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
論文紹介 Hardness-Aware Deep Metric Learning [CVPR ...
Search
hyodo
June 10, 2019
Technology
0
520
論文紹介 Hardness-Aware Deep Metric Learning [CVPR 2019]
研究室のゼミで"Deep Metric Learning"というタイトルで発表した資料の一部になります。ご指摘や議論等お待ちしております。
Twitter @onysuke
hyodo
June 10, 2019
Tweet
Share
More Decks by hyodo
See All by hyodo
The Impact of Advertising along the Conversion Funnel
onysuke
2
1.6k
Can offline stores drive online sales?
onysuke
0
1.5k
SizeFlags: Reducing Size and Fit Related Returns in Fashion E-Commerce
onysuke
0
880
意思決定のための機械学習
onysuke
1
1k
Mixture of Expertsに関する文献調査
onysuke
1
2k
Other Decks in Technology
See All in Technology
夢の印税生活 / Life on Royalties
tmtms
0
280
見てわかるテスト駆動開発
recruitengineers
PRO
4
280
Goでマークダウンの独自記法を実装する
lag129
0
210
【5分でわかる】セーフィー エンジニア向け会社紹介
safie_recruit
0
30k
RAID6 を楔形文字で組んで現代人を怖がらせましょう(実装編)
mimifuwa
0
300
LLMエージェント時代に適応した開発フロー
hiragram
1
410
Goss: New Production-Ready Go Binding for Faiss #coefl_go_jp
bengo4com
0
1.1k
AIエージェントの開発に必須な「コンテキスト・エンジニアリング」とは何か──プロンプト・エンジニアリングとの違いを手がかりに考える
masayamoriofficial
0
370
kintone開発チームの紹介
cybozuinsideout
PRO
0
73k
自社製CMSからmicroCMSへのリプレースがプロダクトグロースを加速させた話
nextbeatdev
0
130
我々は雰囲気で仕事をしている / How can we do vibe coding as well
naospon
2
220
「守る」から「進化させる」セキュリティへ ~AWS re:Inforce 2025参加報告~ / AWS re:Inforce 2025 Participation Report
yuj1osm
1
110
Featured
See All Featured
What's in a price? How to price your products and services
michaelherold
246
12k
Fireside Chat
paigeccino
39
3.6k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
283
13k
Facilitating Awesome Meetings
lara
55
6.5k
Raft: Consensus for Rubyists
vanstee
140
7.1k
Why You Should Never Use an ORM
jnunemaker
PRO
59
9.5k
A better future with KSS
kneath
239
17k
For a Future-Friendly Web
brad_frost
179
9.9k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
15
1.6k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
231
53k
[Rails World 2023 - Day 1 Closing Keynote] - The Magic of Rails
eileencodes
36
2.5k
StorybookのUI Testing Handbookを読んだ
zakiyama
30
6k
Transcript
)BSEOFTT"XBSF%FFQ.FUSJD-FBSOJOH $7130SBM 8FO[IBP ;IFOH ;IBPEPOH $IFO +JXFO -V +JF ;IPV
%FQBSUNFOUPG"VUPNBUJPO 5TJOHIVB6OJWFSTJUZ $IJOB FUD 1
֓ཁ 2 ɾ/FHBUJWFTBNQMFͷқΛௐ͢ΔϑϨʔϜϫʔΫ )%.- )BSEOFTT"XBSF%FFQ.FUSJD-FBSOJOH ΛఏҊ /FHBUJWFTBNQMFͷқΛજࡏ্ۭؒͷઢܗิؒʹΑΓௐ ֶशঢ়گʹదͳ͠͞ͷOFHBUJWFαϯϓϧΛੜ͢Δ
എܠ • /FHBUJWFTBNQMJOHॏཁͳ • ఏҊ͞Ε͍ͯΔख๏ͷଟ͘ɼֶशΛଅਐ͢Δ ͠ ͍ /FHBUJWFΛͲ͏બ͢Δ͔ʹযΛ͍͋ͯͯͨ ‑ Ұ෦ͷTBNQMFΛऔΓଓ͚Δ͜ͱʹͳΓɼજࡏۭؒͷେ
ہతͳܗΛଊ͑Δ͜ͱ͕Ͱ͖͍ͯͳ͍ PWFSGJUUJOH 3
4 ఏҊख๏֓આ ᶃ )BSEBXBSFGFBUVSFTZOUIFTJT ΞϯΧʔʹ͚ۙͮͨOFHBUJWF ! Λੜ ᶄ )BSEOFTTBOE-BCFM1SFTFSWJOHGFBUVSFTZOUIFTJT
ੜͨ͠OFHBUJWF ! Λ ͷϥϕϧͱಉ͡ʹͳΔΑ͏ʹඍௐ ᶃ ᶄ ! " = "
5 .BOJGPME $MBTT" ఏҊख๏֓આ ᶃ)BSEBXBSFGFBUVSFTZOUIFTJT .BOJGPME $MBTT#
6 .BOJGPME $MBTT" : → GFBUVSFTQBDF͔Β FNCFEEJOHTQBDF NFUSJDTQBDF ʹࣹӨ ఏҊख๏֓આ
ᶃ)BSEBXBSFGFBUVSFTZOUIFTJT .BOJGPME $MBTT#
7 .BOJGPME $MBTT" & ! = + " ! −
" ∈ [0,1] ҎԼͷઢܗิؒʹΑΓ ʹ͚ۙͮͨΑΓ͍͠ ̂ Λੜ ఏҊख๏֓આ ᶃ)BSEBXBSFGFBUVSFTZOUIFTJT .BOJGPME $MBTT#
8 Hard-aware feature .BOJGPME $MBTT" l% !ͱ!͕ಉϥϕϧz อূ͞Ε͍ͯͳ͍ ˣ !ͱಉϥϕϧʹ
ͳΔΑ͏ͳ( !ΛϚοϓ ఏҊख๏֓આ ᶄ)BSEOFTTBOE-BCFM1SFTFSWJOHGFBUVSFTZOUIFTJT : → .BOJGPME $MBTT#
ఏҊϑϨʔϜϫʔΫ )%.- 9 : → : → .FUSJDOFUXPSL "VHNFOUFS HLP(Hardness-and-Label-Preserving)
Generator Network "VHNFOUFS )-1(FOFSBUPS/FUXPSL
"VHNFOUFS 10 : → : → .FUSJDOFUXPSL "VHNFOUPS & !
