データ分析コンテストの 勝者解答から学ぶ

by @smly

Slide 1

Slide 1 text

σʔλ෼ੳίϯςετͷ  উऀղ౴͔ΒֶͿ ,PIFJ0[BLJ !TNMZ 3FDSVJU5FDIOPMPHJFT "EWBODFE5FDIOPMPHZ-BC "5- εςΞϥϘਓ޻஌ೳγϯϙδ΢Ϝ!ઍ༿޻ۀେֶ

Slide 2

Slide 2 text

ࣗݾ঺հ ϦΫϧʔτςΫϊϩδʔζ"5-ͷ4S4PGUXBSF&OHJOFFSͰ͢ɽ ˝,BHHMFΦϑΟεͰͷू߹ࣸਅ αϯϑϥϯγεί ɾ,BHHMFSྺ೥ (SBOENBTUFS )JHIFTUSBOLUI   ɾ,BHHMF 5PQpOJTIFTY 1SJ[FY   ɾ5PQ$PEFS.BSBUIPO.BUDI 8*/4   ɾ"$.,%% ,%%$VQTUQSJ[FXJOOFS ˝ϦΫϧʔτςΫϊϩδʔζ"5- ޿ඌ ˞લճͷߨԋ͔Β̏ݸ૿͑·ͨ͠

Slide 3

Slide 3 text

ࠓ೔ͷ࿩ w ͜͜਺೥ͷ,BHHMFͱσʔλ෼ੳίϯςετ w ͲͷΑ͏ʹͯ͠σʔλ෼ੳίϯςετͰউ͔ͭ w ࠷ۙͷίϯςετʹ͓͚Δղ๏ͷղઆ ࠓճ͸Ұൠ࿦ΑΓ΍΍֤࿦ʹ౿ΈࠐΜͩ࿩Λͯ͠Έ·͢ɻ ⼀一般⼀一般応⽤用 ͕͜͜ຊ୊

Slide 4

Slide 4 text

ࠓ೔ͷ࿩ w ͜͜਺೥ͷ,BHHMFͱσʔλ෼ੳίϯςετ w ͲͷΑ͏ʹͯ͠σʔλ෼ੳίϯςετͰউ͔ͭ w ࠷ۙͷίϯςετʹ͓͚Δղ๏ͷղઆ ࠓճ͸Ұൠ࿦ΑΓ΍΍֤࿦ʹ౿ΈࠐΜͩ࿩Λͯ͠Έ·͢ɻ

Slide 5

Slide 5 text

ৼΓฦΓɿ//๖ժɾπʔϧ։ൃՃ೤ w ೥͝Ζ͸4UF⒎FO3FOEMFͷMJC'. 'BDUPSJ[BUJPO .BDIJOFT<> ΍3JF+PIOTPOͷ3(' 3FHVMBSJ[FE(SFFEZ 'PSFTU<> ͳͲɺΞϧΰϦζϜͷ৽ن։ൃʹΑͬͯଞͷࠩΛ͚ͭ ͯউར͢Δ৔໘͕Կ౓͔͋ͬͨɻ NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017- w ػցֶश΍σʔλϚΠχϯάΛ׆༻͢Δਓ޻͕૿Ճɺ৽͍͠ಓ۩͕ ࣍ʑʹੜ·Εͨɻ-BTBHOF DYYOFU LFSBT 9(#PPTUͳͲɻ։ൃ ऀຊਓ͕ίϯςετΛ௨ͯ͡ιϑτ΢ΣΞͷϕϯνϚʔΫΛߦͬͨ Γએ఻͢ΔͳͲ͍ͯͨ͠ɻ

Slide 6

Slide 6 text

ৼΓฦΓɿ//๖ժɾπʔϧ։ൃՃ೤ w ೥ͷ*NBHF/FU *-473$ ʹ͓͚ΔϒϨʔΫεϧʔ͸େ͖ͳি ܸͰ͸͕͋ͬͨɺ·ͩ(16Λѻ͏؀ڥΛ੔උͨ͠Γ$6%"Λ࢖ͬ ͨϓϩάϥϜΛ,BHHMFͰ࢖͏Ϣʔβʔ͸গ਺೿Ͱ͋ͬͨɻ w ೥ʹ͸+PC4BMBSZ1SFEJDUJPO$IBMMFOHFͰ)JOUPOͷݚڀࣨ ͷ0#͕εύʔεಛ௃ྔΛೖྗͱͨ͠.-1Λ࣮૷ͯ͠ଞΛѹ౗ɻ NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017-

Slide 7

Slide 7 text

ৼΓฦΓɿϑϥϯέϯγϡλΠϯ w େྔʹϞσϧΛ࡞ΓελοΩϯά͢Δͱ͍͏ख๏͕ྲྀߦΔɻ༩͑Β ΕͨࢦඪͰগ͠Ͱ΋্ճ͍ͬͯΕ͹উར͢Δͱ͍͏ίϯςετͷ࢓ ૊Έ্ɺτοϓ૪͍͕ᷮࠩͰ͋Ε͹ϞσϧΛ૿΍͢͜ͱʹΠϯηϯ ςΟϒ͕ੜ·ΕΔɻ NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017- 出典：http://blog.kaggle.com/2016/04/08/homesite-quote-conversion-winners-write-up-1st-place-kazanova-faron-clobber/ w ϞσϧɺϞσϧͱΞϯα ϯϒϧʹ࢖͏Ϟσϧ਺ʹࡍݶ͕ͳ͘ ͳΓɺGSBOLFOTUFJOFOTFNCMFͳͲ ͱᎏ᎐͞ΕΔ͜ͱ΋ɻ

Slide 8

Slide 8 text

ৼΓฦΓɿϑϥϯέϯγϡλΠϯ ҰԠϑΥϩʔ͢Δͱɺ,BHHMF͸ίϯςετೖ৆ऀʹιϦϡʔγϣϯ ಺Ͱͷ஌ݟΛυΩϡϝϯτͱͯ͠༻ҙͤ͞ΔʢͲͷΑ͏ͳಛ௃ྔΛ࢖ ͍ɺؾ෇͖͕͔͋ͬͨͳͲʣɻ ΞϯαϯϒϧͰ͕ࠩͭ͘ίϯςετ͸ɺͦ΋ͦ΋λεΫઃܭ͕୯७͢ ͗Δͱ͔ɺ޻෉ͷ༨஍͕ͳ͍ͱ͔ɺଞͱࠩΛ͚ͭΔ͜ͱ͕೉͍͠ɻί ϯςετʹ޲͍͍ͯͳ͍͜ͱ͕ࠜຊతͳ໰୊Ͱ͋Δͱࢥ͏ɻ NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017-

