生成系AIの実応用に向けて

by LINE Developers

Slide 1

Slide 1 text

:PTIJIJTB*+*3* ੜ੒"*ͷ࣮༻ʹ޲͚ͯ "*EFW$7- EJSFDUPS -*/&$PSQPSBUJPO

Slide 2

Slide 2 text

:PTIJIJTB*KJSJ 1I% -*/&גࣜձࣾ σʔλαΠΤϯεηϯλʔ "*%FWࣨ ࣨ௕ɺ$7-Ϛωʔδϟʔ > ઐ໳ɿίϯϐϡʔλϏδϣϯɾϩϘςΟΫεɺͦΕΒΛࢧ͑Δػցֶश > झຯɿ > 0VUEPPSొࢁɾεΩʔɾୌ८ΓɾࣸਅࡱӨɾόΠΫτϥΠΞϧɾɾɾ > *OEPPSϐΞϊԋ૗ɾྺ࢙ɾᗉ੡ɾίʔώʔᖿઝɾञΛᅂΉ > ೥ΦϜϩϯೖࣾ > إͷݕग़ೝࣝͷσδΧϝɾܞଳి࿩ɺ؂ࢹΧϝϥԠ༻ > ෺ମݕग़ɾŤŞƄŸƃũŖŢŔƃɾ0$3ͷ'"޲͚঎඼Խ > ͠ͳ΍͔ͳ੍ޚΛ࣮ݱ͢Δࣗ཯ιϑτϩϘοτݚڀਪਐ > Ϧαʔνϕϯνϟʔ্ཱͪ͛ 0.30/4*/*$9 > ೥-*/&ೖࣾ > $PNQVUFS7JTJPO-BCͷ্ཱͪ͛ɺ"*։ൃࣨͷ૊৫Խ

Slide 3

Slide 3 text

ຊ೔ͷߨԋͷ಺༰ • ੜ੒"*ͱͦͷՄೳੑ • -*/&Ͱͷ"* • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

Slide 4

Slide 4 text

ຊ೔ͷߨԋͷ಺༰ • ੜ੒"*ͱͦͷՄೳੑ • -*/&Ͱͷ"* • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

Slide 5

Slide 5 text

ʮਪ࿦ʯͷࣗಈԽͷਐల ػցֶशϕʔε `T ϧʔϧϕʔε ؍࡯΍஌ݟΛݩʹ ౷ܭϞσϧߏங σʔλ͔Β ༧ଌϞσϧΛ௚઀ֶश ը૾ɾݴޠͳͲ ଟ࣍ݩσʔλΛѻ͍ͮΒ͘௿ਫ਼౓ ղऍੑɾઆ໌ੑ͕௿͘ ܭࢉ͕๲େ ղऍੑɾઆ໌ੑ͕ߴ͘ ܭࢉ΋গͳΊ ը૾ɾݴޠͳͲ ଟ࣍ݩσʔλ͕ѻ͑ߴਫ਼౓ *GUIFOϧʔϧ ϕΠζ౷ܭ χϡʔϥϧωοτϫʔΫ ಛ௃ ܽ఺ ར఺ ୅දख๏ WT

Slide 6

Slide 6 text

"*ֵ໋ɿػց͕ʮਪ࿦ʯͰਓΛ௒͑Δ࣌୅ $IBU(15 ήʔϜʢғޟʣͷੈքͰ ਓΛ௒͑ͨʂ ʮݴޠʯʮࢥߟʯͳͲʮੜ੒ʯͰ ਓΛ௒͑ͭͭ͋Δʁ ๲େͳعේσʔλͰֶश 8FC্ͷ๲େͳ஌ࣝ ίϯϐϡʔλಉ࢜Ͱରઓ͠ ࣗ཯ֶश ਓͷϑΟʔυόοΫͰֶश "MQIB(P ܭࢉػਐԽ

Slide 7

Slide 7 text

ͩΕ΋ѻ͑Δʮݴޠʯʮը૾ʯΛೖྗ͠ ͩΕ΋͕Θ͔Δʮݴޠʯ΍ʮը૾ʯɾʮԻ੠ʯͱ͍͏ܗͰग़ྗͰ͖Δ "* ੜ੒"*ͱ͸ʁ 7 "* ਓ͕͍ͯ͠Δ࡞ۀʹ ͍͍ۙͮͯΔ

Slide 8

Slide 8 text

ͩΕ΋ѻ͑Δʮݴޠʯʮը૾ʯΛೖྗ͠ ೝࣝ݁Ռɾ൑ఆ݁ՌΛग़ྗ ैདྷͷ"* 8 "* ਓ͕͍ͯ͠Δ࡞ۀʹ ͍͍ۙͮͯΔ :FT/P.BZCF ΫϥεϥϕϧͳͲ جຊతʹ࣍ݩͷग़ྗ

