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: Cyclic Program Synthesis
Search
Idein
June 08, 2022
Research
0
1.6k
PLDI '21論文読み会: Cyclic Program Synthesis
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
1.9k
PLDI '21論文読み会: AKG: Automatic Kernel Generation for Neural Processing Units using Polyhedral Transformations
ideininc
0
1.6k
PLDI '21論文読み会: Specification Synthesis with Constrainted Horn Clauses
ideininc
0
1.5k
PLDI '21論文読み会: High Performance Correctly Rounded Math Libraries for 32-bit Floating Point Representations
ideininc
0
1.5k
PLDI '21論文読み会: Quantum Abstract Interpretation
ideininc
0
1.5k
PLDI '21論文読み会: Provable Repair of Deep Neural Networks
ideininc
2
1.8k
会社紹介資料/Idein株式会社
ideininc
0
49k
Other Decks in Research
See All in Research
[CV勉強会@関東 CVPR2025] VLM自動運転model S4-Driver
shinkyoto
2
670
Vision and LanguageからのEmbodied AIとAI for Science
yushiku
PRO
1
590
EOGS: Gaussian Splatting for Efficient Satellite Image Photogrammetry
satai
4
820
Learning to (Learn at Test Time): RNNs with Expressive Hidden States
kurita
1
290
AIスパコン「さくらONE」のLLM学習ベンチマークによる性能評価 / SAKURAONE LLM Training Benchmarking
yuukit
2
850
機械学習と数理最適化の融合 (MOAI) による革新
mickey_kubo
1
420
なめらかなシステムと運用維持の終わらぬ未来 / dicomo2025_coherently_fittable_system
monochromegane
0
5.1k
LLM-jp-3 and beyond: Training Large Language Models
odashi
1
620
Minimax and Bayes Optimal Best-arm Identification: Adaptive Experimental Design for Treatment Choice
masakat0
0
190
論文読み会 SNLP2025 Learning Dynamics of LLM Finetuning. In: ICLR 2025
s_mizuki_nlp
0
340
2025/7/5 応用音響研究会招待講演@北海道大学
takuma_okamoto
1
230