Slide 9

Slide 9 text

ৼΓฦΓɿ//࣌୅ɾେن໛σʔλ w (16͕ݚڀऀ΍։ൃऀͳͲʹ޿͘ීٴ͠ɺ%-ؔ࿈ͷݚڀ։ൃͱ ͦͷ੒Ռ෺ͷϦϦʔεαΠΫϧ͕ߴ଎ʹͳΓɺଟ͘ͷλεΫͰԠ༻ ՄೳͰ͋Δͱೝࣝ͞ΕΔΑ͏ʹͳΔɻڝٕऀͷղ๏΋όϦΤʔγϣ ϯ͕๛͔ʹͳͬͨɻ w σʔληοτ΋େن໛ͳϚϧςΟϝσΟΞσʔλʢ৴߸ɺԻ੠ɺը ૾ɺಈըʣ͕૿͑ɺ5#ͷίϯςετσʔλΛμ΢ϯϩʔυ͢Δ ৔໘΋Ͱ͖ͨɻ NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017-

Slide 10

Slide 10 text

ৼΓฦΓɿ,FSOFMɾڧԽֶशίϯϖ w ܭࢉࢿݯͱਫ਼౓ͷτϨʔυΦϑ͕஫໨͞Ε͸͡ΊΔɻ5XP4JHNB .FSDBSJͳͲͷܭࢉࢿݯ΍࣮ߦ؀ڥʹ੍໿ΛՃ͑ͨίϯςετ͕ొ ৔ɻ΄͔5PQ$PEFS։࠵΋4QBDF/FU$IBMMFOHF΋ܭࢉࢿݯͷ੍ ໿ΛՃ͍͑ͯΔ w ೥ʹ͸ڧԽֶशͳͲͷίϯςετ΋ϓϥϯʹ͋Δͱද໌͞Εɺ ϢχʔΫͳλεΫͷίϯςετ͕૿͍͑ͯ͘ͱࢥΘΕΔɻ ग़యɿz3FWJFXJOHBOE1SFWJFXJOHCMPHLBHHMFDPNSFWJFXJOHBOEQSFWJFXJOH NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017-

Slide 11

Slide 11 text

ݴ͍͔ͨͬͨ͜ͱΛ੔ཧ͢Δͱ w Ξϯαϯϒϧͷڝ͍߹͍͸ෆໟ͕ͩɺ໰୊ઃఆʹ΋ݪҼ͕͋Δɻ w //ͷݚڀ։ൃ͕׆ൃʹͳ͓͔ͬͨ͛Ͱղ๏΋όϦΤʔγϣϯ͕๛ ෋ʹͳͬͨɻ w ܭࢉࢿݯͷ੍໿ΛೖΕͨΓɺڧԽֶशΛ୊ࡐʹ͢ΔͳͲมԽΛଓ͚ ͍ͯΔɻ NN 萌芽・  ツール開発過熱 2012-2014 フランケンシュタイン  アンサンブル 2014-2016? NN 時代・  ⼤大規模MMデータ 2016- Kernel コンペ・  強化学習？ 2017-

Slide 12

Slide 12 text

Slide 13

Slide 13 text

Ͳ͏΍ͬͨΒউͯΔͷ͔ େࡶ೺ʹߟ͑Δͱࠓ΋ੲ΋ඞཁͳ͜ͱ͸มΘΒͳ͍ɻ ɹ&%" ୳ࡧతσʔλ෼ੳ ɹ7BMJEBUJPO ద੾ͳϞσϧධՁ ɹ4VSWFSZ ݚڀ੒Ռ΍աڈղ౴͔ΒֶͿ

Slide 14

Slide 14 text

୳ࡧతσʔλ෼ੳ &%" ༷ʑͳ੾ΓޱͰσʔλΛݟͯɺՄࢹԽͯ͠ɺԾઆΛཱͯɺ໰୊ͷղ͖ํ Λߟ͑Δɻͱͯ΋ॏཁͳϑΣʔζɻ ظ଴͢Δ੒Ռɿ ɾυϝΠϯ஌ࣝ΍σʔλͷύλʔϯ͔ΒϞσϧɾղ๏Λઃܭ͢Δ ɾλεΫഎܠ͔ΒσʔλͷنଇੑΛਪଌͯ͠ಛ௃ྔΛ࡞ΔͳͲ 集計（集約）と可視化を道具として、アイディアを産む

Slide 15

Slide 15 text

ूܭɾू໿ॲཧɺՄࢹԽʹΑΔσʔλཧղ w খن໛ͳߏ଄ԽσʔλͰ͋Ε͹&YDFMͷϐϘοτςʔϒϧ΋༗༻ w ը૾σʔλͰ͋Ε͹ύλʔϯ͝ͱʹάϧʔϓԽͯ͠ฒ΂ͨΓɺ  ༧ଌϞσϧͷϩε͕େ͖͍ࣄྫɾখ͍͞ࣄྫΛ၆ᛌͯ͠ΈΔ w ඞཁʹԠͯ͡ՄࢹԽπʔϧΛࣗ࡞͢Δ ग़యɿIUUQTXXXLBHHMFDPNDUJUBOJDEJTDVTTJPO

Slide 16

Slide 16 text

σʔλΛՄࢹԽ͢ΔɿԾઆݕূͱΞΠσΟΞ w ಓ۩ͱͯ͠ՄࢹԽΛ׆༻͢Δ w ՄࢹԽˠΞΠσΟΞʢ࣭໰ʹͭͳ͕ΔύλʔϯΛ୳͢ʣ w ΞΠσΟΞˠՄࢹԽʢԾઆݕূʣ ग़యɿIUUQTXXXDPVSTFSBPSHMFBSODPNQFUJUJWFEBUBTDJFODF

Slide 17

Slide 17 text

σʔλΛՄࢹԽ͢ΔɿΤϥʔ෼ੳ Ӵ੕ը૾ͷTFNBOUJDTFHNFOUBUJPOͰ༧ଌϚεΫͷΤϯίʔσΟϯά ʹόά͕͋Δ͜ͱʹؾͮ͘͜ͱ͕ग़དྷͨέʔεɻ σʔλग़యɿ%TUM4BUFMMJUF*NBHFSZ'FBUVSF%FUFDUJPOzIUUQTXXXLBHHMFDPNDETUMTBUFMMJUFJNBHFSZGFBUVSFEFUFDUJPO #FGPSF "GUFS େ͖ͳ-PTTʹͳΔ