Slide 9

Slide 9 text

ドメインを選ばない⼤量データで学習させた多⽬的なモデル⽣成AIを可能とする基盤モデル υ ϝ Π ϯ λεΫ λεΫಛԽܕ"* λεΫಛԽܕ"* ج൫Ϟσϧ

Slide 10

Slide 10 text

ج൫Ϟσϧʢ'PVOEBUJPONPEFMʣʹΑΔ൚༻Խ XFC ,OPX MFEHF CBTF 1VSDIBTF SFDPSE 1VSDIBTF SFDPSE 5SBOT BDUJPO SFDPSE 4QFFDI EPDT "E CBOOFS 4UJDLFS 4FBSDI 2" 4IPQQJOH FYQFSJFODF DIBUCPU $POW X TUJDLFST 4$. %JBMPHVF 0$3 "E PQUJNJ[F ʜ ʜ 4FBSDI NPEFM 2" NPEFM 3FDPN .PEFM %JBMPHVF .PEFM %FNBOE 1SFE NPEFM 4QFFDI 3FDPH NPEFM 0$3 NPEFM &GGFDU 1SFE NPEFM 4UJDLFS 3FDPN NPEFM ʜ ϚϧνυϝΠϯσʔλ 4FBSDI 2" ʜ 0$3 "E PQUJNJ[F ʜ 0OFNPEFM 'PVOEBUJPONPEFM ೖ ྗ ޙ ஈ λ ε Ϋ ॲ ཧ ैདྷ λεΫຖʹϞσϧߏஙɺܾΊͨλεΫͷΈʹར༻ ج൫Ϟσϧ υϝΠϯλεΫʹґଘͤͣ࢖͑Δ ͭͷٕज़ֵ৽

Slide 11

Slide 11 text

୯७ʹ"UUFOUJPO''/ϨΠϠʔΛॏͶΔ͚ͩͰਫ਼౓্͕͕Δʂ ͦΕ·Ͱ͸3//΍-45.ͳͲͰ૚͝ͱʹ޻෉͕ඞཁ ̍ͭ໨ͷٕज़ֵ৽ɿߏ଄ͷ୯७Խʢ5SBOTGPSNFSʢʣʣ Ϟσϧߏ଄ΤϯδχΞϦϯάͰͷ ࠩҟԽͷऴᖼʁ ͨͩ͠ɺ·ͩγϯάϧλεΫϞσϧ ·ͩ·ͩ൚༻Խ͕ඞཁͳঢ়گͩͬͨ ൚༻Խʹ޲͚ͯͷ՝୊͸ ֶश͢Δσʔλྔ ʴ ͦͷͨΊͷΞϊςʔγϣϯ

Slide 12

Slide 12 text

̎ͭ໨ͷٕज़ֵ৽ɿڭࢣͳ͠ࣄલֶश #&35 (15 Inclusive multi-modal data Search QA … OCR Ad optimize … Pre-trained model Small data Fine tuning GT XFC ,OPX MFEHF CBTF 1VSDIBTF SFDPSE 1VSDIBTF SFDPSE 5SBOT BDUJPO SFDPSE 4QFFDI EPDT "E CBOOFS 4UJDLFS 4FBSDI 2" 4IPQQJOH FYQFSJFODF DIBUCPU $POW X TUJDLFST 4$. %JBMPHVF 0$3 "E PQUJNJ[F ʜ ʜ 4FBSDI NPEFM 2" NPEFM 3FDPN .PEFM %JBMPHVF .PEFM %FNBOE 1SFE NPEFM 4QFFDI 3FDPH NPEFM 0$3 NPEFM &GGFDU 1SFE NPEFM 4UJDLFS 3FDPN NPEFM ʜ ೖ ྗ ޙ ஈ λ ε Ϋ ॲ ཧ ैདྷ ैདྷ͸λεΫ͝ͱʹೖྗͱਖ਼ղ͕ඞཁ ڭࢣͳ͠ࣄલֶश ਖ਼ղ͕ཁΒͳֶ͍शํ๏

Slide 13

Slide 13 text

݀ຒΊ໰୊ͱ઀ଓ໰୊Λɺʢ΄΅ʣແݶʹ࡞Γग़͠ɺֶशͤ͞Δʢࣗݾڭࢣֶशʣ #&35ͰఏҊ͞Εͨɺڭࢣͳ͠ࣄલֶशͷΠϝʔδ ੨ۭจݿʮۜՏమಓͷ໷ʯΑΓ

Slide 14

Slide 14 text

ハハ積もれば価格が⼭となる塵も積もれば⽂脈次に来る⾔葉 ? ⾔語モデルとは、与えられた⽂章の次に来る⾔葉を当てるよう学習させたAIモデルその性能を最⼤限引き出すため、モデルサイズ（パラメータ）を１０億以上に拡⼤ #&35ͰఏҊ͞Εͨɺڭࢣͳ͠ࣄલֶशͷΠϝʔδ ͜ΕΛݴޠੜ੒ʹԠ༻͢Δͱɾɾɾʢ(15ʣ

Slide 15

Slide 15 text

4DBMJOH-BXT<,BQMBO BS9JW >ʹΑΕ͹ɺσʔληοτ΍ύϥϝʔλΛେ͖͘͢Ε͹ͲΜͲΜਫ਼౓͕ྑ͘ͳΔʂ εέʔϦϯά๏ଇ Kaplan+, Scaling Laws for Neural Language Models, arXiv 2020より抜粋

Slide 16

Slide 16 text

େن໛ݴޠϞσϧͷ։ൃڝ૪ ։ൃݩ ΞϧΰϦζϜ ύϥϝʔλ਺ ݴޠ 0QFO"* (15 3BEGPSE . &OHMJTI (15 3BEGPSE # &OHMJTI (15 #SPXO # &OHMJTI (PPHMF #&35 %FWMJO # &OHMJTI NBOZ+1WBSJBOUTBWBJMBCMF 5 3BGGFM 9VF # &OHMJTI 4XJUDI5SBOTGPSNFS 'FEVT # MBOHVBHFT 1B-. .JDSPTPGU .5/-( # &OHMJTI -*/& )ZQFS$-07" ,JN d# # +BQBOFTF /"7&3 )ZQFS$-07" ,JN d# # ,PSFBO .FUB 015 # &OHMJTI

