Upgrade to Pro — share decks privately, control downloads, hide ads and more …

이미지까지 이해하는 Multimodal LLM의 학습 방법 밝혀내기

kakao
November 01, 2024

이미지까지 이해하는 Multimodal LLM의 학습 방법 밝혀내기

#Kanana #MultiModal

최근 GPT-4v, GPT-4o와 같은 멀티모달 LLM에 대한 관심이 급증하면서 관련 연구들이 쏟아지고 있지만, 이 모델들을 구체적으로 어떻게 학습시키고 최적화하는지에 대한 정보는 희소한 상황입니다.
본 발표에서는 CVPR 2024에 하이라이트 논문으로 선정된 카카오의 자체 멀티모달 LLM을 소개하고, 이후 최근까지의 개선 과정을 통해 얻은 인사이트와 노하우를 공유합니다.

발표자 : edwin.ai
카나나 알파 조직의 Multimodal LLM Core 팀원 에드윈입니다. 기존의 텍스트 모달리티만을 다루는 LLM에서 더 나아가 시각, 음성을 포함한 다양한 모달리티를 이해하고 반응할 수 있도록 하는 방법에 대해 연구하고 있습니다.

kakao

November 01, 2024
Tweet

More Decks by kakao

Other Decks in Programming

Transcript

  1. द੼ U ө૑ ઱য૓ ױযٜਸ ӝ߈ਵ۽ ׮਺ U  ױযܳ

    ৘ஏೞب۾ ೟णػ BVUPSFHSFTTJWFNPEFM -BOHVBHF.PEFM -. 𝑚𝑎𝑥 $ !"# $ log 𝑝 𝑤!%& 𝑤'! )
  2. द੼ U ө૑ ઱য૓ ױযٜਸ ӝ߈ਵ۽ ׮਺ U  ױযܳ

    ৘ஏೞب۾ ೟णػ BVUPSFHSFTTJWFNPEFM -BOHVBHF.PEFM -. t߄աաח ઱۽ ޖट ࢝੉ঠ u t߄աաח ઱۽ ֢ۆ࢝ ੑפ׮u 𝑚𝑎𝑥 $ !"# $ log 𝑝 𝑤!%& 𝑤'! )
  3. ೟ण ؘ੉ఠ৬ ݽ؛ ௼ӝ ҙ੼ীࢲ੄ TDBMFVQ -BSHF-BOHVBHF.PEFM --. t 

       j׮਺ী ৢ ं੗ח ޖ঺ੋо u t੉ ࣻৌ਷ п ं੗о ੗োࣻ੄ ઁғੋ ಁఢਸ ٮܵפ׮ ٮۄࢲju
  4. .VMUJNPEBM--. ׮ܲ NPEBMJUZ ৘੉޷૑ ৬ U द੼ө૑ ઱য૓ ױযٜਸ ӝ߈ਵ۽

    ׮਺ ױযܳ ৘ஏ 𝑚𝑎𝑥 $ !"# $ log 𝑝 𝑤!%& 𝑣, 𝑤'! ) .--.
  5. .VMUJNPEBM--. ׮ܲ NPEBMJUZ ৘੉޷૑ ৬ U द੼ө૑ ઱য૓ ױযٜਸ ӝ߈ਵ۽

    ׮਺ ױযܳ ৘ஏ 𝑚𝑎𝑥 $ !"# $ log 𝑝 𝑤!%& 𝑣, 𝑤'! ) tࢎ૓ী ੓ח زޛী ؀೧ ࢸݺ೧઻u tࢎ૓ী ੓ח زޛ਷ ೒ۄ߁Ҋ۽ ju MLLM
  6. .VMUJNPEBM-BSHF-BOHVBHF.PEFMT .--.T    'MBNJOHP #-*1 ,PTNPT 1B-.& --B.""EBQUFS

    *OTUSVDU#-*1 .JOJ(15 --B7" $IBNFMFPO 7JTJPO--. -ZOY 2XFO7- (15W ,PTNPT )POFZCFF --B7" /&Y5(15 9DPNQPTFS $PH7-. (FNJOJ 9DPNQPTFS 2XFO7-.BY --B7"/F95 (FNJOJ .. $PH7-. 2XFO7- 1BMJ(FNNB --B7"0OFWJTJPO $MBVEF /7-. (15P
  7. .VMUJNPEBM-BSHF-BOHVBHF.PEFMT .--.T    'MBNJOHP #-*1 ,PTNPT 1B-.& --B.""EBQUFS

