problem problem NN adeshpande3-github-io-4431 238 16 . . . adeshpande3-github-io-4431 239 1 The the DT adeshpande3-github-io-4431 239 2 entity entity NN adeshpande3-github-io-4431 239 3 in in IN adeshpande3-github-io-4431 239 4 traditional traditional JJ adeshpande3-github-io-4431 239 5 CNN CNN NNP adeshpande3-github-io-4431 239 6 models model NNS adeshpande3-github-io-4431 239 7 that that WDT adeshpande3-github-io-4431 239 8 dealt deal VBD adeshpande3-github-io-4431 239 9 with with IN adeshpande3-github-io-4431 239 10 spatial spatial JJ adeshpande3-github-io-4431 239 11 invariance invariance NN adeshpande3-github-io-4431 239 12 was be VBD adeshpande3-github-io-4431 239 13 the the DT adeshpande3-github-io-4431 239 14 maxpooling maxpooling NN adeshpande3-github-io-4431 239 15 layer layer NN adeshpande3-github-io-4431 239 16 . . . adeshpande3-github-io-4431 240 1 The the DT adeshpande3-github-io-4431 240 2 intuitive intuitive JJ adeshpande3-github-io-4431 240 3 reasoning reasoning NN adeshpande3-github-io-4431 240 4 behind behind IN adeshpande3-github-io-4431 240 5 this this DT adeshpande3-github-io-4431 240 6 layer layer NN adeshpande3-github-io-4431 240 7 was be VBD adeshpande3-github-io-4431 240 8 that that IN adeshpande3-github-io-4431 240 9 once once RB adeshpande3-github-io-4431 240 10 we -PRON- PRP adeshpande3-github-io-4431 240 11 know know VBP adeshpande3-github-io-4431 240 12 that that IN adeshpande3-github-io-4431 240 13 a a DT adeshpande3-github-io-4431 240 14 specific specific JJ adeshpande3-github-io-4431 240 15 feature feature NN adeshpande3-github-io-4431 240 16 is be VBZ adeshpande3-github-io-4431 240 17 in in IN adeshpande3-github-io-4431 240 18 the the DT adeshpande3-github-io-4431 240 19 original original JJ adeshpande3-github-io-4431 240 20 input input NN adeshpande3-github-io-4431 240 21 volume volume NN adeshpande3-github-io-4431 240 22 ( ( -LRB- adeshpande3-github-io-4431 240 23 wherever wherever WRB adeshpande3-github-io-4431 240 24 there there EX adeshpande3-github-io-4431 240 25 are be VBP adeshpande3-github-io-4431 240 26 high high JJ adeshpande3-github-io-4431 240 27 activation activation NN adeshpande3-github-io-4431 240 28 values value NNS adeshpande3-github-io-4431 240 29 ) ) -RRB- adeshpande3-github-io-4431 240 30 , , , adeshpande3-github-io-4431 240 31 it -PRON- PRP adeshpande3-github-io-4431 240 32 ’s ’ VBZ adeshpande3-github-io-4431 240 33 exact exact JJ adeshpande3-github-io-4431 240 34 location location NN adeshpande3-github-io-4431 240 35 is be VBZ adeshpande3-github-io-4431 240 36 not not RB adeshpande3-github-io-4431 240 37 as as RB adeshpande3-github-io-4431 240 38 important important JJ adeshpande3-github-io-4431 240 39 as as IN adeshpande3-github-io-4431 240 40 its -PRON- PRP$ adeshpande3-github-io-4431 240 41 relative relative JJ adeshpande3-github-io-4431 240 42 location location NN adeshpande3-github-io-4431 240 43 to to IN adeshpande3-github-io-4431 240 44 other other JJ adeshpande3-github-io-4431 240 45 features feature NNS adeshpande3-github-io-4431 240 46 . . . adeshpande3-github-io-4431 241 1 This this DT adeshpande3-github-io-4431 241 2 new new JJ adeshpande3-github-io-4431 241 3 spatial spatial JJ adeshpande3-github-io-4431 241 4 transformer transformer NN adeshpande3-github-io-4431 241 5 is be VBZ adeshpande3-github-io-4431 241 6 dynamic dynamic JJ adeshpande3-github-io-4431 241 7 in in IN adeshpande3-github-io-4431 