Slide 18

Slide 18 text

ద੾ͳϞσϧධՁ 1VCMJD-#Λ࢖͏΂͖͔ w σʔληοτͷ෼ׂ͝ͱʹద੾ͳ7BMJEBUJPOηοτΛ༻ҙ͢Δ w 1VCMJD-#͸ςετࣄྫ͕গͳ͍ͱ͖͸ಛʹؾΛ͚ͭΔ ,BHHMFͷϑϨʔϜϫʔΫʹ͓͚ΔաֶशΛߟ͑Δͱɺཧ࿦తʹ͸ʮ෼ׂ ͷׂ߹ʯͰ͸ͳ͘ʮςετࣄྫ਺ʯʹӨڹ͞ΕΔ໰୊Ͱ͋Δ͸ͣ SFG<> ग़యl5JQTBOEUSJDLTUPXJOLBHHMFEBUBTDJFODFDPNQFUJUJPOT lIUUQTXXXTMJEFTIBSFOFU%BSJVT#BSVBVTLBTUJQTBOEUSJDLTUPXJOLBHHMFEBUBTDJFODFDPNQFUJUJPOT

Slide 19

Slide 19 text

աֶशͷ௚ײతͳྫɿ#PPTUJOH"UUBDL ୯७Խͨ͠/ݸͷςετࣄྫͷೋ஋෼ྨ໰୊Λߟ͑Δɿ 正解予測 1 予測 2 ࣍ϕΫτϧΛ౰ͯΔɽධՁࢦඪ͸Τϥʔ཰ͱ͢Δɽ y 2 {0, 1}N sH(yi) yi 2 {0, 1}N ͋Δ༧ଌ ͷ1VCMJD-#είΞΛ ͱ͢Δɽ N 1 0 0 1 0 0 0 1 y = 1 1 1 0 0 1 1 1 y1 = 1 0 1 1 1 1 0 0 y2 = Public LB  Score 1 Public LB  Score 2 sH(y1) = 0.75 sH(y2) = 0.25 Public Private

Slide 20

Slide 20 text

#PPTUJOH"UUBDL<#MVN)BSEU`> Algorithm (Boosting Attack): 正解 1 0 0 1 0 0 0 1 y = 0 0 1 0 0 0 0 1 1 1 1 0 1 0 1 1 0 1 0 1 0 1 0 0 1 0 0 1 0 0 0 1 1 0 0 1 0 1 1 1 0 0 0 0 1 1 1 0 ランダムに予測のベクトルを作成する Public Private 1 0 1 1 1 1 0 0 1 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 0 0 1 1 0 0 1 1 0 0 1 0 0 0 1 1 1 1 0 1 0 0 0 0 ϥϯμϜͳ༧ଌ͔Β1VCMJD4DPSFͷྑ͍݁ՌΛબ୒͢Δ

Slide 21

Slide 21 text

#PPTUJOH"UUBDL<#MVN)BSEU`> 正解 1 0 0 1 0 0 0 1 y = 1 0 1 1 1 1 0 0 1 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 1 1 1 1 1 0 1 0 0 0 0 㱺 majority  voting 1 0 0 1 0 1 1 1 Public LB  Score sH(ˆ y) = 0.0 Algorithm (Boosting Attack): Public LB スコアの良いベクトル sH (yi) < 0.5 だけを選ぶ Public Private ϥϯμϜͳ༧ଌ͔Β1VCMJD4DPSFͷྑ͍݁ՌΛબ୒͢Δ

Slide 22

Slide 22 text

#PPTUJOH"UUBDL<#MVN)BSEU`> 正解 1 0 0 1 0 0 0 1 y = 1 0 1 1 1 1 0 0 1 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 1 1 1 1 1 0 1 0 0 0 0 㱺 majority  voting 1 0 0 1 0 1 1 1 Public LB  Score sH(ˆ y) = 0.0 Private LB Score が  良くなる保証は  何もない !! Algorithm (Boosting Attack): Public Private Public LB スコアの良いベクトル sH (yi) < 0.5 だけを選ぶ ϥϯμϜͳ༧ଌ͔Β1VCMJD4DPSFͷྑ͍݁ՌΛબ୒͢Δ

Slide 23

Slide 23 text

#PPTUJOH"UUBDL<#MVN)BSEU`> ࠷ऴతͳॱҐΛܾఆ͢Δ1SJWBUF-#είΞ͸վળ͠ͳ͍ɻςετࣄྫ਺͕ଟ͚ Ε͹ɺ1VCMJDͷΤϥʔ཰ͷ෼ࢄ͸খ͘͞ͳΔͷͰဃ཭΋খ͘͞ͳΔɻ エラー率（低いほど良い）予測結果の提出回数最終順位の評価対象であるPrivate LB は改善しないテスト事例例数が少ない  場合は特に注意する

Slide 24

Slide 24 text

αʔϕΠɿ෼໺ಠಛͷΞϓϩʔνΛֶͿ ˙੺֎෼ޫ๏ʹΑΔ౔৕ௐࠪͷ؍ଌσʔλʹର͢Δલॲཧྫ  ʢ"GSJDB4PJM1SPQFSUZ1SFEJDUJPO$IBMMFOHFʣɿ 4BWJU[LZ(PMBZpMUFS $POUJOVVN3FNPWBM %JTDSFUFXBWFMFU USBOTGPSNT 'JSTU%FSJWBUJWFT 6OTVQFSWJTFE'FBUVSF4FMFDUJPO -PH USBOTGPSNʜ ˙೴࣓ਤه࿥ͷલॲཧɾಛ௃ྔநग़ %FD.FH$IBMMFOHF ɿ PSEFS#VUUFSXPSUICBOEQBTTpMUFS 4QBUJBM'JMUFSJOH 3JFNBOOJBO.FBO 5BOHFOU4QBDF.BQQJOHʜ IUUQTXXXLBHHMFDPNDBGTJTTPJMQSPQFSUJFTEJTDVTTJPO IUUQTHJUIVCDPNBMFYBOESFCBSBDIBOU%FD.FH  IUUQTXFCBSDIJWFPSHXFCIUUQBMFYBOESFCBSBDIBOUPSHXQDPOUFOUVQMPBETEPDVNFOUBUJPOQEG

