Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
[email protected]
参加レポ
Search
kzmssk
March 26, 2020
Research
1.2k
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
[email protected]
参加レポ
アメリカ人工知能学会の国際会議AAAI2020のレポートです。聴講した中で面白いと思った研究を紹介しています。
kzmssk
March 26, 2020
More Decks by kzmssk
See All by kzmssk
SAM3を用いたコマ・吹き出しの 領域検出と分割構造からの読み順推定
kzmssk
0
190
実例から見るLLMのマンガ理解:実務VQAタスクによる長期的文脈と視覚情報の定性評価
kzmssk
0
190
Dwangoでの漫画データ活用〜漫画理解と動画作成〜@コミック工学シンポジウム2025
kzmssk
0
410
マールタスの仕組みを解説
kzmssk
0
3.9k
強化学習による3Dキャラクターの モーション生成@SIGGRAPH ASIA2018
kzmssk
0
1.5k
Montezuma’s Revenge Solved by Go-Explore, a New Algorithm for Hard-Exploration Problems
kzmssk
0
480
深層学習を用いた3D仮想物理環境における キャラクターの動的な行動生成
kzmssk
0
370
Other Decks in Research
See All in Research
第64回CV・PRML勉強会 論文紹介:Linguistic Priors for Visual Decoupling: Towards Symmetric Vision-Brain Alignment
sokikatayama
0
250
GLIM とMegaParticles:正規分布近似の限界とタイトカップリング&パーティクルフィルタの進展 / GLIM and MegaParticles : Progress of the distribution representation in SLAM
koide3
0
950
視覚若手の会LENSって何??
mickey_0226
0
350
VLMの推論を高速化する視覚トークン削減の仕組み
tattaka
2
370
某助成金プロジェクト採択に向けて企業研究所のアウトリーチ専任者がやったこと
afroscript
0
200
Karkada さんの論文 × 2 の紹介: (1) Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models, (2) Symmetry in language statistics shapes the geometry of model representations
eumesy
PRO
1
850
Bilan travail en élevage : première restitution des enquêtes 2025-2026
institutdelelevage
PRO
0
210