    *OTUSVDU#-*1 .JOJ(15 --B7" $IBNFMFPO 7JTJPO--. -ZOY 2XFO7- (15W ,PTNPT )POFZCFF --B7" /&Y5(15 9DPNQPTFS $PH7-. (FNJOJ 9DPNQPTFS 2XFO7-.BY --B7"/F95 (FNJOJ .. $PH7-. 2XFO7- 1BMJ(FNNB --B7"0OFWJTJPO $MBVEF /7-. (15P
  8. *OTVGGJDJFOUUSBJOJOHSFDJQFTj    Flamingo BLIP2 Kosmos-1 PaLM-E LLaMA-Adapter InstructBLIP

    MiniGPT-4 LLaVA Chameleon VisionLLM Lynx Qwen-VL GPT-4v Kosmos-2 Honeybee LLaVA1.5 NExT-GPT Xcomposer CogVLM Gemini Xcomposer2 Qwen-VL-Max LLaVA-NeXT Gemini1.5 MM1 CogVLM2 Qwen2-VL PaliGemma LLaVA-Onevision Claude 3.5 NVLM GPT-4o ݆਷ ݽ؛ٜ੉ ॔ই૑Ҋ ੓঻૑݅ ҳઑա ೟ण ߑߨۿ ҙ੼ীࢲ੄ ׮নೠ ఐ࢝਷ ࠗ઒೮؍ दӝ
  9. )PX ই ӒѢ $-*1-JOFBS7JDVOB ҳઑ ӝ߈ਵ۽ 72"  (2" $$.

    5FYU$BQTjؘ੉ఠ۽ ೟णदఃҊ MFBSOJOHSBUFח j ೞݶ ظਃ_ …
  10. ݽ؛ ҳઑ ೟ण ߑߨ tݽ؛ ҳઑীࢲ ޥо ֬஖Ҋ ੓ח Ѫ਷

    হա u tযڃ ؘ੉ఠٜਸ о૑Ҋ যڌѱ ೟णदெঠ ೞ૑ u
  11. ݽ؛ ҳઑ ೟ण ߑߨ tݽ؛ ҳઑীࢲ ޥо ֬஖Ҋ ੓ח Ѫ਷

    হա u tযڃ ؘ੉ఠٜਸ о૑Ҋ যڌѱ ೟णदெঠ ೞ૑ u
  12. Basic architecture of MLLMs Vision Encoder Large Language Model Projector

    $-*1 4JH-*1 .P$P %*/0 4". j --B." 7JDVOB 2XFO *OUFSO-. j
  13. #BTJDBSDIJUFDUVSFPG.--.T 7JTJPO &ODPEFS -BSHF -BOHVBHF.PEFM 1SPKFDUPS t೒ۄ߁Ҋח ؀୓۽ju t੉ زޛ੄

    ࢲध૑ח u 1SPKFDUPS QMBZTBDSVDJBMSPMFJOCSJEHJOH QSFUSBJOFEWJTJPO FODPEFSTXJUI--.T
  14. $VSSFOUQSPKFDUPST-JOFBSWT3FTBNQMFS 7JTJPO &ODPEFS "CTUSBDUPS 3FTBNQMFS 7JTJPO &ODPEFS -JOFBS ✘ 'MFYJCJMJUZ

    0OFUPPOF NBQQJOH ✓ 'MFYJCJMJUZ "CTUSBDUFE NBQQJOH 3FTBNQMFS 7JTVBMGFBUVSFT $SPTT "UUFOUJPO -FBSOBCMF RVFSJFT 7JTVBMUPLFOT -JOFBS 1SPKFDUPS
  15. $VSSFOUQSPKFDUPST-JOFBSWT3FTBNQMFS 7JTJPO &ODPEFS 7JTJPO &ODPEFS -JOFBS 3FTBNQMFS 7JTVBMGFBUVSFT $SPTT "UUFOUJPO