Slide 17

Slide 17 text

̎ͭ໨ͷٕज़ֵ৽ɿڭࢣͳ͠ࣄલֶश #&35 (15 Inclusive multi-modal data Search QA … OCR Ad optimize … Pre-trained model Small data Fine tuning GT ڭࢣͳ͠ࣄલֶश ਖ਼ղ͕ཁΒͳֶ͍शํ๏ λεΫదԠͷͨΊͷ ϑΝΠϯνϡʔχϯά͕ඞཁ

Slide 18

Slide 18 text

େن໛ͳࣄલֶशϞσϧʹλεΫࢦࣔ͢Δ͜ͱͰɺଟ͘ͷλεΫΛ௥Ճֶशͳ࣮͘ݱͰ͖Δ͜ͱ͕ൃݟ͞Εͨʂ ̏ͭ໨ͷٕज़ֵ৽ɿϓϩϯϓςΟϯά (15 ೖ ྗ ޙ ஈ λ ε Ϋ ॲ ཧ *ODMVTJWFNVMUJNPEBMEBUB 4FBSDI 2" ʜ 0$3 "E PQUJNJ[F ʜ 1SFUSBJOFE 'PVOEBUJPONPEFM 'FXTIPUQSPNQUGPSUBTLTQFDJGJDBUJPO *ODMVTJWF NVMUJNPEBMEBUB 4FBSDI 2" ʜ 0$3 "E PQUJNJ[F ʜ 1SFUSBJOFE NPEFM 4NBMMEBUB 'JOF UVOJOH (5 ैདྷ λεΫ͝ͱͷϑΝΠϯνϡʔχϯά͕ඞཁ ϓϩϯϓςΟϯά λεΫࢦࣔͰԿͰ΋Ͱ͖Δʂ

Slide 19

Slide 19 text

ϓϩϯϓςΟϯάͷ6* ղઆ͔Βആ۟Λੜ੒͠·͢ɻ */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟ Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟ ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠ ͍ΔقޠͰ͢ɻ 065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬ ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ 065ؓ͞΍ؠʹ͠Έೖઊͷ੠ */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠ ͍Δɻ 065 λεΫ એݴ

Slide 20

Slide 20 text

Slide 21

Slide 21 text

Slide 22

Slide 22 text

Slide 23

Slide 23 text

Slide 24

Slide 24 text

Slide 25

Slide 25 text

ϓϩϯϓςΟϯάʹΑΔ঎඼આ໌จͷੜ੒

Slide 26

Slide 26 text

%#ͷεέʔϧ֦ு΋ॏཁ͕ͩɺਓͷհࡏ΋ॏཁͰ͋Δ͜ͱ͕ঃʑʹ൑໌ "*Λ͞ΒʹҭͯΔʂ 26 5-%3EBUBTFU SFEEJUDPNΑΓసࡌ .BTLFEMBOHVBHFNPEFMMJOH

Slide 27

Slide 27 text

࣌୅ʹٯߦ͢ΔΑ͏͕ͩɺڭࢣ͋ΓֶशΛ࢖ͬͯϑΝΠϯνϡʔχϯά͢Δ "*ΛҭͯΔ 'JOFUVOFEMBOHVBHFNPEFMTBSF[FSPTIPUMFBSOFST<8FJ *$-3>ΑΓసࡌ

Slide 28

Slide 28 text

Ұ෦ͷύϥϝʔλͷΈϑΝΠϯνϡʔχϯά "*ΛҭͯΔ 28 -P3"5SBOTGPSNFSͷ֤ϨΠϠʹ͓͍ͯɺ"EBQUPSNPEVMF ΛCZQBTTͤ͞ɺͦΕΛUVOJOH͢Δ͜ͱͰগ਺σʔλͰ΋աద ߹ͳ͘ɺ͔ͭޮ཰Α͘దԠͤ͞ΒΕΔख๏ )V -03"-083"/,"%"15"5*0/0'-"3(&-"/(6"(& .0%&-4 *$-3 ΑΓసࡌ

Slide 29

Slide 29 text

ਓͷհࡏΛ͏·͘׆༻Ͱ͖ΔֶशΞϧΰϦζϜͱͯ͠ͷ3- "*ΛҭͯΔ • ใुؔ਺Λਓͷ؆୯ͳϑΟʔυόοΫʹ ج͍ͮͯϑΟοςΟϯά • ݁Ռతʹେྔͷσʔλʹεέʔϧͤ͞Δ ͜ͱ͕Մೳ • ͜ΕʹΑΓෳࡶͳใुؔ਺΋දݱͰ͖Δ • ใु੒ܗʢSFXBSETIBQJOHʣͷҰͭͱߟ ͑ΒΕ͍ͯΔɻ • ౰࣌͸ɺ"$ 5310ͳͲͷڧԽֶशΛ༻ ͍ͯɺϩϘςΟΫεͱήʔϜͷλεΫʹ ׆༻ %FFQSFJOGPSDFNFOUMFBSOJOHGSPNIVNBOQSFGFSFODFT<$ISJTUJBOP BS9JW>ΑΓసࡌ

Slide 30

Slide 30 text

*OTUSVDU(15 <0VZBO .BS> 5SBJOJOHMBOHVBHFNPEFMTUPGPMMPXJOTUSVDUJPOTXJUIIVNBOGFFECBDL<0VZBOH >ΑΓసࡌ

Slide 31

Slide 31 text

ͦΕͧΕͷεςοϓʹඞཁͳσʔλྔ *OTUSVDU(15 <0VZBO .BS> 31 4VQFSWJTFEGJOFUVOJOH 3FXBSENPEFM 3-XJUI )VNBO'FFECBDL ,QSPNQUT IVNBOMBCFMFS ,QSPNQUT IVNBOMBCFMFS ,QSPNQUT

