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
宝くじ仮説の話を少し / LT Lottery Ticket Hypothesis
Search
SKA’s Web
November 14, 2021
Programming
0
360
宝くじ仮説の話を少し / LT Lottery Ticket Hypothesis
Python機械学習勉強会 in 新潟のLTの資料です。深層学習の「宝くじ仮説」について少しだけ話します。
SKA’s Web
November 14, 2021
Tweet
Share
More Decks by SKA’s Web
See All by SKA’s Web
.NETでOBS Studio操作してみたけど…… / Operating OBS Studio by .NET
skasweb
0
26
学習しやすいネットワークガチャ? 深層学習の「宝くじ仮説」 【VRアカデミア クリスマスリレー】 / VR Lottery Ticket Hypothesis
skasweb
1
480
自然言語処理の基本 / Basic of Natural Language Processing
skasweb
0
290
AIとクラスター分析について自作アイテムを用いて解説してみた 【VRアカデミア3周年カウントダウン】 / Explaining AI and Cluster Analysis Using My Item
skasweb
0
270
Other Decks in Programming
See All in Programming
ChatGPT とつくる PHP で OS 実装
memory1994
PRO
2
110
今年一番支援させていただいたのは認証系サービスでした
satoshi256kbyte
1
260
range over funcの使い道と非同期N+1リゾルバーの夢 / about a range over func
mackee
0
110
MCP with Cloudflare Workers
yusukebe
2
220
20年もののレガシープロダクトに 0からPHPStanを入れるまで / phpcon2024
hirobe1999
0
500
KubeCon + CloudNativeCon NA 2024 Overviewat Kubernetes Meetup Tokyo #68 / amsy810_k8sjp68
masayaaoyama
0
260
Recoilを剥がしている話
kirik
5
6.8k
nekko cloudにおけるProxmox VE利用事例
irumaru
3
440
Effective Signals in Angular 19+: Rules and Helpers
manfredsteyer
PRO
0
110
CQRS+ES の力を使って効果を感じる / Feel the effects of using the power of CQRS+ES
seike460
PRO
0
140
Keeping it Ruby: Why Your Product Needs a Ruby SDK - RubyWorld 2024
envek
0
190
生成AIでGitHubソースコード取得して仕様書を作成
shukob
0
480
Featured
See All Featured
XXLCSS - How to scale CSS and keep your sanity
sugarenia
247
1.3M
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
5
450
Docker and Python
trallard
42
3.1k
What's in a price? How to price your products and services
michaelherold
243
12k
Adopting Sorbet at Scale
ufuk
73
9.1k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
32
2.7k
Building Adaptive Systems
keathley
38
2.3k
We Have a Design System, Now What?
morganepeng
51
7.3k
Visualization
eitanlees
146
15k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
26
1.5k
Building Your Own Lightsaber
phodgson
103
6.1k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
127
18k
Transcript
͍ۙͷͰʮๅ͘͡ԾઆʯͷΛগ͠ 1ZUIPOػցֶशษڧձ JO৽ׁ ৽ׁݝ࢈75VCFS 4,"`T8FC CJO 1
ࢿྉͷϦϯΫͱࢀߟจݙ 2 IUUQTTQFBLFSEFDLDPNTLBTXFCMUMPUUFSZUJDLFUIZQPUIFTJT ࢀߟจݙ ɾIUUQTEFFQTRVBSFKQMPUUFSZUJDLFU ɾIUUQTMBCPSPBJBDUJWJUZDPMVNOFOHJOFFSๅ͘͡Ծઆ ݩจ ɾJonathan Frankle, Michael
Carbin, “The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks”, cs.LG : 1803.03635, ICLR 2019 ɾ5JBOMPOH$IFO +POBUIBO'SBOLMF 4IJZV $IBOH 4JKJB -JV :BOH;IBOH ;IBOHZBOH 8BOH .JDIBFM$BSCJO l5IF-PUUFSZ5JDLFU)ZQPUIFTJTGPS1SFUSBJOFE#&35/FUXPSLTz
5XJUUFS࿈ܞ 3 #pyml_niigata ͰͭͿ͘ͱʜʜ
࢈ۀ 4 ɾ ڊେ ͳϞσϧʢ#&35ɺ5ɺFUDʣ͕ͳͥਫ਼ྑ͍ͷ͔ʁ ɾύϥϝʔλʔ͕ଟ͍΄ͲΨνϟ͕͋ͨΔʂʁ ɾԠ༻͢Δͱ&EHF"*Ͱߴਫ਼ͳϞσϧΛಈ͔ͤΔʁ
ڊେͳϞσϧ 5
ڊେͳϞσϧ 6 ɾ͜ΜͳʹେྔͷύϥϝʔλʔΛ͏·͘ߋ৽͢Δ͜ͱͰ͖Δͷʁ ɾաֶश͠ͳ͍ͷʁ ๅ͘͡Ծઆ
ๅ͘͡Ծઆ 7 ɾχϡʔϥϧωοτϫʔΫॳظͷॏΈ͕ཚ ɾ࠷ॳ͔Β͍͍ײ͡ʹֶशՄೳͳ෦ωοτϫʔΫ͕͋Δ ɾ͋ͨΓ͘͡ωοτϫʔΫΛத৺ʹֶश͞ΕΔ ɾͣΕ͘͡ωοτϫʔΫࢬמΓͯ͠ਫ਼΄΅མͪͳ͍ ɾ#&35Ͱׂʙׂͷলύϥϝʔλʔ͕Մೳʢʁʣ ɾߴʹֶशɺਪ͕Ͱ͖Δ ɾ&EHF"*Ͱಈ͔ͤΔʁ
Ϟσϧͷݮ 8 ʢ·ͩௐ͍ͯ·ͤΜ PS[ ʣ ʢʣʹಈըߘ͠·͢ ʢࠓ͙͢Γ͍ͨΜ͚ͩͲͬͯਓࢀߟจݙݩจΛಡΜͰΈ͍ͯͩ͘͞ʣ
·ͱΊ 9 ɾχϡʔϥϧωοτϫʔΫͷॳظͷॏΈཚ ɾ͋ͨΓωοτϫʔΫͱͣΕωοτϫʔΫ͕͋Δ ɾ࿈ΨνϟͰͨΒͳ͍ʁ ԯ࿈͘Β͍Ε͚ͬ͜͏ͨΔΑʂ ɾ͍ͬͯ͏ͷ͕࠷ۙͷ ڊେϞσϧʢ#&35ɺ5ɺFUDʣ ɾͣΕωοτϫʔΫΛࢬמΓ͢Εগͳ͍ετϨʔδͰߴʹਪͰ͖Δ ɾ͡Ό͋Ͳ͏Δͷʁ
ˠ ·ͨࠓͳʂ
ࢿྉͷϦϯΫͱࢀߟจݙ 10 IUUQTTQFBLFSEFDLDPNTLBTXFC΄͛΄͛ ࢀߟจݙ ɾIUUQTEFFQTRVBSFKQMPUUFSZUJDLFU ɾIUUQTMBCPSPBJBDUJWJUZDPMVNOFOHJOFFSๅ͘͡Ծઆ ݩจ ɾJonathan Frankle, Michael
Carbin, “The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks”, cs.LG : 1803.03635, ICLR 2019 ɾ5JBOMPOH$IFO +POBUIBO'SBOLMF 4IJZV $IBOH 4JKJB -JV :BOH;IBOH ;IBOHZBOH 8BOH .JDIBFM$BSCJO l5IF-PUUFSZ5JDLFU)ZQPUIFTJTGPS1SFUSBJOFE#&35/FUXPSLTz