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
20190223_nlpaperchallenge_CV_4.3to5.5
Search
yoppe
February 23, 2019
Science
2
830
20190223_nlpaperchallenge_CV_4.3to5.5
Presentation at
https://nlpaper-challenge.connpass.com/event/118557/
.
yoppe
February 23, 2019
Tweet
Share
More Decks by yoppe
See All by yoppe
20211023_recsys2021_paper_reading_YoheiKikuta
diracdiego
2
480
20201121_oldpaperreading_computing_machinery_and_intelligence
diracdiego
0
160
20200906_ACL2020_metric_for_ordinal_classification_YoheiKikuta
diracdiego
1
1.3k
20191102_ACL2019_adversarial_examples_in_NLP_YoheiKIKUTA
diracdiego
2
1.4k
20180701_CVPR2018_reading_YoheiKIKUTA
diracdiego
3
1.2k
20180414_WSDM2018_reading_YoheiKIKUTA
diracdiego
0
710
20180306_NIPS2017_DeepLearning
diracdiego
4
5.9k
20180215_MLKitchen7_YoheiKIKUTA
diracdiego
0
440
20180210_Cookpad_TechConf2018_YoheiKIKUTA
diracdiego
5
1.2k
Other Decks in Science
See All in Science
Gemini Prompt Engineering: Practical Techniques for Tangible AI Outcomes
mfonobong
2
150
科学で迫る勝敗の法則(電気学会・SICE若手セミナー講演 2024年12月) / The principle of victory discovered by science (Lecture for young academists in IEEJ-SICE))
konakalab
0
120
IWASAKI Hideo
genomethica
0
130
局所保存性・相似変換対称性を満たす機械学習モデルによる数値流体力学
yellowshippo
1
300
My Favourite Book in 2024: Get Rid of Your Japanese Accent
lagenorhynque
1
110
Symfony Console Facelift
chalasr
2
460
academist Prize 4期生 研究トーク延長戦!「美は世界を救う」っていうけど、どうやって?
jimpe_hitsuwari
0
160
コンピュータビジョンによるロボットの視覚と判断:宇宙空間での適応と課題
hf149
1
300
Agent開発フレームワークのOverviewとW&B Weaveとのインテグレーション
