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
技術力で世界と戦う機械学習コンペティション「Kaggle」の魅力 / Attractiven...
Search
Shotaro Ishihara
September 07, 2019
Technology
3k
3
Share
技術力で世界と戦う 機械学習コンペティション「Kaggle」の魅力 / Attractiveness of Kaggle
AIchi勉強会での登壇資料
https://connpass.com/event/134720/
Shotaro Ishihara
September 07, 2019
More Decks by Shotaro Ishihara
See All by Shotaro Ishihara
[ACL 2026 Demo] Fast-MIA: Efficient and Scalable Membership Inference for LLMs
upura
0
19
Fast-MIA: Efficient and Scalable Membership Inference for LLMs
upura
0
9
JAPAN AI CUP Prediction Tutorial
upura
2
1.1k
情報技術の社会実装に向けた応用と課題:ニュースメディアの事例から / appmech-jsce 2025
upura
0
370
日本語新聞記事を用いた大規模言語モデルの暗記定量化 / LLMC2025
upura
0
640
Quantifying Memorization in Continual Pre-training with Japanese General or Industry-Specific Corpora
upura
1
110
JOAI2025講評 / joai2025-review
upura
0
1.6k
AI エージェントを活用した研究再現性の自動定量評価 / scisci2025
upura
1
240
JSAI2025 企画セッション「人工知能とコンペティション」/ jsai2025-competition
upura
0
130
Other Decks in Technology
See All in Technology
AI時代に、 データアナリストがデータエンジニアに異動して
jackojacko_
0
240
Every Conversation Counts
kawaguti
PRO
0
130
AIエージェントの支払い基盤 AgentCore Payments概要
kmiya84377
1
140
Fabric MCPの紹介と使い分け
ryomaru0825
1
150
"うちにはまだ早い"は本当? ─ 小さく始めるPlatform Engineering入門
harukasakihara
1
230
Oracle Cloud Infrastructure presents managed, serverless MCP Servers for Oracle AI Database
thatjeffsmith
0
130
Shiny New Tools Won't Fix Your Problem
trishagee
1
110
フロントエンドの相手が変わった - AIが加わったWebの新しいインターフェース設計
azukiazusa1
33