Slide 32

Slide 32 text

"#ςετܗࣜͰධՁͤ͞Δ͜ͱͰɺ3-)'༻ͷ৽ͨͳֶशσʔλΛूΊΔ )VNBOBOOPUBUJPOͷ6* 32

Slide 33

Slide 33 text

$IBU(15 ਓͷ໛ൣճ౴Λֶश͠ɺैདྷΑΓ΋๬·͍͠ճ౴͕Ͱ͖Δ 33 (15 検索質問回答 … 推薦対話 … Foundation model ؒҧͬͨ͜ͱΛ ಊʑͱ౴͑Δ ภݟ͕໰୊ൃݴ͕ ؚ·ΕΔճ౴Λ͢Δ $IBU(15 Tuned model ࣭໰ճ౴ ճ౴ධՁ ਓͷ࣭໰ճ౴΍ճ౴ධՁʹج͖ͮɺϞσϧΛ͞Βʹ܇࿅ ͢Δ͜ͱͰɺ๬·͍͠ճ౴ΛಘΒΕΔΑ͏ʹͨ͠ ճ౴ਫ਼౓͕޲্ ෆద੾ճ౴ͷ௿ݮ ݴ͍ճ͠ɾදݱͷվળ %FFQTQFFE .JDSPTPGU ͳͲͰ࣮૷ɿ IUUQTXXXBYJPO[POFEFFQTQFFEFYUSFNFTDBMFNPEFMUSBJOJOHGPSFWFSZPOF

Slide 34

Slide 34 text

ࢹ֮৘ใͱݴޠ৘ใΛೖྗͱͯ͠ɺݴޠΛग़ྗɻਤදೖͷࢼݧ໰୊Ͱ΋ߴಘ఺Λ࣮ݱ (15 $IBU(15 QMVT 34

Slide 35

Slide 35 text

ΞϝϦΧͷ༷ʑͳࢼݧͰɺਓؒͷ্Ґͷ੒੷Λ࢒͍ͯ͠Δ "*ͷ࣮ྗɿςΩετϕʔεͰ͸طʹਓؒΛ௒͑ͭͭ͋Δ 35 6OJGPSN#BS&YBN ౷Ұ࢘๏ࢼݧ 4"53FBEJOH8SJUJOH େֶਐֶదੑࢼݧ ϦʔσΟϯά ˍϥΠςΟϯά 4"5.BUI େֶਐֶదੑࢼݧ ਺ֶ (3&2VBOUJUBUJWF େֶӃڞ௨ࢼݧ ਺ֶ (3&7FSCBM େֶӃڞ௨ࢼݧ ݴޠ QFSDFOUJMF

Slide 36

Slide 36 text

ਓؒͰ͸͙͢ʹ࡞Εͳ͍จষΛॻ͚Δ "*ͷ࣮ྗɿςΩετϕʔεͰ͸طʹਓؒΛ௒͑ͭͭ͋Δ

Slide 37

Slide 37 text

แׅతͳϞσϧͱͦΕʹΑΔଟ༷ͳʢແݶͷʁʣλεΫͷ࣮ݱɺ͜ΕʹΑΓਓΛ௒͑ΒΕ ΔՄೳੑ͕ग़͖ͯͨʂ ج൫Ϟσϧ͕ࣔࠦ͢Δ͜ͱ υ ϝ Π ϯ λεΫ λεΫಛԽܕ"* λεΫಛԽܕ"* ج൫Ϟσϧ ͜ ͷ ෯ ͕ ڭ ࢣ ͳ ͠ ࣄ લ ֶ श Ͱ Մ ೳ ʹ ϓϩϯϓςΟϯά ڭ ࢣ ͳ ͠ ࣄ લ ֶ श ࣗ ݾ ڭ ࢣ ֶ श

Slide 38

Slide 38 text

%"--&<3BNFTI BS9JW > $-*1<3BEGPSE BS9JW > ϚϧνϞʔμϧͳੜ੒ 38 3BEGPSE -FBSOJOH5SBOTGFSBCMF7JTVBM.PEFMT'SPN/BUVSBM-BOHVBHF4VQFSWJTJPO BS9JW 3BNFTI )JFSBSDIJDBM5FYU$POEJUJPOBM*NBHF(FOFSBUJPOXJUI$-*1-BUFOUT BS9JW ΑΓసࡌ 画像と、テキストのエンコーディングの対から、対象学習を⾏うことで、双⽅の共通の埋め込み空間を学習

Slide 39

Slide 39 text

௿࣍ݩۭؒͰ %JGGVTJPONPEFMΛֶशਪ࿦͢Δ͜ͱͰɺলϦιʔε͔ͭߴ଎ͳֶशਪ࿦͕Մೳ 4UBCMFEJGGVTJPO<3PNCBDI "QS> 359 (#ϏσΦϝϞϦʔ Ҏ্ఔ౓ͷ(16͕͋Ε͹ࣗ୐Ͱ΋ಈ͔͢͜ͱ͕Ͱ͖ΔΑ͏ʹ ͳͬͨ IUUQTBSYJWPSHBCTΑΓసࡌ ίʔυɿ IUUQTHJUIVCDPN$PNQ7JTTUBCMFEJGGVTJPO IUUQTIVHHJOHGBDFDPTQBDFTTUBCJMJUZBJTUBCMFEJGGVTJPO

