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
Deep Learning Book 10その2 / deep learning book 1...
Search
himkt
January 29, 2018
Research
2
200
Deep Learning Book 10その2 / deep learning book 10 vol2
himkt
January 29, 2018
Tweet
Share
More Decks by himkt
See All by himkt
Linformer: paper reading
himkt
0
530
RoBERTa: paper reading
himkt
1
350
NLP SoTA 勉強会 / ner_2019
himkt
2
1.4k
自然言語処理 @ クックパッド / nlp at cookpad
himkt
1
520
Interpretable Machine Learning 6.3 - Prototypes and Criticisms
himkt
2
170
ニューラル固有表現抽出 / Neural Named Entity Recognition
himkt
3
750
ニューラル固有表現抽出器を実装してみる / PyNER
himkt
6
2.1k
Spacyでお手軽NLP / NLP with spacy
himkt
0
1k
ふわふわ系列ラベリング / ner 2018
himkt
5
850
Other Decks in Research
See All in Research
在庫管理のための機械学習と最適化の融合
mickey_kubo
3
1.1k
SSII2025 [SS1] レンズレスカメラ
ssii
PRO
2
1.1k
[論文紹介] Intuitive Fine-Tuning
ryou0634
0
110
A scalable, annual aboveground biomass product for monitoring carbon impacts of ecosystem restoration projects
satai
3
220
Creation and environmental applications of 15-year daily inundation and vegetation maps for Siberia by integrating satellite and meteorological datasets
satai
3
260
Towards a More Efficient Reasoning LLM: AIMO2 Solution Summary and Introduction to Fast-Math Models
analokmaus
2
780
とあるSREの博士「過程」 / A Certain SRE’s Ph.D. Journey
yuukit
10
4.2k
AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data
satai
1
190
MetaEarth: A Generative Foundation Model for Global-Scale Remote Sensing Image Generation
satai
4
180
Streamlit 総合解説 ~ PythonistaのためのWebアプリ開発 ~
mickey_kubo
2
1.4k
SNLP2025:Can Language Models Reason about Individualistic Human Values and Preferences?
yukizenimoto
0
120
