Slide 1

Slide 1 text

࣌ܥྻղੳ 5BJDIJ .VSBZBNB େࡕେֶ ࢈ۀՊֶݚڀॴ 2022/09/28 AI⼈材養成プログラム

Slide 2

Slide 2 text

2 l࣌ܥྻzͱ͸ʁ

Slide 3

Slide 3 text

3 ࣌ܥྻͱ͸ ࣌ܥྻ UJNFTFSJFT ࣌ؒͱͱ΋ʹෆنଇʹมಈ͢Δݱ৅Λɺ࿈ଓతʹ؍ଌͯ͠ಘΒΕͨ ஋ͷܥྻ 𝑡؍ଌͨ࣌͠ࠁ 𝑦! ࣌ࠁ 𝑡 ʹ͓͚Δ؍ଌ஋ 𝑦" , … , 𝑦# ࣌ܥྻ ଟ͘ͷ࣌ؒͱͱ΋ʹมԽ͍ͯ͘͠σʔλ͸ɺ͜ͷΑ͏ͳදهͰදݱ ͢Δ͜ͱ͕Մೳ 𝑇 𝑌 ྫυϧԁ 2022年7⽉1⽇ 𝑦! = 135

Slide 4

Slide 4 text

4 ࣌ܥྻͱ͸ ೔ܦฏۉגՁ $07*%ͷײછऀ਺ ೥ฏۉؾԹͷมԽ

Slide 5

Slide 5 text

5 ࣌ܥྻͱ͸ .PUJPO$BQUVSFͷྫ From: Matsubara, Yasuko, and Yasushi Sakurai. "Regime shifts in streams: Real-time forecasting of co- evolving time sequences." Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016.

Slide 6

Slide 6 text

6 ࣌ܥྻͱ͸ l ࣌ܥྻ༧ଌকདྷͷมಈΛਪଌ l ࣌ܥྻ෼ྨ༩͑ΒΕͨ࣌ܥྻΛछྨ͝ͱʹ෼ྨ l ҟৗݕ஌ มԽݕ஌࣌ܥྻͷதʹଘࡏ͢ΔҟৗΛൃݟ l ՄࢹԽ ಛ௃நग़ਓ͕ؒղऍՄೳͳܗʹ࣌ܥྻΛม׵ l ੍ޚૢ࡞Մೳͳม਺ΛมԽͤ͞ɺ໨తม਺Λ੍ޚ FUD ࣌ܥྻΛѻͬͨ՝୊ λεΫ

Slide 7

Slide 7 text

7 ࣌ܥྻͱ͸ l ࣌ܥྻ༧ଌকདྷͷมಈΛਪଌ l ࣌ܥྻ෼ྨ༩͑ΒΕͨ࣌ܥྻΛछྨ͝ͱʹ෼ྨ l ҟৗݕ஌ มԽݕ஌࣌ܥྻͷதʹଘࡏ͢ΔҟৗΛൃݟ l ՄࢹԽ ಛ௃நग़ਓ͕ؒղऍՄೳͳܗʹ࣌ܥྻΛม׵ l ੍ޚૢ࡞Մೳͳม਺ΛมԽͤ͞ɺ໨తม਺Λ੍ޚ FUD ࣌ܥྻΛѻͬͨ՝୊ λεΫ

Slide 8

Slide 8 text

8 ࣌ܥྻͱ͸ ࣌ܥྻ༧ଌ ? ݱࡏ কདྷ ֶश

Slide 9

Slide 9 text

9 ࣌ܥྻͱ͸ ࣌ܥྻ༧ଌ ݱࡏ কདྷ ֶश

Slide 10

Slide 10 text

10 ࣌ܥྻͱ͸ ࣌ܥྻ෼ྨ from: https://medium.com/@hassanismailfawaz/deep-learning-for-time- series-classification-a-brief-overview-73b58767ed0f ?

Slide 11

Slide 11 text

11 ࣌ܥྻͱ͸ ҟৗݕ஌ มԽݕ஌ From: https://neptune.ai/blog/anomaly-detection-in-time-series ҟৗʂ ҟৗʂ

Slide 12

Slide 12 text

12 ࣌ܥྻͱ͸ ಛ௃நग़ %FDPNQPTJUJPO ྫ࣌ܥྻ෼ղ ྫߦྻ෼ղ from: https://www.statsmodels.org/dev/generated/statsmodels.tsa.seasonal.seasonal_decompose.html from: https://www.researchgate.net/figure/Matrix-decomposition-vs-tensor-decomposition- a-low-rank-matrix-decomposition_fig4_359367892

Slide 13

Slide 13 text

13 ࣌ܥྻͱ͸ ੍ޚ from: https://xtech.nikkei.com/atcl/nxt/news/18/08031/

Slide 14

Slide 14 text

14 ࣌ܥྻͱ͸ ओͳϞσϦϯάϑϨʔϜϫʔΫ モデル σʔλ ܦݧత஌ࣝ ཧ࿦ ࣌ܥྻ༧ଌ ࣌ܥྻ෼ྨ ҟৗ஋ݕ஌

Slide 15

Slide 15 text

15 ࠓճͷେ·͔ͳ֓ཁ l ౷ܭతϞσϧ l ࣌ܥྻͷੑ࣭ l ݹయత࣌ܥྻϞσϧ l Ϟσϧͷਪఆɾબ୒ l ਂ૚ֶशϕʔεϞσϧ l ۙ೥ͷਂ૚ֶशͷൃల l ࣌ܥྻ༧ଌ ਂ૚ֶश ౷ܭతϞσϧ ਂ૚ֶशϕʔεϞσϧ

Slide 16

Slide 16 text

16 ࣌ܥྻͷੑ࣭

Slide 17

Slide 17 text

17 ࣌ܥྻͷੑ࣭ l ౳ִؒ࣌ܥྻ l αϯϓϦϯάظ͕ؒҰఆ l ݚڀͳͲͰѻ͏ݚڀͷ΄ͱΜͲ͸͜Εʹ౰ͨΔ ܽଛ஋ิ׬ͳͲͰ౳ִؒʹ͢Δ৔߹΋ l ෆ౳ִؒ࣌ܥྻ l Πϕϯτ࣌ܥྻͳͲ l αϯϓϦϯάظ͕ؒૄΒ ౳ִؒ࣌ܥྻ ෆ౳ִؒ࣌ܥྻ ࣌ؒ 猫かわいい 1PTU 35 35

Slide 18

Slide 18 text

18 ࣌ܥྻͷੑ࣭ l Ұมྔ࣌ܥྻ l छྨͷܥྻʹΑͬͯߏ੒ l ଟมྔ࣌ܥྻ l ෳ਺ͷܥྻʹΑͬͯߏ੒ Ұมྔ࣌ܥྻ ଟมྔ࣌ܥྻ

Slide 19

Slide 19 text

19 ࣌ܥྻͷੑ࣭ l ૬Ճ ฏۉ l ෼ࢄ l ฏۉ͔Βͷ͹Β͖ͭ۩߹Λࣔ͢ࢦඪ l ฏۉ஋͔Βͷࠩ෼Ͱ͋ΔภࠩΛೋ৐ͷฏۉ l ڞ෼ࢄ l ڞ෼ࢄ͕େ͖͍ͱਖ਼ͷڧ͍ؔ܎Λ࣋ͭͳͲɺ ૊ͷରԠ͢Δσʔλͷؔ܎ Λࣔ͢ ఆৗੑඇఆৗੑ 𝐸(𝑦) = µ = 1 𝑛 ) 𝑦! 𝜎" = 𝑉𝐴𝑅 𝑦! = 1 𝑛 ) 𝑦! − µ " Cov(𝑥, 𝑦) = 𝜎#$ = E[ 𝑥 − µ# (𝑦 − µ$ )]

Slide 20

Slide 20 text

20 ࣌ܥྻͷੑ࣭ l ڧఆৗੑ TUSJDUTUBUJPOBSJUZ l ೚ҙͷtͱ𝑘 ೚ҙͷ࣌ؒࠩ ʹରͯ͠ɺ σʔλΛੜ੒͢Δ֬཰աఔ f(𝑦!, … , 𝑦!+,)͕ৗʹಉҰͷಉ࣌෼෍Λ࣋ͭ l ֬཰෼෍͕࣌ؒมԽ͠ͳ͍ l ྫαΠίϩͷग़໨ͷܥྻ΍ίΠϯτεͷ݁ՌͳͲ ఆৗੑඇఆৗੑ f 𝑦% , … , 𝑦& = f 𝑦' , … , 𝑦'(&

Slide 21

Slide 21 text

21 ࣌ܥྻͷੑ࣭ l ऑఆৗੑ XFBLTUBUJPOBSJUZ l ظ଴஋ͱࣗݾڞ෼ࢄ͕࣌఺ʹґଘ͠ͳ͍ l ೚ҙͷtͱ𝑘 ೚ҙͷ࣌ؒࠩ ʹରͯ͠ҎԼ͕੒Γཱͭ ࣌఺ʹґଘ͠ͳ͍ l ࣍ͷϞʔϝϯτ·Ͱ͕ෆม l Ұൠతʹɺʮఆৗੑʯͱݺ͹ΕΔ࣌ɺͪ͜ΒΛࢦ͢ࣄ͕ଟ͍ l ࣗݾ૬͕ؔLͱͱ΋ʹࢦ਺తʹݮਰ ఆৗੑඇఆৗੑ E 𝑦* = E 𝑦*(& , VAR 𝑦* = VAR 𝑦*(& Cov 𝑦* , 𝑦*(& = 𝛾& > 0