Slide 40

Slide 40 text

ੜ੒͞ΕΔը૾͕ಉ͡ʹͳΔΑ͏ʹɺೖྗจࣈྻͷ&NCFEEJOHΛ࠷దԽ ৽ͨͳʮ࡞෩ʯΛ֮͑ͤ͞Δɿ5FYUVSBMJOWFSTJPO<(BM > (BM "O*NBHFJT8PSUI 0OF8PSE1FSTPOBMJ[JOH5FYUUP*NBHF(FOFSBUJPOVTJOH5FYUVBM*OWFSTJPO BS9JW ΑΓసࡌ

Slide 41

Slide 41 text

ͨͬͨ਺ຕͷը૾Ͱաద߹͢Δ͜ͱͳ͘ϑΝΠϯνϡʔχϯάɺࣗݾੜ੒ͨ͠αϯϓϧΛ༻͍Δ͜ͱ Ͱաֶशͱ֓೦γϑτΛ๷ࢭ ৽ͨͳର৅Λ֮͑ͤ͞Δɿ%SFBN#PPUI <3VJ[ "VH> IUUQTBSYJWPSHBCT ΑΓసࡌ ίʔυɿ IUUQTHJUIVCDPNTNZESFBNCPPUIHVJ

Slide 42

Slide 42 text

গ਺ͷը૾Ͱաద߹͢Δ͜ͱͳ͘ϑΝΠϯνϡʔχϯά ͞·͟·ͳը૾ϓϩϯϓτͷ׆༻ɿ$POUSPM/FU<;IBOH > 42 基本的な考え⽅はLoRAなどに使われるAdaptorと同じ考え⽅、0で初期化するZeroConvが特徴的 Zhang+, Adding Conditional Control to Text-to-Image Diffusion Models, arXiv, 2023より転載

Slide 43

Slide 43 text

'MBNJOHP<"MBZSBD BS9JW > ϚϧνϞʔμϧ"*ʹ͓͚ΔϓϩϯϓςΟϯά "MBZSBD 'MBNJOHPB7JTVBM-BOHVBHF.PEFMGPS'FX4IPU-FBSOJOH BS9JW ը૾ɾςΩετͷฒͼ͔Βɺ࣍ͷฒͼΛ༧ଌ͢ΔֶशΛ ͓ͯ͘͜͠ͱͰ৽ͨͳೖྗʹର͢Δద੾ͳ౴͑Λਪ࿦

Slide 44

Slide 44 text

(BUPBHFOFSBMJTUBHFOU<3FFE BS9JW > ͞ΒʹϚϧνϞʔμϧʹɾɾɾ 44 3FFE "(FOFSBMJTU"HFOU BS9JW ΑΓసࡌ ೚ҙͷλεΫʹ͓͚Δ؍ଌͱΞΫγϣϯͷܥྻΛਖ਼نԽ͠τʔΫϯԽͯ͠ɺݴޠͱಉ͡Α͏ʹ༩͑ɺ 5SBOTGPSNFSߏ଄ͷϞσϧʹ͓͍ͯɺ࣍ͷτʔΫϯΛ༧ଌ͢ΔΑ͏ֶश

Slide 45

Slide 45 text

ྫɿࣸਅͷͲ͕͓͔͍͔͜͠Λཧղ͠ݴޠԽ ը૾Λཧղɾղऍ͢ΔೳྗͰ΋ਓʹฒΜͩʁ 45

Slide 46

Slide 46 text

ྫɿྫྷଂݿͷத਎ͷࣸਅΛݩʹɺͦΕΒͰԿ͕࡞ΕΔ͔Λਪન ը૾Λཧղɾղऍ͢ΔೳྗͰ΋ਓʹฒΜͩʁ 46

Slide 47

Slide 47 text

47 ྫɿςΩετͷࢦࣔΛཧղͯ͠ɺෳࡶͳΠϥετΛੜ੒ -BTDPBJ ը૾Λඳ͘ೳྗͰ΋ਓʹฒΜͩʁ

Slide 48

Slide 48 text

ಛఆͷϞμϦςΟ͔ΒϚϧνϞʔμϧ΁ɻͦΕʹ൐͍ͲΜͳλεΫ͕ߟ͑ΒΕΔ͔ʁ ͲΜͲΜͱυϝΠϯͱλεΫ͸޿͕Δɾɾ λεΫ ϓϩϯϓςΟϯά ڭ ࢣ ͳ ͠ ࣄ લ ֶ श ࣗ ݾ ڭ ࢣ ֶ श υ ϝ Π ϯ ϚϧνϞʔμϧԽ Ϛ ϧ ν Ϟ ồ μ ϧ Խ

Slide 49

Slide 49 text

൚༻ਓ޻஌ೳ "(*BSUJGJDJBMHFOFSBMJOUFMMJHFODF <(VCSVE > 49 "(* ਓؒϨϕϧͷ൚༻తͳ஌ೳ ঢ়گมԽʹରͯ͠ॊೈ /BSSPX"* ಛఆͷػೳɾঢ়گΛ૝ఆ͠ σʔλʹ࠷దԽͨ͠ػց ϓϩάϥϜͨ͠Ҏ্ͷ͜ͱ͕Ͱ͖Δɻ ಺ૠతʹߟ͍͑ͯΔ͕ɺͦͷൣғ͕ਓஐ Λ௒͍͑ͯΔ ڭ͑ͨ͜ͱ͔͠΍Βͳ͍ɻ ڭ͍͑ͯͳ͍͜ͱ͸Ͱ͖ͳ͍ɻ ਓͷΑ͏ʹҰͭͷ಄೴Ͱঢ়گʹ߹Θ͍ͤͯΖ͍Ζͳग़ྗ͕Մೳ

