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
[Journal club] Scalable Diffusion Models with T...
Search
Semantic Machine Intelligence Lab., Keio Univ.
PRO
July 22, 2024
Technology
0
86
[Journal club] Scalable Diffusion Models with Transformers
Semantic Machine Intelligence Lab., Keio Univ.
PRO
July 22, 2024
Tweet
Share
More Decks by Semantic Machine Intelligence Lab., Keio Univ.
See All by Semantic Machine Intelligence Lab., Keio Univ.
[Journal club] Model Alignment as Prospect Theoretic Optimization
keio_smilab
PRO
0
55
[Journal club] DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models
keio_smilab
PRO
0
11
[Journal club] LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders
keio_smilab
PRO
0
49
Will multimodal language processing change the world?
keio_smilab
PRO
3
480
[Journal club] MOKA: Open-Vocabulary Robotic Manipulation through Mark-Based Visual Prompting
keio_smilab
PRO
0
110
[Journal club] Seeing the Unseen: Visual Common Sense for Semantic Placement
keio_smilab
PRO
0
100
[Journal club] Language-Embedded Gaussian Splats (LEGS): Incrementally Building Room-Scale Representations with a Mobile Robot
keio_smilab
PRO
0
110
[Journal club] RAM: Retrieval-Based Affordance Transfer for Generalizable Zero-Shot Robotic Manipulation
keio_smilab
PRO
1
140
[Journal club] Simplified State Space Layers for Sequence Modeling
keio_smilab
PRO
0
130
Other Decks in Technology
See All in Technology
データエンジニアリング領域におけるDuckDBのユースケース
chanyou0311
6
1.9k
JEDAI Meetup! Databricks AI/BI概要
databricksjapan
0
300
AIエージェント元年
shukob
0
140
役員・マネージャー・著者・エンジニアそれぞれの立場から見たAWS認定資格
nrinetcom
PRO
1
1.6k
遷移の高速化 ヤフートップの試行錯誤
narirou
5
840
OpenID Connect for Identity Assurance の概要と翻訳版のご紹介 / 20250219-BizDay17-OIDC4IDA-Intro
oidfj
0
460
システム・ML活用を広げるdbtのデータモデリング / Expanding System & ML Use with dbt Modeling
i125
1
310
短縮URLをお手軽に導入しよう
nakasho
0
130
Windows の新しい管理者保護モード
murachiakira
0
190
NFV基盤のOpenStack更新 ~9世代バージョンアップへの挑戦~
vtj
0
320
大規模アジャイルフレームワークから学ぶエンジニアマネジメントの本質
staka121