Slide 22

Slide 22 text

22 ࣌ܥྻͷੑ࣭ l ऑఆৗੑ XFBLTUBUJPOBSJUZ l ظ଴஋ͱࣗݾڞ෼ࢄ͕࣌఺ʹґଘ͠ͳ͍ l ೚ҙͷtͱ𝑘 ೚ҙͷ࣌ؒࠩ ʹରͯ͠ҎԼ͕੒Γཱͭ ࣌఺ʹґଘ͠ͳ͍ l ࣍ͷϞʔϝϯτ·Ͱ͕ෆม l Ұൠతʹɺʮఆৗੑʯͱݺ͹ΕΔ࣌ɺͪ͜ΒΛࢦ͢ࣄ͕ଟ͍ l ࣗݾ૬͕ؔLͱͱ΋ʹࢦ਺తʹݮਰ ఆৗੑඇఆৗੑ E 𝑦* = E 𝑦*(& , VAR 𝑦* = VAR 𝑦*(& Cov 𝑦* , 𝑦*(& = 𝛾& > 0 ଟ͘ͷ࣌ܥྻ౷ܭతϞσϧ͕ఆৗੑΛલఏͱ͍ͯ͠Δͨ Ίɺ෼ੳ͢Δ࣌ܥྻ͕ఆৗͰ͋Δͱخ͍͠ ఆৗੑΛલఏͱ͠ͳ͍ͱɺฏۉ΍෼ࢄ͕ҰఆͰແ͘ͳΔͨ Ίɺ෼ੳ͕ؤ݈Ͱͳ͘ͳΔ

Slide 23

Slide 23 text

23 ࣌ܥྻͷੑ࣭ l ऑఆৗ ਖ਼ن෼෍ 㱺 ڧఆৗ l ฏۉɺ෼ࢄɺڞ෼ࢄ͕ଘࡏ͢Δڧఆৗ 㱺 ऑఆৗ ˎ ڧఆৗΛຬ͔ͨ͢Βͱ͍ͬͯɺඞͣ͠΋ऑఆৗͷ৚݅Λຬͨ͢ ͱ͸ݶΒͳ͍ FH ฏۉɺ෼ࢄ͕ଘࡏ͠ͳ͍෼෍ͳͲ ίʔγʔ෼෍ͳͲͷ੄͕ॏ͘ɾ޿͍෼෍ ఆৗੑඇఆৗੑ

Slide 24

Slide 24 text

24 ࣌ܥྻͷੑ࣭ l ଟ͘ͷϞσϦϯάͰ͸ɺೖྗ࣌ܥྻ͕ʮઢܗɾఆৗɾਖ਼نੑʯͳ ͲΛԾఆͯ͠ϞσϦϯά l ϞσϦϯάΛ༰қʹ͢ΔͨΊʹೖྗܥྻΛม׵͠ɺύϥϝʔλਪ ఆΛ༰қʹ͢Δ͜ͱ͕໨త l ର਺ม׵ l ֊ࠩม׵ l ฏ׈Խ FUDʜ ఆৗੑ΁ͷม׵

Slide 25

Slide 25 text

25 ࣌ܥྻͷੑ࣭ l ର਺ม׵ l σʔλมಈͷܹ͍࣌͠ܥྻΛɺτϨϯυΛอͬͨ··খ͞ͳ஋ʹม׵͢Δ ख๏ l ਖ਼ͷΈͷσʔλΛෛͷ஋ΛͱΔΑ͏ʹม׵Մೳ ఆৗੑ΁ͷม׵ log 𝑦* = 𝑧* from: https://datascienceplus.com /time-series-analysis-in-r- part-2-time-series- transformations/

Slide 26

Slide 26 text

26 ࣌ܥྻͷੑ࣭ l #PY$PYม׵ l σʔλͷ෼෍Λਖ਼ن෼෍ʹ͚ۙͮΔͨΊͷ΂͖৐ܕม׵ l 𝜆͸ύϥϝʔλͰɺ 𝜆ʹΑͬͯ ม׵ͷܗ͕ࣜมΘΔ ఆৗੑ΁ͷม׵ 𝑧* = B $!+% , (𝜆 ≠ 0) ln 𝑦 (𝜆 = 0) from: http://www.kmdatascience.com/2017/07/box-cox-transformations-in-python.html

Slide 27

Slide 27 text

27 ࣌ܥྻͷੑ࣭ l ϩδοτม׵ l ͔Β·Ͱͷ஋ΛͱΔׂ߹֬཰ͷσʔλΛਖ਼ͱෛͷແݶେͷ஋ʹม׵͢ Δ͜ͱͰ࣌ܥྻϞσϦϯάΛ༰қʹ͢Δม׵ ఆৗੑ΁ͷม׵ 𝑧* = log $" %+$" 0,1 ↔ (−∞, ∞)

Slide 28

Slide 28 text

28 ࣌ܥྻͷੑ࣭ l ֊ ֊ࠩม׵ l τϨϯυ੒෼ΛऔΓআ͖ɺฏۉʹؔͯ͠ఆৗੑΛຬͨ͢Α͏ʹม׵ ఆৗੑ΁ͷม׵ 𝑧* = 𝑦* − 𝑦*+% from: https://datascienceplus.com/time-series-analysis-in-r-part-2-time-series-transformations/

Slide 29

Slide 29 text

29 ࣌ܥྻͷੑ࣭ l قઅ ֊ࠩม׵ l قઅ੒෼ʹΑΔӨڹपظੑΛऔΓআ͖ɺฏۉʹؔͯ͠ఆৗੑΛຬͨ͢Α ͏ʹม׵ l 𝑠 ͸पظ ఆৗੑ΁ͷม׵ 𝑧* = 𝑦* − 𝑦*+- = (1 − 𝐵-)𝑦* from: https://datascienceplus.com/time-series-analysis-in-r-part-2-time-series-transformations/

Slide 30

Slide 30 text

30 ࣌ܥྻͷੑ࣭ l ฏ׈Խ l ϊΠζ΍୹ظతͳมಈΛআ͘ޮՌΛظ଴ͯ͠ɺ׈Β͔ʹͳΔΑ͏ʹม׵ l ୅දྫͱͯ͠ɺҠಈฏۉϑΟϧλ .PWJOH"WFSBHF'JMUFS ఆৗੑ΁ͷม׵ 𝑧* = 1 M ) ./0 1+% 𝑦. from: https://machinelearningmastery.com/moving-average-smoothing-for-time-series- forecasting-python/

Slide 31

Slide 31 text

31 ࣌ܥྻͷੑ࣭ l ૬ؔ l पظੑ ࣌ܥྻͷ܏޲ͷཧղ

Slide 32

Slide 32 text

32 ࣌ܥྻͷੑ࣭ l ࣗݾڞ෼ࢄؔ਺ l ऑఆৗաఔͷࣗݾڞ෼ࢄΛԾఆ l ࣗݾڞ෼ࢄؔ਺͸ҎԼͷܗͰఆٛ͞ΕΔ l ࣗݾڞ෼ࢄؔ਺͸ۮؔ਺Cov 𝑦!, 𝑦!+, = 𝛾, = 𝛾-, = Cov 𝑦!, 𝑦!-, l ࣗݾڞ෼ࢄؔ਺ͷઈର஋ͷ࠷େ͸𝛾. 𝛾. ≥ 𝛾, l ࣗݾڞ෼ࢄؔ਺ͷਤࣔͱͯ͠ίϨϩάϥϜ͕༻͍ΒΕΔ ࣌ܥྻͷ܏޲ͷཧղ Cov 𝑦% , 𝑦& = Cov 𝑦* , 𝑦*(& = 𝛾& 𝑘͸ϥάΛද͢

Slide 33

Slide 33 text

33 ࣌ܥྻͷੑ࣭ l ࣗݾ૬ؔؔ਺ l ऑఆৗաఔͷࣗݾ૬ؔΛԾఆ l ҎԼͷܗͰఆٛ͞ΕΔ ࣌ܥྻͷ܏޲ͷཧղ Cor 𝑦* , 𝑦*(& = 234 $",$"#$ 678 $" 678($"#$) = ;$ ;% = 𝑅& 𝑘͸ϥάΛද͢

Slide 34

Slide 34 text

34 ࣌ܥྻͷੑ࣭ l ࣗݾ૬ؔؔ਺ͷίϨϩάϥϜ l ܥྻ͕ఆৗͳΒɺࣗݾ૬ؔؔ਺͕͸ز Կڃ਺తʹʹऩଋ͢Δ͜ͱΛࣔ͢ l पظੑͷ͋ΔܥྻͳͲͷൃݟ΋༰қʹ ࣌ܥྻͷ܏޲ͷཧղ from: 実証のための計量時系列分析