Slide 50

Slide 50 text

ੜ੒"*ͷ༻్Ձ஋͸ओʹछྨ 50 ޮ཰Խ ίϯςϯπੜ੒ ύʔιφϧԽ

Slide 51

Slide 51 text

ੜ੒ܕݕࡧɺݕࡧ݁ՌΛҰͭͣͭग़͢ͷͰ͸ͳ͘ɺ·ͱΊͨϨϙʔτΛදࣔ ৽ͨͳݕࡧύϥμΠϜ 51 .JDSPTPGU#JOH &EHF ੜ੒ܕݕࡧ ޮ཰Խ

Slide 52

Slide 52 text

$FSUBJOMZɿΧελϚʔαʔϏεͷࣗಈԽ ⼈⼒によるカスタマーサービスを代替 52 ޮ཰Խ

Slide 53

Slide 53 text

.JDSPTPGUɿϩϘοτͷ੍ޚίʔυ࡞੒ͷࣗಈԽ ςΩετͰࢦࣔΛग़ͤ͹ɺੜ੒"*͕ࣗಈͰίʔυʹม׵ 53 ޮ཰Խ

Slide 54

Slide 54 text

-*/&ɿ޿ࠂίϐʔ࡞੒ͷࣗಈԽ ਓྗ࡞ۀͷࣗಈԽʹΑΓੜ࢈ੑΛେ͖͘վળ 54 ίϯςϯπ ੜ੒

Slide 55

Slide 55 text

ը૾ʴΩϟονίϐʔʴϨΠΞ΢τͷੜ੒ʹΑΔ޿ࠂແݶੜ੒ ి௨ɿ޿ࠂੜ੒΁ͷԠ༻ Φ Ϧ Τ ϯ ೖ ྗ ి௨ใΑΓసࡌ IUUQTEFOUTVIPDPNBSUJDMFT ੜ ੒ ύ ϥ ϝ ồ λ ग़ ྗ Ϩ Π Ξ ΢ τ ੜ ੒ Ω Ỿ ỽ ν ੜ ੒ ը ૾ ੜ ੒ ޮ Ռ ༧ ଌ ίϯςϯπ ੜ੒

Slide 56

Slide 56 text

NJHOɿݐஙઃܭۀ຿ͷޮ཰Խ 56 キーワードや写真をインプット⽣成AIがデザインイメージをアウトプット ίϯςϯπ ੜ੒ “リビング“ ＋

Slide 57

Slide 57 text

&YQFEJBɿཱྀఔܾఆΛαϙʔτ ཱྀߦʹؔ͢Δ࣭໰ʹ౴͑Δ͜ͱͰɺސ٬ຬ଍౓ͱച্Λ޲্ 57 ύʔιφϧԽ

Slide 58

Slide 58 text

ੜ੒"*ͷԠ༻Ωϟϯόε ΤϯλʔςΠϯϝϯ τ ϚʔέςΟϯά Πϯλʔωοτ ڭҭ ҩྍ෱ࢱ ϑΝογϣϯɾσβ Πϯ Ұൠۀ຿ ޮ ཰ Խ ৘ใݕࡧ ը૾ɾԻָݕࡧ ৘ใਪનɺ޿ࠂ഑৴ ݕࡧɺϨϙʔςΟϯ άɺཱྀߦܦ࿏ɾҿ৯ ళɾߪങఏҊ ڭࡐࢧԉɺ"*νϡʔ λʔͳͲ ݸਓϨϕϧͰͷॳظ ਍அࢧԉ ίʔσΟϯάࢧԉ ৘ใՃ޻ ฤۂɺ੾Γൈ͖ಈը ੜ੒ɺτϨΠϥʔੜ ੒ ৘ใΩϡϨʔγϣϯ จॻɾϝʔϧ౳࡞੒ ࢧԉɾਤද࡞੒ ίϛϡχέʔ γϣϯ ΧελϚʔαϙʔτɺ νϟοτϘοτ ձ࿩ࢧԉ จॻɾϝʔϧ౳࡞੒ ࢧԉ ΧελϚʔαϙʔτ ί ϯ ς ϯ π ੜ ੒ ίϯςϯπੜ੒ өըɾ57٭ຊੜ੒ɺ ಈըɾָۂੜ੒ɺ࡞ ࢺ࡞ۂࢧԉɺήʔϜ ΩϟϥΫλʔɾγφ ϦΦੜ੒ɺΞχϝੜ ੒ ঎඼ɾ޿ࠂจͷੜ੒ɺ ޿ࠂੜ੒αΠτ࡞੒ɺ ϒϩάɾ4/4هࣄੜ ੒ɺ αΠτੜ੒ɺϒϩ άɾ4/4هࣄੜ੒ɺ ֶशίϯςϯπੜ੒ σβΠϯҊੜ੒ɺ"* Ϟσϧੜ੒ Πϥετ࡞੒ɺϏδ ωεઓཱུҊɾΞΠ σΟΤʔγϣϯࢧԉɺ ύ ồ ι φ ϧ Խ ਪનɾ൑அࢧԉ ύʔιφϥΠζϓϨ ΠϦετ ޿ࠂ৹ࠪࣗಈԽ ޿ࠂޮՌ࠷దԽɿૌ ٻϙΠϯτɾλʔ ήοτ౳ೖΕΕ͹ɺ ഔମɾΫϦΤΠςΟ ϒؚΊਪન 5SJQΞγετɿཱྀఔ ߏஙˍ༧໿ ࠾఺ɾධՁࢧԉ ڭҭϓϩάϥϜߏ ஙɿڭࡐબఆɾ೉қ ౓બఆɾɾɾ ਍அࢧԉ ݈߁ɾάϧϝΞγε τɿݙཱˍ഑ୡ༧໿ɺ ӡಈϝχϡʔߏங όʔνϟϧτϥΠϯ Φϯɺίʔσਪન ཁ໿ɺςετࢧԉ ϨϙʔςΟϯά τϨϯυϨϙʔτ ·ͱΊهࣄੜ੒ ใࠂॻ࡞੒ࢧԉ ใࠂॻ࡞੒ࢧԉɺ࿦ จɾಛڐࣥචࢧԉ ੜ੒ܕݕࡧɿݕࡧ݁ Ռͷ·ͱΊ