PRO
2
150
Helm , Kustomize に代わる !? 次世代 k8s パッケージマネージャー Glasskube 入門 / glasskube-entry
parupappa2929
0
290
Featured
See All Featured
Designing for humans not robots
tammielis
250
25k
A Philosophy of Restraint
colly
203
16k
[RailsConf 2023] Rails as a piece of cake
palkan
53
5.3k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
33
2.8k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
28
9.3k
Building Better People: How to give real-time feedback that sticks.
wjessup
367
19k
Building Applications with DynamoDB
mza
93
6.2k
Testing 201, or: Great Expectations
jmmastey
42
7.2k
4 Signs Your Business is Dying
shpigford
182
22k
Why Our Code Smells
bkeepers
PRO
336
57k
Stop Working from a Prison Cell
hatefulcrawdad
267
20k
No one is an island. Learnings from fostering a developers community.
thoeni
21
3.1k
Transcript
4DBMBCMF%JGGVTJPO.PEFMTXJUI 5SBOTGPSNFST ܚጯٛक़େֶ ਿӜ໌ݚڀࣨ#ീౡେ 8JMMJBN1FFCMFT 4BJOJOH9JF 6$#FSLFMFZ /FX:PSL6OJWFSTJUZ *$$7 8JMMJBN1FFCMFT
4BJOJOH9JFl4DBMBCMF%JGGVTJPO.PEFMTXJUI5SBOTGPSNFSTzJO*$$7 QQ
എܠɿ֦ࢄϞσϧʹΑΔಈը૾ੜ FH 4PSB ͷൃల IUUQTXXXZPVUVCFDPNXBUDI W),Z%"1/@
ؔ࿈ݚڀɿ֦ࢄϞσϧͷόοΫϘʔϯͱͯ͠6/FU͕ଟ༻ • 6/FUͷ.VMUJTDBMFTLJQDPOOFDUJPOTˠ ෆཁͳܭࢉࢿݯͷ༻ 手法 概要 DALL-E 2 [Ramesh+,
22] CLIPを用いてテキストと画像のAlignmentを行う Stable Diffusion [Rombach+, CVPR22] 潜在拡散モデル 6/FU<3POOFCFSHFS .*$$"*> 4UBCMF%JGGVTJPO<3PNCBDI $713>
ఏҊख๏ɿ%JGGVTJPO5SBOTGPSNFS %J5 • જࡏ֦ࢄϞσϧ -%. <3PNCBDI $713>Λϕʔεʹߏங • 7JTJPO5SBOTGPSNFS 7J5
<%PTPWJUTLJZ *$-3>ػߏΛಋೖ • $POEJUJPOJOHʹΑΔ݅ใͷೖྗ
ఏҊख๏ ɿજࡏ֦ࢄϞσϧͱֶͯ͠शͤ͞Δ • ߴ࣍ݩͷըૉۭؒͰ֦ࢄϞσϧΛֶश ͤ͞Δ͜ͱܭࢉྔతʹࠔ • -%.ͱֶͯ͠शͤ͞Δ͜ͱͰ ܭࢉྔΛݮ
• ըૉۭؒͷ֦ࢄϞσϧͰ͋Δ "%.<%IBSJXBM /FVS*14> ͷͷͷ(GMPQTͰֶशՄೳ
ఏҊख๏ ɿೖྗϊΠζΛQBUDIʹղ • "VUPFODPEFS͔ΒಘΒΕͨ /PJTFE-BUFOU YY Λ 7J5ͱಉ༷ʹE࣍ݩͷτʔΫϯ5ʹม •
1BUDIαΠζQΛʹ͢Δͱ5ഒ ʹͳΓUSBOTGPSNFSͷ (GMPQTগͳ͘ͱഒҎ্
ఏҊख๏ ɿ͖݅ೖྗ $POEJUJPOJOH ͷॲཧ • ͖֦݅ࢄϞσϧͰϊΠζΛؚΉը૾ͱͱʹՃใ͕Ճ͑ΒΕΔ FH UJNFTUFQɼΫϥεϥϕϧɼࣗવݴޠ FUD
• ຊݚڀͰ͜ΕΒͷ͖݅ೖྗΛॲཧ͢ΔͨΊʹҎԼͷͭͷҟͳΔઃܭΛఏҊ • *ODPOUFYUDPOEJUJPOJOH • $SPTT"UUFOUJPOCMPDL • "EBQUJWFMBZFSOPSN BEB-/ CMPDL • BEB-/;FSPCMPDL
ఏҊख๏ ɿBEB-/;FSPCMPDL • 7J5ͷTFMGBUUFOUJPOCMPDLʹରͯ͠"EB-/ػߏΛಋೖ • "EB-/ͷεέʔϧ ͓Αͼ ࠩଓͷલͷεέʔϧ