Slide 25

Slide 25 text

αʔϕΠɿਂ૚ֶशͷ࣮ફతͳݚڀ੒Ռ ࢖͑ͦ͏ͳ΋ͷ͸νΣοΫ͢Δɻ̍͜͜ʙ̎೥ͷ੒Ռ͕࢖ΘΕΔ͜ͱ΋ ௝͘͠ͳ͍ɻຊεϥΠυͰ঺հ͢Δ্Ґղ๏͕ࢀߟʹͨ͠จݙͷྫɿ [Zhang+ ’17] "mixup: Beyond Empirical Risk Minimization", https:// arxiv.org/abs/1710.09412 [Miech+ '17] "Learnable pooling with Context Gating for video classiﬁcation", https://arxiv.org/abs/1706.06905 [Tian+ '16] "Detecting Text in Natural Image with Connectionist Text Proposal Network”, In Proc. of ECCV ‘16 https://arxiv.org/abs/1609.03605 [Qi+ '17] "PointNet: Deep Learning on Point Sets for 3D Classiﬁcation and Segmentation”, In Proc. of the CVPR ‘17 https://arxiv.org/abs/1612.00593 [Ma+ '17] "Multi-View Deep Learning for Consistent Semantic Mapping with RGB-D Cameras”, In Proc. of the IROS ‘17 https://arxiv.org/abs/1703.08866

Slide 26

Slide 26 text

Slide 27

Slide 27 text

࠷ۙͷίϯςετʹ͓͚Δղ๏ͷղઆ ̎ͭͷίϯςετʹ͓͚Δղ๏Λ঺հ͢Δɻ ~ 2018/01/23 ~ 2017/12/14

Slide 28

Slide 28 text

࠷ۙͷίϯςετʹ͓͚Δղ๏ͷղઆ ̎ͭͷίϯςετʹ͓͚Δղ๏Λ঺հ͢Δɻ ~ 2017/12/14

Slide 29

Slide 29 text

γϣοϐϯάαΠτͷը૾͔ΒଟΫϥε෼ྨ λεΫɿෳ਺ͷը૾͕ఴ͑ΒΕͨ঎඼*%͝ͱʹଟΫϥε෼ྨ ධՁࢦඪɿ"DDVSBDZ ঎඼*%͝ͱʹ̍ʙ̐ຕͷෳ਺ը૾͕༩͑ΒΔɻ ͜ͷෳ਺ը૾͔ΒΫϥεʹ෼ྨ͢Δ ग़యɿIUUQTXXXDEJTDPVOUDPNFMFDUSPNFOBHFSQFUJUEFKFVOFSDBGFQIJMJQTTFOTFPPSJHJOBMIEBSHFOUGQIJIEIUNM

Slide 30

Slide 30 text

σʔληοτͷجຊ৘ใΛ͓͑͞Δ Ϋϥε਺͸Ϋϥεʢ֊૚ͷΧςΰϦ͕ఆٛ͞Ε͍ͯΔʣ ɾඇৗʹTLFXͳΫϥε෼෍Ͱ͋Δ ɾԾઆɿ঎඼*%͸ར༻Ͱ͖ͳ͍͔ʁ ɾԾઆɿ܇࿅ࣄྫͱςετࣄྫͰಉ͡ը૾͕ଘࡏ͠ͳ͍͔ʁ جຊํ਑ͱͯ͠$//ϞσϧΛը૾୯ҐͰೖྗֶͯ͠श͠ɺ ঎඼͝ͱʹ݁ՌΛฏۉͰू໿ͯ͠༧ଌ͢Δɻ ܇࿅ࣄྫ (# ঎඼ F ςετࣄྫ (# ঎඼ F

Slide 31

Slide 31 text

Ծઆɿ঎඼*%͸ར༻Ͱ͖ͳ͍͔ʁ ݁࿦ɿಛఆͷΫϥε͚ͩ঎඼*%ʹنଇੑ͕͋Δɻ ϥϕϧͷݱΕͳ͍঎඼*%஋ҬΛϒϥοΫϦετͱͯ͠ɺ  ผͷϥϕϧΛ౴͑ΔώϡʔϦεςΟΫεΛ࢖͏ɻ ܇࿅ࣄྫશମͷ෼෍ සग़൪ͷΫϥεͷ෼෍ ঎඼*%ͷঢॱʢCJOOJOHQFSʣˠ සग़൪ͷΫϥεͷ෼෍ සग़൪ͷΫϥεͷ෼෍

Slide 32

Slide 32 text

Ծઆɿ܇࿅ςετࣄྫͰಉ͡ը૾͕ଘࡏ .%Λ͢΂ͯͷը૾Ͱܭࢉͨ݁͠Ռɺ঎඼ʹରԠ͢Δ͢΂ͯͷը૾͕׬શʹҰ க͢Δ৔߹ͱɺҰ෦ͷը૾͚ͩ׬શʹҰக͢Δ৔߹͕͋Δɻ )BTIJOHʹΑΔྨࣅը૾΋ಉ༷ͷ݁Ռɻ 類類似画像同⼀一画像

Slide 33

Slide 33 text

$//Ϟσϧ (16࢖༻཰Λҙࣝ͠ͳ͍ͱɺ͍ͭ·Ͱܦͬͯ΋܇࿅͕ऴΘΒͳ͍ɻ #40/ CJOBSZGPSNBU.POHP%# Ͱ֨ೲ͞Ε͍ͯΔσʔλΛޮ཰తʹ 73".ʹઈ͑ؒͳ͘సૹͯ͠$6%"ʹܭࢉ͚ͤͭͮ͞Δඞཁ͕͋Δɻ $16ϘτϧωοΫ(16ຕ਺Λ૿΍͢ͱҰ౓ʹॲཧͰ͖Δࣄ ྫ਺͕૿͑Δɻ͢Δͱࠓ౓͸%BUBBVHNFOUBUJPOͳͲͷը૾ ॲཧͷܭࢉίετ͕૿͑ΔɻϚϧνεϨουͰσʔλΛॲཧ͢ ΔͳͲͷ޻෉͕ඞཁͱͳΔ *0ϘτϧωοΫେ༰ྔ44%Λߪೖ͢Δ͔ɺ3". UNQGT ʹ #40/ϑΝΠϧΛσϓϩΠ͢ΔͳͲͯ͠ղফ͢Δඞཁ͕͋Δ

Slide 34

Slide 34 text

watch -n5 nvidia-smi͢ΔͳͲͯ͠໨ࢹͰ֬ೝɻ ܭଌ΍֬ೝΛ͠ͳ͍ͱɺͳ͔ͳ͔ؾ͕ͭ͘͜ͱ͕ग़དྷͳ͍ɻ ϓϩάϥϜͰ*0 $16 (16ͦΕͧΕʹ͔͔࣮ͭߦ࣌ؒΛܭଌ͢Δͱߋʹྑ͍ɻ