Slide 35

Slide 35 text

35 ࣌ܥྻͷੑ࣭ ࣌ܥྻͷ܏޲ͷཧղ from: https://elf-c.he.u-tokyo.ac.jp/courses/379

Slide 36

Slide 36 text

36 ࣌ܥྻͷੑ࣭ ࣌ܥྻͷ܏޲ͷཧղ from: https://elf-c.he.u-tokyo.ac.jp/courses/379

Slide 37

Slide 37 text

37 ࣌ܥྻͷੑ࣭ l εϖΫτϧղੳ l ࣌ܥྻσʔλΛߏ੒प೾੒෼ʹ෼ղ͠ɺ֤प೾਺ͱΤωϧΪʔͷؔ܎Λऔ Γग़ͨ͢Ίͷख๏ l Α͘༻͍ΒΕΔͷ͸ɺ཭ࢄϑʔϦΤม׵ ''5 l ෳࡶͳपظΛऔΓग़͢͜ͱΛՄೳʹ͢Δɻपظੑͱ͸𝑝͕पظΛࣔ࣌͢ʹ 𝑦! = 𝑦!-/ ͕੒Γཱͭ͜ͱ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ from: https://qiita.com/AnchorB lues/items/5497ee68c3a3 d64875d4 FFT 振動数 振幅

Slide 38

Slide 38 text

38 ࣌ܥྻͷੑ࣭ l ϑʔϦΤม׵ l प೾਺𝒇ɼपظؔ਺y(t ͱͨ͠ͱ͖ҎԼͷܗʹͳΔ l ෳ਺ͷ೾ͰܥྻΛઆ໌Ͱ͖ΔΑ͏ʹ෼ղ l ٯϑʔϦΤม׵ l ೾਺𝟐𝝅𝒇ͷਖ਼ݭ೾ʹରͯ͠ॏΈΛֻ͚ͯɺੵ෼͢Δ͜ͱͰݩͷܥྻʹ໭͢ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ 𝒀 𝒇 = 2 "# # 𝒚(𝒕)𝒆"𝟐𝝅𝒊𝒕𝒇𝒅𝒕 ∗ ∫ "# # 𝒚 𝒕 𝒅𝒕 < ∞Λຬͨ͢ 𝒚 𝒕 = 2 "# # 𝒀(𝒇)𝒆𝟐𝝅𝒊𝒕𝒇𝒅𝒇

Slide 39

Slide 39 text

39 ࣌ܥྻͷੑ࣭ l ϑʔϦΤม׵ l प೾਺𝒇ɼपظؔ਺y(t ͱͨ͠ͱ͖ҎԼͷܗʹͳΔ l ෳ਺ͷ೾ͰܥྻΛઆ໌Ͱ͖ΔΑ͏ʹ෼ղ l ٯϑʔϦΤม׵ l ೾਺𝟐𝝅𝒇ͷਖ਼ݭ೾ʹରͯ͠ॏΈΛֻ͚ͯɺੵ෼͢Δ͜ͱͰݩͷܥྻʹ໭͢ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ ∗ ∫ "# # 𝒚 𝒕 𝒅𝒕 < ∞Λຬͨ͢ ॏΈਖ਼ݭ೾ 𝒀 𝒇 = 2 "# # 𝒚(𝒕)𝒆"𝟐𝝅𝒊𝒕𝒇𝒅𝒕 𝒚 𝒕 = 2 "# # 𝒀(𝒇)𝒆𝟐𝝅𝒊𝒕𝒇𝒅𝒇

Slide 40

Slide 40 text

40 ࣌ܥྻͷੑ࣭ l ύϫʔεϖΫτϧ l ϑʔϦΤ੒෼𝒀(𝒇)͸೾Ͱ͋Δ𝒆𝒊𝒕𝟐𝝅𝒇ͷৼΕ෯Λද͢ 㱺 𝒀(𝒇) 𝟐͸ΤωϧΪʔ l ܥྻͷதʹؚ·ΕΔप೾਺ͱͦͷύϫʔΛՄࢹԽ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ 𝒚 𝒕 = 2 "# # 𝒀(𝒇)𝒆𝟐𝝅𝒊𝒕𝒇𝒅𝒇 ॏΈਖ਼ݭ೾

Slide 41

Slide 41 text

41 ࣌ܥྻͷੑ࣭ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ 𝑓(𝑡) = B 1, 0 < 𝑡 < 𝜏 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 ೖྗܥྻ 𝐹 𝜔 = 2 "# # 𝑓(𝑡)𝑒")*!𝑑𝑡 = 2 + , 𝑓(𝑡)𝑒")*!𝑑𝑡 = 1 −𝑗𝑤 𝑒")*! + , = 1 𝑗𝑤 1 − 𝑒")*, = -!"#$/& ). 𝑒)*,/0 − 𝑒")*,/0 = 𝑒")*,/0 0 * sin *, 0 ϑʔϦΤม׵ 𝐹 𝜔 = 𝑒")*,/0 2 𝜔 sin 𝜔𝜏 2 = 2 𝜔 sin 𝜔𝜏 2 ৼΕ෯εϖΫτϧ ( 𝐹 𝜔 = 𝐹 𝜔 𝑒)1(*) ) ύϫʔεϖΫτϧ 𝐹 𝜔 0 = 2 𝜔 sin 𝜔𝜏 2 0 = 4 𝜔0 sin0 𝜔𝜏 2

Slide 42

Slide 42 text

42 ࣌ܥྻͷੑ࣭ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ 𝑓(𝑡) = B 1, 0 < 𝑡 < 𝜏 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 ೖྗܥྻ 𝐹 𝜔 = 2 "# # 𝑓(𝑡)𝑒")*!𝑑𝑡 = 2 + , 𝑓(𝑡)𝑒")*!𝑑𝑡 = 1 −𝑗𝑤 𝑒")*! + , = 1 𝑗𝑤 1 − 𝑒")*, = -!"#$/& ). 𝑒)*,/0 − 𝑒")*,/0 = 𝑒")*,/0 0 * sin *, 0 ϑʔϦΤม׵ 𝐹 𝜔 = 𝑒")*,/0 2 𝜔 sin 𝜔𝜏 2 = 2 𝜔 sin 𝜔𝜏 2 ৼΕ෯εϖΫτϧ ( 𝐹 𝜔 = 𝐹 𝜔 𝑒)1(*) ) ύϫʔεϖΫτϧ 𝐹 𝜔 0 = 2 𝜔 sin 𝜔𝜏 2 0 = 4 𝜔0 sin0 𝜔𝜏 2

Slide 43

Slide 43 text

43 ࣌ܥྻͷੑ࣭ l ύϫʔεϖΫτϧ l ϑʔϦΤ੒෼𝒀(𝒇)͸೾Ͱ͋Δ𝒆𝒊𝒕𝟐𝝅𝒇ͷৼΕ෯Λද͢ 㱺 𝒀(𝒇) 𝟐͸ΤωϧΪʔ l ܥྻͷதʹؚ·ΕΔप೾਺ͱͦͷύϫʔΛՄࢹԽ l ΢Οʔφʔɾώϯνϯ 8JFOFS,IJOUDIOF ͷެࣜࣗݾڞ෼ࢄؔ਺𝛾, ͷϑʔ ϦΤม׵͸ύϫʔεϖΫτϧʹͳΔ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ 𝒚 𝒕 = 2 "# # 𝒀(𝒇)𝒆𝟐𝝅𝒊𝒕𝒇𝒅𝒇 ॏΈਖ਼ݭ೾ Cov 𝑦", 𝑦, = 𝐶ov 𝑦!, 𝑦!+, = 𝛾, 𝑘͸ϥάΛද͢ 𝑌 𝑓 = ? !"#$ $ 𝛾!𝑒#%&'!( − 1 2 ≤ 𝑓 ≤ 1 2

Slide 44

Slide 44 text

44 ࣌ܥྻͷੑ࣭ l ύϫʔεϖΫτϧ l ۮؔ਺ͱίαΠϯม׵ͷੑ࣭͔Β l ྫ୯ৼಈ ࣌ܥྻͷ܏޲ͷཧղεϖΫτϧղੳ 𝑌 𝑓 = ? !"#$ $ 𝛾!𝑒#%&'!( = 𝛾) + 2 ? !"* $ 𝛾!cos (2𝜋𝑘𝑓) 𝛾! = 𝑎% 2 cos(2𝜋𝑓)𝑘) 𝑌 𝑓 = ? !"#$ $ cos(2𝜋𝑓)𝑘)cos(2𝜋𝑓𝑘) = 1 2 ? !"#$ $ cos 2𝜋𝑘 𝑓) − 𝑓 − cos 2𝜋𝑘 𝑓) + 𝑓 = M * %+, 𝑓 = 𝑓) 0, 𝑓 ≠ 𝑓)

Slide 45

Slide 45 text

45 ࣌ܥྻϞσϧ

Slide 46

Slide 46 text

46 ࣌ܥྻϞσϧ ࣌ܥྻϞσϧͷߏங ⼊⼒特徴 𝑦!-W … , 𝑦!-" ͳͲ ホワイトノイズ 𝜀! 時系列モデル 時系列 𝑦!