11k
AIが盛んな時代に 技術記事を書き始めて起きた私の中での小さな変化
peintangos
0
360
20260428_Product Management Summit_Loglass_JoeHirose
loglassjoe
4
7.3k
多角的な視点から見たAGI
terisuke
0
120
ハーネスエンジニアリング入門
hatyibei
0
110
Featured
See All Featured
Beyond borders and beyond the search box: How to win the global "messy middle" with AI-driven SEO
davidcarrasco
3
120
Prompt Engineering for Job Search
mfonobong
0
290
Large-scale JavaScript Application Architecture
addyosmani
515
110k
The Cult of Friendly URLs
andyhume
79
6.9k
Balancing Empowerment & Direction
lara
6
1.1k
Max Prin - Stacking Signals: How International SEO Comes Together (And Falls Apart)
techseoconnect
PRO
0
160
HU Berlin: Industrial-Strength Natural Language Processing with spaCy and Prodigy
inesmontani
PRO
0
370
The innovator’s Mindset - Leading Through an Era of Exponential Change - McGill University 2025
jdejongh
PRO
1
160
Breaking role norms: Why Content Design is so much more than writing copy - Taylor Woolridge
uxyall
0
270
Practical Tips for Bootstrapping Information Extraction Pipelines
honnibal
25
1.9k
The Power of CSS Pseudo Elements
geoffreycrofte
82
6.2k
AI in Enterprises - Java and Open Source to the Rescue
ivargrimstad
0
1.3k
Transcript
ٕज़ྗͰੈքͱઓ͏ ػցֶशίϯϖςΟγϣϯ ʮ,BHHMFʯͷັྗ V !VQVSB ݄ "JDIJษڧձ
ͬͨ͜ͱ͋Δ 2
ฉ͍ͨ͜ͱ͋Δ 3
ຊͷ • ਓೳ "* ͷதͷʮػցֶशʯͱʁ • ػցֶशͷதͷʮڭࢣ͋Γֶशʯͱʁ • ػցֶशίϯϖςΟγϣϯʮ,BHHMFʯͱʁ
ࣗݾհʛV • ͷࣄۀձࣾͰσʔλαΠΤϯςΟετ • ,BHHMF.BTUFS • ࠷ߴϥϯΫҐɺҐೖܦݧ͋Γ • Ѫݝग़
• ౦ւߴߍˠ౦ژେֶˠݱ৬ • ࡢ൩౦ژͰొஃˠࠓேʹ໊ݹʹҠಈ
ࣗݾհʛV
ຊͷ • ਓೳ "* ͷதͷʮػցֶशʯͱʁ • ػցֶशͷதͷʮڭࢣ͋Γֶशʯͱʁ • ػցֶशίϯϖςΟγϣϯʮ,BHHMFʯͱʁ