Slide 59

Slide 59 text

ຊ೔ͷߨԋͷ಺༰ • ੜ੒"*ͱͦͷՄೳੑ • -*/&Ͱͷ"* • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

Slide 60

Slide 60 text

-*/&ʹ͓͚Δ"*ݚڀ /-1 "41 $7- 5SVTUXPSUIZ "* .VMUJNPEBM"*

Slide 61

Slide 61 text

1VCMJDBUJPOTBUUIFUPQUJFSDPOGFSFODFT 3%BDDPNQMJTINFOUT ': $POGFSFODFT *$"441 */5&341&&$) 8"41"" #JH%BUB *$"441 &64*1$0 */5&341&&$) %$"4& "14*1" $713 51%1 '-*$.- -%3$ *$"441 *$3" *6* *$%& *$$7 "14*1" */5&341&&$) *&&&54141- *$-3 *$"441 */5&341&&$) 4*(.0% "14*1" *+$7 4-5 7-%# #JH%BUB 8"$7 """* *$"441 4-5

Slide 62

Slide 62 text

-*/&"*ͷ3%7JTJPO $POTFSWBUJWF %JTSVQUJWF 5JNF *OUFSBDUJWFWJSUVBM FYQFSJFODF "VUPOPNPVT"* XPSLGMPX %JHJUBM.F .F"7"5"3 %JHJUBM*EFOUJUZ #FUUFS$BSF 5SVTUXPSUIZ"* "*'BJSOFTT &YQMBJOBCMF"* %BSL%BUB 0NOJQPUFOU"* (JHBOUJD-BOHVBHFNPEFM 6OMBCFMFE%BUB %BUB.BSLFUQMBDF (FOFSBUJWF*OUFMMJHFODF /FX&EVDBUJPO %FQFOEBCMF455 1SJWBDZQSFTFSWJOH 4FBN%JTDSJNJOBUPS -*/&ͷ࣋ͭଟ༷ͳ"*ཁૉٕज़Λجʹɺ෼໺ԣஅతͳٕज़Λ஫ྗٕज़ͱͯ͠શମײΛ࣋ͬͯݚڀ։ൃ͍ͯ͠·͢

Slide 63

Slide 63 text

63 OPUPOMZ$PHOJUJPO CVU $SFBUJWJUZTVQQPSU

Slide 64

Slide 64 text

世界でも⼤規模⾔語モデルの開発は限られ⽇本ではHyperCLOVAが最⼤ ˞͜͜Ͱ͸#ఔ౓Ҏ্Λେن໛ͱදݱ Alibaba Tongyi Qianwen Cohere.ai Large LM DeepMind Gopher, Chinchilla, Flamingo Amazon Amazon Titan Baidu Ernie Bot Google T5, PaLM/PaLM-E, LaMDA/Bard MS-Nvidia Megatron-Turing NLG Anthropic Claude Kakaobrain koGPT3, Coyo LINE/WMJ HyperCLOVA NAVER HyperCLOVA OpenAI GPT/ChatGPT EleutherAI pythia Meta/Stanford/UCB OPT/LLaMA/Alpaca/Koala AI21 Labs Jurassic-1

Slide 65

Slide 65 text

จॻ࡞੒ࢧԉ ཁ݅ͷΈͷϝʔϧͰ࡞จ จॻߍਖ਼ɾ੔ܗɾλΠτϧੜ੒ σδλϧϚʔέςΟϯάࢧԉ Ωϟονίϐʔ࡞੒ࢧԉ େྔͷ঎඼ʹର͢Δ঎඼આ໌ ΧελϚέΞʔࢧԉ ސ٬Ұ࣍αϙʔτɾৼΓ෼͚ ίʔϧηϯλͷ͞ΒͳΔࣗಈԽ ৘ใऩूࢧԉ େྔͷσʔλʹج࣭ͮ͘໰ճ౴ จॻͷཁ໿ɾϨϙʔτੜ੒ݕࡧ ج൫Ϟσϧͷ׆༻ ͦͷଞ ίʔσΟϯάࢧԉɾจ๏νΣοΫɾ຋༁ɾɾɾ

Slide 66

Slide 66 text

-*/&ʹ͓͚Δج൫Ϟσϧߏங %FDPEFS

Slide 67

Slide 67 text

޿ࠂΩϟονίϐʔੜ੒ ΩʔϫʔυೖྗͷΈͰɺ؆қʹΩϟονΛෳ਺ੜ੒

Slide 68

Slide 68 text

ୈ̑ճ ର࿩γεςϜ ϥΠϒίϯϖςΟγϣϯ オープントラックシチュエーショントラック IUUQTTJUFTHPPHMFDPNWJFXETMD&#&&$ BVUIVTFS