Slide 35

Slide 35 text

$//Ϟσϧ ֶशʹ͍ͭͯɻը૾ຕ਺͕ඇৗʹଟ͍ͷͰɺҰճͷࢼߦʹҰिؒ଴ͨ͞ΕΔ͜ ͱ΋͋Γɺඇৗʹ೉ّͨ͠ͱ͜Ζɻ ɾ܇࿅ࣄྫ͕*NBHF/FUΑΓଟ͍ͷͰɺQSFUSBJOFENPEFMͷֶशͨ͠ಛ௃ ϚοϓΛյͭ͢΋ΓͰେ͖ͳMFBSOJOHSBUFF͔Β4(%Ͱֶश ɾे෼ʹWBMJEBUJPOTFUͰͷMPHMPTT͕ανͬͨͱ൑அͰ͖Δ·ͰMFBSOJOH SBUFΛԼ͛ͳ͍ɻMFBSOJOHSBUFΛ͙͢ʹԼ͛Δͱޙ൒ͷվળ͕খ͍͞ ɾ࣮ݧ৚͕݅ಉ͡Ͱ͋Ε͹ɺ࣌ؒઅ໿ͷͨΊ FQPDIͰ࣮ݧΛଧͪ੾Δ

Slide 36

Slide 36 text

$//Ϟσϧ ˙νϟϯωϧ$// dຕͱը૾ͷຕ਺͕ݶΒΕ͍ͯͨͨΊɺDIBOOFMTͷ DIBOOFMTΛೖྗͱ͢Δ$//ϞσϧΛ࡞੒ͨ͠ɻ ݁Ռ͸͋·Γྑ͘ͳ͔͕ͬͨɺΞϯαϯϒϧͰͷEJWFSTJUZΛ্͛ Δ͜ͱʹ͸੒ޭͯ͠είΞΛগ͚ͩ͠ԡ্͛ͨ͠ɻ ˙ઐ໳Ո$// Τϥʔ෼ੳͨ͠ͱ͜Ζɺ͍͔ͭ͘ͷେΧςΰϦͰਫ਼౓͕ѱ͍ͨΊɺ ͜ΕΒʹಛՁͨ͠Ϟσϧʢઐ໳ՈϞσϧʣΛ༻ҙɻҰ෦ͷ༧ଌ݁Ռ Λஔ͖׵͑ͨɻ

Slide 37

Slide 37 text

4JOHMFNPEFMT #FTUTJOHMFNPEFM͸1VCMJD-#ͰҐ νʔϜத ɻ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ ResNet101-12ch ResNet101 Inception V3 DPN-92 InceptionRes V2 Expert

Slide 38

Slide 38 text

Ξϯαϯϒϧ ॏΈ෇͖ฏۉ ςετࣄྫ਺͕े෼ʹେ͖ͷͰɺ1VCMJD-#Λ֬ೝͭͭ͠ௐ੔ɻ Ξϯαϯϒϧ1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ 1VCMJD-# Ґ૬౰ ResNet101-12ch ResNet101 Inception V3 InceptionRes V2 DPN-92 Expert

Slide 39

Slide 39 text

UIϞσϧͷΞϯαϯϒϧЋ ಉҰ.%ͷ܇࿅ࣄྫΛ࢖͍ɺ$//ϞσϧͷTPGUNBYΛิਖ਼͢Δɿ

Slide 40

Slide 40 text

UIϞσϧͷΞϯαϯϒϧЋ (1) 訓練事例例の  正例例の割合 (2) Softmax (1), (2) の重み付き平均とみなす ಉҰ.%ͷ܇࿅ࣄྫΛ࢖͍ɺ$//ϞσϧͷTPGUNBYΛิਖ਼͢Δɿ

Slide 41

Slide 41 text

UIϞσϧͷΞϯαϯϒϧЋ Ґ૬౰ ˠ Ґ૬౰ ʴΫϥε෼෍͔Βܭࢉͨ͠ϒϥοΫϦετͷώϡʔϦεςΟΫεΛద༻ (1) 訓練事例例の  正例例の割合 (2) Softmax (1), (2) の重み付き平均とみなす ಉҰ.%ͷ܇࿅ࣄྫΛ࢖͍ɺ$//ϞσϧͷTPGUNBYΛิਖ਼͢Δɿ

Slide 42

Slide 42 text

UIಛ௃நग़ͯ͠ू໿ˍߋʹֶश ϕʔεϞσϧͷCPUUMFOFDLCMPDLͷग़ྗΛ࢖ͬͯ঎඼୯Ґʹू໿ˍֶ शɻ0$3 &"45<>$3//<> ͯ͠ςΩετ৘ใΛநग़ɻ ͜ΕΒΛೖྗͱͯ͠.-1 3// /FU7MBE<>ͳͲͷτοϓϨϕϧϞ σϧΛֶशɻ࠷ޙʹϕʔεϞσϧͱτοϓϨϕϧϞσϧͰΞϯαϯϒϧɻ w ϕʔεϞσϧ͸*ODFQUJPO3FTOFUW 3FTOFU 4&*ODFQUJPO7 9DFQUJPO w ̐ຕͷը૾Λೖྗͯ͠ಘͨ#PUUMFOFDLGFBUVSFTΛ঎඼͝ͱʹάϧʔϓͯ͠  τοϓϨϕϧϞσϧͷೖྗͱ͢Δ w /FU7MBEͷ1PPMJOHֶशͱ(BUJOH͸ɺաڈίϯϖl:PVUVCF.,BHHMF -BSHF4DBMF7JEFPVOEFSTUBEJOHDIBMMFOHFzͷղ๏Ͱ΋࢖ΘΕɺ-061&ͱ͠ ͯπʔϧϘοΫε͕ެ։͞Ε͍ͯΔ ग़యɿIUUQTXXXLBHHMFDPNDDEJTDPVOUJNBHFDMBTTJpDBUJPODIBMMFOHFEJTDVTTJPO

Slide 43

Slide 43 text

UIೋ஋෼ྨ໰୊ʹม׵ ͋Γ͑ͦ͏ͳ༧ଌΫϥεͷީิlQPTTJCMF@DMBTT@JEzͱ͍͏ಛ௃ྔΛ࡞Γɺ  ଟΫϥε෼ྨ໰୊Λೋ஋෼ྨ໰୊ʹม׵ɻͦͯ͠9(#PPTUͳͲͰֶशɻ w.%ϨϕϧͰҰக͢Δը૾ͷ࠷සग़Ϋϥε*%Λ  ΧςΰϦΧϧಛ௃ྔͱͨ͠ w.%ϨϕϧͰҰக͢Δը૾ͷॏෳ਺Λྔతಛ௃ྔͱͨ͠ ෳ਺ຕͷը૾Λ঎඼͝ͱʹ·ͱΊΔू໿ͷϋϯυϦϯάͱɺ .%ϨϕϧͷҰகͷ৘ใΛϞσϧʹ૊ΈࠐΜͰ͍Δɻ ग़యɿIUUQTXXXLBHHMFDPNDDEJTDPVOUJNBHFDMBTTJpDBUJPODIBMMFOHFEJTDVTTJPO