241 8 a a DT adeshpande3-github-io-4431 241 9 way way NN adeshpande3-github-io-4431 241 10 that that IN adeshpande3-github-io-4431 241 11 it -PRON- PRP adeshpande3-github-io-4431 241 12 will will MD adeshpande3-github-io-4431 241 13 produce produce VB adeshpande3-github-io-4431 241 14 different different JJ adeshpande3-github-io-4431 241 15 behavior behavior NN adeshpande3-github-io-4431 241 16 ( ( -LRB- adeshpande3-github-io-4431 241 17 different different JJ adeshpande3-github-io-4431 241 18 distortions distortion NNS adeshpande3-github-io-4431 241 19 / / SYM adeshpande3-github-io-4431 241 20 transformations transformation NNS adeshpande3-github-io-4431 241 21 ) ) -RRB- adeshpande3-github-io-4431 241 22 for for IN adeshpande3-github-io-4431 241 23 each each DT adeshpande3-github-io-4431 241 24 input input NN adeshpande3-github-io-4431 241 25 image image NN adeshpande3-github-io-4431 241 26 . . . adeshpande3-github-io-4431 242 1 It -PRON- PRP adeshpande3-github-io-4431 242 2 ’s ’ VBZ adeshpande3-github-io-4431 242 3 not not RB adeshpande3-github-io-4431 242 4 just just RB adeshpande3-github-io-4431 242 5 as as RB adeshpande3-github-io-4431 242 6 simple simple JJ adeshpande3-github-io-4431 242 7 and and CC adeshpande3-github-io-4431 242 8 pre pre JJ adeshpande3-github-io-4431 242 9 - - VBN adeshpande3-github-io-4431 242 10 defined define VBN adeshpande3-github-io-4431 242 11 as as IN adeshpande3-github-io-4431 242 12 a a DT adeshpande3-github-io-4431 242 13 traditional traditional JJ adeshpande3-github-io-4431 242 14 maxpool maxpool NN adeshpande3-github-io-4431 242 15 . . . adeshpande3-github-io-4431 243 1 Let let VB adeshpande3-github-io-4431 243 2 ’s -PRON- PRP adeshpande3-github-io-4431 243 3 take take VB adeshpande3-github-io-4431 243 4 look look NN adeshpande3-github-io-4431 243 5 at at IN adeshpande3-github-io-4431 243 6 how how WRB adeshpande3-github-io-4431 243 7 this this DT adeshpande3-github-io-4431 243 8 transformer transformer NN adeshpande3-github-io-4431 243 9 module module NN adeshpande3-github-io-4431 243 10 works work VBZ adeshpande3-github-io-4431 243 11 . . . adeshpande3-github-io-4431 244 1 The the DT adeshpande3-github-io-4431 244 2 module module NN adeshpande3-github-io-4431 244 3 consists consist VBZ adeshpande3-github-io-4431 244 4 of of IN adeshpande3-github-io-4431 244 5 : : : adeshpande3-github-io-4431 244 6 A a DT adeshpande3-github-io-4431 244 7 localization localization NN adeshpande3-github-io-4431 244 8 network network NN adeshpande3-github-io-4431 244 9 which which WDT adeshpande3-github-io-4431 244 10 takes take VBZ adeshpande3-github-io-4431 244 11 in in RP adeshpande3-github-io-4431 244 12 the the DT adeshpande3-github-io-4431 244 13 input input NN adeshpande3-github-io-4431 244 14 volume volume NN adeshpande3-github-io-4431 244 15 and and CC adeshpande3-github-io-4431 244 16 outputs output NNS adeshpande3-github-io-4431 244 17 parameters parameter NNS adeshpande3-github-io-4431 244 18 of of IN adeshpande3-github-io-4431 244 19 the the DT adeshpande3-github-io-4431 244 20 spatial spatial JJ adeshpande3-github-io-4431 244 21 transformation transformation NN adeshpande3-github-io-4431 244 22 that that WDT adeshpande3-github-io-4431 244 23 should should MD adeshpande3-github-io-4431 244 24 be be VB adeshpande3-github-io-4431 244 25 applied apply VBN adeshpande3-github-io-4431 244 26 . . . adeshpande3-github-io-4431 