Slide 47

Slide 47 text

47 ࣌ܥྻϞσϧ l ࣌ؒతʹ૬͕ؔແ͍࣌ܥྻ𝜀! l ࣗݾ૬ؔؔ਺΋ l ֬཰తͳ͹Β͖ͭΛදݱ l FH ಠཱʹੜ੒ͨ͠ਖ਼نཚ਺ྻͳͲ ന৭ࡶԻ ϗϫΠτϊΠζ 𝐸 𝜀* = 0, 𝑉𝐴𝑅 𝜀* = 𝜎", 𝑅. = 0 𝑗͸ϥάΛද͢ from: https://machinelearningmastery.com/white-noise-time-series-python/

Slide 48

Slide 48 text

48 ࣌ܥྻϞσϧ l 𝑝࣍ͷઢܗࠩ෼ํఔࣜʹϗϫΠτϊΠζΛՃ͑ͨϞσϧ"3 𝒑 l 𝑝࣍"3Ϟσϧͱݺ͹ΕΔ l աڈͷܥྻ 𝑦!-" , … 𝑦!-/ ͷೖྗͱϗϫΠτϊΠζͰݱࡏͷܥྻΛճؼ͢Δ Ϟσϧ ࣗݾճؼաఔ "3Ϟσϧ 𝑦* = ) !/% P 𝑎! 𝑦*+! + 𝜀*

Slide 49

Slide 49 text

49 ࣌ܥྻϞσϧ l ௚લͷ𝒒 ݸͷϗϫΠτϊΠζͷՃॏ࿨Ͱఆٛ͞ΕΔ𝒒 ࣍."Ϟσϧ." 𝒒 l աڈͷܥྻ 𝑦!-", … 𝑦!-/ ʹ༩͑ΒΕΔϗϫΠτϊΠζ ϥϯμϜγϣοΫ ͕࣌ܥྻͷকདྷͷ஋ʹ௚઀఻ൖ͢Δ͜ͱΛϞσϦϯά l ϗϫΠτϊΠζͱಉ༷ʹࣗݾڞ෼ࢄ΍ظ଴஋͸࣌఺ʹґଘͤͣɺऑఆৗੑ Λຬͨ͢ Ҡಈฏۉաఔ ."Ϟσϧ 𝑦* = 𝜇 + ) !/0 Q 𝑏! 𝜀*+!

Slide 50

Slide 50 text

50 ࣌ܥྻϞσϧ l ௚લͷ𝒑 ݸͷܥྻͱ௚લͷ𝒒 ݸͷϗϫΠτϊΠζͷՃॏ࿨ʹΑͬͯߏ੒͞ ΕΔϞσϧ"3." 𝒑, 𝒒) l "3Ϟσϧ΋."Ϟσϧ΋"3."Ϟσϧͷಛघͳྫ l "3Ϟσϧ𝒒 = 0 l ."Ϟσϧ 𝒑 = 0 l ϗϫΠτϊΠζ𝒑 = 0, 𝒒 = 0 l ϗϫΠτϊΠζ𝜀! ͸աڈͷܥྻͱແ૬ؔ𝔼 𝜀! , 𝑦!-X = 0, j > 0 ࣗݾճؼҠಈฏۉաఔ "3."Ϟσϧ 𝑦* = ) !/% P 𝑎! 𝑦*+! + ) !/% Q 𝑏! 𝜀*+! + 𝜀*

Slide 51

Slide 51 text

51 ࣌ܥྻϞσϧ l -BHPQFSBUPSΛ༻͍Δ͜ͱͰ"3." 𝒑, 𝒒)ϞσϧΛҎԼͷܗʹॻ͖׵͑Δ ͜ͱ͕Ͱ͖Δ ࣗݾճؼҠಈฏۉաఔ "3."Ϟσϧ 𝑦* = ) !/% P 𝑎! 𝑦*+! + ) !/% Q 𝑏! 𝜀*+! + 𝜀* 𝐿-BHPQFSBUPS 𝐿% 𝑦* = 𝑦*+% , 𝐿& 𝑦* = 𝑦*+& , (1 − ) !/% P 𝑎! 𝐿!)𝑦* = (1 − ) !/% Q 𝑏! 𝐿!)𝜀* = 𝑎(𝐿) = 𝑏(𝐿) "30QFSBUPS ."0QFSBUPS 𝑎(𝐿)𝑦* = 𝑏(𝐿)𝜀*

Slide 52

Slide 52 text

52 ࣌ܥྻϞσϧ ࣗݾճؼҠಈฏۉաఔ "3."Ϟσϧ (1 − ) !/% P 𝑎! 𝐿!)𝑦* = (1 − ) !/% Q 𝑏! 𝐿!)𝜀* = 𝑎(𝐿) = 𝑏(𝐿) 𝑎(𝐿)𝑦* = 𝑏(𝐿)𝜀* 𝑦* = 𝑎(𝐿)+%𝑏(𝐿)𝜀* 𝑦* = ) ./0 T 𝑔. 𝐿.𝜀* = ) ./0 T 𝑔. 𝜀*+.

Slide 53

Slide 53 text

53 ࣌ܥྻϞσϧ ࣗݾճؼҠಈฏۉաఔ "3."Ϟσϧ (1 − ) !/% P 𝑎! 𝐿!)𝑦* = (1 − ) !/% Q 𝑏! 𝐿!)𝜀* = 𝑎(𝐿) = 𝑏(𝐿) 𝑎(𝐿)𝑦* = 𝑏(𝐿)𝜀* 𝑦* = 𝑎(𝐿)+%𝑏(𝐿)𝜀* 𝑦* = ) ./0 T 𝑔. 𝐿.𝜀* = ) ./0 T 𝑔. 𝜀*+. "3."Ϟσϧ͸ແݶ࣍ͷ."Ϟσϧ Ͱදݱ͕Մೳ Πϯύϧε Ԡ౴ؔ਺

Slide 54

Slide 54 text

54 ࣌ܥྻϞσϧ ΠϯύϧεԠ౴ؔ਺ *3' 𝑦* = ) ./0 T 𝑔. 𝜀*+. l ༩͑ΒΕͨϊΠζ Πϯύϧε ͕𝑗ظޙʹ࣌ܥྻʹͲΕ͘Β͍ӨڹΛ༩͑ Δ͔ʢӨڹྗ͕ͲΕ͚ͩ࢒Δ͔ʣΛࣔͨؔ͠਺ 𝑔0 = 1, 𝑔! = ) ./% P 𝑎. 𝑔!+. − 𝑏! *3'ͷࢉग़ from: https://alexchinco.com/impulse-response-functions-for-vars/

Slide 55

Slide 55 text

55 ࣌ܥྻϞσϧ l ࣗݾ૬ؔؔ਺ BVUPDPSSFMBUJPOGVODUJPO"$' l ภࣗݾ૬ؔؔ਺ BVUPDPSSFMBUJPOGVODUJPO1"$' l ϥά𝑘ͷࣗݾ૬ؔΛߟྀ͢Δࡍʹɺϥά1, … 𝑘 − 1·Ͱͷܥྻͷࣗݾ૬ؔͷ Өڹ΋ଘࡏ͢Δ l ͜Ε·ͰͷϥάͷӨڹΛআڈͨ͠૬ؔؔ਺ ࣗݾ૬ؔؔ਺ "$' ͱภࣗݾ૬ؔؔ਺ 1"$' Cor 𝑦* , 𝑦*(& = 234 $",$"#$ 678 $" 678($"#$) = ;$ ;% = 𝑅& 𝑘͸ϥάΛද͢ PACF 𝑦* , 𝑦*(& = Cov 𝑦*(& 𝑦*(&+% , … , 𝑦*(% , 𝑦* 𝑦*(&+% , … , 𝑦*(% 𝑉𝑎𝑟 𝑦* 𝑦*(&+% , … , 𝑦*(% 𝑉𝑎𝑟 𝑦*(& 𝑦*(&+% , … , 𝑦*(%

Slide 56

Slide 56 text

56 ࣌ܥྻϞσϧ l ࣗݾ૬ؔؔ਺ BVUPDPSSFMBUJPOGVODUJPO"$' l ภࣗݾ૬ؔؔ਺ BVUPDPSSFMBUJPOGVODUJPO1"$' l ϥά𝑘ͷࣗݾ૬ؔΛߟྀ͢Δࡍʹɺϥά1, … 𝑘 − 1·Ͱͷܥྻͷࣗݾ૬ؔͷ Өڹ΋ଘࡏ͢Δ l ͜Ε·ͰͷϥάͷӨڹΛআڈͨ͠૬ؔؔ਺ ࣗݾ૬ؔؔ਺ "$' ͱภࣗݾ૬ؔؔ਺ 1"$' Cor 𝑦* , 𝑦*(& = 234 $",$"#$ 678 $" 678($"#$) = ;$ ;% = 𝑅& 𝑘͸ϥάΛද͢ PACF 𝑦* , 𝑦*(& = Cov 𝑦*(& 𝑦*(&+% , … , 𝑦*(% , 𝑦* 𝑦*(&+% , … , 𝑦*(% 𝑉𝑎𝑟 𝑦* 𝑦*(&+% , … , 𝑦*(% 𝑉𝑎𝑟 𝑦*(& 𝑦*(&+% , … , 𝑦*(%

Slide 57

Slide 57 text

57 ࣌ܥྻϞσϧ l "3Ϟσϧͷ"$'ͱ1"$' ࣗݾ૬ؔؔ਺ "$' ͱภࣗݾ૬ؔؔ਺ 1"$'

Slide 58

Slide 58 text

58 ࣌ܥྻϞσϧ l ."Ϟσϧͷ"$'ͱ1"$' ࣗݾ૬ؔؔ਺ "$' ͱภࣗݾ૬ؔؔ਺ 1"$'

Slide 59

Slide 59 text

59 ࣌ܥྻϞσϧ l "3."Ϟσϧͷ"$'ͱ1"$' ࣗݾ૬ؔؔ਺ "$' ͱภࣗݾ૬ؔؔ਺ 1"$'