siyoo
0
320
機械学習 - 授業概要
trycycle
PRO
0
230
データベース08: 実体関連モデルとは?
trycycle
PRO
0
930
オンプレミス環境にKubernetesを構築する
koukimiura
0
330
Featured
See All Featured
Music & Morning Musume
bryan
46
6.7k
Why You Should Never Use an ORM
jnunemaker
PRO
59
9.5k
Designing for humans not robots
tammielis
253
25k
Speed Design
sergeychernyshev
32
1.1k
Large-scale JavaScript Application Architecture
addyosmani
512
110k
Done Done
chrislema
185
16k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
46
7.6k
A better future with KSS
kneath
239
17k
Building an army of robots
kneath
306
46k
Stop Working from a Prison Cell
hatefulcrawdad
271
21k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
50
5.5k
Keith and Marios Guide to Fast Websites
keithpitt
411
22k
Transcript
ୈ2ճ nlpaper.challenge NLP/CV ަྲྀษڧձ ը૾ೝࣝ ୈ4.3ষ~ୈ5.5ষ 20190223 Yohei KIKUTA
ࣗݾհ • Twitter ID @yohei_kikuta • Resume • ࠷ۙୀ৬ͯ͠ແ৬ʹͳΓ·ͨ͠ ϒϩάΤϯτϦ
ແ৬ͱͯ͠Έ͍ͨਓ͓͕͚͍ͩ͘͞ʂ ΫοΫύουྑ͍ձࣾͳͷͰస৬͍ͨ͠ਓͥͻͲ͏ͧʂ
ଟ༷ମֶश͚ͩৄ͘͠Γ·͢ • ը૾ೝࣝ ຊͷ 4.3~5.5 Λൃද • ہॴಛʹ͔ؔͯ͠ͳΓߴີͰ࣌ؒʹ͖͠Εͳ͍ ଟ༷ମֶश ΧʔωϧؔۙࣅʢಛʹͦͷͨΊͷجૅ͕ࣝॏ͍ʣ
ͦͷ΄͔༷ʑͳ • શମͷ·ͱΊͱུ֓Λہॴಛͷ෦Λܰ͘આ໌ͨ͠ޙɺଟ ༷ମֶशΛৄ͘͠આ໌ Q. ͳͥଟ༷ମֶशͳͷ͔ʁ A. ͕ࣗগ͠ਅ໘ʹษڧͯ͠Έ͔͔ͨͬͨΒ
4.3~4.6 ͷ·ͱΊ ہॴಛͰࣦΘΕۭͨؒใΛ༩ͨ͠Γ͢Δʢ4.6ʣ
5.1~5.5 ͷ·ͱΊ ʢ֬తʣޯ߱Լ๏χϡʔτϯ๏Ͱύϥϝλ w Λֶश
ہॴಛΛ༻͍ͨ࠶ߏங ͜Ε݁ہͲͷล͕ଟ༷ମͳͷ͔ʁ ͱ͍͏͕ฉ͖͑ͯͦ͜͏͕ͩɺͦΕޙͰগ͠ৄ͘͠આ໌
ಛࣸ૾ʹΑΔΧʔωϧͷۙࣅ ػցֶशʹଌ͍Δͷʁͱ͍͏ٙΛ࣋ͭਓ Mercer’s theorem ͱ͔ Bochner’s theorem ͱ͔Λݟ·͠ΐ͏
ۭؒใͷ׆༻ • ϓʔϦϯάʹΑࣦͬͯΘΕΔҐஔใΛखͰՃ͑Δ • ہॴهड़ࢠͷҐஔεέʔϧΛՃ͑Δ • ہॴهड़ࢠͷࣗݾ૬ؔߦྻͷཁૉΛՃ͑Δ • spatial pyramid
ϓʔϦϯάޙͷಛʹۭؒใΛ༩
ଟ༷ମֶश ଟ༷ମֶशʹͯ͠গ͠ৄ͘͠આ໌͢Δʢݸਓతڵຯʣ આ໌͢ΔͷҎԼͷτϐοΫɿ • ͦͦଟ༷ମͱʁʢֶతʹݫີͳ͠ͳ͍ʣ • ଟ༷ମֶशʹࢸΔϞνϕʔγϣϯ • Ұͭͷྫͱͯ͠ہॴ࠲ඪ coding
