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
城ヶ崎美嘉で学ぶRNNLM
Search
Kento Nozawa
June 05, 2016
Programming
2
3k
城ヶ崎美嘉で学ぶRNNLM
オタク機械学習勉強会#0 のLT
Kento Nozawa
June 05, 2016
Tweet
Share
More Decks by Kento Nozawa
See All by Kento Nozawa
Analysis on Negative Sample Size in Contrastive Unsupervised Representation Learning
nzw0301
0
160
[IJCAI-ECAI 2022] Evaluation Methods for Representation Learning: A Survey
nzw0301
0
610
[NeurIPS Japan meetup 2021 talk] Understanding Negative Samples in Instance Discriminative Self-supervised Representation Learning
nzw0301
0
190
[IBIS2021] 対照的自己教師付き表現学習おける負例数の解析
nzw0301
0
180
Understanding Negative Samples in Instance Discriminative Self-supervised Representation Learning
nzw0301
0
480
Introduction of PAC-Bayes and its Application for Contrastive Unsupervised Representation Learning
nzw0301
2
810
NLP Tutorial; word representation learning
nzw0301
0
210
Analyzing Centralities of Embedded Nodes
nzw0301
0
170
Paper Reading: Noise-Contrastive Estimation of Unnormalized Statistical Models, with Applications to Natural Image Statistics
nzw0301
2
1.2k
Other Decks in Programming
See All in Programming
「テストは愚直&&網羅的に書くほどよい」という誤解 / Test Smarter, Not Harder
munetoshi
0
200
Goで作る、開発・CI環境
sin392
0
260
Porting a visionOS App to Android XR
akkeylab
0
680
TypeScriptでDXを上げろ! Hono編
yusukebe
3
770
AWS Summit Japan 2024と2025の比較/はじめてのKiro、今あなたは岐路に立つ
satoshi256kbyte
0
120
チームで開発し事業を加速するための"良い"設計の考え方 @ サポーターズCoLab 2025-07-08
agatan
1
470
MDN Web Docs に日本語翻訳でコントリビュートしたくなる
ohmori_yusuke
1
130
テストから始めるAgentic Coding 〜Claude Codeと共に行うTDD〜 / Agentic Coding starts with testing
rkaga
15
5.6k
新メンバーも今日から大活躍!SREが支えるスケールし続ける組織のオンボーディング
honmarkhunt
5
8.7k
テスト駆動Kaggle
isax1015
1
630
ご注文の差分はこちらですか? 〜 AWS CDK のいろいろな差分検出と安全なデプロイ
konokenj
3
580
The Modern View Layer Rails Deserves: A Vision For 2025 And Beyond @ RailsConf 2025, Philadelphia, PA
marcoroth
2
730
Featured
See All Featured
How GitHub (no longer) Works
holman
314
140k
Faster Mobile Websites
deanohume
308
31k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
50
5.5k
10 Git Anti Patterns You Should be Aware of
lemiorhan
PRO