Slide 60

Slide 60 text

60 ࣌ܥྻϞσϧ l "3Ϟσϧͷࣗݾ૬ؔͷ௕͞͸ແݶ͕ͩɺภࣗݾ૬ؔͷ௕͞͸"3 ͷ܎਺ l ."Ϟσϧͷภࣗݾ૬ؔͷ௕͞͸ແݶ͕ͩɺࣗݾ૬ؔͷ௕͞͸." ͷ܎਺ l "3."Ϟσϧ͸ࣗݾ૬ؔɺภࣗݾ૬ؔͱ΋ʹݮਰ͍͕ͯ͘͠Ұఆ ʹଘࡏ͢Δ ࣗݾ૬ؔؔ਺ "$' ͱภࣗݾ૬ؔؔ਺ 1"$'

Slide 61

Slide 61 text

61 ࣌ܥྻϞσϧ ύϫʔεϖΫτϧͱ"3."Ϟσϧ from: https://pyspectrum.readthedocs.io/en/latest/ref_param.html

Slide 62

Slide 62 text

62 ࣌ܥྻϞσϧ l ฏۉʹؔͯ͠ඇఆৗͳ࣌ܥྻΛL֊ࠩ෼Λߦ͍ɺ"3."ϞσϧΛϑΟο ςΟϯάͨ͠Ϟσϧ"3*." 𝒑, 𝒌, 𝒒) l 𝒌 = 𝟎ͷ࣌"3."Ϟσϧ ࣗݾճؼ࿨෼Ҡಈฏۉաఔ "3*."Ϟσϧ 𝑧* = ) !/% P 𝑎! 𝑧*+! + ) !/% Q 𝑏! 𝜀*+! + 𝜀* ∆𝑦* = 𝑦* − 𝑦*+% 𝑧* = ∆&𝑦* L֊ࠩ෼ΛऔΔ͜ͱͰ ఆৗͳ࣌ܥྻʹม׵

Slide 63

Slide 63 text

63 ࣌ܥྻϞσϧ l قઅੑ֊ࠩΛߟྀͨ͠"3*."Ϟσϧͱɺقઅੑ֊ࠩΛߟྀ͠ͳ͍௨ৗͷ "3*."Ϟσϧͷͭͷ੒෼Λ૊Έ߹Θͨ͠Ϟσϧ 4"3*.""3*." 𝒑, 𝒌, 𝒒) º 𝑷, 𝑲, 𝑸)𝑺 4FBTPOBM"3*."Ϟσϧ 4"3*." 𝑧* U = ) !/% P 𝑎! U𝑧*+! U + ) !/% Q 𝑏! U𝜀*+! + 𝜀* 𝑧* U = 𝑦* − 𝑦*+U 𝑆͸Ұपظͷ௕͞

Slide 64

Slide 64 text

64 ࣌ܥྻϞσϧ l ྫ𝑆 = 12ͷ4"3*."Ϟσϧ 4"3*.""3*." 𝟏, 𝟎, 𝟎) º 𝟏, 𝟎, 𝟎)𝟏𝟐 4FBTPOBM"3*."Ϟσϧ 4"3*." 1 − 𝑎% 𝐿% 1 − 𝑎% U𝐿%" 𝑦* = 𝜀* 𝑦* = 𝑎% 𝑦*+% + 𝑎% U𝑦*+%" − 𝑎% 𝑎% U𝑦*+%V + 𝜀* 𝐿-BHPQFSBUPS 𝐿% 𝑦* = 𝑦*+% , 𝐿& 𝑦* = 𝑦*+& , قઅੑ ඇقઅੑ

Slide 65

Slide 65 text

65 ࣌ܥྻϞσϧ 4FBTPOBM"3*."Ϟσϧ 4"3*." Kumar Barik, Aditya, et al. "Analysis of GHI Forecasting Using Seasonal ARIMA." Data Management, Analytics and Innovation. Springer, Singapore, 2021. 55-69.

Slide 66

Slide 66 text

66 ࣌ܥྻϞσϧ l ࣌ܥྻΛτϨϯυɺقઅੑͳͲͷෳ਺ͷߏ੒ཁૉʹ෼ղ͢Δख๏ l "EEJUJWFEFDPNQPTJUJPO l .VMUJQMJDBUJWFEFDPNQPTJUJPO l ୅දతͳख๏ l X11 decomposition l STL decopostion l SEATS decompositon ࣌ܥྻ෼ղ 5JNFTFSJFTEFDPNQPTJUJPO 𝑦! = 𝑆! ×𝑇! ×𝑅! , log 𝑦! = log 𝑆! + log 𝑇! + log 𝑅! 𝑦! = 𝑆! + 𝑇! + 𝑅!

Slide 67

Slide 67 text

67 ࣌ܥྻϞσϧ ࣌ܥྻ෼ղ 5JNFTFSJFTEFDPNQPTJUJPO from: https://otexts.com/fpp2/components.htmlå

Slide 68

Slide 68 text

68 ࣌ܥྻϞσϧ l ࣌఺𝑡ʹ͓͚Δ𝑛ݸͷมྔ͔ΒͳΔܥྻσʔλΛϕΫτϧͱͯ͠ଊ͑Δ l ϗϫΠτϊΠζ΋ಉ༷ʹϕΫτϧͱͯ͠ଊ͑Δ l 7"3 Q Ϟσϧ͸ҎԼͷࣜͱͳΔ ϕΫτϧࣗݾճؼϞσϧ 7"3Ϟσϧ 𝑌* = (𝑦%,* , 𝑦",* , … , 𝑦',* )W 𝔼 𝜺𝒕 = [0, … , 0]W, 𝔼 𝜺𝒕 𝜺𝒕+𝒌 = [ 𝚺 (𝑘 = 0) 𝟎, (𝑘 ≠ 0) 𝑌* = ) !/% P 𝑨𝒊 𝑌*+! + 𝜺𝒕 𝑨𝒊 = 𝑎! (1,1) ⋯ 𝑎! (1, 𝑛) ⋮ ⋱ ⋮ 𝑎! (𝑛, 1) ⋯ 𝑎! (𝑛, 𝑛)

Slide 69

Slide 69 text

69 ࣌ܥྻϞσϧ άϨϯδϟʔҼՌ (SBOHFS$BVTBMJUZ l ࣌ܥྻؒಉ࢜ͷҼՌؔ܎Λਪఆ͢ΔͨΊͷํ๏Ͱɺܥྻಉ࢜ͷ޲ ͖Λൃݟ͢Δख๏ l 𝑦%,* ͷ༧ଌʹ͓͍ͯ 𝑦%,* ͷաڈͷ஋ʹج͍ͮͨ༧ଌΑΓ΋ɺ 𝑦%,* , 𝑦" * ͷաڈͷ஋Λ༻͍ͨ༧ଌʹΑͬͯ.4&͕খ͘͞ͳΔ৔߹ 𝒚𝟐 𝒕 ͔Β 𝒚𝟏,𝒕 ΁ͷάϨϯδϟʔҼՌੑ͕ଘࡏ͢Δͱ͢Δ 𝑦% 𝑦"

Slide 70

Slide 70 text

70 ࣌ܥྻϞσϧ άϨϯδϟʔҼՌ (SBOHFS$BVTBMJUZ l άϨϯδϟʔҼՌͷൃݟʹ͸'౷ܭྔΛ༻͍Δͷ͕Ұൠత 𝑦% ͷϞσϧʹΑΔ࢒ࠩฏํ࿨𝑆𝑆𝑅0 𝑦% , 𝑦" ͷϞσϧʹΑΔ࢒ࠩฏํ࿨𝑆𝑆𝑅% '౷ܭྔͷࢉग़𝐹 ≡ (UU]% +UU]&)/8 UU]&/(W +'P +%) '౷ܭྔͷ஋ʹج͍ͮͯάϨϯδϟʔҼՌ ͕ଘࡏ͢Δ͔Ͳ͏͔Λݕఆ

Slide 71

Slide 71 text

71 ࣌ܥྻϞσϧ l ඇఆৗ࣌ܥྻΛϑΟοςΟϯά͢ΔͨΊͷϞσϦϯάख๏ͷ̍ͭ ہॴఆৗ"3Ϟσϧ ϨδʔϜ෼ׂ Matsubara, Yasuko, Yasushi Sakurai, and Christos Faloutsos. "Autoplait: Automatic mining of co-evolving time sequences." Proceedings of the 2014 ACM SIGMOD international conference on Management of data. 2014.

