Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
PLDI '21論文読み会: Provable Repair of Deep Neural N...
Search
Idein
June 08, 2022
Research
2
1.6k
PLDI '21論文読み会: Provable Repair of Deep Neural Networks
Idein
June 08, 2022
Tweet
Share
More Decks by Idein
See All by Idein
PLDI '21論文読み会: DNNFusion: Accelerating Deep Neural Networks Execution with Advanced Operator Fusion
ideininc
1
1.6k
PLDI '21論文読み会: AKG: Automatic Kernel Generation for Neural Processing Units using Polyhedral Transformations
ideininc
0
1.5k
PLDI '21論文読み会: Specification Synthesis with Constrainted Horn Clauses
ideininc
0
1.4k
PLDI '21論文読み会: Cyclic Program Synthesis
ideininc
0
1.4k
PLDI '21論文読み会: High Performance Correctly Rounded Math Libraries for 32-bit Floating Point Representations
ideininc
0
1.4k
PLDI '21論文読み会: Quantum Abstract Interpretation
ideininc
0
1.4k
Idein会社紹介資料(積極採用中)
ideininc
0
33k
Other Decks in Research
See All in Research
言語と数理の交差点:テキストの埋め込みと構造のモデル化 (IBIS 2024 チュートリアル)
yukiar
3
730
[CV勉強会@関東 CVPR2024] Visual Layout Composer: Image-Vector Dual Diffusion Model for Design Layout Generation / kantocv 61th CVPR 2024
shunk031
1
450
熊本から日本の都市交通政策を立て直す~「車1割削減、渋滞半減、公共交通2倍」の実現へ~@公共交通マーケティング研究会リスタートセミナー
trafficbrain
0
140
20240918 交通くまもとーく 未来の鉄道網編(こねくま)
trafficbrain
0
230
129 2 th
0325
0
240
Introducing Research Units of Matsuo-Iwasawa Laboratory
matsuolab
0
920
大規模言語モデルを用いた日本語視覚言語モデルの評価方法とベースラインモデルの提案 【MIRU 2024】
kentosasaki
2
520
言語処理学会30周年記念事業留学支援交流会@YANS2024:「学生のための短期留学」
a1da4
1
240
TransformerによるBEV Perception
hf149
1
430
20240820: Minimum Bayes Risk Decoding for High-Quality Text Generation Beyond High-Probability Text
de9uch1
0
120
12