ਓೳ "* • େ͖͘ྲྀʢ͜ͷྨࣗମʹ͕ٞ͋Δʣ ਓؒͷೳͦͷͷΛͭػցΛ࡞Ζ͏ͱ͢Δཱ ਓ͕ؒೳΛͬͯ͢Δ͜ͱΛػցʹͤ͞Α͏ͱ͢ Δཱ
• ࣮ࡍͷݚڀ΄ͱΜͲޙऀͷཱʢͳͷʹ͔ͬͯ Β͔ͣલऀͷҙຯͰ૽ཱ͗ͯΔք۾͕͋Δʣ ਓೳֶձਓೳͬͯԿʁ IUUQTXXXBJHBLLBJPSKQXIBUTBJ"*XIBUTIUNM
ػցֶश • ίϯϐϡʔλʹώτͷΑ͏ͳֶशೳྗΛ֫ಘ ͤ͞ΔͨΊͷٕज़ͷ૯শ • ਓೳ ∋ ػցֶश •
࠷ۙͷਓೳ ػցֶश ਿࢁক ʰΠϥετͰֶͿػցֶशʱ ߨஊࣾ
ຊͷ • ਓೳ "* ͷதͷʮػցֶशʯͱʁ • ػցֶशͷதͷʮڭࢣ͋Γֶशʯͱʁ • ػցֶशίϯϖςΟγϣϯʮ,BHHMFʯͱʁ
ػցֶशͷྨ • ֶशํ๏ʹґͬͯେ͖͘ྨ • ڭࢣ͋Γֶश • ڭࢣͳֶ͠श • ڧԽֶश
ਿࢁক ʰΠϥετͰֶͿػցֶशʱ ߨஊࣾ
ڭࢣ͋Γֶश • ίϯϐϡʔλʹͱ͑ͷରΛ͍͔ͭ͘ ڭ͑Δ͜ͱʹΑΓɼڭΘ͍ͬͯͳ͍ʹ ਖ਼͘͠ճͰ͖Δ൚ԽೳྗΛίϯϐϡʔ λʹ֫ಘͤ͞Δ ਿࢁক ʰΠϥετͰֶͿػցֶशʱ ߨஊࣾ
ڭࢣ͋Γֶशͷ۩ମྫ • खॻ͖จࣈೝࣝɼԻೝࣝɼը૾ೝࣝɼ໎ ϝʔϧͷࣗಈྨͳͲ
ֶशͷΞϧΰϦζϜ • ͍Ζ͍ΖͳΞϧΰϦζϜ͕։ൃ͞Ε͍ͯΔ • ͦͷҰ͕ͭɺσΟʔϓϥʔχϯά
࡞ͬͯΈͨʢσϞʣ • ʮ໊ݹ౦ژʯͷը૾ྨΞϓϦ • ωοτͰऩूͨ͠ը૾ͦΕͧΕຕ • 7((ʢσΟʔϓϥʔχϯάʣΛར༻ IUUQTXXXNEQJDPNIUN
࡞ͬͯΈͨʢσϞʣ IUUQTVQVSBIBUFOBCMPHDPN FOUSZ
ڭࢣͳֶ͠श • ͑ͷ͔͍ͬͯͳ͍σʔλͷू߹͔Βɺ ༗ӹͳࣝΛ֫ಘ͠Α͏ͱ͢Δ • ੈͷதʹਖ਼ղϥϕϧͷͳ͍σʔλͷํ͕ ѹతʹଟ͍ • ۩ମྫɿΫϥελϦϯάҟৗݕ
ਿࢁক ʰΠϥετͰֶͿػցֶशʱ ߨஊࣾ
ڧԽֶश • ڭࢣ͋Γֶशͱಉ͘͡൚ԽੑೳΛίϯϐϡʔ λʹ֫ಘͤ͞Δ͜ͱ͕ඪ • ͑Λڭ͑ΔΘΓʹɼ༧ଌͨ͑͠ͷ ྑ͞ΛධՁ • ධՁ͕࠷ߴ·ΔΑ͏ʹֶश͢Δ
ਿࢁক ʰΠϥετͰֶͿػցֶशʱ ߨஊࣾ
ڧԽֶशͷ۩ମྫ • ϩϘοτͷࣗಈ੍ޚɼίϯϐϡʔλήʔϜͳͲ • ໌֬ʹਖ਼ղσʔλ͕༩͑ͮΒ͍߹ʹ༻͍ΒΕΔ • ྫɿғޟͰʮͱ͋Δ൫໘ͰԿΛଧͭͷ͕ਖ਼ղʯ͔ ݴ͍Δͷ͍͠ ˠͱ͋ΔखΛଧͬͨΒউͬͨෛ͚ͨͱ͍͏ධՁ
Λͱʹֶश͍ͤͯ͘͜͞ͱͰ͖ͦ͏
ຊͷ • ਓೳ "* ͷதͷʮػցֶशʯͱʁ • ػցֶशͷதͷʮڭࢣ͋Γֶशʯͱʁ • ػցֶशίϯϖςΟγϣϯʮ,BHHMFʯͱʁ
,BHHMFͱ • ओʹڭࢣ͋Γֶश͕ରͷػցֶशϞσϧͷ ੑೳΛڝ͏ίϯϖςΟγϣϯ • (PPHMFࡿԼͷ,BHHMF͕ओ࠵ • ੈքதͷσʔλαΠΤϯςΟετ͕ू͏
(PPHMFτϨϯυ • ۙɺຊͰਓؾ͕ߴ·͍ͬͯΔ
,BHHMFͱ IUUQTXXXTMJEFTIBSFOFU)BSBEB,FJEFWTVNJTVNNFS
,BHHMFϥϯΩϯά IUUQTXXXLBHHMFDPNTJTIJIBSB • ֤ίϯϖͷ্Ґ ˠۚɿ ۜɿ ಔɿ • ۚݸۜݸҎ্
ˠ,BHHMF .BTUFS
,FSOFM%JTDVTTJPO • ,FSOFMʢ࣮ͷެ։ɺࢲ͜Μͳ࣮ͩʂʣ • %JTDVTTJPOʢ͜͏ΔͱείΞ্͕͕Δʂʣ • ੈքதͷࢀՃऀ͕࢛࣌த͍ٞͯ͠Δ • ˞ίϯϖͷॱҐ͚ͩͰͳ͘ɺ,FSOFM
%JTDVTTJPOͰͦΕͧΕϙΠϯτ͕͘