Slide 72

Slide 72 text

72 ࣌ܥྻϞσϧ l ֤۠෼͸ఆৗͱԾఆͯ͠ɺ֤۠෼͝ͱ ʹ"3ϞσϧͳͲͷ࣌ܥྻϞσϧΛద༻ l ۠෼෼͚ͷมԽ఺͸"*$΍.%-ͳͲΛ ༻͍Δ͜ͱͰൃݟ ہॴఆৗ"3Ϟσϧ ϨδʔϜ෼ׂ

Slide 73

Slide 73 text

73 ࣌ܥྻϞσϧ l ࣮ࡍͷঢ়ଶΛද͢ม਺ͱ࣮ࡍʹ؍ଌͰ͖Δม਺͕ҟͳΔΑ͏ͳܥΛ਺ࣜͰ දݱͨ͠ϑϨʔϜϫʔΫ l "3ϞσϧͳͲͷઢܗ࣌ܥྻϞσϧ΋औΓѻ͑ΔΑ͏ͳҰൠతͳϑϨʔϜ ϫʔΫ l ඇఆৗϞσϧ΍ϕΠζϞσϧΛදݱ͢Δ͜ͱ͕Մೳ ঢ়ଶۭؒϞσϧ

Slide 74

Slide 74 text

74 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑥! = 𝐹! 𝑥!$# + 𝐺! 𝑣!$# 𝑦! = 𝐻! 𝑥! + 𝑤! γεςϜϞσϧ ؍ଌϞσϧ 𝑦* 𝑛࣍ݩ࣌ܥྻ 𝑥* 𝑚࣍ݩঢ়ଶϕΫτϧ 𝑣* 𝑘࣍ݩγεςϜϊΠζ 𝑤* 𝑛࣍ݩ؍ଌϊΠζ 𝑣* ~ 𝑁 0, 𝑄* , 𝑤* ~ 𝑁 0, 𝑅*

Slide 75

Slide 75 text

75 ࣌ܥྻϞσϧ "3ϞσϧΛঢ়ଶۭؒϞσϧͰදݱ 𝑥! = 𝐹! 𝑥!$# + 𝐺! 𝑣!$# 𝑦! = 𝐻! 𝑥! + 𝑤! γεςϜϞσϧ ؍ଌϞσϧ 𝑦* = ) !/% P 𝑎! 𝑦*+! + 𝜀* 𝜀* ~ N(0, 𝜎") "3Ϟσϧ 𝑦! 𝑦!-" ⋮ 𝑦!-/+" = 𝑎4 𝑎0 1 ⋯ 𝑎5 ⋮ ⋱ ⋮ ⋯ 1 𝑦!-" 𝑦!-Z ⋮ 𝑦!-/ + 1 0 ⋮ 0 𝜀! 𝑦* = 1 0 … 0 𝑦* 𝑦*+% ⋮ 𝑦*+P(% γεςϜϞσϧ ؍ଌϞσϧ

Slide 76

Slide 76 text

76 ࣌ܥྻϞσϧ "3ϞσϧΛঢ়ଶۭؒϞσϧͰදݱ 𝑥! = 𝐹! 𝑥!$# + 𝐺! 𝑣!$# 𝑦! = 𝐻! 𝑥! + 𝑤! γεςϜϞσϧ ؍ଌϞσϧ 𝑦* = ) !/% P 𝑎! 𝑦*+! + 𝜀* 𝜀* ~ N(0, 𝜎") "3Ϟσϧ 𝑦! 𝑦!-" ⋮ 𝑦!-/+" = 𝑎4 𝑎0 1 ⋯ 𝑎5 ⋮ ⋱ ⋮ ⋯ 1 𝑦!-" 𝑦!-Z ⋮ 𝑦!-/ + 1 0 ⋮ 0 𝜀! 𝑦* = 1 0 … 0 𝑦* 𝑦*+% ⋮ 𝑦*+P(% 𝐹! 𝐺! 𝐻! 𝑥!

Slide 77

Slide 77 text

77 ࣌ܥྻϞσϧ "3ϞσϧΛঢ়ଶۭؒϞσϧͰදݱ 𝑥! = 𝐹! 𝑥!$# + 𝐺! 𝑣!$# 𝑦! = 𝐻! 𝑥! + 𝑤! γεςϜϞσϧ ؍ଌϞσϧ 𝑦* = ) !/% P 𝑎! 𝑦*+! + 𝜀* 𝜀* ~ N(0, 𝜎") "3Ϟσϧ 𝑦! 𝑦!-" ⋮ 𝑦!-/+" = 𝑎4 𝑎0 1 ⋯ 𝑎5 ⋮ ⋱ ⋮ ⋯ 1 𝑦!-" 𝑦!-Z ⋮ 𝑦!-/ + 1 0 ⋮ 0 𝜀! 𝑦* = 1 0 … 0 𝑦* 𝑦*+% ⋮ 𝑦*+P(% 𝐹! 𝐺! 𝐻! 𝑥! "3Ϟσϧ͸࣌ෆมͰ؍ଌϊΠζ 𝑤! = 0ͷ ঢ়ଶۭؒϞσϧͰදݱՄೳ

Slide 78

Slide 78 text

78 ࣌ܥྻϞσϧ "3."ϞσϧΛঢ়ଶۭؒϞσϧͰදݱ 𝜀!~ N(0, 𝜎Z) "3."Ϟσϧ 𝑦! _ 𝑦!+"|!-" ⋮ _ 𝑦!+,-"|!-" = 𝑎" 1 𝑎Z ⋯ ⋮ ⋱ 1 𝑎, ⋯ 𝑦!-" _ 𝑦!|!-Z ⋮ _ 𝑦!+,-Z|!-Z + 1 𝑏" ⋮ 𝑏,-" 𝜀! 𝑦* = 1 0 … 0 𝑦* m 𝑦*(%|*+% ⋮ m 𝑦*(&+%|*+% 𝐹! 𝐺! 𝐻! 𝑥! 𝑦! = b \]" / 𝑎\𝑦!-\ + b \]" ^ 𝑏\𝜀!-\ + 𝜀! _ 𝑦!+X|!-" = b \]X+" / 𝑎\ 𝑦!+X-\ + b \]X ^ 𝑏\ 𝜀!+X-\ 𝑘 = max(𝑝, 𝑞 + 1) 𝑥!-"

Slide 79

Slide 79 text

79 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧͷਪఆ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑦% 𝑥% 𝑣% ʜ σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝(𝑥! |𝑦& )

Slide 80

Slide 80 text

80 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧͷਪఆ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑦% 𝑥% 𝑣% ʜ σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝(𝑥! |𝑦& ) 𝑗 < 𝑡ͷ৔߹ ༧ଌ

Slide 81

Slide 81 text

81 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧͷਪఆ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑦% 𝑥% 𝑣% ʜ σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝(𝑥! |𝑦& ) 𝑗 = 𝑡ͷ৔߹ ϑΟϧλ

Slide 82

Slide 82 text

82 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧͷਪఆ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑦% 𝑥% 𝑣% ʜ σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝(𝑥! |𝑦& ) 𝑗 > 𝑡ͷ৔߹ ฏ׈Խ

Slide 83

Slide 83 text

83 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧͷਪఆ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑦% 𝑥% 𝑣% ʜ σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝(𝑥! |𝑦& ) ࣌ܥྻϞσϧͷ໬౓ܭࢉ΍௕ظ༧ଌɺܽଛ஋ͷิ׬ ͳͲͷλεΫ͕͜ͷϑϨʔϜϫʔΫͰ࣮ݱͰ͖Δ

Slide 84

Slide 84 text

84 ࣌ܥྻϞσϧ ઢܗɾΨ΢εܕঢ়ଶۭؒϞσϧ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& )

Slide 85

Slide 85 text

85 ࣌ܥྻϞσϧ ઢܗɾΨ΢εܕঢ়ଶۭؒϞσϧ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) ΧϧϚϯϑΟϧλΛར༻

Slide 86

Slide 86 text

86 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ༧ଌ 𝑥!|!$# = 𝐹! 𝑥!$#|!$# 𝑉!|!$# = 𝐹! 𝑉!$#|!$# 𝐹! ( + 𝐺! 𝑄! 𝐺! ( 𝑥!$# 𝐻!

Slide 87

Slide 87 text

87 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ༧ଌ 𝑥!|!$# = 𝐹! 𝑥!$#|!$# 𝑉!|!$# = 𝐹! 𝑉!$#|!$# 𝐹! ( + 𝐺! 𝑄! 𝐺! ( 𝑥!$# 𝐻! ͭલͷঢ়ଶ 𝑡 − 1 ͔Βݱࡏͷঢ়ଶ 𝑡 Λਪఆ 𝑥!|!$# = 𝔼 𝑥! 𝑦!$# ] = 𝔼 𝐹𝑥!$# + 𝐺! 𝑣! 𝑦!$# ] = 𝐹𝔼 𝑥!$# 𝑦!$# ] = 𝐹𝑥!$#|!$#

Slide 88

Slide 88 text

88 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ༧ଌ 𝑥!|!$# = 𝐹! 𝑥!$#|!$# 𝑉!|!$# = 𝐹! 𝑉!$#|!$# 𝐹! ( + 𝐺! 𝑄! 𝐺! ( 𝑥!$# 𝐻! ͭલͷঢ়ଶޡࠩ ͔Βݱࡏͷঢ়ଶޡࠩΛਪఆ 𝑉!|!$# = 𝔼[ 𝑥!$ 𝑥!|!$# ' (𝑥!$ 𝑥!|!$# )] = 𝔼[(𝐹 𝑥!$#$ 𝑥!$#|!$# + 𝐺𝑣! )' (𝐹 𝑥!$#$ 𝑥!$#|!$# + 𝐺𝑣! )] = 𝐹𝔼 𝑥!$#$ 𝑥!$#|!$# ' 𝑥!$#$ 𝑥!$#|!$# 𝐹' + 𝐺𝔼 𝑣! '𝑣! 𝐺

Slide 89

Slide 89 text

89 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ϑΟϧλ 𝑥!$# K! = 𝑉!|!$# 𝐻! ((𝐻! 𝑉!|!$# 𝐻! ( + 𝑅! )$# 𝑉!|! = (𝐼 − 𝐾! 𝐻! )𝑉!|!$# 𝐻! 𝑥!|! = 𝑥!|!$# + 𝐾! (𝑦! − 𝐻! 𝑥!|!$# )