Slide 69

Slide 69 text

%JBMPHVF3PCPU$PNQFUJUJPO *304ͷXPSLTIPQͱͯ͠։࠵͞Εͨର࿩ϩϘοτνϟϨϯδ

Slide 70

Slide 70 text

70 9"* 1SJWBDZDBSF UP $PMMBCPSBUJWF"* &UIJDBM'BJS"*

Slide 71

Slide 71 text

ج൫Ϟσϧʹ͓͚Δμ΢ϯαΠυݒ೦

Slide 72

Slide 72 text

"*ʹ͸ΨʔυϨʔϧ͕ඞཁ

Slide 73

Slide 73 text

/(ϫʔυʴ༗֐දݱ൑ఆ

Slide 74

Slide 74 text

)BMMVDJOBUJPOͷ௿ݮ

Slide 75

Slide 75 text

)BMMVDJOBUJPOͷ௿ݮ

Slide 76

Slide 76 text

5SVTUXPSUIZ"*ʹ޲͚ͯ What is Trustworthy AI? Expert Quality Explainability Transparency Confidentiality Fairness Harmless Robustness Compliance Trustworthy AI

Slide 77

Slide 77 text

&UIJDT3BEBS

Slide 78

Slide 78 text

ʢࢀߟʣޮ཰తͳςετํ๏ ҰํͰɺݴޠϞσϧ͸ྙཧతʹ໰୊ͷ͋ΔൃݴΛ͢ΔةݥੑΛ࣋ͭ ԥΔͧʂ YYY!NBJMDPN ·Ͱ࿈བྷͯ͠Ͷɻ ߈ܸతൃݴ ࿈བྷ৘ใͷ๫࿐

Slide 79

Slide 79 text

ఏҊख๏ɿ*UFSBUJWF4UPDIBTUJD'FXTIPU(FOFSBUJPO 3FE-. 5BSHFU-. 3FE$MG ߈ܸ ग़ྗ ௨஌

Slide 80

Slide 80 text

ఏҊख๏ɿ*UFSBUJWF4UPDIBTUJD'FXTIPU(FOFSBUJPO 3FE-. UFTUDBTF UFTUDBTF -JTUPGRVFTUJPOT \UFTUDBTF^ ϓϩϯϓτ 5BSHFU-. 3FE$MG ؂ࢹ ຒࠐ ೖྗ ςετέʔε Ϩεϙϯε ௥Ճ ςετέʔεϦετ

Slide 81

Slide 81 text

4BZVSJ • ୅ͷঁੑͷ੠ • গ͠໌Δ͘ɺਓؒΒ͍͠཈༲ %BJLJ • ୅ͷஉੑͷ੠ • མͪண͍ͨ੠ /BPNJ@# • /BPNJͷൃ࿩ελΠϧ֦ு • ঎඼΍ϨγϐͳͲΛ͓͢͢Ί͢Δςϯ γϣϯߴ͍ελΠϧ /BPNJ@" • ୅ͷঁੑͷ੠ • ໌Δ͞ͱམͪண͖ͷ͋Δ੠ • ਓؒΒ͍͠཈༲ 5PNPLP • ୅ͷঁੑͷ੠ • গ͠໌Δ͘ɺਓؒΒ͍͠཈༲ ଞʹ΋ੜ੒"*͍Ζ͍Ζ

Slide 82

Slide 82 text

$0OUSPMMBCMF )JHIRVBMJUZ "OEFYQ3FTT*WF554 ໌Δ͍ ҉͍ 😀 😄 🙂 😐 😢 😰 😥 ײ৘Λॊೈʹ੍ޚՄೳͳԻ੠߹੒Λ࣮ݱ ײ৘ͷ͜΋ͬͨԻ੠߹੒

Slide 83

Slide 83 text

Slide 84

Slide 84 text

Slide 85

Slide 85 text

ࠓ·Ͱ࣮ݱ͖ͯͨ͠ιϦϡʔγϣϯ ஫ʣݱࡏ͸ؔ࿈ձࣾͰ͋ΔϫʔΫεϞόΠϧδϟύϯࣾΑΓఏڙத

Slide 86

Slide 86 text

ຊ೔ͷߨԋͷ಺༰ • ੜ੒"*ͱͦͷՄೳੑ • -*/&Ͱͷ"* • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

Slide 87

Slide 87 text

ࢲ͕ߟ͑ͨάϥϯυνϟϨϯδ "*ͰղܾͰ͖͍ͯͳ͍໰୊ 87 • ίϯύΫτͳ൚༻ਓ޻஌ೳ • লΤωਪ࿦ɾলΤωֶशʢେن໛ͳܭࢉثΛඞཁͱ ͠ͳ͍ʣ • ώτͷ೴ͷΑ͏ͳػೳ෼ԽΛ࣮ݱ͢ΔͨΊͷҰൠత ͳख๏ • ϩδΧϧͳਪ࿦ࢉज़ԋࢉɾ࿦ཧԋࢉͳͲ • ϖϧιφɾݸੑɾҰ؏ੑɿֶशͱهԱͷ౷߹ʹؔ͢ ΔҰൠతͳख๏ • ໨త΍໨ඪΛ࡞Γग़͢ɺαϒΰʔϧઃఆͳͲ • ৽ͨͳυϝΠϯ΁ͷదԠʢ[FSPGFXTIPUʣ • Ԡ༻Ϩϕϧͷ՝୊ • "*ྙཧɾݖར໰୊ɾϋϧγωʔγϣϯ $IBU(15ͷߟ͑ΔάϥϯυνϟϨϯδ

Slide 88

Slide 88 text

No content