Slide 44

Slide 44 text

OE.-1aXHBUFETDBMJOH w CPUUMFOFDLCMPDLͷDPOWYY͔Βը૾෼YΛநग़ w ֤TUSFBNT͔Βಛ௃ΛDPODBUͨ͠YΛೖྗͱͯ͠.-1Ͱֶश w .VMUJWJFXQPPMJOH<>ʹ͍͕ۙɺσʔλ͕େ͖͍ͷͰֶश͸TUBHFʹ෼ׂ qJQʹΑΔσʔλ֦ு͕͋ΔͷͰ  4&3FT/F9U͸PVUQVUEJNY͕̎ͭͰYY ग़యIUUQTXXXLBHHMFDPNDDEJTDPVOUJNBHFDMBTTJpDBUJPODIBMMFOHFEJTDVTTJPO

Slide 45

Slide 45 text

OE.-1aXHBUFETDBMJOH w தؒ੒Ռ෺ɿը૾ຕ਺Yಛ௃਺Yɻ͜ΕʹHBUFETDBMJOH.-1ద༻ w ը૾ʹ͖ͭͷEJN͔Β4&4DBMFNPEVMFͰը૾Λબ୒ͯ͠Ճࢉ͢Δ w ը૾ͷ͏ͪຕ͚͕ͩॏཁͱ͍͏ࣄྫ͕͋ΔͷͰ༗ޮͦ͏ SFG1PJOU/FU<> ग़యIUUQTXXXLBHHMFDPNDDEJTDPVOUJNBHFDMBTTJpDBUJPODIBMMFOHFEJTDVTTJPO EJN EJN EJN EJN EJN ঎඼ը૾ Y 4DBMJOH ը૾બ୒ "%% } + ○ * * * * .-1΁GPSXBSE

Slide 46

Slide 46 text

Slide 47

Slide 47 text

OE.-1aXHBUFETDBMJOH w ΄͔ʹ΋ɺதؒ૚ͷDPODBU෦෼Ͱ.JYVQBVHNFOUBUJPO<>͢ΔϞσϧͳͲ w ͦΕͧΕඍົʹҟͳΔτοϓϨϕϧϞσϧܭ̐ݸΛΞϯαϯϒϧ ग़యIUUQTXXXLBHHMFDPNDDEJTDPVOUJNBHFDMBTTJpDBUJPODIBMMFOHFEJTDVTTJPO

Slide 48

Slide 48 text

TUදݱྗΛ্͛ΔͨΊʹBSDIมߋ ࠷ॳ͸খ͞ͳ3FT/FUͰ࣮ݧΛ͘Γฦͯ͠Ϟσϧͷݕূɻ ଟ͘ͷ*NBHF/FUQSFUSBJOFE͸Ϋϥεͷઃܭ͚ͩͲࠓճ͸Ϋϥε਺͸ ΫϥεɻΞʔΩςΫνϟͷදݱྗʹݶք͕͋Γɺ͜ΕΛղফ͢ΔมߋΛ௥Ճɻ νϟϯωϧ਺Λ૿΍ͨ͢ΊʹYDPOWMBZFSΛ௥Ճͯ͠νϟϯωϧ਺Λ ͔Βʹ૿΍͢ɻ7((OFUʹश͍GDMBZFS΋௥Ճ͢Δɻ YDPOW GD GD #FGPSF "GUFS ௥Ճ ௥Ճ มߋ

Slide 49

Slide 49 text

TUෳ਺ը૾Λ࿈݁ͯ͠୯Ұը૾ʹ ෳ਺ը૾ͷऔΓѻ͍ɿ࿈݁ͯ͠Ұͭͷը૾ʹ͢Δ ঎඼ຕͷ෦෼ηοτ ঎඼ຕͷ෦෼ηοτ ঎඼ຕͷ෦෼ηοτ ঎඼͝ͱͷը૾ຕ਺ͰσʔληοτΛ෼ׂʢ̐௨Γʣɻը૾ຕ਺͝ͱͷ σʔληοτΛ࡞੒ͯ͠ɺ୯Ұը૾Ͱֶशͨ͠Ϟσϧ͔ΒpOFUVOJOH

Slide 50

Slide 50 text

TU0$3ͯ͠NVMUJJOQVU$// $51/<>ͰςΩετྖҬͷநग़ɺ$3//<>ͰςΩετͷநग़Λߦ͏ɻ.VMUJ JOQVU$//Ͱ$3//GFBUVSFTͱ3FT/FUͷ'$GFBUVSFTΛ࿈݁͢Δɻ   தؒϑΝΠϧ͸VJOUʹྔࢠԽʢͯ͠DBTUʣͯ͠TQBSTFNBUSJYͱͯ͠  อଘ͢Δɻ.#͕.#ҎԼʹͳΔɻ $51/ $3// ରԠ͢Δຕ਺ͷ  ઐ༻3FT/FU .VMUJJOQVU $// ༧ଌ ೖྗը૾ <>%FUFDUJOH5FYUJO/BUVSBM*NBHFXJUI$POOFDUJPOJTU5FYU1SPQPTBM/FUXPSL  <>"O&OEUP&OE5SBJOBCMF/FVSBM/FUXPSLGPS*NBHFCBTFE4FRVFODF3FDPHOJUJPOBOE*UT"QQMJDBUJPOUP4DFOF5FYU3FDPHOJUJPO

Slide 51

Slide 51 text

࠷ۙͷίϯςετʹ͓͚Δղ๏ͷղઆ ̎ͭͷίϯςετʹ͓͚Δղ๏Λ঺հ͢Δɻ ~ 2018/01/23

Slide 52

Slide 52 text

4UBUPJM$$03&*DFCFSH λεΫɿ߹੒։ޱϨʔμʔ 4"3 ͷ؍ଌσʔλ͔Βණࢁ͔ધഫ͔ೋ஋෼ྨ͢Δ ධՁࢦඪɿ-PHMPTT ภ޲೾ͷೖࣹ֯౓ʢྔతม਺ʣ ภ޲೾ͷ؍ଌ஋ ภ޲೾ͷ؍ଌ஋ ණࢁ͔ધ͔ʢ໨ඪม਺ʣ