Mamba-in-Mamba: Centralized Mamba-Cross-Scan in Tokenized Mamba Model for Hyperspectral Image Classification
satai
3
210
Featured
See All Featured
Designing for humans not robots
tammielis
254
26k
The Cost Of JavaScript in 2023
addyosmani
55
9.3k
Docker and Python
trallard
46
3.7k
Principles of Awesome APIs and How to Build Them.
keavy
127
17k
Visualization
eitanlees
150
16k
A better future with KSS
kneath
239
18k
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
46
2.6k
The Invisible Side of Design
smashingmag
302
51k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
359
30k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
31
9.8k
The Straight Up "How To Draw Better" Workshop
denniskardys
239
140k
Practical Tips for Bootstrapping Information Extraction Pipelines
honnibal
25
1.6k
Transcript
୲ߐޱ $ZDMJD1SPHSBN4ZOUIFTJT 1-%*ಡΈձ!*EFJO
EFEVDUJWFQSPHSBNTZOUIFTJT ༩͑ΒΕ༷ͨΛຬͨ͢ϓϩάϥϜΛࣗಈͰߏ͢ΔλεΫ ਖ਼͕͠͞ূ໌͞Ε͍ͯΔ
ݚڀ֓ཁ ࠶ؼతͳิॿؔΛؚΉIFBQNBOJQVMBUJOHQSPHSBNͷࣗಈੜ w $ZDMJD1SPPGͷٕज़ΛϓϩάϥϜ߹ʹԠ༻
എܠ
ϓϩάϥϜͷ༷ 4FQBSBUJPOMPHJD w ϙΠϯλɺࢀরͳͲΛ͏ϓϩάϥϜͷঢ়ଶมԽΛදݱͰ͖Δཧ w ࠶ؼతͳड़ޠΛͬͯɺ༷ʑͳσʔλߏΛදݱͰ͖Δ
طଘݚڀ4VT-JD 4USVDUVSJOHUIFTZOUIFTJTPGIFBQNBOJQVMBUJOHQSPHSBNT<101-> w 4FQBSBUJPOMPHJDΛ༷ʹ$MJLFͳ࠶ؼؔΛੜ w 4ZOUIFUJD4FQBSBUJPO-PHJD ͱ͍͏ϓϩάϥϜੜ༻ͷਪମܥΛ࡞ SSL ˎϒϥβ্ͰࢼͤΔ
ɹIUUQDPNDPNDTBJMNJUFEVDPNDPN4V4-JL
طଘݚڀ4VT-JD ͷݶք ิॿ͖ؔϓϩάϥϜΛੜͰ͖ͳ͍ ྫҎԼͷੜࣦഊ͢Δ ˎMJTUBQQFOEΛߦ͏࠶ؼతͳ͕ؔඞཁʹͳΔҝ Ұൠతʹɺิॿؔͷ༷ΛࣗಈͰݟ͚ͭΔͷ͍͠ ɹূ໌୳ࡧͷ؍Ͱɺؼೲ๏͕ճΔΑ͏ʹదʹิΛੜ͢Δ͜ͱʹରԠ
ఏҊख๏
ఏҊख๏ w DZDMJDQSPPGΛ༻͍ͯ ͱ͍͏ਪମܥΛߟҊ w 4VT-J,ͷ Λ֦ு w ্Ͱূ໌Λ୳ࡧ͠ɺϓϩάϥϜΛੜ͢Δ w
ิॿ͖ؔϓϩάϥϜͷੜ͕Մೳʹ SSL↺ SSL SSL↺
$ZDMJDQSPPG "O*OUSPEVDUJPOUP$ZDMJD1SPPGTΑΓ IUUQXXXDTVDMBDVLTUB ff +#SPUIFSTUPOTMJEFT 1"3*4@'-P$@@@QBSUQEG ॥Λڐ͢ূ໌ମܥ ূ໌ͷ-FBG෦ͱผͷOPEFͷzCBDLMJOLz͕࡞ΕΔ w ॥ͷํʹ੍Λ՝͢