1FU'JOEFSίϯϖ
1FU'JOEFSίϯϖ IUUQTXXXLBHHMFDPNDQFUGJOEFSBEPQUJPOQSFEJDUJPO • 1FU'JOEFSNZ"EPQUJPO1SFEJDUJPO • ݄ʙ݄ʹ։࠵ • ʮػցֶशΛ༻͍ͯɺϚϨʔγΞͷϖοτ γϣοϓͰͷݘɾೣ͕Ҿ͖औΒΕΔૣ͞Λ
༧ଌ͢Δʯͱ͍͏͓
ར༻Ͱ͖Δσʔλ • ϖοτͷը૾ܗࣜσʔλ • આ໌จϖοτͷ໊લͳͲͷςΩετܗࣜσʔλ • ɾମॏɾଐੑͳͲͷςʔϒϧܗࣜσʔλ • ඞཁʹԠͯ͡֎෦σʔλར༻Մೳ
ྨ • ݘɾೣ͕Ҿ͖औΒΕΔ͞ • ϦετΞοϓͷʢʣ • ϦετΞοϓͷʙ
• ϦετΞοϓͷʙ • ϦετΞοϓͷʙ • Ҿ͖औΒΕͳ͍
.PEFM1JQFMJOF IUUQTXXXLBHHMFDPNDQFUGJOEFSBEPQUJPOQSFEJDUJPOEJTDVTTJPO
,BHHMFཱ͕ͭ ͱʹ͔ࣗ͘ͰखΛಈ͔ͤΔΑ͏ʹͳΔ ࣗࣗͷ٬؍తࢦඪͷҰͭʹ ࣾ֎ͷਓ͕૿͑Δ
ࣗͰखΛಈ͔͢ • είΞΛ্͛Δʹʮ%PFWFSZUIJOHʯ • ػցֶशۀͰඞͣ͠Θͳ͍͕ ಛྔΛ࡞Δ෦ʢσʔλͷՄࢹԽɾܗɾूܭ ͳͲʣɺ௨ৗͷۀʹਂؔ͘ΘΔ • ,BHHMFͷίʔυΛίϐϖ͢ΔͷͰࣄ͕͘ͳΔ
• σʔλʹର͢Δצॴʹͭ͘ʢ"*࠷ڧͰͳ͍ʣ
٬؍తࢦඪ • ࣾ֎ͷਓ͔Βʮ͜͏͍͏ͷ͕͖ͳͭʯ ͱೝͯ͠Β͑Δʢָͦ͠͏ͳ͕͋Δͱ ༠ͬͯΒ͑Δ͜ͱʣ • ࣾ֎ͷਓͱൺֱͯࣗ͘͠ͷ࣮ྗ͕͔Δ ͷͰɺ͞ΒͳΔษֶͷϞνϕʔγϣϯʹ
ࣾ֎ͷਓ • LBHHMFSKBTMBDLʢਓҎ্ʣ ˠσʔλੳʹؔ͢Δݟͷڞ༗ • ,BHHMF5PLZP.FFUVQ ˠ Ͱొஃ
• ͦͷଞɺTMBDLUXJUUFSͳͲͰࠃࡍަྲྀ
8JOOFS`T$BMM
ຊͰ,BHHMFºاۀ • ࣾ,BHHMF੍ʢ%F/"ʣ ˠ࠾༻ɾϒϥϯυઓུͷҰͰ͋Δ • ,BHHMFͰͷίϯϖ։࠵ʢϦΫϧʔτɺϝϧΧϦʣ ˠϒϥϯυઓུɺ༏উϞσϧͷ׆༻ • ,BHHMFҎ֎ͷίϯϖʢ4JHOBUFɺΦϯαΠτʣ
ˠ࠾༻ɾϒϥϯυઓུɺ༏উϞσϧͷ׆༻
%F/"ͷ,BHHMF੍ IUUQTEFOBBJLBHHMF
ຊاۀͷ։࠵࣮ ϦΫϧʔτ • Ϩετϥϯͷདྷऀ༧ଌ IUUQTXXXLBHHMFDPNDSFDSVJUSFTUBVSBOUWJTJUPSGPSFDBTUJOH ϝϧΧϦ • ग़Ձ֨ͷਪఆ IUUQTXXXLBHHMFDPNDNFSDBSJQSJDFTVHHFTUJPODIBMMFOHF
ϝϧΧϦͷࣄྫ IUUQTUFDINFSDBSJDPNFOUSZ
։࠵අ༻ɿສԁʁ IUUQTXXXLBHHMFDPNTUBUJDTMJEFTNFFULBHHMFQEG
ϗετ͚ϦϯΫू • IUUQTXXXLBHHMFDPNTUBUJDTMJEFT NFFULBHHMFQEG • IUUQTXXXLBHHMFDPNIPTUJOH JORVJSZ • IUUQTXXXLBHHMFDPNIPTUCVTJOFTT
4JHOBUF • ຊ൛ͷ,BHHMFʁ IUUQTTJHOBUFKQ
ΦϯαΠτ • ΠϕϯτܗࣜͰձʹࢀՃऀΛूΊΔܗࣜ • ύɾϦʔάɺ"CFNB57ɺ%F/" ͳͲଟ IUUQTEBUBTIJQKQQMN
·ͱΊ • ਓೳ "* ͷதͷʮػցֶशʯͱʁ • ػցֶशͷதͷʮڭࢣ͋Γֶशʯͱʁ • ػցֶशίϯϖςΟγϣϯʮ,BHHMFʯͱʁ
,BHHMFͬͯΈ͍ͨਓ