最先端NLP勉強会2026 論文紹介:Reasoning with Sampling: Your Base Model is Smarter Than You Think (ICLR 2026 paper)
kogoro
4
710
OWASP AISVS - C7
shiell
2
830
CDCL を用いた MILP の厳密解法
imai448
0
300
量子サマースクール2026「量子計算機アーキテクチャ分野の概観」
youten622
1
970
ros2-perf-multihost: 分散システムにおける客観的なアーキテクチャ評価フレームワーク
takasehideki
0
290
Featured
See All Featured
Bioeconomy Workshop: Dr. Julius Ecuru, Opportunities for a Bioeconomy in West Africa
akademiya2063
PRO
1
400
[SF Ruby Conf 2025] Rails X
palkan
3
1.4k
Building Experiences: Design Systems, User Experience, and Full Site Editing
marktimemedia
1
620
Thoughts on Productivity
jonyablonski
76
5.4k
Reflections from 52 weeks, 52 projects
jeffersonlam
356
22k
How Software Deployment tools have changed in the past 20 years
geshan
2
35k
[Rails World 2026] Durable orchestration on Rails: from continuation to workflow
palkan
1
730
GitHub's CSS Performance
jonrohan
1033
470k
B2B Lead Gen: Tactics, Traps & Triumph
marketingsoph
1
280
Fireside Chat
paigeccino
43
4.1k
Navigating the Design Leadership Dip - Product Design Week Design Leaders+ Conference 2024
apolaine
2
470
Agile Actions for Facilitating Distributed Teams - ADO2019
mkilby
0
300
Transcript
[email protected]
ࢀՃϨϙ %XBOHP.FEJB7JMMBHF ࠤʑɹҰຏ
4QFDJBMFWFOUGFBUVSJOHUIF5VSJOH"XBSE8JOOFST :PTIVB#FOHJPਓؒͷແҙࣝతͳػೳʢʹݱࡏͷ"*ʣ͔ΒҙࣝతͳػೳΛ࣋ͬͨγεςϜʹ͚ͯ (FP⒎SFZ&)JOUPOɿ$BQTVMF/FUͷ :BOO-F$VO4FMGTVQFSWJTFEMFBSOJOHJTGVUVSFʢະདྷ༧ଌͷʣ "*)JTUPSZ1BOFM"EWBODJOH"*CZ1MBZJOH(BNFT ήʔϜ͕"*ٕज़ͷൃలʹͲ͏د༩͖ͯͨ͠ͷ͔ Πϕϯτ