ɹɹʮ॥ͷதͰূ໌͕lਐΜͰ͍ΔzʯΈ͍ͨͳ੍ ˎࣗ༝ʹ॥ΛڐͤԿͰূ໌Ͱ͖ͯ͠·͏ w ূ໌୳ࡧͱͷ૬ੑ͕ྑ͍෦͕͋Δ w POEFNBOEʹؼೲ๏͕ճͤΔ
SSL↺ w ͷ֦ு w ҎԼͷ߲̐ؔΛಋग़͢Δূ໌ମܥ SSL Γ; 𝒫 ⇝ 𝒬
∣ c ࣄલ݅ ࣄޙ݅ มڥ ͔ Ͱଋറ ∀ ∃ ϓϩάϥϜ w ྫ w ߏจৄࡉ จͷ'JHVSF ∀r, s, x∃y; {r ↦ x * tree(x, s)} ⇝ {r ↦ y * sll(y, s)} ∣ fl BUUFO S
ͷಋग़نଇൈਮ SSL↺ ʹಉ༷ͳͷ͋Δͷ SSL
ͷಋग़نଇൈਮ SSL↺ ಠࣗͷͷ ؔݺͼग़͠पΓ ͜Εؔݺͼग़͠ͷ४උͷͨΊͷSVMFɹ 8SJUFϧʔϧ Λಋ༷͘ʹ͔͑͠ͳ͍ͷͰඞཁ Q ؔݺͼग़͠ ࣗମ͕ੜͷରʹͳ͍ͬͯΔ
f(¯ x)
USFF'MBUUFOͷಋग़ w fl BUUFOͷ࠶ؼݺͼग़͠ʹରԠ w ิॿؔBQQFOEͷ࠶ؼʹରԠ
w BQQFOEҎԼͷΑ͏ʹൃݟ͞ΕΔ ೋͭͷ Λͨ͠ޙ B Λ͞Βʹల։3FBEͨ͠ޙ C C ͔Β B ͷ$BMM͕ద༻Ͱ͖Δ͜ͱʹؾ͘ D ͷ༷ʹQSPDΛૠೖ͠CBDLMJOLΛ࡞Δ flatten() sll(yl, sl )
ؔݺͼग़͠ͷੜ ީิΛ fi Y͞Ε্ͨͰɺҎԼͷඇܾఆੑ͕͋Δ w ೖ w Γύʔτ ͷબ
w TFUVQ෦ ͷੜ σ R c1 ީิͱͳΔCBDLMJOLઌ DPNQBOJPO શͯΛࢼ͢
ؔݺͼग़͠ͷੜ S` Y`GSFTIͳ ଋറ͞Εͨม ∃ ͷTVCTFU QSFIFBQ͕ͬͯྑ͍WFSTJPOͷFNQنଇ Ͱผ్ੜΛղ͘ SSL↺
ྫࠨਤͷ ͷ෦ͷੜ ͷಋग़͔Β FNQنଇͰͬͨ෦͕ TFUVQϓϩάϥϜ ͕ಘΒΕΔ ∃ σ R c
ධՁ
࣮ݧ݁Ռ w ࠶ؼతͳิॿ͖ؔͷϓϩάϥϜ w ࠨਤͷͱҎ֎ w zDPNQMFYUFSNJOBUJPONFUSJDzΛඞཁͱ ͢Δͷ w ͱ
w ૬ޓ࠶ؼతͳϓϩάϥϜ w w ͱ͕૬ޓ࠶ؼʹͳΔͷਓؒͷ ײ֮ͱͣΕΔʣ ༷ʑͳlෳࡶͳzϓϩάϥϜͷੜʹޭ
ϕϯνϚʔΫྫMJTUJOQMBDFTPSU
ϕϯνϚʔΫྫMJTUJOQMBDFTPSU ༨ஊTPSUͷఆ͕ٛΑ͔͘Βͳ͍ ˣͳͲͰμϝͳͷ͔ 5FDIOJDBMQBQFSͷ"QQFOEJOH$ʹҎԼͷهࡌ
ϕϯνϚʔΫྫVOJRVFMJTUJOUFSTFDUJPO ҎԼͷΑ͏ͳTJNQMFͳ༷Ͱੜʹࣦഊ͢Δ ҎԼͷΑ͏ͳFMFN͕ؔੜͰ͖Εྑ͍͕ɺ ख๏ͷݶքΛ͑Δ ͜ͷ࣌ͰɺZEFTUSVDUͯ͠ΔͷͰYΛSʹՃ͑ͯྑ͍͔Δखஈ͕ͳ͍ ҎԼͷΑ͏ʹ୳ࡧ͢Δ͕ɺɺ
ϕϯνϚʔΫྫVOJRVFMJTUJOUFSTFDUJPO ZΛEFTUSVDU͠ͳ͍Α͏QPTUDPOEJUJPOʹڧΊΕޭ͢Δ ͕ɺҎԼͷΑ͏ͳzී௨ͷz࣮ੜ͞Εͳ͍ʢͰ͖ͳ͍ʣ ͜͜Ͱิॿؔͷ༷ͷQSFDPOEJUJPOͷ݅Λ؇ΊΔඞཁ͕͋ΓɺఏҊख๏ͰͰ͖ͳ͍ ΘΓʹɺzී௨ͱఔԕ͍zϓϩάϥϜ͕ੜ͞ΕΔ ஶऀୡͰ͢ΒผͷϓϩάϥϜݕূػΛͬͯਖ਼͠͞Λ͔֬ΊΔఔ
MJNJUBUJPO w ੜͰ͖Δิॿؔʹ੍ݶ͋Γ w ৽ͨͳEBUBߏΛ͏ͷɺBDDVNVMBUPSͳͲՃҾΛ͏Α͏ͳͷੜ Ͱ͖ͳ͍ w ิॿؔͷ༷ΛదʹzҰൠԽzͰ͖ͳ͍ʢJOUFSTFDUͷྫʹ༷͋ͬͨʹʣ w ੜίʔυͷύϑΥʔϚϯεอূ͞Εͳ͍
w MPPQͷαϙʔτ͕ͳ͍ w ඌ࠶ؼΛੜ͢ΔΈΛ࡞Εྑ͍͕ɺదͳBDDVNVMBUPSΛੜ͢Δඞཁ͕ ͋Δͷ͕͍͠
·ͱΊ w ࠶ؼతͳิॿؔΛؚΉIFBQNBOJQVMBUJOHQSPHSBNͷࣗಈੜ w $ZDMJD1SPPGͷٕज़ΛϓϩάϥϜ߹ʹԠ༻ w ϓϩάϥϜ߹πʔϧʢ$ZQSFTTʣΛ࣮ w ૬ޓ࠶ؼؚΉิॿ͖ؔϓϩάϥϜͷੜʹޭ