    -FBSOBCMF RVFSJFT 7JTVBMUPLFOT ✘ 'MFYJCJMJUZ ✓ -PDBMJUZ ✓ 'MFYJCJMJUZ ✘ -PDBMJUZ "CTUSBDUPS 3FTBNQMFS -JOFBS 1SPKFDUPS
  16. 4PMVUJPOJOKFDUJOHMPDBMJUZUPBCTUSBDUPST 5XPEFTJHOQSJODJQMFT  'MFYJCJMJUZ JONBOBHJOHUIFOVNCFSPGWJTVBMUPLFOT  1SFTFSWBUJPOPGMPDBMDPOUFYUTJOWJTVBMGFBUVSFT (a) Linear Projector

    ✗ Flexibility ✓ Locality Preservation Linear Projector (b) Abstractor ✓ Flexibility ✗ Locality Preservation (c) Locality-enhanced Abstractor ✓ Flexibility ✓ Locality Preservation Vision Encoder Projector Large Language Model Visual Tokens Text Tokens Visual Features Resampler C-Abstractor D-Abstractor  
  17. $PNQBSJTPOXJUIQSFWJPVTQSPKFDUPST &GGFDUJWFOFTTPGPVSMPDBMJUZFOIBODFEQSPKFDUPS  #FUUFSTQBUJBMVOEFSTUBOEJOHDBQBCJMJUZ  #FUUFSQFSGPSNBODFFGGJDJFODZCBMBODF 1SPKFDUPS . TTUFQ ..&

    ..# 4&&% "WH 104 43 0- 13 43 *- -JOFBS   6OBWBJMBCMFEVFUPJOGMFYJCJMJUZ  3FTBNQMFS          $"CTUSBDUPS          -JOFBS          3FTBNQMFS          $"CTUSBDUPS         
  18. $PNQBSJTPOXJUIQSFWJPVTQSPKFDUPST &GGFDUJWFOFTTPGPVSMPDBMJUZFOIBODFEQSPKFDUPS  #FUUFSTQBUJBMVOEFSTUBOEJOHDBQBCJMJUZ  #FUUFSQFSGPSNBODFFGGJDJFODZCBMBODF 1SPKFDUPS . TTUFQ ..&

    ..# 4&&% "WH 104 43 0- 13 43 *- -JOFBS   6OBWBJMBCMFEVFUPJOGMFYJCJMJUZ  3FTBNQMFS          $"CTUSBDUPS          -JOFBS          3FTBNQMFS          $"CTUSBDUPS         
  19. $PNQBSJTPOXJUIQSFWJPVTQSPKFDUPST &GGFDUJWFOFTTPGPVSMPDBMJUZFOIBODFEQSPKFDUPS  #FUUFSTQBUJBMVOEFSTUBOEJOHDBQBCJMJUZ  #FUUFSQFSGPSNBODFFGGJDJFODZCBMBODF 1SPKFDUPS . TTUFQ ..&

    ..# 4&&% "WH 104 43 0- 13 43 *- -JOFBS   6OBWBJMBCMFEVFUPJOGMFYJCJMJUZ  3FTBNQMFS          $"CTUSBDUPS          -JOFBS          3FTBNQMFS          $"CTUSBDUPS         
  20. $PNQBSJTPOXJUIQSFWJPVTQSPKFDUPST &GGFDUJWFOFTTPGPVSMPDBMJUZFOIBODFEQSPKFDUPS  #FUUFSTQBUJBMVOEFSTUBOEJOHDBQBCJMJUZ  #FUUFSQFSGPSNBODFFGGJDJFODZCBMBODF 1SPKFDUPS . TTUFQ ..&

    ..# 4&&% "WH 104 43 0- 13 43 *- -JOFBS   6OBWBJMBCMFEVFUPJOGMFYJCJMJUZ  3FTBNQMFS          $"CTUSBDUPS          -JOFBS          3FTBNQMFS          $"CTUSBDUPS         
  21. ݽ؛ ҳઑ ೟ण ߑߨ tݽ؛ ҳઑীࢲ ޥо ֬஖Ҋ ੓ח Ѫ਷