Adaptive Experimental Design for Efficient Average Treatment Effect Estimation and Treatment Choice
masakat0
0
110
Featured
See All Featured
The MySQL Ecosystem @ GitHub 2015
samlambert
251
13k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
358
30k
Site-Speed That Sticks
csswizardry
10
810
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
PRO
187
54k
RailsConf 2023
tenderlove
30
1.2k
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
8
910
Faster Mobile Websites
deanohume
309
31k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
jQuery: Nuts, Bolts and Bling
dougneiner
64
7.9k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
139
34k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
46
7.6k
Being A Developer After 40
akosma
90
590k
Transcript
&DIP4UBUF/FUXPSLT&YQMJDJU.FNPSZ IJNLU!य़ΤϦΞ DEEP LEARNING BOOK 4FRVFODF.PEFMJOH3FDVSSFOUBOE3FDVSTJWF/FUT
&DIP4UBUF/FUXPSLT w 3//ʹֶ͓͍ͯश͕େมͳύϥϝʔλ w ӅΕӅΕ SFDVSSFOUXFJHIUT w ೖྗӅΕ JOQVUXFJHIUT
w &DIP4UBUF/FUXPSL w ӅΕӅΕॏΈΛݻఆ w ֶश͢Δͷʜ w ೖྗӅΕ JOQVUXFJHIUT w ӅΕग़ྗ PVUQVUXFJHIUT
&DIP4UBUF/FUXPSLT IUUQXXXTDIPMBSQFEJBPSHBSUJDMF&DIP@TUBUF@OFUXPSL
,FSOFMNBDIJOFͱͷྨࣅੑ w Χʔωϧ͕ͬͯΔ͜ͱͬͯʁ w ҙͷ͞ͷܥྻΛݻఆͷϕΫτϧࣸ͢ w ݻఆͷϕΫτϧΛ༻͍ͯྨث͕Λղ͘ w ͜ͷܗͷ߹ɼֶशͷج४ͷઃܭ͕༰қͰ͋Δ w
ग़ྗઢܗճؼͷ߹.4&ͰֶशͰ͖Δ w &4/TೖྗΛԿΒ͔ͷϕΫτϧʹࣸ͢ૢ࡞Λ͍ͯ͠Δ w தͷॏΈݻఆ͍ͯ͠Δ ͍͔ʹաڈͷใΛ๛ʹؚΉදݱ͕ಘΒΕΔ ॏΈΛઃఆ͢ΕΑ͍͔ʁ શવҙຯ͕Θ͔Βͣʜ 3//ΛಈతγεςϜͱΈͳ͢ γεςϜ͕҆ఆ͢ΔΑ͏ͳॏΈΛઃఆ͢Δ
-FBLZ6OJUTBOE0UIFS4USBUFHJFTGPS.VMUJQMF5JNF4DBMF w աڈͷใΛ͑ΔͨΊͷ "EEJOH4LJQ$POOFDUJPOTUISPVHI5JNF w ޯͷফࣦͷ͕͘ͳΔ w രൃݩͷ3//ͱಉ͡Ͱൃੜ͢Δ
-FBLZ6OJUTBOEB4QFDUSVNPG%J⒎FSFOU5JNF4DBMFT w աڈͷใΛͲͷఔ͔͢Λ੍ޚ͢Δ 3FNPWJOH$POOFDUJPOT w ͍࣌ࠁͰͷґଘΛ͍࣌ࠁͰͷґଘʹஔ͖͑Δ
-FBLZ6OJUT w աڈͷใΛͲͷ͘Β͍͔͢Λௐ͢Δ w ҠಈฏۉͷΑ͏ͳ;Δ·͍Λ͢Δ w Ћ͕େ͖͍ ʹ͍ۙ աڈͷใΛΑΓอଘ͢Δ w
