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
Optimization as a Model for Few-Shot Learning -...
Search
Hokuto Kagaya
June 16, 2017
Science
2
1.3k
Optimization as a Model for Few-Shot Learning - ICLR 2017 reading seminar
Hokuto Kagaya
June 16, 2017
Tweet
Share
More Decks by Hokuto Kagaya
See All by Hokuto Kagaya
LINE Bot ✕ Chainer - low-carb recipe bot -
hokkun
0
1.1k
Other Decks in Science
See All in Science
Boil Order
uni_of_nomi
0
120
ウェーブレットおきもち講座
aikiriao
1
790
位相的データ解析とその応用例
brainpadpr
1
610
統計的因果探索の方法
sshimizu2006
1
1.2k
マテリアルズ・インフォマティクスの先端で起きていること / What's Happening at the Cutting Edge of Materials Informatics
snhryt
1
130
白金鉱業Meetup Vol.15 DMLによる条件付処置効果の推定_sotaroIZUMI_20240919
brainpadpr
1
490
HAS Dark Site Orientation
astronomyhouston
0
5.3k
作業領域内の障害物を回避可能なバイナリマニピュレータの設計 / Design of binary manipulator avoiding obstacles in workspace
konakalab
0
160
2024-06-16-pydata_london
sofievl
0
520
Презентация программы магистратуры СПбГУ "Искусственный интеллект и наука о данных"
dscs
0
390
Snowflakeによる統合バイオインフォマティクス
ktatsuya
0
490
Science of Scienceおよび科学計量学に関する研究論文の俯瞰可視化_ポスター版
hayataka88
0
130
Featured
See All Featured
Making Projects Easy
brettharned
115
5.9k
Happy Clients
brianwarren
98
6.7k
Writing Fast Ruby
sferik
627
61k
Product Roadmaps are Hard
iamctodd
PRO
49
11k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
28
2k
A better future with KSS
kneath
238
17k
BBQ
matthewcrist
85
9.3k
How to Think Like a Performance Engineer
csswizardry
20
1.1k
The Power of CSS Pseudo Elements
geoffreycrofte
73
5.3k
Intergalactic Javascript Robots from Outer Space
tanoku
269
27k
Scaling GitHub
holman
458
140k
GraphQLの誤解/rethinking-graphql
sonatard
67
10k
Transcript
!%F/" 4IJCVZB)JLBSJF )PLVUP ,BHBZB !@IPLLVO@ 0QUJNJ[BUJPOBTB.PEFM GPS'FX4IPU-FBSOJOH
5-%3 • 1VSQPTF • #FUUFSJOGFSFODFGPSGFXTIPUPOFTIPUMFBSOJOH QSPCMFN • .FUIPE • .FUBMFBSOJOHCBTFEPO-45.PGEFFQOFVSBM
OFUXPSL • 3FTVMU • $PNQFUJUJWFXJUIEFFQNFUSJDMFBSOJOHUFDIOJRVFT
#BDLHSPVOE • 8IZEFFQMFBSOJOHTVDDFFEFE • NBDIJOFQPXFS • BNPVOUPGEBUB • -BSHF%BUBTFU
• *NBHF/FU *NBHF • .JDSPTPGU$0$0$BQUJPOT *NBHF$BQUJPO • :PV5VCF. 7JEFP • 8JLJ5FYU 5FYU
#BDLHSPVOE • )PXFWFS *ONBOZGJFMET UPDPMMFDUBMBSHF BNPVOUPGUSBJOJOHTBNQMFTJT • EJGGJDVMU •
&Y'JOFHSBJOFESFDPHOJUJPO DBS CJSE GPPE • UJNFDPOTVNJOH • TDSBQJOH DSBXMJOH BOOPUBUJOHʜ • "DUVBMMZIVNBOCFJOHTDBOHFOFSBMJ[FVTJOH GFXTBNQMFTPGUBSHFUT
1SPCMFN1VSQPTF • )PXDBOXFBDRVJSFHFOFSBMJ[FENPEFMVTJOH GFXTBNQMFTBOEBTFUOVNCFSPGVQEBUFT • &YJTUFEHSBEJFOUCBTFEUSBJOJOHBMHPSJUIN 4(% "%".