0325
0
190
20240918 交通くまもとーく 未来の鉄道網編(太田恒平)
trafficbrain
0
220
Featured
See All Featured
Principles of Awesome APIs and How to Build Them.
keavy
126
17k
How To Stay Up To Date on Web Technology
chriscoyier
788
250k
Become a Pro
speakerdeck
PRO
25
5k
The Invisible Side of Design
smashingmag
298
50k
VelocityConf: Rendering Performance Case Studies
addyosmani
325
24k
Producing Creativity
orderedlist
PRO
341
39k
Measuring & Analyzing Core Web Vitals
bluesmoon
4
120
The World Runs on Bad Software
bkeepers
PRO
65
11k
How to train your dragon (web standard)
notwaldorf
88
5.7k
GitHub's CSS Performance
jonrohan
1030
460k
Easily Structure & Communicate Ideas using Wireframe
afnizarnur
191
16k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
232
17k
Transcript
1SPWBCMF3FQBJSPG%FFQ /FVSBM/FUXPSLT 1-%*ಡΈձ!*EFJO
Ϟνϕʔγϣϯ w %//͕TBGFUZDSJUJDBMͳྖҬʹΘΕग़͠͡Ίͨ w %//Ϟσϧ͕࣋ͭzVOTBGFͳڍಈͷݕग़ʹର͢Δಈػͱͳ͖ͬͯͨɽ w ྫ "EWFSTBSJBM&YBNQMFT w zVOTBGFͳڍಈz͕ݕग़͞Εͨޙͷɼ%//Ϟσϧͷzमਖ਼͕͍ͨ͠
w ͜ͷจɼ%//ੑ࣭ͷྑ͍ύονΛͯ ΒΕ ΔΈͱͦͷύον ͷߏ๏ʹ͍ͭͯͷఏҊ
طଘͷमਖ਼ํ๏ GJOFUVOJOH w 'JOFUVOJOH࣮ੈքͷӡ༻γφϦΦʹద߹͍͠ w ඇޮ w ݩʑͷσʔληοτ͕ࠓೖखͰ͖Δঢ়ଶɾ͔݅Θ͔Βͳ͍ w ৽ͨͳόά
҆શͰͳ͍ڍಈ ͷࠞೖ w ϋΠύϥνϡʔϯ͕ηϯγςΟϒ w ESBXEPXO fi OFUVOJOHʹΑΔ ݩʑຬͨ͞Ε͍ͯͨഺͷੑ࣭ͷݮগ
lमਖ਼zΞϧΰϦζϜʹٻΊΒΕΔੑ࣭ w ޮೳमਖ਼ͨ͠%//Ͱ͋Δ/`मਖ਼લͷ%///͕VOTBGFͳڍಈΛ͢Δೖྗσ ʔληοτ9ͷೖྗʹର͠ɼظ௨ΓৼΔ͏Α͏ʹमਖ਼͞ΕΔ w ൚Խੑೳ/9ͷೖྗʹࣅͨೖྗʹରͯ͠ɼظ௨ΓৼΔ͏Α͏ʹम ਖ਼͞ΕΔ w ہॴੑ/9ͷೖྗʹࣅͯͳ͍ೖྗʹରͯ͠ɼ/ͱಉ༷ʹৼΔ͏Α͏ मਖ਼͞ΕΔ
w ޮ/ʹର͠/`͕͢Έ͔ʹٻΊΒΕΔ
ఏҊख๏ w %FDPVQMFE%//T %%//T ΞʔΩςΫνϟͷಋೖ w ݩͷ%//ʹύον༻αϒܭࢉάϥϑΛޙ͚ͨ͠%// w 1SPWBCMF1PJOUXJTF3FQBJS w
%%//T্Ͱͷ༗ݶʹର͢Δࢦఆ୯ʹର͢ΔॏΈमਖ਼ w 1SPWBCMF1PMZUPQF3FQBJS w ˢͷ༗ݶΛತଟ໘ମͷू߹ແݶʹ֦ுͨ͠ͷ
1JFDFXJTFMJOFBS 18- ؔ w ͋ΔؔGͰ w EPNBJOΛ༗ݶݸͷଟ໘ମʹׂͰ͖Δ w ׂ֤ͨ͠EPNBJOຖʹGͱՁͳΞϑΟϯมG`͕ଘࡏ͢Δ w
ྫ3F-6-FBLZ3F-6ɼ3F-6ɼ$MJQ͔͠ΘΕ͍ͯͳ͍%//