Slide 90

Slide 90 text

90 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ϑΟϧλ 𝑥!$# K! = 𝑉!|!$# 𝐻! ((𝐻! 𝑉!|!$# 𝐻! ( + 𝑅! )$# 𝑉!|! = (𝐼 − 𝐾! 𝐻! )𝑉!|!$# 𝐻! 𝑥!|! = 𝑥!|!$# + 𝐾! (𝑦! − 𝐻! 𝑥!|!$# ) 𝐾* ΧϧϚϯήΠϯͱݺ͹ΕΔ΋ͷΛࢉग़ 𝑤* 𝑛࣍ݩ؍ଌϊΠζ 𝑤* ~ 𝑁 0, 𝑅*

Slide 91

Slide 91 text

91 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ϑΟϧλ 𝑥!$# K! = 𝑉!|!$# 𝐻! ((𝐻! 𝑉!|!$# 𝐻! ( + 𝑅! )$# 𝑉!|! = (𝐼 − 𝐾! 𝐻! )𝑉!|!$# 𝐻! 𝑥!|! = 𝑥!|!$# + 𝐾! (𝑦! − 𝐻! 𝑥!|!$# ) ਪఆޡࠩͱ؍ଌϊΠζ 𝑹𝒕 ͷ࿨ʹ͋ͨΔ෦෼ ΧϧϚϯήΠϯ͸શମͷޡࠩʹ͓͚Δਪఆޡࠩͷׂ ߹ʹ૬౰

Slide 92

Slide 92 text

92 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ϑΟϧλ 𝑥!$# K! = 𝑉!|!$# 𝐻! ((𝐻! 𝑉!|!$# 𝐻! ( + 𝑅! )$# 𝑉!|! = (𝐼 − 𝐾! 𝐻! )𝑉!|!$# 𝐻! 𝑥!|! = 𝑥!|!$# + 𝐾! (𝑦! − 𝐻! 𝑥!|!$# ) ؍ଌ஋ͱࣄલʹਪఆ͞Εͨঢ়ଶͱͷࠩ

Slide 93

Slide 93 text

93 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ϑΟϧλ 𝑥!$# K! = 𝑉!|!$# 𝐻! ((𝐻! 𝑉!|!$# 𝐻! ( + 𝑅! )$# 𝑉!|! = (𝐼 − 𝐾! 𝐻! )𝑉!|!$# 𝐻! 𝑥!|! = 𝑥!|!$# + 𝐾! (𝑦! − 𝐻! 𝑥!|!$# ) ΧϧϚϯήΠϯ͸ॏΈͷ໾ׂͰ؍ଌޡࠩͱਪఆޡࠩ ͷόϥϯεΛऔΓঢ়ଶ 𝑥* Λߋ৽

Slide 94

Slide 94 text

94 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ϑΟϧλ 𝑥!$# K! = 𝑉!|!$# 𝐻! ((𝐻! 𝑉!|!$# 𝐻! ( + 𝑅! )$# 𝑉!|! = (𝐼 − 𝐾! 𝐻! )𝑉!|!$# 𝐻! 𝑥!|! = 𝑥!|!$# + 𝐾! (𝑦! − 𝐻! 𝑥!|!$# ) ؍ଌޡ͕ࠩେ͖͍ͱڞ෼ࢄ͕େ͖͘ͳΔΑ͏ʹߋ৽

Slide 95

Slide 95 text

95 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλͷஞ࣍ਪఆ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 96

Slide 96 text

96 ࣌ܥྻϞσϧ ઢܗɾΨ΢εܕঢ়ଶۭؒϞσϧ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) ΧϧϚϯϑΟϧλΛར༻

Slide 97

Slide 97 text

97 ࣌ܥྻϞσϧ ௕ظ༧ଌΛ͍ͨ͠৔߹ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) l ௕ظ༧ଌ 𝑥!")|!$# = 𝐹! 𝑥!$#")|!$# 𝑉!")|!$# = 𝐹! 𝑉!")|!")$# 𝐹! ( + 𝐺! 𝑄! 𝐺! ( 𝑥!$# 𝐻! ϑΟϧλͷॲཧΛলུ͢Δ͜ͱͰ࣮ݱ

Slide 98

Slide 98 text

98 ࣌ܥྻϞσϧ ௕ظ༧ଌΛ͍ͨ͠৔߹ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 99

Slide 99 text

99 ࣌ܥྻϞσϧ ঢ়ଶۭؒϞσϧͷਪఆ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑦% 𝑥% 𝑣% ʜ σʔλʹج͍ͮͯঢ়ଶ 𝑥* Λਪఆ͍ͨ͠ 𝑝(𝑥! |𝑦& ) 𝑗 > 𝑡ͷ৔߹ ฏ׈Խ

Slide 100

Slide 100 text

100 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ ฏ׈Խ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) 𝑥!$# 𝐻! l ฏ׈Խ 𝑥!|% = 𝑥!|! + 𝐴! (𝑥!"#|( − 𝑥!"#|! ) 𝑉!|( = 𝑉!|! + 𝐴! (𝑉!"#|( − 𝑉!"#|! )𝐴! ( 𝐴! = 𝑉!|! 𝐹!"# ( 𝑉!"#|! $#

Slide 101

Slide 101 text

101 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ ฏ׈Խ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) 𝑥!$# 𝐻! l ฏ׈Խ 3BVDI5VOH4USJFCFM TNPPUIFS 𝑥!|% = 𝑥!|! + 𝐴! (𝑥!"#|( − 𝑥!"#|! ) 𝑉!|( = 𝑉!|! + 𝐴! (𝑉!"#|( − 𝑉!"#|! )𝐴! ( 𝐴! = 𝑉!|! 𝐹!"# ( 𝑉!"#|! $# 𝐴* ฏۉԽརಘ

Slide 102

Slide 102 text

102 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ ฏ׈Խ 𝑦! 𝑥! 𝑣! 𝑤! 𝐹! 𝐺! 𝑝 𝑥! 𝑦& ~ N(𝑥!|& , 𝑉!|& ) 𝑥!$# 𝐻! l ฏ׈Խ 3BVDI5VOH4USJFCFM TNPPUIFS 𝑥!|% = 𝑥!|! + 𝐴! (𝑥!"#|( − 𝑥!"#|! ) 𝑉!|( = 𝑉!|! + 𝐴! (𝑉!"#|( − 𝑉!"#|! )𝐴! ( 𝐴! = 𝑉!|! 𝐹!"# ( 𝑉!"#|! $# ฏۉԽརಘ͕ঢ়ଶߋ৽ʹ͓͚ΔॏΈͷ໾ׂ

Slide 103

Slide 103 text

103 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ ฏ׈Խ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 104

Slide 104 text

104 ࣌ܥྻϞσϧ ඇઢܗɾඇΨ΢εܕঢ়ଶۭؒϞσϧ l ઢܗɾΨ΢εܕঢ়ଶۭؒϞσϧ Ͱ͸ରॲͰ͖ͳ͍࣌ܥྻͷ܏޲ ΋ଟ͘ଘࡏ l ߏ଄มԽ l ҟৗ஋ l ඇରশ෼෍ l ඇઢܗੑ

Slide 105

Slide 105 text

105 ࣌ܥྻϞσϧ ඇઢܗɾඇΨ΢εܕঢ়ଶۭؒϞσϧ 𝑦! 𝑦!"# 𝑦!$# 𝑥!$# 𝑥! 𝑥!"# 𝑣!"# 𝑣! 𝑣!$# 𝑤!$# 𝑤! 𝑤!"# 𝐹! 𝐺! 𝐻! 𝑥! = 𝐹! 𝑥!$# + 𝐺! 𝑣! 𝑦! = 𝐻! 𝑥! + 𝑤! ઢܗɾΨ΢εܕ ඇઢܗɾඇΨ΢εܕ 𝑥! = 𝑓(𝑥!$# , 𝑣! ) 𝑦! = ℎ(𝑥! ) + 𝑤! 𝑣! ~𝑞 𝑣 , 𝑤! ~𝑟(𝑤)

Slide 106

Slide 106 text

106 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ l ༧ଌ l ϑΟϧλ l ฏ׈Խ 𝑝 𝑥! 𝑌!$# ) = ; $* * 𝑝 𝑥! 𝑥!$# )𝑝 𝑥!$# 𝑌!$# )𝑑𝑥!$# 𝑥! = 𝑓(𝑥!$# , 𝑣! ) 𝑦! = ℎ(𝑥! ) + 𝑤! 𝑝 𝑥! 𝑌! ) = 𝑝 𝑦! 𝑥! )𝑝 𝑥! 𝑌!$# ) 𝑝 𝑦! 𝑌!–# ) 𝑝 𝑥! 𝑌( ) = 𝑝 𝑥! 𝑌! ) ∫ $* * , -!"# -!), -!"# /$) , -!"# /!) 𝑑𝑥!"#