245 1 The the DT adeshpande3-github-io-4431 245 2 parameters parameter NNS adeshpande3-github-io-4431 245 3 , , , adeshpande3-github-io-4431 245 4 or or CC adeshpande3-github-io-4431 245 5 theta theta NN adeshpande3-github-io-4431 245 6 , , , adeshpande3-github-io-4431 245 7 can can MD adeshpande3-github-io-4431 245 8 be be VB adeshpande3-github-io-4431 245 9 6 6 CD adeshpande3-github-io-4431 245 10 dimensional dimensional JJ adeshpande3-github-io-4431 245 11 for for IN adeshpande3-github-io-4431 245 12 an an DT adeshpande3-github-io-4431 245 13 affine affine NN adeshpande3-github-io-4431 245 14 transformation transformation NN adeshpande3-github-io-4431 245 15 . . . adeshpande3-github-io-4431 246 1 The the DT adeshpande3-github-io-4431 246 2 creation creation NN adeshpande3-github-io-4431 246 3 of of IN adeshpande3-github-io-4431 246 4 a a DT adeshpande3-github-io-4431 246 5 sampling sampling NN adeshpande3-github-io-4431 246 6 grid grid NN adeshpande3-github-io-4431 246 7 that that WDT adeshpande3-github-io-4431 246 8 is be VBZ adeshpande3-github-io-4431 246 9 the the DT adeshpande3-github-io-4431 246 10 result result NN adeshpande3-github-io-4431 246 11 of of IN adeshpande3-github-io-4431 246 12 warping warp VBG adeshpande3-github-io-4431 246 13 the the DT adeshpande3-github-io-4431 246 14 regular regular JJ adeshpande3-github-io-4431 246 15 grid grid NN adeshpande3-github-io-4431 246 16 with with IN adeshpande3-github-io-4431 246 17 the the DT adeshpande3-github-io-4431 246 18 affine affine NN adeshpande3-github-io-4431 246 19 transformation transformation NN adeshpande3-github-io-4431 246 20 ( ( -LRB- adeshpande3-github-io-4431 246 21 theta theta NN adeshpande3-github-io-4431 246 22 ) ) -RRB- adeshpande3-github-io-4431 246 23 created create VBN adeshpande3-github-io-4431 246 24 in in IN adeshpande3-github-io-4431 246 25 the the DT adeshpande3-github-io-4431 246 26 localization localization NN adeshpande3-github-io-4431 246 27 network network NN adeshpande3-github-io-4431 246 28 . . . adeshpande3-github-io-4431 247 1 A a DT adeshpande3-github-io-4431 247 2 sampler sampler NN adeshpande3-github-io-4431 247 3 whose whose WP$ adeshpande3-github-io-4431 247 4 purpose purpose NN adeshpande3-github-io-4431 247 5 is be VBZ adeshpande3-github-io-4431 247 6 to to TO adeshpande3-github-io-4431 247 7 perform perform VB adeshpande3-github-io-4431 247 8 a a DT adeshpande3-github-io-4431 247 9 warping warping NN adeshpande3-github-io-4431 247 10 of of IN adeshpande3-github-io-4431 247 11 the the DT adeshpande3-github-io-4431 247 12 input input NN adeshpande3-github-io-4431 247 13 feature feature NN adeshpande3-github-io-4431 247 14 map map NN adeshpande3-github-io-4431 247 15 . . . adeshpande3-github-io-4431 248 1 This this DT adeshpande3-github-io-4431 248 2 module module NN adeshpande3-github-io-4431 248 3 can can MD adeshpande3-github-io-4431 248 4 be be VB adeshpande3-github-io-4431 248 5 dropped drop VBN adeshpande3-github-io-4431 248 6 into into IN adeshpande3-github-io-4431 248 7 a a DT adeshpande3-github-io-4431 248 8 CNN CNN NNP adeshpande3-github-io-4431 248 9 at at IN adeshpande3-github-io-4431 248 10 any any DT adeshpande3-github-io-4431 248 11 point point NN adeshpande3-github-io-4431 248 12 and and CC adeshpande3-github-io-4431 248 13 basically basically RB adeshpande3-github-io-4431 248 14 helps help VBZ adeshpande3-github-io-4431 248 15 the the DT