Λ-1ʹؼண͍ͤͨ͞ɼ͕ w %//ඇઢܗͳ׆ੑԽؔΛ࣋ͭ w Γ͍ͨ͜ͱ 1PJOUXJTF3FQBJSͷϕʔεͱͳΔΞΠσΞ w 18-%//EPNBJOΛదʹׂ͢Δͱɼͦͷׂͨ͠ྖҬຖʹઢܗʹ ѻ͑Δ
w मਖ਼͍ͨ͠ೖྗσʔληοτ9தͷ֤͕ଐ͢ΔઢܗྖҬʹର͠ɼظ͢Δ ݁ՌͱͷؒͰ݅ࣜͱ͠ɼॏΈΛ࠷దԽ͢Δ-1Λղ͘
Λ-1ʹؼணͤ͞ΒΕͳ͍ɼ୯७ʹ w ॏΈΛม͑ΔͱEPNBJOͷׂ·ͨมΘͬͯ͠·͏ͨΊ w Մೳͳͷ׆ੑԽؔΛ࣋ͨͳ͍࠷ऴͷΈ w தؒʹΔͱμϝ
ׂ͕มΘͬͯ͠·͏ྫ Z3F-6 Y 3F-6 Y 3F-6 Y w Γ͍ͨमਖ਼ w
Yͷͱ͖㱡Z㱡 w Yͷͱ͖㱡Z㱡 w YˠIͷॏΈΛʹमਖ਼͢Δ ਤC ্ ͱׂ ਤCԼͷ3(#۠ؒ ม Θͬͯ͠·͍ظ௨ΓʹͳΒͳ ͍
%FDPVQMFE%//T w ݩͷ%//͔ΒՁͳͷΛػցతʹ ߏͰ͖Δɼͭͷαϒάϥϑ͔Βͳ Δߏͷ%// w "DUJWBUJPODIBOOFM w 7BMVFDIBOOFM w
ׂΛม͑ͣʹॏΈͷมߋ͚ͩө͞ ͤΒΕΔ
%FDPVQMFE%//T "DUJWBUJPODIBOOFM w ׂͷҡ࣋ͷͨΊ w ೖྗ͔Βݩͷ%//Ͱܭࢉ͠ɼमਖ਼ର ͷ݁ՌΛ7BMVFDIBOOFMͷରԠ ʹ͓͚Δܭࢉʹөͤ͞Δ w ͦΕҎ֎ݩͷ%//ͱҰॹ
%FDPVQMFE%//T 7BMVFDIBOOFM w -1ͷ݁ՌΛద༻͢Δଆ w "DUJWBUJPODIBOOFMͷରԠʹ͓͚ Δ݁ՌΛҎͬͯ׆ੑԽؔΛઢܗؔ Ͱஔ͖͑ͨͷ w 3F-6ͳΒPS
w Ұൠత 18-Ͱͳ͍׆ੑԽؔͷ ߹ ʹ+BDPCJBOͰ͍͍
%FDPVQMFE%//Tͷಛ w ྆$IBOOFMͷ֤ʹݩͷ%//ͷରԠͱಉ͡ॏΈ͕ೖ͍ͬͯΔͳΒɼ֤ೖ ྗʹ͓͍ͯݩͷ%//ͱ%%//֎ԆతʹՁͳؔ w ҙͷೖྗΛݻఆ͢Δͱ7BMVF$IBOOFMத͋ΔͷॏΈͷมߋʹରͯ͠ ઢܗ w ݩͷ%//͕18-ͷ߹ɼ%%//ͷ7BMVF$IBOOFMதͷॏΈͷมߋݩͷ 18-%//ͷઢܗྖҬΛม͑ͯ͠·͏͜ͱ͕ͳ͍
1SPWBCMF1PJOUXJTF3FQBJS w %//͔ΒՁͳ%%//Λߏ w %%//ʹରͯ͠ݩʑ%//ͷॏΈʹରͯ͠Γ͔ͨͬͨΑ͏ ʹ-1Ͱ7BMVFDIBOOFMͷॏΈΛʹ͍ͭͯղ͖ɼߋ৽͢Δ
मਖ਼ͷ֦ு w 1PJOUXJTF3FQBJSͰ༗ݶݸͷͷೖྗʹରͯͩͬͨ͠ w Yͷͱ͖㱡Z㱡 w Yͷͱ͖㱡Z㱡 w ೖྗΛͰͳ͘༗ݶݸͷ ತͳ
QPMZUPQF ͷVOJPO ͳྖҬͰදݱ͞ΕΔແݶ ʹ֦ு͍ͨ͠ w 㱡Y㱡ͷͱ͖㱡Z㱡
मਖ਼ͷ֦ுͷ1PJOUXJTF3FQBJSͷؼண w 1PMZUPQFͷྖҬͱEPNBJOׂͨ͠ઢܗྖҬͷJOUFSTFDUJPOΛऔΔ w ઢܗྖҬΛٻΊͯ͏ͷͰ18-%//ʹର͔ͯ͑͠͠ͳ͍ w *OUFSTFDUJPOͷ֤͚ͩΛݟ੍ͯʹ͢Ε͍͍ w ༗ݶݸͷʹର͢Δ੍ʹؼண͞ΕΔ w
1SPWBCMF1PJOUXJTF3FQBJSͰղ͚Δ
1SPWBCMF1PMZUPQF3FQBJS w Γ͍ͨमਖ਼ ࠶ܝ w 㱡Y㱡ͷͱ͖㱡Z㱡 w JOUFSTFDUJPOͷ֤͚ͩΛݟ֤ͯͰͷ੍ʹ͢Δ w
Y ͷͰ㱡Z㱡 w 1SPWBCMF1PJOUXJTF3FQBJS͢Δ
࣮ݧ݁Ռ 4RVFF[F/FU *NBHF/FU /BUVSBM"EWFSTBSJBM&YBNQMFT w 13͕ຊख๏Ͱ%SBXEPXO͕ϕετͷͷͷɼ'5͕'JOFUVOJOHɼ .'5 NPEJ fi FE
ͷΈͷ w && ffi DBDZ %%SBXEPXO 55JNF w 13 '5Ͱ&
·ͱΊ w ܇࿅ࡁΈ%//ϞσϧΛύονमਖ਼Ͱ͖Δख๏ΛఏҊ w ܇࿅࣌σʔληοτ͕ෆཁ w ESBXEPXOͳͲͷѱӨڹ͕গͳ͍ w -1Ͱޮྑ͘ղ͚Δ