ͷվળͷจΛհ [140] K. Yu, T. Zhang. Improved local coordinate coding using local tangents. In ICML, 2010.
ଟ༷ମͱʁ ہॴతʹ Euclid ۭؒͰهड़Ͱ͖ΔਤΛషΓ߹ΘͤͯදͤΔ ྫʣද໘Λߟ͑Δ ※ Φ ͱ Θ ͚ͩͰ࠲ඪܥషΓ͖Εͳ͍ʢશͳٿͰͳ͍ʣ
ہॴతʢզʑͷৗεέʔϧʣʹ Euclid ࠲ඪͰهड़Ͱ͖Δ ͦͷ࠲ඪΛషΓ߹ΘͤΕද໘શମΛΧόʔͰ͖Δ ͦͷΑ͏ͳషΓ߹ΘͤͷใͰ 3 ࣍ݩͷٿͷใٞՄ
ଟ༷ମͱʁ σʔλͷॅΉۭؒಉ͡Α͏ͳͷͩͱ૾͞ΕΔ ͜ͷΑ͏ͳঢ়گͰɺྫ͑ೋؒͷڑΛଌΔͱ͖ʹߴ࣍ݩۭ ؒͰͷ Euclid ڑෆద͔͠Εͳ͍ → ଟ༷ମʹԊͬͯڑΛଌΔํ͕ੑ࣭Λଊ͑ΒΕͦ͏ʢଌઢʣ → ใزԿʹ͓͚Δࣗવޯ๏ͳͲ͜ΕΛ͍ͬͯΔ
ଟ༷ମֶशʹࢸΔಈػ • ࣍ݩͷढ͍ͷճආ ྫʣσʔλେྔʹ͋Ε kNN Ͱ͍͍͡ΌΜ → μϝͰ͢ • σʔλͷ༗͢ΔಛΛΑΓྑ͘Ѳ
ઌ΄ͲͷྫͷΑ͏ʹ୯ͳΔ Euclid ڑ͕ෆద͔͠Εͳ͍ ྫʣࣗવޯ๏ɺt-SNE • ༷ʑͳΞϓϩʔν͔Β৽ͨͳݟ͕ಘΒΕΔ͔ ඍزԿతͳΞϓϩʔνʢہॴతʣˠ େମͷͬͪ͜ Ґ૬زԿతͳΞϓϩʔνʢେҬతʣˠ ࠓ৮Εͳ͍
ʢิʣ࣍ݩͷढ͍ • ಛྔͷ࣍ݩ͕େ͖͗͢Δ߹ Ϟσϧ͕ෳࡶ͗ͯ͢దʹֶशͰ͖ͳ͍ • σʔλۭؒͷ࣍ݩ͕େ͖͗͢Δ߹ ྫͱͯ͠ΫϥελϦϯάΛߟ͑Δ ٿ໘ूதݱʹΑΓɺҟͳΔؒͷڑ͕͘͠ͳ͍ͬͯ͘ ݁Ռͱͯ͠ A
͔Βݟͯ B C ΄΅ಉ͡Ͱ۠ผෆՄ ʢ࣍ݩͷढ͍ͷ ͡Ίͯͷύλʔϯೝࣝ ͕ৄ͍͠ʣ
ଟ༷ମֶशͷྫ ہॴ࠲ඪ coding Λվળ͢Δ͜ͱΛߟ͑Δ K. Yu, T. Zhang. Improved local
coordinate coding using local tangents. In ICML, 2010. ࠷ऴతͳඪσʔλ͕͢ଟ༷ମͷࡏతͳ࣍ݩʢPCA Ͱٻ ΊΔʣΛߟྀ͠ɺͦͷใΛͬͯ coding Λิਖ਼
४උ
ہॴ࠲ඪ coding ʹΑΔۙࣅ
(4.79) ࣜͷূ໌
ہॴ࠲ඪ coding ͷֶश๏
֦ுہॴ࠲ඪ coding
σʔλଟ༷ମͷఆٛ
ہॴಛ coding with u ূ໌ུʢઌ΄Ͳͷূ໌͕͔ͬͯΕ͘͠ͳ͍ʣ u (local) PCA ͳͲͰٻΊɺm
खͰܾΊΔύϥϝλ c(M) ؔʹ͓͚ΔϦϓγοπఆʹ૬ ॏཁͳͷ c(M) ͕খ͍͞߹ʹ bound ͕Ωπ͘ͳΔͱ͍͏͜ͱ → ͜Εখ͞ͳྖҬͰσʔλଟ༷ମ͕ flat ͳߏͰ͋Δ߹ → “ఆੑత” ͳԾఆͰ flat Ͱͳͯ͘༗ޮͳ߹͋ΓಘΔ
ֶशΞϧΰϦζϜ จͷΞϧΰϦζϜࡌͤΔʢs, m hyperparameterʣ
࣮ݧ݁ՌɿMNIST Cross validation ʹΑΓ s = 0, m = 64
୯७ͳہॴಛ coding ͱൺΔͱ anchor ͕গͳͯ͘ྑ͍
࣮ݧ݁ՌɿCIFAR10 Cross validation ʹΑΓ s = 10, m = 256
ͪ͜Β anchor ͕গͳͯ͘ྑ͍݁ՌΛ͍ࣔͯ͠Δ
ଟ༷ମֶशʹཱͭ • σʔλͷߏΛ͏·͘ଊ͑ͯ༗༻ͳಛྔΛ࡞Εͨ • ہॴੑʹΛ͠ͳ͕Βߴ࣍ݩใΛ࣍ݩͰۙࣅ • ࣮ࡍʹࣝผੑೳ্͕ • σΟʔϓϥʔχϯάͰଟ༷ମֶशڵຯਂ͍τϐοΫ ใزԿ
ϦʔϚϯଟ༷ମ্ͰֶशΛఆࣜԽ …
4.3~4.6 ͷ·ͱΊʢ࠶ܝʣ ہॴಛͰࣦΘΕۭͨؒใΛ༩ͨ͠Γ͢Δʢ4.6ʣ