adeshpande3-github-io-4431 248 16 network network NN adeshpande3-github-io-4431 248 17 learn learn VB adeshpande3-github-io-4431 248 18 how how WRB adeshpande3-github-io-4431 248 19 to to TO adeshpande3-github-io-4431 248 20 transform transform VB adeshpande3-github-io-4431 248 21 feature feature NN adeshpande3-github-io-4431 248 22 maps map NNS adeshpande3-github-io-4431 248 23 in in IN adeshpande3-github-io-4431 248 24 a a DT adeshpande3-github-io-4431 248 25 way way NN adeshpande3-github-io-4431 248 26 that that WDT adeshpande3-github-io-4431 248 27 minimizes minimize VBZ adeshpande3-github-io-4431 248 28 the the DT adeshpande3-github-io-4431 248 29 cost cost NN adeshpande3-github-io-4431 248 30 function function NN adeshpande3-github-io-4431 248 31 during during IN adeshpande3-github-io-4431 248 32 training training NN adeshpande3-github-io-4431 248 33 . . . adeshpande3-github-io-4431 249 1 Why why WRB adeshpande3-github-io-4431 249 2 It -PRON- PRP adeshpande3-github-io-4431 249 3 ’s ’ VBZ adeshpande3-github-io-4431 249 4 Important important JJ adeshpande3-github-io-4431 249 5                                 _SP adeshpande3-github-io-4431 249 6 This this DT adeshpande3-github-io-4431 249 7 paper paper NN adeshpande3-github-io-4431 249 8 caught catch VBD adeshpande3-github-io-4431 249 9 my -PRON- PRP$ adeshpande3-github-io-4431 249 10 eye eye NN adeshpande3-github-io-4431 249 11 for for IN adeshpande3-github-io-4431 249 12 the the DT adeshpande3-github-io-4431 249 13 main main JJ adeshpande3-github-io-4431 249 14 reason reason NN adeshpande3-github-io-4431 249 15 that that WDT adeshpande3-github-io-4431 249 16 improvements improvement VBZ adeshpande3-github-io-4431 249 17 in in IN adeshpande3-github-io-4431 249 18 CNNs cnn NNS adeshpande3-github-io-4431 249 19 do do VBP adeshpande3-github-io-4431 249 20 n’t not RB adeshpande3-github-io-4431 249 21 necessarily necessarily RB adeshpande3-github-io-4431 249 22 have have VB adeshpande3-github-io-4431 249 23 to to TO adeshpande3-github-io-4431 249 24 come come VB adeshpande3-github-io-4431 249 25 from from IN adeshpande3-github-io-4431 249 26 drastic drastic JJ adeshpande3-github-io-4431 249 27 changes change NNS adeshpande3-github-io-4431 249 28 in in IN adeshpande3-github-io-4431 249 29 network network NN adeshpande3-github-io-4431 249 30 architecture architecture NN adeshpande3-github-io-4431 249 31 . . . adeshpande3-github-io-4431 250 1 We -PRON- PRP adeshpande3-github-io-4431 250 2 do do VBP adeshpande3-github-io-4431 250 3 n’t not RB adeshpande3-github-io-4431 250 4 need need VB adeshpande3-github-io-4431 250 5 to to TO adeshpande3-github-io-4431 250 6 create create VB adeshpande3-github-io-4431 250 7 the the DT adeshpande3-github-io-4431 250 8 next next JJ adeshpande3-github-io-4431 250 9 ResNet ResNet NNP adeshpande3-github-io-4431 250 10 or or CC adeshpande3-github-io-4431 250 11 Inception Inception NNP adeshpande3-github-io-4431 250 12 module module NN adeshpande3-github-io-4431 250 13 . . . adeshpande3-github-io-4431 251 1 This this DT adeshpande3-github-io-4431 251 2 paper paper NN adeshpande3-github-io-4431 251 3 implements implement VBZ adeshpande3-github-io-4431 251 4 the the DT adeshpande3-github-io-4431 251 5 simple simple JJ adeshpande3-github-io-4431 251 6 idea idea NN adeshpande3-github-io-4431 251 7 of of IN adeshpande3-github-io-4431 251 8 making make VBG adeshpande3-github-io-4431 251 9 affine