%FFQ.JOEͷݚڀऀͨͪʹΑΔXPSLTIPQɻگɻ "UBSJ)&"%"UBSJ)VNBO&ZF5SBDLJOHBOE%FNPOTUSBUJPO%BUBTFU "UBSJͷήʔϜΛਓ͕ؒϓϨΠ͢ΔࡍͷࢹઢσʔλΛऩूͨ͠ݚڀ ࢹઢσʔλ3-ΤʔδΣϯτͷֶशੑೳ্ʹཱͭ͜ͱ͕Θ͔ͬͨ *OWJUFE5BMLCZ(FSSZ5FTBVSP 5%(BNNPO ڧԽֶशͰόοΫΪϟϞϯΛͬͨਓʣ ྺ࢙ͷ FHڧԽֶशΛ࠷ॳʹͬͨͷ୭͔ʁ$MBVEF4IBOOPJO
*OWJUFE5BMLCZ.JDIBFM#PXMJOH ଟਓڠྗήʔϜʮ)BOBCJʯͷઆ໌ 5IFPSZPG.JOE͕ඞཁͳήʔϜσβΠϯͩͱ͍͏ओு *OWJUFE5BMLCZ%BWJE4JMWFS NV;FSPͷ 2"ʮ࿈ଓߦಈͷ֦ுՄೳʁʯˠʮ.$54Λ࿈ଓߦಈʹ֦ுͰ͖Εग़དྷͦ͏ʯ 8PSLTIPQ3FJOGPSDFNFOU-FBSOJOHJO(BNFT [1]
໘ന͍ͱࢥͬͨݚڀͨͪ
ɿ%FFQ2-FBSOJOHͷֶश͕ෆ҆ఆ Ξϓϩʔνͱ࣮ ෳͷ5FBDIFSํࡦΛ༻ҙͦ͠ΕͧΕ͕5%ֶश 3FQMBZ#V⒎FSڞ༗ 4UVEFOUํࡦʹ5FBDIFS͔Βৠཹ ࣮ݧ "5"3*ͷͭͷήʔϜ 5FBDIFSͭΑΓෳͷํ͕ྑ͍݁Ռ ײ 0⒎QPMJDZͷಛΛ׆͔ͨ͠ݚڀ
5FBDIFSํࡦΛͦΕͧΕҧ͏ͷʹม͑ͨΒͲ͏ͳΔ͔ؾʹͳͬͨ &OTFNCMF1PMJDZ%JTUJMMBUJPOJO%FFQ3FJOGPSDFNFOU-FBSOJOH จ Fig.1 ΑΓ [2]
ɿ&YQFSJFODF3FQMBZʹ͓͚Δഁ໓త٫ Ξϓϩʔνͱ࣮ ͨ͘͞Μ๚ΕΔঢ়ଶΛΑΓ༏ઌతʹαϯϓϦϯά͢Δ ๚සʢಛʹ࿈ଓঢ়ଶͷ߹ʹʣܭࢉࠔ 3FQMBZCV⒎FS͔Βͨ͘͞ΜऔΔˠPOQPMJDZͷঢ়ଶͱͷڑ͕͍ۙUPQ,ݸΛͬͯํࡦߋ৽ ڑίαΠϯྨࣅೋޡࠩ ࣮ݧ .V+P$P "UBSJ ੑೳগ͠ྑ͘ͳΔ͕ɺ࠷దԽʹ͔͔Δ͕࣌ؒ༨ܭʹ͔͔ΔʢαϯϓϧΛͨ͘͞Μࣺ͍ͯͯΔͷͰʣ
ײ อकతͳํ๏ 1SJPSJUZΛ5%ޡࠩҎ֎ͷํ๏Ͱܭࢉ͢Δͷྑ͍ΞΠσΞ "UUFOUJWF&YQFSJFODF3FQMBZ [3]
ɿ4BNQMFDPNQMFYJUZ ҰఆͷੑೳΛୡ͢ΔͨΊʹඞཁͳαϯϓϧʣͷߴ͞ ਓؒͷσʔλͳͲͷσϞϯετϨʔγϣϯΛ͏ͱऩଋૣ͍͕σϞΛ͑ΒΕͳ͍ Ξϓϩʔνͱ࣮ɿ ༧ଌͨ͠Ձʹ৴པੑ͕ͳ͍ͱ͖σϞϯετϨʔγϣϯͷ༧ଌΛ͏ʢΞυόΠεʣ ৴པੑՁͷࢄͰଌΔ ࣮ݧ (SJEXPSME 1POH "5"3*
1POHͷ߹ࣄલֶशͨ͠ڧ͍BHFOUΛEFNPOTUSBUPSʹ͢Δ ͲͪΒΞυόΠεͳ͠ɺϥϯμϜͳΞυόΠεͷ߹ΑΓੑೳ্͕ ײ 2"ͰʮֶशରͷํࡦΞυόΠεΛˋड͚ೖΕΔͨΊσϞҎ্ͷੑೳ͕ग़ͳ͍ͷͰͳ͍͔ʯ ͱ͍͏͕ٞ͋ͬͨͷ͕ڵຯਂ͔ͬͨ ऩଋ͖ͯͨ͠ΒٯʹΞυόΠεΛཪΔํ๏Λࢼ͢Α͏ͳώϡʔϦεςΟοΫಋೖͰ͖ͦ͏ 6ODFSUBJOUZ"XBSF"DUJPO"EWJTJOHGPS%FFQ3FJOGPSDFNFOU-FBSOJOH"HFOUT6ODFSUBJOUZ "XBSF"DUJPO"EWJTJOHGPS%FFQ3FJOGPSDFNFOU-FBSOJOH"HFOUT [4]