656
60k
The Straight Up "How To Draw Better" Workshop
denniskardys
235
140k
Speed Design
sergeychernyshev
32
1k
Imperfection Machines: The Place of Print at Facebook
scottboms
267
13k
Documentation Writing (for coders)
carmenintech
72
4.9k
Put a Button on it: Removing Barriers to Going Fast.
kastner
60
3.9k
Product Roadmaps are Hard
iamctodd
PRO
54
11k
The Cult of Friendly URLs
andyhume
79
6.5k
Visualization
eitanlees
146
16k
Transcript
ϲ࡚ඒՅ Λը૾ݕࡧ͓ͯͪ͠Լ͍͞
ϲ࡚ඒՅͰֶͿ RNNLM 2016/6/5 ΦλΫػցֶशษڧձ #0 @nzw0301
Ϟνϕʔγϣϯ ϲ࡚ඒՅͷηϦϑੜ
Recurrent Neural Network Language Model • ηϦϑੜ: લ·Ͱͷ୯ޠ͔Β࣍ͷ1୯ޠΛ༧ଌ͠ଓ͚Δ • ྫɿΊΔΊΔʜᣦՅʹϝʔϧૹ৴ͬ˒
• ୯ޠׂ: <BOS> ΊΔΊΔʜᣦՅʹϝʔϧૹ৴ͬ˒&04 • ֶश: Q ΊΔΊΔc#04 ͱ͔ Q ᣦՅc<BOS>, ΊΔΊΔ ʜ
RNNLMͷߏ ޠኮV࣍ݩͷϕΫτϧ softmax ؔ 1ͭલͷதؒͷϕΫτϧ RNNͷ༝ԑ h࣍ݩͷதؒ
p(ΊΔΊΔ|<BOS>) ͷܭࢉྫɿೖྗ w #04ͷPOFPG,දݱΛೖྗ w ࣍ݩͰີͳϕΫτϧʹม <BOS> ΊΔΊΔ 0 B
B B B B @ 0 1 0 . . . 0 1 C C C C C A
p(ΊΔΊΔ|<BOS>) ͷܭࢉྫɿதؒ • ີͳϕΫτϧΛதؒʹ͢ • ଟύʔηϓτϩϯͱಉ͡ <BOS> ΊΔΊΔ
p(ΊΔΊΔ|<BOS>) ͷܭࢉྫɿग़ྗ • ग़ྗʹதؒͷϕΫτϧΛ͢ • ݱࡏͷதؒͷΛอ࣋ <BOS> ΊΔΊΔ
p(ΊΔΊΔ|<BOS>) ͷܭࢉྫɿॏΈߋ৽ • SoftmaxؔͰ֬Λܭࢉ • Backpropagation Ͱ ΊΔΊΔ ͷ͕֬େ͖͘ͳΔΑ͏ʹߋ৽ <BOS>
ΊΔΊΔ
p(ʜc#04 ΊΔΊΔ) ͷܭࢉྫɿೖྗ ૄΊΔΊΔϕΫτϧΛೖྗ͠ɼີͳΊΔΊΔϕΫτϧʹม p(ΊΔΊΔ|<BOS>)Ͱܭࢉͨ͠தؒͷϕΫτϧ ʜ ΊΔΊΔ 0 B B
B B B B B B B B @ 0 . . . 0 1 0 . . . 0 1 C C C C C C C C C C A
p(ʜc#04 ΊΔΊΔ) ͷܭࢉྫɿதؒ ີͳΊΔΊΔϕΫτϧͱલʹܭࢉͨ͠தؒͷϕΫτϧΛதؒ p(ΊΔΊΔ|<BOS>)Ͱܭࢉͨ͠தؒͷϕΫτϧ ʜ ΊΔΊΔ
p(ʜc#04 ΊΔΊΔ) ͷܭࢉྫɿग़ྗ • ग़ྗʹதؒͷϕΫτϧΛͯ͠ɼݱࡏͷதؒͷϕΫτϧΛอ࣋ p(ʜ|<BOS>, ΊΔΊΔ)Ͱܭࢉͨ͠தؒͷϕΫτϧ ʜ ΊΔΊΔ
p(ʜc#04 ΊΔΊΔ) ͷܭࢉྫɿॏΈߋ৽ • SoftmaxؔͰ֬Λܭࢉ • Backpropagation Ͱ ʜ ͷ͕֬େ͖͘ͳΔΑ͏ʹߋ৽
ʜ ΊΔΊΔ
࣮ݧ
࣮ݧ֓ཁ • SCRNΛ༻ • LSTM GRU ΛΘͳ͍ • Keras
Ͱ࣮ • લॲཧ • ܗଶૉղੳͤͣʹจࣈ୯ҐͰֶश • /。|★|?|!|♪/ ͰηϦϑΛׂ • 900ηϦϑ (Վࢺ) Λ༻ • ϞόϚε • σϨες • TOKIMEKIΤεΧϨʔτ
݁Ռ
10epochޙɿϓϩσϡʔαʔͷҰ෦͕ͱΕͯΔ ϓϩσϩσϡʔͯͳͪʙʹෲΞλ γ΄ϡʔαʔΒతͳʔɺͨ͜ͳ
40epochޙɿΪϟϧޠʁ ϓϩσϡʔαʔʹ͍ͪΌΜɺ ݟ͘ͳ͍ʔ͘ͱԿߴͩ͠ʔͬ̇
80epochޙɿݺΕͨؾ͕ͨ͠ ϓϩσϡʔαʔ!
“<BOS> ϓ” ͔Β࠷ਪఆɿϧʔϓ ϓϩσϡʔαʔɺΞλγͷ͜ͱ͔Βɺ ϓϩσϡʔαʔɺΞλγͷ͜ͱ
ϥϯμϜʹηϦϑੜ
ॴײ • ηϦϑΛͲ͜ͰΔ͖͔ • ྫɿ͝Μʹ͢Δ?͓෩࿊ʹ͢Δ?…͜ΕͪΐͬͱϕλͬΆ͍ͳ͊ • ? Ͱ۠Δ͖͔൱͔ • …લޙͲͬͪͰ۠Δ͔൱͔ʁͦΕͱͳ͘͢ʁ
• ήʔϜը໘ͷͨΊ͔1ηϦϑܥྻ͕΄΅Ұఆʢֶͼʣ
ࢀߟจݙͳͲ • http://keras.io/ • DLͷϥΠϒϥϦ • ָ͍͢͝ʹॻ͚Δ • Mikolov at.el.
Recurrent neural network based language model. 2010. • RNNͷը૾͜ͷจͷͷΛ༻ • Mikolov at.el Learning Longer Memory in Recurrent Neural Networks. 2014. • ࠓճ༻ͨ͠Ϟσϧ