    হա u tযڃ ؘ੉ఠٜਸ о૑Ҋ যڌѱ ೟णदெঠ ೞ૑ u
  22. 7JTVBMJOTUSVDUJPOUVOJOHJOTUSVDUJ[BUJPO $BQUJPOt5IFSFJTBHSPVQPGNFOBSPVOEB USVDLu t5IFDPMPSPGBWJOUBHFUSVDLJTju j 0CKFDUT5SVDL<   > j

    72"t8IBUJTUIFDPMPSPGUIFUSVDL 3FEu  t)PXNBOZQFPQMFBSFUIFSF u'JWFu j (15 )VNBO)FMQNFXSJUFBOBEWFSUJTFNFOU GPSUIFWJOUBHFSFEUSVDL "*$MBTTJDT3FE7JOUBHF1JDLVQ 5SVDLm "5JNFMFTT#FBVUZ-PPLJOHUP PXOBQJFDFPGBVUPNPUJWFIJTUPSZ 5IJT NFUJDVMPVTMZSFTUPSFETWJOUBHFSFE QJDLVQUSVDLJTj 1SFEFGJOFE UFNQMBUFT )VNBO"OTXFSVTJOHBTJOHMFXPSEPS QISBTF8IBUDPMPSJTUIFUSVDLPOUIFMFGU "*3FE )VNBO)PXNBOZQFPQMFBSFUIFSF "*GJWF 5FNQMBUFCBTFE (15BTTJTUFE 
  23. 7JTVBMJOTUSVDUJPOUVOJOHJOTUSVDUJ[BUJPO 7BSJPVTWJTJPOMBOHVBHFEBUBTFUTBSFVTFEGPSUIFUFNQMBUFCBTFEWJTVBMJOTUSVDUJPO 5BTL %BUBTFUT TBNQMFT $BQUJPOJOH #MJQ$BQ'JMU $0:0. . 72"

    0QFO 72"W (2" 0$372" 743 . 72" .$ 4DJFODF2" "0,72" . 3&$ 3FG$0$0 3FG$0$0 3FG$0$0H 7JTVBM(FOPNF . *OTUSVDUJPO --B7"L 4IBSF(15 .
  24. )JEEFOSFDJQF EBUBTFUNJYUVSF (15BTTJTUFE %WT% à IFMQGVMGPSEFTDSJQUJWFSFTQPOTFT 5FNQMBUFCBTFE %WT% à IFMQGVMJOQFSDFQUVBMVOEFSTUBOEJOH

    *OTUSVDUJPO5BTL5ZQF .--.#FODINBSL 5FNQMBUFCBTFE (15BTTJTUFE .VMUJQMFDIPJDF #JOBSZZFTOP (15FWBM 72" 0QFO 72" .$ 3&$ $BQ 7*OTU 5*OTU ..# 4&&%J ..&Q ..& --B7"X % ✓ ✓ ✓ ✓ ✓ ✓      % ✓ ✓ ✓ ✓        ­ %     ✓ ✓  ­  ­  ­  ­ 
  25. )JEEFOSFDJQF EBUBTFUNJYUVSF %JWFSTJUZPGEBUBTFUTBOEUBTLTJTJNQPSUBOUGPSMFBSOJOHQSPGJDJFODZJOSFMBUFEUBTLT *OTUSVDUJPO5BTL5ZQF .--.#FODINBSL 5FNQMBUFCBTFE (15BTTJTUFE .VMUJQMFDIPJDF #JOBSZZFTOP (15FWBM

    72" 0QFO 72" .$ 3&$ $BQ 7*OTU 5*OTU ..# 4&&%J ..&Q ..& --B7"X % ✓ ✓ ✓ ✓ ✓ ✓      % ✓ ✓ ✓ ✓ ✓ ✓  ­   ­  ­  ­ %  ✓ ✓ ✓ ✓ ✓   ­  ­  ­  % ✓  ✓ ✓ ✓ ✓  ­  ­   
  26. )JEEFOSFDJQF EBUBTFUCBMBODF  %BUBTFUMFWFMVOJGPSNTBNQMJOHJTFGGFDUJWF  .BOVBMUVOJOHPGTBNQMJOHXFJHIUTJTUIFCFTU .JYUVSFUZQF ..# 4&&%J ..&Q