ɿ4FRVFOUJBMͳܭࢉάϥϑΛޮతʹ࣮ߦͰ͖ΔΑ͏ʹෳͷ$16ʹׂΓ͍ͯͨ ྫɿ5FOTPSqPXͷܭࢉάϥϑͷ࣮ߦσόΠεׂΓͯ εϧʔϓοτ͕࠷େԽ͢ΔΑ͏ͳׂΓͯΛ୳ࡧ͢Δͷ/1ࠔ Ξϓϩʔνͱ࣮ɿ ڧԽֶशͱͯ͠ଊ͑ΔɿσόΠεʢ$16 ͷׂΓͯΛߦಈɺঢ়ଶΛάϥϑɺใुΛεϧʔϓοτ (SBQIDPOWͨ͠ಛྔΛ-45.ͰΈࠐΜͩϞσϧΛ3&*/'03$&Ͱ࠷దԽ γϛϡϨʔλʔΛͬͯεϧʔϓοτΛܭࢉ ࣮ݧɿγϛϡϨʔλʔͰධՁɻܶతʹվળ͠ͳ͔͕ͬͨˋͷྫͰ)FVSJTUJDΑΓྑ͍݁Ռ ײɿ࣮ݧͰ༻ͨ͠άϥϑ͕؆୯͗͢Δؾ͕͢Δ͕ɺγϛϡϨʔλͷਫ਼͕Α͚Εྑͦ͞͏
(FOFSBMJ[BCMF3FTPVSDF"MMPDBUJPOJO4USFBN1SPDFTTJOHWJB%FFQ3FJOGPSDFNFOU-FBSOJOH [5]
ɿϚϧνΤʔδΣϯτͷܥʹ͓͚ΔΤʔδΣϯτؒͷΠϯλϥΫγϣϯΛϞσϦϯά͍ͨ͠ ΠϯλϥΫγϣϯ0 /?E Ͱେม Ξϓϩʔνͱ࣮ ΤʔδΣϯτͷঢ়ଶભҠͱଞΤʔδΣϯτͷग़ྗʢߦಈʣΛͦΕͧΕχϡʔϥϧωοτͰϞσϦϯά ঢ়ଶΨεʹै͏ͱԾఆ͢Δ ΠϯλϥΫγϣϯग़ྗʹʮͲͷΤʔδΣϯτʹରͯ͠ΑΓಇ͖͔͚Δ͔ʯͷ֬Λ͔͚ͨઢܗͱͯ͠දݱ "UUFOUJPOʹͳ͍ͬͯΔɻ0 /
ͷܭࢉྔɻ ࣮ݧ όωͰܨ͕࣭ͬͨɺόεέοτબखͷಈ͖ͳͲͷσʔλΛϞσϦϯάͨ͠ࡍͷ༧ଌޡࠩΛධՁ γϯϓϧͳ-45.ͳͲʹൺͯྑ͍݁Ռ ײ ֬తͳঢ়ଶભҠΛಋೖ͢Δ͜ͱʹΑͬͯγεςϜͷྗֶతͳੑ࣭͕Ͳ͏มΘΔͷ͔͕ؾʹͳͬͨ (FOFSBUJWF"UUFOUJPO/FUXPSLTGPS.VMUJ"HFOU#FIBWJPSBM.PEFMJOH [6]
ϊʔυ͕ؒ౸ୡՄೳ͔Ͳ͏͔͕࣌ؒมԽ͢Δಈతͳແࢦੑάϥϑ Ұఆ࣌ؒʹ9͔Β౸ୡՄೳͰ͋Δͷू߹ʢSFBDIBCJMJUZTFUTʣΛߟ͑Δ άϥϑߏͷΑΓʮࣗવʯͰʮJOGSBTUSVDUVSFGSJFOEMZʯͳૢ࡞ʹΑͬͯSFBDIBCJMJUZTFUT Λ࠷దԽ͍ͨ͠ Ξϓϩʔνͱ࣮ ఏҊૢ࡞ɿ.FSHJOHʢपลͷΤοδͷ౸ୡՄೳ࣌ؒΛϚʔδ͢Δʣ ఏҊૢ࡞ɿ%FMBZJOHʢಛఆͷΤοδͷ౸ୡՄೳ࣌ؒʹΕΛೖΕΔʣ ࣮ݧ .FSHJOH
%FMBZJOHڞʹ/1ࠔͰ͋Δ͜ͱΛূ໌ͨ͠ ͨͩ͠%FMBZJOHͷΕʹ੍ݶ͕ͳ͚ΕۙࣅΞϧΰϦζϜ͕ଘࡏ ײ ఏىతͳจɻಈతͳάϥϑྺ࢙͕ઙ͍ɻ ੩తͳάϥϑʹൺ͍͕ͯ͠ɺϒϥβ্ͷϖʔδભҠϢʔβʔྲྀಈͷղੳͳͲԠ༻ઌ ͨ͘͞Μ͋Γͦ͏ 0QUJNJ[JOH3FBDIBCJMJUZ4FUTJO5FNQPSBM(SBQITCZ%FMBZJOH จ Fig.6 ΑΓ [7]
γʔέϯγϟϧͳϨίϝϯυͰ"UUFOUJPO͖ͷϞσϧ͕ߴ͍ੑೳΛग़͍ͯ͠Δ ͔͠͠طଘख๏Ͱ"UUFOUJPOͷߏ͕λεΫґଘʹͳ͍ͬͯΔ Ξϓϩʔνͱ࣮ "UUFOUJPOϕΫτϧΛϢʔβʔɺΞΠςϜɺΞΠςϜͷڞىߦྻʢΞΠςϜ"Λબͨ͠Ϣʔβʔ#બ͢ΔʣΛͬ ͯٻΊΔ "UUFOUJPOਖ਼نʢ4LFXOPSNBMEJTUSJCVUJPOʣʹै͏ͱԾఆɻ3FQBSBNFUFSJ[BUJPOUSJDL͏ ΞΠςϜͷڞىߦྻΞΠςϜͷϖΞසɺΞΠςϜͷFNCFEEJOHɺϢʔβͷFNCFEEJOHͷͭͷઢܗͰύϥϝλϥ Πζ ࣮ݧ
ଞͷطଘख๏ʹൺͯ༏Ґͳ݁Ռ ֶशʹΑͬͯಘΒΕͨڞىߦྻදݱͰಉ͡өըγϦʔζ͕ߴ͍ʹͳΔ͜ͱ͕؍͞Εͨ ײ "UUFOUJPOʹΛԾఆ͢Δํ๏ଞͷλεΫʹԠ༻Ͱ͖ͦ͏ 4FRVFOUJBM3FDPNNFOEBUJPOXJUI3FMBUJPO"XBSF,FSOFMJ[FE4FMG"UUFOUJPO [8]