"EB(MBE EPFTOPUGJUUIFQSPCMFNXJUIB TFUOVNCFSPGQBSBNFUFSVQEBUFT • *OPUIFSTJNQMFXPSET BVUIPSTXBOUUPGJOE HPPEJOJUJBMQBSBNFUFSTPG// • DG SFWJFXDPNNFOUTJUJTNVDICFUUFSUPCFBCMFUP GJOEBSDIJUFDUVSBMQBSBNFUFSTPG//
1SPCMFN1VSQPTF • )PX • .FUBMFBSOJOH • -FBSOJOHUPMFBSO5SBJOMFBSOFSJUTFMG • "WBSJFUZPGNFUBMFBSOJOH
• 5SBOTGFSMFBSOJOH • 6TFUIFFYQFSJFODFTPGEJGGFSFOUEPNBJO • 1PQVMBSJOUIFGJFMEPGJNBHFDMBTTJGJDBUJPO FTQFDJBMMZGPS GJOFHSBJOFEWJTVBMDMBTTJGJDBUJPO • &OTFNCMFDMBTTJGJFS • DPNCJOFNVMUJQMFDMBTTJGJFS 5IJTBSUJDMFJTWFSZHPPEUPVOEFSTUBOENFUBMFBSOJOH IUUQIUUQXXXTDIPMBSQFEJBPSHBSUJDMF.FUBMFBSOJOH
1SPQPTFENFUIPE • -45.CBTFENFUBMFBSOJOH
1SFSFRVJTJUFT • 8IBUJT-45. • -POHUJNF4IPSU5FSN.FNPSZ • ࣌ܥྻΛѻ͍͍ͨɺͰޡ͕ࠩൃࢄফࣦͪ͠Ό͏ • աڈͷσʔλͷॏΈΛ̍ʹͯ͠Εͳ͍Α͏ʹ্ͨ͠ Ͱɺબతʹೖྗग़ྗΛߦ͏Α͏ʹͨ͠
b • ͔͠͠ٸܹͳঢ়گͷมԽʢʁʣʹରԠͰ͖ͳ͔ͬͨͷ Ͱɺ٫ήʔτΛઃஔ͢Δ͜ͱͰબతʹաڈͷσʔ λͷهԱΛফڈͰ͖ΔΑ͏ʹͨ͠ ` • ࢀߟʢຊޠʣ • IUUQRJJUBDPNU@4JHOVMMJUFNTCCFCGEC
%BUB4FQBSBUJPO • NFUBUSBJOEBUBTFU • NFUBUFTUEBUBTFU • NFUBTBNQMF UBSHFU USBJOJOH
TBNQMFT UBSHFU UFTUJOH TBNQMFT POFNFUBTBNQMF BLBFQJTPEF
1SPQPTFE.FUIPE " = "$% + )*+, ℒ" " =
" ⨀"$% + " ⨀̅" XIFSF " = (6 7 ? + b9 ) " = ; 7 ? + b< XIFSF ? DVSSFOUHSBEJFOUT DVSSFOUMPTT QSFWJPVT QSFWJPVTJUTFMG , /PSNBM4(% .FUBQIPS OPUDPOTUBOU UP FTDBQFGSPNCBE MPDBMPQUJNB
1SPQPTFE.FUIPE ←meta learner‘s iteration ←learner‘s iteration (meta) loss
value is computed by final state of LSTM (= parameters of target model) and "?@"’s data and labels.