    "WH 1FSEBUBTFU     1FSUBTL     1FSTBNQMFL     1FSEBUBTFUUVOFE    
  27. )JEEFOSFDJQF UFNQMBUFEFTJHO  %BUBTFUMFWFM GJOF UBTLMFWFM DPBSTF  0OFUFNQMBUFQFSEBUBTFUJTTVGGJDJFOU (SBOVMBSJUZ

    %JWFSTJUZ ..# 4&&%J ..&Q "WH 'JOF 4JOHMF     $PBSTF 4JOHMF     'JOF .VMUJ     'JOF .VMUJ GMJQ    
  28. )JEEFOSFDJQF UFNQMBUFEFTJHO  %BUBTFUMFWFM GJOF UBTLMFWFM DPBSTF  0OFUFNQMBUFQFSEBUBTFUJTTVGGJDJFOU Granularity

    Diversity MMB SEED-i MME-p Avg Fine Single     Coarse Single     Fine Multi     Fine Multi + flip    
  29. )JEEFOSFDJQF UFNQMBUFEFTJHO  %BUBTFUMFWFM GJOF UBTLMFWFM DPBSTF  0OFUFNQMBUFQFSEBUBTFUJTTVGGJDJFOU (SBOVMBSJUZ

    %JWFSTJUZ ..# 4&&%J ..&Q "WH 'JOF 4JOHMF     $PBSTF 4JOHMF     'JOF .VMUJ     'JOF .VMUJ GMJQ    
  30. $PNQBSJTPOXJUIQSFWJPVT4P5" Method LLM Projector Vision Enc. Res. MMB MME-p MME

    SEED-i LLaVA-w "QQSPBDIFTVTJOH#--. MiniGPT-4 7JDVOB# 3FTBNQMFS &7"$-*17J5(       InstructBLIP 7JDVOB# 2GPSNFS &7"$-*17J5(       IDEFICS --B."# 'MBNJOHP 0QFO$-*1 7J5)       Qwen-VL-Chat 2XFO# 3FTBNQMFS 0QFO$-*1 7J5CJH(       LLaVA-1.5 7JDVOB# -JOFBS $-*17J5-       Honeybee 7JDVOB# $"CT $-*17J5-       %"CT      "QQSPBDIFTVTJOH#--. MiniGPT-4 7JDVOB# 3FTBNQMFS &7"$-*17J5(       BLIP-2 7JDVOB# 2GPSNFS &7"$-*17J5(       InstructBLIP 7JDVOB# 2GPSNFS &7"$-*17J5(       LLaVA-1.5 7JDVOB# -JOFBS $-*17J5-       Honeybee 7JDVOB# $"CT $-*17J5-       %"CT     
  31. $VSSFOUQSPHSFTT,BOBOBW ୭Ӕ .--.ٜ੄ ౟۪٘ܳ ࢓ಝࠁݶ  ؊਌ ׮নೞҊ ௾ झாੌ੄

    ؘ੉ఠࣇ ೟ण à NBDIJOFHFOFSBUFEؘ੉ఠ ഝਊ ߂ ௬ܻ౭ ஶ౟܀੉ ҙѤ  ؊਌ ௾ ೧࢚ب੄ ੑ۱ ੉޷૑ ഝਊà --.੄ ੑ۱ਵ۽ ٜযоח WJTVBMUPLFOࣻܳ ബҗ੸ਵ۽ ઴ੌ ࣻ ੓যঠ ೣ )POFZCFF .PSFEBUBTFUT #JHHFSSFTPMVUJPO "EWBODFESFDJQF j ,BOBOBW
  32. $VSSFOUQSPHSFTT,BOBOBW ୭Ӕ .--.ٜ੄ ౟۪٘ܳ ࢓ಝࠁݶ  ؊਌ ׮নೞҊ ௾ झாੌ੄

    ؘ੉ఠࣇ ೟ण à NBDIJOFHFOFSBUFEؘ੉ఠ ഝਊ ߂ ௬ܻ౭ ஶ౟܀੉ ҙѤ  ؊਌ ௾ ೧࢚ب੄ ੑ۱ ੉޷૑ ഝਊà --.੄ ੑ۱ਵ۽ ٜযоח WJTVBMUPLFOࣻܳ ബҗ੸ਵ۽ ઴ੌ ࣻ ੓যঠ ೣ ..# $IBSU2" %PD72" *OGP72" .BUI7JTUB 4&&%J ...6 (157        (FNJOJ1SP        )POFZCFF        ,BOBOBW       
  33. ੉ ੉޷૑ח ৈ۞ թࢿ੉ ࡈр ௿ېध ౟۟ ઱ਤী ݽৈ ੓ח