Slide 53

Slide 53 text

氷⼭山船舶

Slide 54

Slide 54 text

ͲͪΒ͕ྲྀණʁ

Slide 55

Slide 55 text

氷⼭山氷⼭山船舶 ࠔ೉ͳࣄྫ͕͋Δɻ աֶशʹ஫ҙ͕ඞཁɻ

Slide 56

Slide 56 text

σʔλʹର͢ΔཧղΛਂΊΔ ߹੒։ޱϨʔμʔ͸ೖࣹ֯ʢ*ODJEFODFBOHMFʣΛௐ੔ͯ͠؍ଌ͢ΔͷͰɺ  ΄΅ಉ஋ͷάϧʔϓ͕Ͱ͖Δͱߟ͑Δ͜ͱ͕ग़དྷΔɻ 1 2 3 4 5 6 ic=44.1923 ic=44.1925 ic=44.1919 ic=41.4321 ic=41.4315 ic=41.4317 ग़యʢࠨʣɿIUUQXXXBTDDTBHDDBFOHTBUFMMJUFTSBEBSTBUDPNQPOFOUTBTQ  ʢӈʣ*NBHFSZ*#$"0 -BOETBU$PQFSOJDVT .BQEBUB(PPHMF ʢ˞ૉਓʹΑΔߟ࡯ͳͷͰɺؒҧ͍͕͋Δ͔΋͠Ε·ͤΜʣ

Slide 57

Slide 57 text

σʔληοτը૾͸͋Β͔͡ΊDSPQͨ͠΋ͷ $SPQલͷΠϝʔδਤ ༩͑ΒΕͨը૾ ic=44.1923 ic=44.1925 ic=44.1919 ic=41.4321 ic=41.4315 ic=41.4317 ߹੒։ޱϨʔμʔͷ؍ଌྖҬͷΠϝʔδਤ ೖࣹ֯ ͜ͷྖҬ͸͢΂ͯ ೖࣹ֯ Ծઆɿೖࣹ֯ͷগ਺දܻࣔ͢΂͕ͯ׬શҰக͢Δ৔߹ɺ ݩʑ͕ಉ͡؍ଌྖҬ˙͔Β੾ΓऔͬͨDSPQ͞Εͨը૾˙Ͱ༩͑ΒΕΔ ༩͑ΒΕͨ؍ଌσʔλ͸ɺધഫ͋Δ͍͸ණࢁ͕த৺Β͖͠΋ͷ͕ ͍ࣸͬͯΔɻ͋Β͔͡ΊDSPQ͞Εͨը૾ɻ

Slide 58

Slide 58 text

Ծઆೖࣹ͕֯Ұக͢Δάϧʔϓ͸ಉ͡ϥϕϧ ೖࣹ͕֯׬શҰக͢Δࣄྫάϧʔϓ͸ɺ΄΅͓ͳ͡ϥϕϧʹͳΔɻ എܠΛߟ͑Δͱଥ౰ͳ݁Ռɻ ԫ৭ʹ͢΂ͯණࢁ ΦϨϯδʹ͢΂ͯધഫ ੺ʹණࢁͱધഫ͕͍ࠞͬͯ͟Δ

Slide 59

Slide 59 text

ώϡʔϦεςΟΫεͰ܇࿅ࣄྫΛ௥Ճ ʮֶशͨ͠$//Ϟσϧʯͷ"DDVSBDZΑΓ΋  ʮ܇࿅ࣄྫʹ͋Δɺ׬શʹಉҰͷೖࣹ֯ͷ؍ଌࣄྫʯ  Λ୳ͯ͠࠷සग़ϥϕϧΛ౴͑ͨ΄͏͕"DDVSBDZྑ͍ ɻ ܇࿅ࣄྫ਺ɿˠ ςετࣄྫ਺ɿˠ ࣮ࡍʹ͸μϛʔը૾Ͱਫ૿͠͞Ε͍ͯΔͷͰධՁͷର৅ͱͳΔςετࣄྫ͸ ͱͯ΋গͳ͍ɻςετࣄྫ΋গͳ͍ͷͰɺ1VCMJD-#Λࢀߟʹ͢Δͷ͸ةݥ ܇࿅ࣄྫ͕গͳ͍ͷͰɺ্ͷ࠷සग़ϥϕϧΛ͢΂ͯ܇࿅ࣄྫʹ௥Ճͨ͠ɻ

Slide 60

Slide 60 text

$//Ϟσϧͷ࡞੒ w ը૾͚ͩΛ࢖ͬͨϞσϧΛ༻ҙ͢Δʢྔతม਺Λಉ࣌ʹֶश͠ͳ͍ʣ w ܇࿅ࣄྫ਺͕গͳ͍ͷͰGPME$7ͰϞσϧΛධՁˍνϡʔχϯά w Ϟσϧ͸૚$POWPMVUJPO-BZFS͚ͩͰे෼ͦ͏  ʢม͑ͯ΋େࠩͳ͍ʣ w QSFUSBJOFENPEFMΛ࢖͏ͱٯʹѱ͘ͳΔ 前処理理 CV Public LB 4L CNN Scaling 0.2212 4L CNN Raw 0.2191 ResNet50 Scaling 0.2206 VGG16 Scaling 0.2107 1VCMJD-#͸֬ೝ͍ͯ͠ͳ͍

Slide 61

Slide 61 text

ೖࣹ֯͝ͱʹਖ਼نԽ ೖࣹ֯͝ͱʹ܏޲͕ผΕ͍ͯΔɻ άϧʔϓԽͯ͠౷ܭྔΛܭࢉͯ͠ɺ  άϧʔϓ͝ͱʹਖ਼نԽ͢Δɻ ੜσʔλ աڈίϯϖͷ༏উऀ͕ໄͷ͋Δը૾Λ ໄআڈ<>ʹΑΓਖ਼نԽͨ͠ղ๏͔Β  ண૝Λಘͨɻਖ਼نԽ͸ॏཁɻ ग़యɿIUUQCMPHLBHHMFDPNQMBOFUVOEFSTUBOEJOHUIFBNB[POGSPNTQBDFTUQMBDFXJOOFSTJOUFSWJFX

Slide 62

Slide 62 text

ੜσʔλ ਖ਼نԽޙ

Slide 63

Slide 63 text

ೖࣹ֯͝ͱͷσʔλਖ਼نԽͰվળ ೖࣹ֯Λ࢖ͬͨσʔλਖ਼نԽΛߦͱ$7վળ %BUB"VHNFOUBUJPO SPU qJQ [PPN ΋վળʹ༗ޮͰ͋ͬͨ 前処理理 CV Public LB 4L CNN Scaling 0.2212 4L CNN Norm+ Scaling 0.1824 4L CNN +DA Norm + Scaling 0.1615 1VCMJD-#͸֬ೝ͍ͯ͠ͳ͍