1SPQPTFE.FUIPE • 'SPNBVUIPS`TTMJEF
1SPQPTFE.FUIPE • 8IBUXJMMCFJNQSPWFEHSBEVBMMZ • 'JSTU-45.QBSBNFUFS BLBNFUBMFBSOFS QBSBNFUFST • UIBUJT
zIPXTIPVMEXFVQEBUFUBSHFUNPEFMT z • 4FDPOE-45.TUBUFT PVUQVUT • 'JOBMA JTTIBSFEBNPOHFBDICBUDI TPMFBSOJOHQSPDFFET SBQJEMZUIBOLTGPSHPPEJOJUJBMJ[BUJPO
0UIFS5PQJDT • DPPSEJOBUFXJTF-45. • 1SFQSPDFTTJOHUP-45.JOQVUT • BCPVUCPUIUPQJDT TFF<"OESZDIPXJD[ /*14> QSFQSPDFTTJOHJTJOBQQFOEJY
• BEKVTUUIFTDBMJOHPGHSBEJFOUTBOEMPTTFT • TFQBSBUFJOGPPGNBHOJUVEFBOETJHO • #BUDIOPSNBMJ[BUJPO • BWPJEzEBUBTFUz FQJTPEF MFWFMMFBLBHFPG JOGPSNBUJPO • 3FMBUFEXPSLNFUSJDMFBSOJOH • FY4JBNFTFOFUXPSL
&WBMVBUJPO.FUIPE • #BTFMJOFOFBSFTUOFJHICPS • NFUBUSBJOUSBJOOFVSBMOFUXPSLVTJOHBMMTBNQMF • NFUBUFTUUSBJOJOHTBNQMFΛ//ʹͿͪ͜Μͩ݁Ռͱ UFTUJOHTBNQMFͷͦΕΛൺֱ • #BTFMJOFGJOFUVOF
• NFUBUSBJOʹՃ͑ͯɺ NFUBWBMJEBUJPO EBUBTFU Λ IZQFS QBSBNFUFSͷ୳ࡧʹ͍ɺ ͷ OFUXPSLΛ GJOFUVOF ͢Δ • #BTFMJOF.BUDIJOHOFUXPSL • ڑֶशͷ405"
&WBMVBUJPO3FTVMU
7JTVBMJ[BUJPOBOE*OTJHIU
7JTVBMJ[BUJPOBOE*OTJHIU • JOQVUHBUFT • EJGGFSFOUBNPOHEBUBTFUT • NFUBMFBSOFSJTO`UTJNQMZMFBSOJOHBGJYFEPQUJNJ[BUJPO TUSBUFHZ • EJGGFSFOUBNPOHUBTLT
• NFUBMFBSOFSIBTVTFEEJGGFSFOUXBZTUPTPMWFFBDI TFUUJOH • GPSHFUHBUFT • TJNQMFEFDBZ • ݁ہ΄ͱΜͲ DPOTUBOU
$PODMVTJPO • 'PVOE-45.CBTFENPEFMUPMFBSOBMFBSOFS XIJDIJTJOTQJSFECZBNFUBQIPSCFUXFFO 4(%VQEBUFTBOE-45. • 5SBJONFUBMFBSOFSUPEJTDPWFS • (PPEJOJUJBMJ[BUJPOPGMFBSOFS
• (PPENFDIBOJTNGPSVQEBUJOHMFBSOFS`T QBSBNFUFST • DPNQFUJUJWFFYQFSJNFOUBMSFTVMUXJUI405" NFUSJDMFBSOJOHNFUIPET
'VUVSFXPSL • GFXTBNQMFTMPUTPGDMBTTFT • NPSFDIBMMFOHJOHTDFOBSJPT • GSPNSFWJFXDPNNFOU • JUJTNVDICFUUFSUPCFBCMFUPGJOEBSDIJUFDUVSBM QBSBNFUFSTPG//
ॴײ • USBOTGFSMFBSOJOHʹ͓͚ΔʮผυϝΠϯͷܦݧ Λ׆͔͢ʯͱ͍͏࡞ۀΛʮ࣌ܥྻͷֶशʯతʹଊ ͑ͯ -45.Ϟσϧͱֶͯ͠शͨ͠ɺͱ͍͏ͷ ࣗવʹࢥ͑ͨ • ͢Ͱʹ͋ͬͨൃʁ࣌ؒͳؔ͘࿈ݚڀ·ͰಡΈࠐΊ ͣɻɻ
• SFWJFXDPNNFOUʹ͋ͬͨɺߏͷ࠷దԽ·Ͱ Ͱ͖Δͱ͘͢͝Αͦ͞͏ͩͱࢥͬͨ • γϯϓϧͳϑΟϧλΛͨ͘͞ΜॏͶΔͱ͍͍ͱ͍͏ ͋Δ͕ɻɻ • ֶ෦࣌ DVEBDPOWOFU Λͬͯͨ͘͞ΜϋΠύύϥ ϝʔλΛࢼͨۤ͠࿑͕ોͬͨ
ଟΘ͔ͬͯͳ͍͜ͱ • ݁ہɺ͜ͷจͰॳΊͯΘ͔ͬͨͷͲ͜ʁ -45.Λ MFBSOJOHUPMFBSOʹͬͨͷଟॳ Ίͯ͡Όͳ͍ʁ • ྫ͑ "OESZDIPXJD[ `
ͰɺޯΛೖྗʹͯ͠ UBSHFUMFBSOFSͷ QBSBNFUFSVQEBUFT Λग़ྗ͢Δ -45.Λֶश • ύϥϝλͦͷͷΛग़ྗͯ͠Δͱ͜Ζʁ