େͳΞΠςϜͷϨίϝϯυͷͨΊʹಛྔϕΫτϧΛྨࣅੑΛอͬͨ· ·࣍ݩݮ͍ͨ͠ -PDBMJUZTFOTJUJWFIBTIJOH -4) 3FDPNNFOEBUJPO3FUSJFWBM ͔͠͠ϋογϡؔΦϑϥΠϯͰ֫ಘ͞ΕΔͷͰΞΠςϜϢʔβʔ͕ม ΘΔͱਫ਼͕Լ͕Δ Ξϓϩʔνͱ࣮ ϋογϡʹϕϧψʔΠΛԾఆ
ϢʔβʔͱΞΠςϜͷಛྔϕΫτϧΛޯ๏Ͱ࠷దԽͯ͠ಘΔ ײ U4/&ʹࣅ͍ͯΔ .FUSJDMFBSOJOHͱͷ͕ؔؾʹͳΔ 4UPDIBTUJDBMMZ3PCVTU1FSTPOBMJ[FE3BOLJOHGPS-4)3FDPNNFOEBUJPO3FUSJFWBM จ Fig.3 ΑΓ [9]
<>;IBOH 3VPIBO $BMFO8BMTIF ;IVPEF-JV -JO(VBO ,BSM4.VMMFS +BLF"8ISJUOFS -VYJO;IBOH .BSZ.)BZIPF BOE
%BOB)#BMMBSEl"UBSJ)&"%"UBSJ)VNBO&ZF5SBDLJOHBOE%FNPOTUSBUJPO%BUBTFUz*O"""*/FX:PSL64" <>4VO :VYJBOHl&OTFNCMF1PMJDZ%JTUJMMBUJPOJO%FFQ3FJOGPSDFNFOU-FBSOJOH6ODFSUBJOUZ3FEVDUJPOWJB3FWFSTF,-z*O """* /FX:PSL64" <>4VO 1FJRVBO 8FOHBOH;IPV BOE)PVRJBOH-Jl"UUFOUJWF&YQFSJFODF3FQMBZz*O"""*/FX:PSL64" <>4JMWB 'FMJQF-FOP%B 1BCMP)FSOBOEF[-FBM #JMBM,BSUBM BOE.BUUIFX&5BZMPSl6ODFSUBJOUZ"XBSF"DUJPO"EWJTJOHGPS%FFQ 3FJOGPSDFNFOU-FBSOJOH"HFOUTz*O"""*/FX:PSL64" <>/J 9JBOH +JOH-J .P:V 8BOH;IPV BOE,VO-VOH8Vl(FOFSBMJ[BCMF3FTPVSDF"MMPDBUJPOJO4USFBN1SPDFTTJOHWJB%FFQ 3FJOGPSDFNFOU-FBSOJOHz*O"""*/FX:PSL64" <>-J (VBOHZV #P+JBOH )BP;IV ;IFOHQJOH$IF BOE:BO-JVl(FOFSBUJWF"UUFOUJPO/FUXPSLTGPS.VMUJ"HFOU#FIBWJPSBM .PEFMJOHz*O"""*/FX:PSL64" <>%FMJHLBT "SHZSJPT BOE*HPS1PUBQPWl0QUJNJ[JOH3FBDIBCJMJUZ4FUTJO5FNQPSBM(SBQITCZ%FMBZJOHz*O"""*/FX:PSL 64" <>%FMJHLBT "SHZSJPT BOE*HPS1PUBQPWl0QUJNJ[JOH3FBDIBCJMJUZ4FUTJO5FNQPSBM(SBQITCZ%FMBZJOHz*O"""*/FX:PSL 64" <>-F %VOH% BOE)BEZ8-BVXl4UPDIBTUJDBMMZ3PCVTU1FSTPOBMJ[FE3BOLJOHGPS-4)3FDPNNFOEBUJPO3FUSJFWBMz*O"""* /FX:PSL64" จݙใ