= + " ! − , "∈ 0,1 … (1) " = + + 1 − # , ! , , ! > # 1 , , ! ≤ # , ∈ 0,1 … (2) ; " ∈ $! $ ," , 1 ͱͯ͠ , ! = ! − ' % ! = + [ , ! + 1 − #] "! $ ," , , ! > # … (3) ' ! = * + [ ! " #!"# , ! + 1 − ! " #!"# $] ! − , ! , , ! > $ ! , , ! ≤ $ … (4) % = 0ͷͱ͖' ! = ͱͳͬͯ͠·͏ʜ ʹ Λೖ͢Δͱ = ! # $%&'ͱͯ͠
"VHNFOUFSֶशঢ়گʹԠͨ͡қͷOFHBUJWFΛੜ ; % # = ' + [ # $
%&'( , # + 1 − # $ %&'( &] # − , # , , # > & # , , # ≤ & … (4) '() ʜͭલͷFQPDIͷ"WFSBHFNFUSJDMPTT FY5SJQMFUMPTT 11 @AB খ େ # $ %&'( 0 1 % ! = + $! $ ," (! − ) % ! = ! % !ͷқ easy hard MPTTͷେ͖͞ ֶशঢ়گ ʹԠͯ͡ੜ͢ΔOFHBUJWFͷқΛௐ
)-1(FOFSBUPS/FUXPSL 12 : → : → "VHNFOUPS HLP(Hardness-and-Label-Preserving) Generator Network
9:; = <:=>; + λ?>@A = − B C + λ?>@A () , ) l% #ͱ#͕ಉϥϕϧzอূ͞Ε͍ͯͳ͍ ⇒ #ͱಉϥϕϧʹͳΔΑ͏ͳE #ΛϚοϓ HFOFSBUPS: → PCKFDUJWFGVODUJPO )-1(FOFSBUPS /FUXPSL &OD %FD ͱͯ͠ͷ੍߲ ݩͷϥϕϧ Λอূ͢Δ
.FUSJDOFUXPSL PCKFDUJWFGVODUJPO .FUSJDOFUXPSL 13 : → : → .FUSJDOFUXPSL "VHNFOUFS
HLP(Hardness-and-Label-Preserving) Generator Network EFGHIJ = ! K L!"#E + 1 − ! K L!"# MNO = ! K L!"#() + 1 − ! K L!"# (; ) NFUSJDMPTT FY5SJQMFUMPTT /QBJSMPTT ݩͷσʔλର ੜͨ͠σʔλର ৴པͰ͖Δ 㱺 ੜͨ͠σʔλର ৴པͰ͖ͳ͍ 㱺 ݩͷσʔλର HFOFSBUPS ͕ ͷNFUSJDMPTTʹॏ͖Λ͓͘
$6#σʔληοτ ௗͷը૾ छྨ ܭ ຕ 5SBJO ຕ छྨ 5FTU
ຕ छྨ 5SBJOͱ5FTUʹಉ͡Ϋϥεͷը૾ଘࡏ͠ͳ͍ 㱺 ;FSPTIPUTFUUJOH 14
࣮ݧઃఆ DMVTUFSJOHSFUSJFWBMUBTL 15 $MVTUFSJOHUBTL ධՁࢦඪ /.* ਖ਼نԽ૬ޓใྔ ' 3FDBMM!, 5FTU
5SBJO Clustering task Retrieval task 3FUSJFWBMUBTL ֤UFTUը૾ RVFSZ ʹରͯ͠ ,ίۙͷΛநग़͠ɼ ಉ͡Ϋϥε͕ଐ͍ͯ͠Ε TDPSFFMTFTDPSF
.FUSJDMPTTͷछྨʹΑΒͣ )%.-Ͱࣝผతͳಛྔ͕ಘΒΕͨ 16
!"#$ ֶ͕शʹ͓͍ͯॏཁͳཁૉͰ͋Δ 17 HFJQO ͳ͠ͰϕʔεϥΠϯΛ্ճΔ 㱺 *+,- ͚ͩͰݱ࣮తͳಛදݱͷϚοϐϯά͕ՄೳͰ͋ͬͨͱߟ͑ΒΕΔ
ΫϥεͷมԽ എܠ ࢹ র໌ FUD ΫϥεؒͷΘ͔ͣͳҧ͍ ௗͷ༷ 18 ʹରॲ