    ݽणਸ ࠁৈસפ׮੉ ౟۟਷ ҟఖ੉ աח ௼܁ Ӓܾҗ ೻٘ۄ੉౟ ӒܻҊ ࣁ۲ػ ٣੗ੋਵ۽ ੋ೧ ֙؀ ژח ֙؀੄ ର۝ਵ۽ ࠁੑפ׮౟۟਷ ੜ ҙܻغҊ ੓ਵݴ ӵՖೠ ࢚క۽ ࠁੑפ׮թࢿٜ਷ ؀ࠗ࠙ ౭࣊எ৬ ୒߄૑ܳ ੑҊ ੓णפ׮ੌࠗח ࢶӖۄझܳ ଱ਊೞҊ ੓Ҋ ೠ ݺ਷ ݽ੗ܳ ॳҊ ੓णפ׮ Ӓٜ਷ ౟۟੄ ٣੗ੋҗ ࢚కী ؀೧ ੉ঠӝೞח Ѫ୊ۢ ࠁ੉ݴ ই݃ب ର۝ী ؀ೠ ҙब੉ա хࢎܳ աఋղҊ ੓ਸ Ѫੑפ׮ߓ҃ীח TVQFSGSFTIۄח р౸੉ ੓ח ࢚੼੉ ੓णפ׮߄ׯ਷ ನ੢غয ੓ਵݴ Ӓܿ੗о ӡѱ ܻ٘ਕઉ ੓য కন੉ ೞטীࢲ ࢚؀੸ਵ۽ ծ਷ пبী ੓਺ਸ աఋշפ׮ 2VBMJUBUJWFFYBNQMFTEFUBJMFEDBQUJPOJOH t੉޷૑ী ؀೧ ੗ࣁ൤ ࢸݺ೧઻u Kanana-v 
  34. #FFG৬ 1PSL੄ ױߔ૕ ബਯ ಣӐ஖ܳ ҅࢑ೞӝ ਤ೧  ݢ੷ пп੄

    чਸ ഛੋ೧ঠ ೤פ׮  #FFG੄ ױߔ૕ ബਯ਷ ੑפ׮  1PSL੄ ױߔ૕ ബਯ਷ ੑפ׮ ಣӐਸ ҳೞӝ ਤ೧ ੉ ف чਸ ؊ೠ റ ۽ աׄפ׮ ಣӐ  #FFG੄ ױߔ૕ ബਯ 1PSL੄ ױߔ૕ ബਯ  ಣӐ     ಣӐ  ٮۄࢲ t#FFGu৬ t1PSLu੄ ױߔ૕ ബਯ ಣӐ஖ח ੑפ׮ 2VBMJUBUJWFFYBNQMFTEFUBJMFEDBQUJPOJOH t#FFG৬ 1PSL੄ ױߔ૕ ബਯ ಣӐ஖ܳ ҅࢑೧઻u Kanana-v 
  35. 𝑥 о о੢ ӟ ߸੄ ӡ੉੉Ҋ ࢖пഋ੉ غӝ ਤೠ ઑѤী

    ੄ೞৈ 𝑥 ਸ ݅઒೧ঠ ೤פ׮ ژೠ كп࢖пഋ੉ غ۰ݶ 𝑥# > 7# + 11#, ∴ 𝑥# > 170ਸ ݅઒೧ঠ ೤פ׮ ف ઑѤਸ ݅઒दఃח ੗োࣻ 𝑥ח    ੉޲۽ ҳೞח ೤਷    ੑפ׮ 2VBMJUBUJWFFYBNQMFTEFUBJMFEDBQUJPOJOH tӒܿਸ ࠁҊ Ӓܿ ࣘ ޙઁ੄ ׹ਸ ݏ୶য ࠁࣁਃu Kanana-v
  36. ࢎ૓ ࣘ ਺ध਷ ӣߏੑפ׮ӣߏਸ ݅٘ח ߑߨ਷ ׮਺җ эणפ׮ वਸ অҊ