Slide 64

Slide 64 text

UIೖࣹ֯ͷ৘ใΛϞσϧʹ૊ΈࠐΉ ྔతม਺ ೖࣹ֯ ͷ৘ใΛϞσϧʹ૊ΈࠐΉͨΊʹɺ$//Ϟσ ϧͷ༧ଌ݁Ռͱྔతม਺ΛTUBDLJOHͨ͠ɻ ݸͷ$//Ϟσϧྔతม਺YͰ-JHIU(#.ͱEFFQCJU (#%5 Ͱֶशɻ࠷ޙʹॏΈ෇͖ฏۉɻ 前処理理 CV Public LB 4L CNN Scaling 0.2212 4L CNN Norm+Scaling 0.1824 4L CNN +DA Norm+Scaling 0.1615 stage2 (LGBM) NA 0.0895 0.0873 stage3 (Average) NA 0.0852 1VCMJD-#OE1SJWBUF-#UI

Slide 65

Slide 65 text

ݸͷ//ϞσϧΛ࡞੒ɻGPME$7Ͱνϡʔχϯάɻ $7Ͱֶशͨ͠ϞσϧͷNFBOQSFEJDUJPOΛಛ௃ྔͱ͢Δ ྔతม਺ ೖࣹ֯ͰάϧʔϓԽͯ͠ूܭ NFBO NFEJBO DPVOU ͯ͠ಛ௃ྔ࡞੒ ೖࣹ֯ͱ ͷNFBOQSFEJDUJPOͰ,//SFHSFTTPSʢྔతม਺ʣ -(#.Λ ͷಛ௃ྔͰֶशɻ߹ܭͰͭͷಛ௃ྔɻ 4. KNNreg 5. LGBM 1. CNNs 2. Mean prediction 3. Agg.  features ೖࣹ֯ ೖྗը૾ UIΑΓগͳ͍Ϟσϧ਺ೖࣹ֯Lۙ๣ ग़యɿIUUQTXXXLBHHMFDPNDTUBUPJMJDFCFSHDMBTTJpFSDIBMMFOHFEJTDVTTJPOɻগ͠؆ུԽͯ͠঺հ

Slide 66

Slide 66 text

Slide 67

Slide 67 text

TUೖࣹ֯ͷద੾ͳϞσϦϯά ೖࣹ֯ͷՄࢹԽ͕伴ɻl7JTVBMJ[JOHJOD@BOHMFXBTUIFLFZpOEJOHUIBU MFEVTEPXOPVSTPMVUJPOQBUIz ˠೖࣹ֯ʹΑͬͯ6OTVQFSWJTFEʹάϧʔϓԽͯ͠ϞσϧΛ࡞Δ ग़యɿIUUQTXXXLBHHMFDPNDTUBUPJMJDFCFSHDMBTTJpFSDIBMMFOHFEJTDVTTJPO άϧʔϓ̍ άϧʔϓ̎

Slide 68

Slide 68 text

࠶ܝɿͲ͏΍ͬͨΒউͯΔͷ͔ େࡶ೺ʹߟ͑Δͱࠓ΋ੲ΋ඞཁͳ͜ͱ͸มΘΒͳ͍ɻ ɹ&%" ୳ࡧతσʔλ෼ੳ ɹ7BMJEBUJPO ద੾ͳϞσϧධՁ ɹ4VSWFSZ ݚڀ੒Ռ΍աڈղ౴͔ΒֶͿ

Slide 69

Slide 69 text

͍͞͝ʹ w ,BHHMFͰ͸ΞϯαϯϒϧҎ֎͕伴ͱͳΔ͜ͱ͕͋Δɻσʔλ͔Βͷ ؾ෇͖ɾ//ͷσβΠϯͳͲɻ//ͷൃల΍ଟ༷ͳσʔλͷ͓͔͛ɻ w ֶशɾٞ࿦ɾήʔϜͱͯ͠ͳͲ༷ʑͳ໨తҙࣝͷࢀՃऀ͕ڞଘ͍ͯ͠ ΔɻڵຯΛ͍͚࣋ͬͯͨͩΔͱ,BHHMFѪ޷ऀͱͯ͠خ͍͠ɻ

Slide 70

Slide 70 text

෇࿥ɿࢀߟจݙ 1. [Blum & Hardt '15] "The Ladder: A Reliable Leaderboard for Machine Learning Competitions", In Proc. of the ICML '15. https://arxiv.org/abs/ 1502.04585 2. [Hardt '17] "Climbing a shaky ladder: Better adaptive risk estimation", https://arxiv.org/abs/1706.02733 3. [Rendle '10] "Factorization machines", In Proc. of the ICDM ’10. 4. [Johnson & Zhang '11] "Learning Nonlinear Functions Using Regularized Greedy Forest", https://arxiv.org/abs/1109.0887 5. [Hu+ ’17] “Squeeze-and-Excitation Networks”, In Proc. of the CVPR ’17. https://arxiv.org/abs/1709.01507 6. [He+ ’09] “Single Image Haze Removal”, In Proc. of the CVPR ’09. http://kaiminghe.com/cvpr09/

Slide 71

Slide 71 text

෇࿥ɿࢀߟจݙ 7. [Zhang+ ’17] "mixup: Beyond Empirical Risk Minimization", https:// arxiv.org/abs/1710.09412 8. [Miech+ '17] "Learnable pooling with Context Gating for video classiﬁcation", https://arxiv.org/abs/1706.06905 9. [Tian+ '16] "Detecting Text in Natural Image with Connectionist Text Proposal Network”, In Proc. of ECCV ‘16 https://arxiv.org/abs/1609.03605 10. [Qi+ '17] "PointNet: Deep Learning on Point Sets for 3D Classiﬁcation and Segmentation”, In Proc. of the CVPR ‘17 https://arxiv.org/abs/ 1612.00593 11. [Ma+ '17] "Multi-View Deep Learning for Consistent Semantic Mapping with RGB-D Cameras”, In Proc. of the IROS ‘17 https://arxiv.org/abs/ 1703.08866 12. [Shi+ '15] "An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition”, https://arxiv.org/abs/1507.05717

Slide 72

Slide 72 text

෇࿥ɿࢀߟจݙ 13. [Arandjelović+ ’16] "NetVLAD: CNN architecture for weakly supervised place recognition”, In Proc. of the CVPR ’16 https://arxiv.org/abs/ 1511.07247 14. [Zhou+ '17] "EAST: An Eﬃcient and Accurate Scene Text Detector", In Proc. of the CVPR’17 https://arxiv.org/abs/1704.03155v2