Ћ͕খ͍͞ ʹ͍ۙ աڈͷใΛ͙͢ʹࣺͯΔ w Ћదʹܾఆ͢ΔϋΠύʔύϥϝʔλ µ(t) ↵µ(t 1) + (1 ↵)v(t)
-POH4IPSU5FSN.FNPSZ w ࣗݾϧʔϓΛಋೖ͢Δ͜ͱͰޯ͕ফ͑ʹ͘͘͢Δ IUUQDPMBIHJUIVCJPQPTUT6OEFSTUBOEJOH-45.T 3// -45.
(BUFE3FDVSSFOU6OJUT w ٙ-45.ෳࡶ͗͢ΔͷͰͳ͍͔ʁ w (36-45.ΑΓߴɾ-45.ͱಉͷੑೳ w ͲͪΒ͕ྑ͍͔λεΫʹΑΔ -45. (36
IUUQTJTBBDDIBOHIBVHJUIVCJP-45.BOE(36'PSNVMB4VNNBSZ
ࣜతʹൺֱ͢ΔʢόΠΞεΛແࢹʣ -45. (36 zt = (xtUz + ht 1Wz)
rt = (xtUr + ht 1Wr) ˜ ht = tanh ⇣ xt + Uh + (rt ht 1)Wh ⌘ ht = (1 zt) ht 1 + zt ˜ ht it = (xtUi + ht 1Wi) ft = (xtUf + ht 1Wf ) ot = (xtUo + ht 1Wg) ˜ Ct = tanh (xtUg + ht 1Wg) Ct = (ft Ct 1 + it ˜ Ct) ht = tanh (Ct) ot (36Ͱೖྗήʔτͱ٫ήʔτ͕౷߹͞Ε͍ͯΔ
0QUJNJ[BUJPOGPS-POH5FSN%FQFOEFODJFT w 3//Λϕʔεͱͨ͠χϡʔϥϧωοτϫʔΫͷඍ w ඇৗʹେ͖ͳΛͱΔPS w ඇৗʹখ͞ͳΛͱΔ w ಛʹɼޯ͕ඇৗʹେ͖ͳͱ͖ʹͲ͏͢Εྑ͍͔ʁ
ޯͷΫϦοϐϯά ޯͷਖ਼نԽ
$MJQQJOH(SBEJFOU w ޯ͕ඇৗʹ େ͖͍cখ͍͞ ͱʁ w ͍͍ͩͨฏΒ͚ͩͲͱ͖Ͳ͖֑͕͋Δ IUUQXXXEFFQMFBSOJOHCPPLPSHMFDUVSF@TMJEFTIUNM
$MJQQJOH(SBEJFOU w ޯ๏ϕʔεͷख๏ʹΑΔͱʜ w ֑ͷपΓͰ͕ਧ͖ඈΜͰ͠·͏ ޯരൃ w ޯ͕େ͖͘ͳΓ͗ͨ͢ΒޯͷϊϧϜͰׂΔ w
ޯΛHͱͯ͠ʜ w WϋΠύʔύϥϝʔλ ࣗવݴޠॲཧͩͱ͕ଟ͍ g ( gv ||g|| (||g|| > v) g (otherwise)
3FHVMBSJ[JOHUP&ODPVSBHF*OGPSNBUJPO'MPX w ਖ਼ଇԽ߲Λಋೖ͢Δ͜ͱͰʮJOGPSNBUJPOqPXʯΛଅਐ w ͜ͷ߲ͷܭࢉ͍͕͠ɼۙࣅ͕ఏҊ͞Ε͍ͯΔ w $MJQQJOHͱΈ߹ΘͤΔ͜ͱͰهԱͰ͖Δڑ͕৳ͼΔ ⌦ =
X t ⇣||(rh(t) L) @h(t) @h(t 1) || ||(rh(t) L)|| 1 ⌘2
&YQMJDJU.FNPSZ w χϡʔϥϧωοτϫʔΫʜ w ҉తͳใͷอ࣋ಘҙ w ໌ࣔతͳใ ࣄ࣮ ͷอ࣋ۤख w
໌ࣔతͳใΛอ࣋͠ɼਪʹ׆༻͢Δߏ ʢϫʔΩϯάϝϞϦͷಋೖʣ w .FNPSZ/FUXPSLT w /FVSBM5VSJOH.BDIJOF
"TDIFNBUJDPGBOFUXPSLXJUIBOFYQMJDJUNFNPSZ IUUQXXXEFFQMFBSOJOHCPPLPSHMFDUVSF@TMJEFTIUNM
"TDIFNBUJDPGBOFUXPSLXJUIBOFYQMJDJUNFNPSZ w ਖ਼֬ͳϝϞϦͷΞυϨεΛग़ྗ͢Δͷ͍͠ w ଟ͘ͷϝϞϦηϧͷॏΈ͖ฏۉΛͱΔ w ॏΈιϑτϚοΫεͳͲͰ࡞Δ ʢͰ͖Δ͚ͩҰՕॴͷϝϞϦΛࢀর͢ΔΑ͏ʹʣ w ϝϞϦηϧεΧϥΑΓϕΫτϧͷํ͕ྑ͍
w ίϯςϯπϕʔεΞυϨογϯά͕ՄೳʹͳΔ w ʮl8FBMMMJWFJOBZFMMPXTVCNBSJOFzΛؚΉՎࢺΛݟ͚ͭΔʯ w ʢϩέʔγϣϯϕʔεΞυϨογϯάͱʁʣ w ʮεϩοτ347ʹ֨ೲ͞Ε͍ͯΔՎࢺΛऔಘ͢Δʯ w ʢΞυϨογϯάΞςϯγϣϯͱಉ͡ܗࣜʣ