    ੸׼۝੄ ޛ۽ ߏਸ ૑਷ റ ҅ۆ ׼Ӕ য়੉ যޗ ೫ ݍ࢓ ױޖ૑ ਋়ਸ ળ࠺೤פ׮ ҅ۆҗ ׼Ӕ਷ ࠃইࢲ ળ࠺ೞҊ য়੉ח ଻ ेয ޛী ׸о نפ׮ যޗҗ ೫ ݍ࢓਷ ࠃইࢲ ળ࠺ೞҊ ਋়਷ ࢕ইࢲ ળ࠺೤פ׮ ӣ ਤী ߏਸ তѱ ಝҊ ળ࠺ೠ ੤ܐٜਸ ৢ۰ࢲ ӣߏਸ ݈ইસפ׮ ӣߏਸ ੜۄࢲ ઁҕೞݶ ؾפ׮ 2VBMJUBUJWFFYBNQMFTEFUBJMFEDBQUJPOJOH tࢎ૓ ࣘ ਺ध੄ ੉ܴҗ ݅٘ח ߨਸ ঌ۰઻u Kanana-v 
  37. 'VUVSFEJSFDUJPOTVQQPSUNVDIMPOHFSDPOUFYUT --.ী ׮নೠ ݽ׳ܻ౭ FH ਺ࢿ ࠺٣য় ܳ ୶оೡࣻ۾ --.੄

    ੑ۱ द௫झ ӡ੉ ߂ ҅࢑۝੉ ௼ѱ ૐо à ؊਌ ബҗ੸ੋ NPEBMJUZBCTUSBDUPS߂ --.ো࢑ োҳ ೙ਃ -BSHF-BOHVBHF.PEFM 4FRVFODF-FOHUI $PNQVUBUJPODPTU
  38. 'VUVSFEJSFDUJPONPSFDPNQMFYWJTVBMSFBTPOJOH ؊਌ ࠂ੟ೞҊ ੉޷૑ܳ ੜ ੉೧೧ঠ ೞח tWJTVBMuSFBTPOJOH מ۱ਸ ؊਌

    ఃਕঠ ೣ 28IBUTIPVMEZPVQVUJO UIFTRVBSFUPHFUB DPSSFDUEJBHSBN 2#JSE#PCCJFKVNQTPOBGFODFGSPNUIFQPTUPO UIFMFGUFOEUPUIFPUIFSFOE&BDIKVNQUBLFTIJN TFDPOET)FNBLFTKVNQTBIFBEBOEUIFO KVNQCBDL5IFOIFBHBJONBLFTKVNQTBIFBE BOEKVNQCBDL BOETPPO*OIPXNBOZTFDPOET DBO#PCCJFHFUGSPNPOFFOEUPUIFPUIFSFOE 24FWFOTUJDLTMJFPOUPQPG FBDIPUIFS4UJDLJTBU UIFCPUUPNUPQ8IJDI TUJDLJTJOUIFNJEEMF   
  39. 2"

  40. 참고 문헌  IUUQTXXXGMJDLSDPNQIPUPTTIBOLBSPOMJOF  IUUQTPVSXPSMEJOEBUBPSHHSBQIFSFOFSHZFGGJDJFODZPGNFBUBOEEBJSZQSPEVDUJPO  IUUQTXXXJOEVTUSZEPDVNFOUTVDTGFEVEPDTJENUHK  $IB

    + ,BOH 8 .VO + 3PI #  )POFZCFF-PDBMJUZFOIBODFEQSPKFDUPSGPSNVMUJNPEBMMMN*O 1SPDFFEJOHTPGUIF*&&&$7'$POGFSFODFPO$PNQVUFS7JTJPOBOE 1BUUFSO3FDPHOJUJPO QQ   IUUQTXXXGMJDLSDPNQIPUPTTJLLJOQIPUPTUSFBN  IUUQTDPNNPOTXJLJNFEJBPSHXJLJ'JMFӣߏKQH  IUUQTDPNNPOTXJLJNFEJBPSHXJLJ'JMF0QFO@GJSF@IZESBOU@  KQH  IUUQTTNBSUEBUBTFUHJUIVCJPTNBSU