Λύϥϝʔλͱͯ͠Ճ ˠ݅ใΛը૾ʹΑΓڧ͘ө • "EB-/;FSPCMPDLͰͦΕΒΛθϩʹॳظԽ ˠֶशͷॳظஈ֊߃ؔʹ͍ۙಇ͖ ˠ ֶशͷ҆ఆԽ
࣮ݧઃఆ • σʔληοτ • $MBTT$POEJUJPOBM*NBHF/FUY Y<%FOH $713> • ΞʔΩςΫνϟ
• 7J5ͱಉ༷ʹͭͷϞσϧͷେ͖͞ 4 # - 9- Λ༻ҙ • QBUDITJ[FQ • %%1.TBNQMJOHTUFQT • ධՁई • '*% T'*% *4 1SFDJTJPO 3FDBMM • (GMPQT • ֶश • 516WQPE #BUDITJ[F
ఆྔత݁Ռɿ6/FUϕʔεͷख๏Λ্ճͬͨ
ఆੑత݁Ռ • 1BUDITJ[FΛখ͘͞ɼϞσϧΛେ͖͘͢ΔͱΑΓࣗવͳը૾͕ग़ྗ͞ΕΔ ˠ%J5Ͱ(GMPQT͕େ͖͍΄Ͳग़ྗը૾ͷ্࣭͕͕Δ
ࢼ͓ΑͼΤϥʔੳ ఆੑత݁Ռ ɿࣦഊྫ • ಛఆͷMBCFMʹରͯ͠ෆࣗવͳը૾͕ੜ͞ΕΔ • ྫɿJOQVUMBCFM UPZQPPEMF %%1.TBNQMJOHTUFQ
• ϥϕϧʹΑͬͯTUFQͰੜը૾͕ෆ҆ఆˠ ਪ࣌ͷTUFQΛಈతʹมߋ
ॴײ • 4USFOHUI • ֦ࢄϞσϧʹUSBOTGPSNFSΛಋೖ • ܭࢉࢿݯͱग़ྗը૾ͷ࣭ʹ͍ͭͯͷߟ • 8FBLOFTT
• ͕ࣜগͳ͔ͬͨ • Τϥʔੳ͕ͳ͍
·ͱΊ • എܠ • ֦ࢄϞσϧʹΑΔಈը૾ੜ FH 4PSB ͷൃల • ֦ࢄϞσϧʹ͓͚ΔUSBOTGPSNFSͷར༻͕গͳ͍
• ఏҊख๏ • USBOTGPSNFSϕʔεͷ֦ࢄϞσϧͰ͋Δ%JGGVTJPO5SBOTGPSNFS %J5 ΛఏҊ • ݁Ռ • %J5εέʔϥϏϦςΟ͕ߴ͘ɼ(GMPQT͕େ͖͍΄Ͳ'*%͕Լ ˠ ܭࢉࢿݯͱग़ྗը૾ͷ࣭ʹڧ͍૬ؔؔ • %J59-Ϟσϧ͕ɼ$MBTT$POEJUJPOBM*NBHF/FUʹ͓͍ͯ ैདྷͷ6/FUϕʔεͷ֦ࢄϞσϧΛ্ճͬͨ
"QQFOEJYɿ%FOPJTJOH%JGGVTJPO1SPCBCJMJTUJD.PEFM %%1. ֶश
"QQFOEJYɿ$MBTTJGJFSGSFFHVJEBODF • ͖֦݅ࢄϞσϧͰΫϥεϥϕϧΛϥϯμϜʹυϩοϓ ˠ αϯϓϦϯάͷਫ਼Λ্ • #BZFTͷఆཧΑΓ • ֦ࢄϞσϧͷग़ྗΛείΞͱͯ͠ղऍ͢Δͱਪఆ͢ΔϊΠζҎԼͷΑ͏ʹͳΔ
TɿΨΠμϯεεέʔϧ
"QQFOEJYɿ*ODPOUFYUDPOEJUJPOJOH • $POEJUJPOJOHͰ݅ͱͯ͠ೖྗ͞ΕͨτʔΫϯΛ ը૾τʔΫϯͷઌ಄ʹՃ • ͜ΕΒͷτʔΫϯը૾τʔΫϯͱಉ༷ʹѻΘΕɺ 7J5ʣʹ͓͚ΔDMTτʔΫϯͱࣅׂͨΛ࣋ͭ
"QQFOEJYɿ$SPTT"UUFOUJPOCMPDL • 4FMG"UUFOUJPOϒϩοΫͷޙʹ$SPTT"UUFOUJPOΛ Ճͨ͠ઃܭ • <7BTXBOJ /*14>-%.ͱྨࣅͨ͠ΞʔΩςΫνϟ
"QQFOEJYɿ%J5CMPDLEFTJHO • %J59-Ϟσϧʹ͓͍ͯBEB-/;FSPΛ༻͍ͨ ߹͕࠷গͳ͍ܭࢉࢿݯͰ࠷ྑ͍ '*%,είΞΛୡ
"QQFOEJYɿ7JTJPO5SBOTGPSNFS <%PTPWJUTLJZ *$-3>
"QQFOEJYɿ*ODFQUJPO4DPSF *4 • *NBHF/FUͰࣄલֶशࡁΈͷ*ODFQUJPOOFUXPSLΛ༻͍ͨධՁࢦඪ • *ODFQUJPOOFUXPSL͕ࣝผ͘͢͠ɼࣝผ͞ΕΔϥϕϧͷଟ༷ੑ͕͋Δ΄Ͳ େ͖͘ͳΔࢦඪ <4[FHFEZ $713>
"QQFOEJYɿ'SFDIFU*ODFQUJPO%JTUBODF '*% • *NBHF/FUͰࣄલֶशࡁΈͷ*ODFQUJPOOFUXPSLΛ༻͍ͨධՁࢦඪ • ੜ͞Εͨը૾ͷಛ͕(5ը૾ͷಛͱͲͷఔ ࣅ͍ͯΔ͔ΛධՁ͢Δࢦඪ • '*%͕খ͍͞΄Ͳੜ͞Εͨը૾ͷ࣭͕(5ը૾ʹ͍ۙͱߟ͑ΒΕΔ
"QQFOEJYɿ1SFDJTJPO3FDBMM • *NBHF/FUͰࣄલֶशࡁΈͷ7((<4JNPOZBO *$-3>Λ༻͍ͯ ಛϕΫτϧू߹ΛಘΔ
"QQFOEJYɿ(GMPQT • 'MPQTɿුಈখԋࢉͷճ • (GMPQT 'MPQT • ը૾ੜλεΫͰΞʔΩςΫνϟͷෳࡶ͞ΛධՁ͢ΔࡍύϥϝʔλΛ༻͍Δͷ ͕Ұൠత
• ੑೳʹେ͖͘Өڹ͢Δը૾ղ૾ΛҰߟྀ͍ͯ͠ͳ͍ • Ϟσϧͷෳࡶ͞Λද͢ࢦඪͱͯ͠ෆेͳ߹͕͋Δ
"QQFOEJYɿఆྔత݁Ռ
"QQFOEJY(GMPQTͱ'*%ͷ૬ؔ • ΑΓଟ͘ͷ(GMPQTΛͭϞσϧ'*%͕͘ͳΔ
"QQFOEJYɿϞσϧαΠζͱύοναΠζͷݕ౼