affine NN adeshpande3-github-io-4431 251 10 transformations transformation NNS adeshpande3-github-io-4431 251 11 to to IN adeshpande3-github-io-4431 251 12 the the DT adeshpande3-github-io-4431 251 13 input input NN adeshpande3-github-io-4431 251 14 image image NN adeshpande3-github-io-4431 251 15 in in IN adeshpande3-github-io-4431 251 16 order order NN adeshpande3-github-io-4431 251 17 to to TO adeshpande3-github-io-4431 251 18 help help VB adeshpande3-github-io-4431 251 19 models model NNS adeshpande3-github-io-4431 251 20 become become VB adeshpande3-github-io-4431 251 21 more more RBR adeshpande3-github-io-4431 251 22 invariant invariant JJ adeshpande3-github-io-4431 251 23 to to IN adeshpande3-github-io-4431 251 24 translation translation NN adeshpande3-github-io-4431 251 25 , , , adeshpande3-github-io-4431 251 26 scale scale NN adeshpande3-github-io-4431 251 27 , , , adeshpande3-github-io-4431 251 28 and and CC adeshpande3-github-io-4431 251 29 rotation rotation NN adeshpande3-github-io-4431 251 30 . . . adeshpande3-github-io-4431 252 1 For for IN adeshpande3-github-io-4431 252 2 those those DT adeshpande3-github-io-4431 252 3 interested interested JJ adeshpande3-github-io-4431 252 4 , , , adeshpande3-github-io-4431 252 5 here here RB adeshpande3-github-io-4431 252 6 is be VBZ adeshpande3-github-io-4431 252 7 a a DT adeshpande3-github-io-4431 252 8 video video NN adeshpande3-github-io-4431 252 9 from from IN adeshpande3-github-io-4431 252 10 Deepmind Deepmind NNP adeshpande3-github-io-4431 252 11 that that WDT adeshpande3-github-io-4431 252 12 has have VBZ adeshpande3-github-io-4431 252 13 a a DT adeshpande3-github-io-4431 252 14 great great JJ adeshpande3-github-io-4431 252 15 animation animation NN adeshpande3-github-io-4431 252 16 of of IN adeshpande3-github-io-4431 252 17 the the DT adeshpande3-github-io-4431 252 18 results result NNS adeshpande3-github-io-4431 252 19 of of IN adeshpande3-github-io-4431 252 20 placing place VBG adeshpande3-github-io-4431 252 21 a a DT adeshpande3-github-io-4431 252 22 Spatial Spatial NNP adeshpande3-github-io-4431 252 23 Transformer Transformer NNP adeshpande3-github-io-4431 252 24 module module NN adeshpande3-github-io-4431 252 25 in in IN adeshpande3-github-io-4431 252 26 a a DT adeshpande3-github-io-4431 252 27 CNN CNN NNP adeshpande3-github-io-4431 252 28 and and CC adeshpande3-github-io-4431 252 29 a a DT adeshpande3-github-io-4431 252 30 good good JJ adeshpande3-github-io-4431 252 31 Quora Quora NNP adeshpande3-github-io-4431 252 32 discussion discussion NN adeshpande3-github-io-4431 252 33 . . . adeshpande3-github-io-4431 253 1 And and CC adeshpande3-github-io-4431 253 2 that that DT adeshpande3-github-io-4431 253 3 ends end VBZ adeshpande3-github-io-4431 253 4 our -PRON- PRP$ adeshpande3-github-io-4431 253 5 3 3 CD adeshpande3-github-io-4431 253 6 part part NN adeshpande3-github-io-4431 253 7 series series NN adeshpande3-github-io-4431 253 8 on on IN adeshpande3-github-io-4431 253 9 ConvNets ConvNets NNP adeshpande3-github-io-4431 253 10 ! ! . adeshpande3-github-io-4431 254 1 Hope hope VBP adeshpande3-github-io-4431 254 2 everyone everyone NN adeshpande3-github-io-4431 254 3 was be VBD adeshpande3-github-io-4431 254 4 able able JJ adeshpande3-github-io-4431 254 5 to to TO adeshpande3-github-io-4431 254 6 follow follow VB adeshpande3-github-io-4431 254 7 along along RB adeshpande3-github-io-4431 254 8 , , , adeshpande3-github-io-4431 254 9 and and CC adeshpande3-github-io-4431 254 10 if if IN adeshpande3-github-io-4431 254 11 you -PRON- PRP adeshpande3-github-io-4431 254 12 feel feel VBP adeshpande3-github-io-4431 254 13 that that IN adeshpande3-github-io-4431 254 14 I -PRON- PRP adeshpande3-github-io-4431 254 15 may may MD adeshpande3-github-io-4431 254 16 have have VB adeshpande3-github-io-4431 254 17 left leave VBN adeshpande3-github-io-4431 254 18 something something NN adeshpande3-github-io-4431 254 19 important important JJ adeshpande3-github-io-4431 254 20 out out RB adeshpande3-github-io-4431 254 21 , , , adeshpande3-github-io-4431 254 22 let let VB adeshpande3-github-io-4431 254 23 me -PRON- PRP adeshpande3-github-io-4431 254 24 know know VB adeshpande3-github-io-4431 254 25 in in IN adeshpande3-github-io-4431 254 26 the the DT adeshpande3-github-io-4431 254 27 comments comment NNS adeshpande3-github-io-4431 254 28 ! ! . adeshpande3-github-io-4431 255 1 If if IN adeshpande3-github-io-4431 255 2 you -PRON- PRP adeshpande3-github-io-4431 255 3 want want VBP adeshpande3-github-io-4431 255 4 more more JJR adeshpande3-github-io-4431 255 5 info info NN adeshpande3-github-io-4431 255 6 on on IN adeshpande3-github-io-4431 255 7 some some DT adeshpande3-github-io-4431 255 8 of of IN adeshpande3-github-io-4431 255 9 these these DT adeshpande3-github-io-4431 255 10 concepts concept NNS adeshpande3-github-io-4431 255 11 , , , adeshpande3-github-io-4431 255 12 I -PRON- PRP adeshpande3-github-io-4431 255 13 once once RB adeshpande3-github-io-4431 255 14 again again RB adeshpande3-github-io-4431 255 15 highly highly RB adeshpande3-github-io-4431 255 16 recommend recommend VBP adeshpande3-github-io-4431 255 17 Stanford Stanford NNP adeshpande3-github-io-4431 255 18 CS CS NNP adeshpande3-github-io-4431 255 19 231n 231n CD adeshpande3-github-io-4431 255 20 lecture lecture NN adeshpande3-github-io-4431 255 21 videos video NNS adeshpande3-github-io-4431 255 22 which which WDT adeshpande3-github-io-4431 255 23 can can MD adeshpande3-github-io-4431 255 24 be be VB adeshpande3-github-io-4431 255 25 found find VBN adeshpande3-github-io-4431 255 26 with with IN adeshpande3-github-io-4431 255 27 a a DT adeshpande3-github-io-4431 255 28 simple simple JJ adeshpande3-github-io-4431 255 29 YouTube YouTube NNP adeshpande3-github-io-4431 255 30 search search NN adeshpande3-github-io-4431 255 31 . . . adeshpande3-github-io-4431 256 1 Dueces duece NNS adeshpande3-github-io-4431 256 2 . . . adeshpande3-github-io-4431 257 1 Sources source NNS adeshpande3-github-io-4431 257 2 Tweet Tweet NNP adeshpande3-github-io-4431 257 3 Written Written NNP adeshpande3-github-io-4431 257 4 on on IN adeshpande3-github-io-4431 257 5 August August NNP adeshpande3-github-io-4431 257 6 24 24 CD adeshpande3-github-io-4431 257 7 , , , adeshpande3-github-io-4431 257 8 2016 2016 CD adeshpande3-github-io-4431 257 9 Please please UH adeshpande3-github-io-4431 257 10 enable enable VB adeshpande3-github-io-4431 257 11 JavaScript JavaScript NNP adeshpande3-github-io-4431 257 12 to to TO adeshpande3-github-io-4431 257 13 view view VB adeshpande3-github-io-4431 257 14 the the DT adeshpande3-github-io-4431 257 15 comments comment NNS adeshpande3-github-io-4431 257 16 powered power VBN adeshpande3-github-io-4431 257 17 by by IN adeshpande3-github-io-4431 257 18 Disqus Disqus NNP adeshpande3-github-io-4431 257 19 . . .