Slide 107

Slide 107 text

107 ࣌ܥྻϞσϧ ΧϧϚϯϑΟϧλ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 108

Slide 108 text

108 ࣌ܥྻϞσϧ ཻࢠϑΟϧλ from: https://www.ieice.org/jpn/books/kaishikiji/2005/200512.pdf l ඇઢܗɾඇΨ΢εঢ়ଶۭؒϞσϧͰ͸ੵ෼ܭࢉ͕ղੳతʹࠔ೉Ͱ ͋Δ͜ͱ͔Β༻͍ΒΕΔख๏ l ղੳతʹදݱͰ͖ͳ͍ಉ࣌෼෍Λʮཻࢠʯ ʹج͍ͮͨཚ਺ϕΫτϧΛੜ੒ɻ ੜ੒ͨ͠ཚ਺ϕΫτϧʹجͮ͘ܦݧ෼෍ ؔ਺ͰٻΊ͍ͨঢ়ଶۭؒͷ෼෍Λۙࣅɻ

Slide 109

Slide 109 text

109 ࣌ܥྻϞσϧ Ϛϧίϑ࿈࠯ϞϯςΧϧϩ๏ .$.$ from: https://www.ieice.org/jpn/books/kaishikiji/2005/200512.pdf l ཚ਺ൃੜΞϧΰϦζϜͷҰछͰɺੜ੒͞Εͨཚ਺Λ༻͍ͯۙࣅܭ ࢉΛߦ͏ɻ l ࣄޙ෼෍ͷܗͦͷ΋ͷΛਪఆ l ঢ়ଶਪఆ͚ͩͰͳ͘ɺϞσϧͷύϥϝλʔ΋ಉ࣌ʹਪఆ͢Δ

Slide 110

Slide 110 text

110 ࣌ܥྻϞσϧ ඇઢܗɾඇΨ΢εܕঢ়ଶۭؒϞσϧͷྫ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 111

Slide 111 text

111 ࣌ܥྻϞσϧ ඇઢܗɾඇΨ΢εܕঢ়ଶۭؒϞσϧͷྫ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 112

Slide 112 text

112 ࣌ܥྻϞσϧ ඇઢܗɾඇΨ΢εܕঢ়ଶۭؒϞσϧͷྫ from: https://elf-c.he.u-tokyo.ac.jp/courses/388

Slide 113

Slide 113 text

113 Ϟσϧͷਪఆɾબ୒

Slide 114

Slide 114 text

114 Ϟσϧͷਪఆɾબ୒ ओͳϞσϦϯάϑϨʔϜϫʔΫ モデル σʔλ ܦݧత஌ࣝ ཧ࿦ ࣌ܥྻ༧ଌ ࣌ܥྻ෼ྨ ҟৗ஋ݕ஌ ͲͷΑ͏ͳϞσϧΛֶश͠࠷ద ͳϞσϧΛબ୒͢Ε͹Α͍͔ʁ

Slide 115

Slide 115 text

115 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ"3Ϟσϧ l "3Ϟσϧ΍7"3Ϟσϧͷύϥϝʔλͷਪఆͱͯ͠:VMF8BMLFS๏ ΍)PVTFIPMEFS๏ɺ1"3$03๏ͳͲ͕ଘࡏ l ࠓճ͸ҰൠతͳճؼϞσϧͰ΋༻͍ΒΕ͍ͯΔ࠷খೋ৐๏ 0-4 ΍ ࠷໬ਪఆ .-& ʹߜͬͯ঺հ

Slide 116

Slide 116 text

116 Ϟσϧͷਪఆɾબ୒ l ߏங͞Εͨؔ਺Ϟσϧ𝑓(ɾ)ͱ؍ଌ஋ͱͷؒͷೋ৐ޡࠩ ࢒ࠩ𝑒! ͷೋ ৐ ͕࠷খͱͳΔΑ͏ͳύϥϝʔλΛਪఆ l ֬཰ͷ֓೦͕ݱΕͳ͍ l ճؼ෼ੳͳͲʹ͓͍ͯɺղੳతʹύϥϝʔλಋग़͕Մೳ Ϟσϧͷਪఆ࠷খೋ৐๏ 0-4 𝐸 = b \]" _ (𝑦\ − 𝑓(𝑥\))Z

Slide 117

Slide 117 text

117 Ϟσϧͷਪఆɾબ୒ l ྫճؼϞσϧ "3 l ͭͷ܎਺Λ࣋ͭճؼϞσϧ l ೋ৐ޡࠩ Ϟσϧͷਪఆ࠷খೋ৐๏ 0-4 𝐰𝐓 = (𝑤0 , 𝑤% ), 𝐱𝐢 b = (1, 𝑥)ͱఆٛ͢Δͱ 𝑓 𝑥! = v 𝑦! = 𝑤0 + 𝑤% 𝑥! = 𝐰𝐓𝐱 v 𝑦% v 𝑦c = 1, 𝑥% 1, 𝑥d 𝑤0 𝑤% v 𝐲 = 𝐗𝐰 ܭըߦྻ E = (𝐲 − 𝐗𝐰)b(𝐲 − 𝐗𝐰)

Slide 118

Slide 118 text

118 Ϟσϧͷਪఆɾબ୒ l ྫճؼϞσϧ "3 l ೋ৐ޡࠩEΛ𝐰Ͱඍ෼ Ϟσϧͷਪఆ࠷খೋ৐๏ 0-4 E = 𝐲 − 𝐗𝐰 b 𝐲 − 𝐗𝐰 = 𝐲𝐓𝐲 − 𝟐𝐰𝐓 𝐗𝐓𝐲 + 𝐰𝐓𝐗𝐓𝐗𝐰 𝜕𝐸 𝜕𝐰 = −𝟐𝐗b𝐲 + 𝐗b𝐗 + 𝐗b𝐗 b 𝐰 = −2𝐗b𝐲 + 𝟐𝐗b𝐗w = 0 w = (𝐗b𝐗)+%𝐗b𝐲

Slide 119

Slide 119 text

119 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ,VMMCBDL-FJCMFS ,- ৘ใྔ l ͭͷ֬཰෼෍ͷࠩҟΛଌఆ͢Δई౓ l ਪఆ͞ΕΔ༧ଌ෼෍͕ਅͷ෼෍Λଊ͍͑ͯΔ͔ΛධՁ l g(x)͕ਅͷ෼෍ f(x)͕Ϟσϧͷ෼෍ͱͨ࣌͠ɺ,-৘ใྔ͸ҎԼͷܗ l D`a (g| 𝑓 = 0ͷ࣌ɺ g x = f(x) l ࡾ֯ෆ౳ࣜ΍ରশੑͳͲͷڑ཭ͷެཧ͸ຬͨ͞ͳ͍ Def (g||𝑓) = ~ +T T g x log g x f x 𝑑𝑥 ∗ Def (g||𝑓) ≥ 0

Slide 120

Slide 120 text

120 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ਖ਼ن෼෍ಉ࢜ͷ,-ڑ཭ D`a(g||𝑓) = m -b b g x log g x 𝑑𝑥 − m -b b g x log 𝑓 x 𝑑𝑥 g x ~ N 𝜇c , 𝜎c Z , f x ~ N 𝜇d , 𝜎d Z m -b b g x log 𝑓 x 𝑑𝑥 = − 1 2 log2π𝜎d Z − 𝜎c Z + 𝜇c − 𝜇d Z 2𝜎d Z m -b b g x log g x 𝑑𝑥 = − 1 2 log2π𝜎c Z − 1 2 正規分布の確率密度関数

Slide 121

Slide 121 text

121 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ Def (g||𝑓) = ~ +T T g x log g x f x 𝑑𝑥 ∗ Def (g||𝑓) ≥ 0 from: https://qiita.com/ceptree/items/9a473b5163d5655420e8

Slide 122

Slide 122 text

122 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ϞσϦϯάʹ͓͚ΔKL৘ใྔͷਪఆ l g(x)͕ਅͷ෼෍ f(x)͕Ϟσϧͷ෼෍ Def (g||f) = ~ +T T g x log g x f x 𝑑𝑥 = ~ +T T g x log g x 𝑑𝑥 − ~ +T T g x log f x 𝑑𝑥

Slide 123

Slide 123 text

123 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ϞσϦϯάʹ͓͚ΔKL৘ใྔͷਪఆ l g(x)͕ਅͷ෼෍ f(x)͕Ϟσϧͷ෼෍ Def (g||f) = ~ +T T g x log g x f x 𝑑𝑥 = ~ +T T g x log g x 𝑑𝑥 − ~ +T T g x log f x 𝑑𝑥 Θ͔Βͳ͍

Slide 124

Slide 124 text

124 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ϞσϦϯάʹ͓͚ΔKL৘ใྔͷਪఆ l g(x)͕ਅͷ෼෍ f(x)͕Ϟσϧͷ෼෍ Def (g||f) = ~ +T T g x log g x f x 𝑑𝑥 = ~ +T T g x log g x 𝑑𝑥 − ~ +T T g x log f x 𝑑𝑥 Θ͔Βͳ͍ ฏۉର਺໬౓ g(x)͕Θ͔Βͳ͍ͨΊະ஌ ͕ͩɺͲͷ f(x) Ͱ΋Ұఆ

Slide 125

Slide 125 text

125 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ϞσϦϯάʹ͓͚ΔKL৘ใྔͷਪఆ l g(x)͕ਅͷ෼෍ f(x)͕Ϟσϧͷ෼෍ Def (g||f) = ~ +T T g x log g x f x 𝑑𝑥 = ~ +T T g x log g x 𝑑𝑥 − ~ +T T g x log f x 𝑑𝑥 Θ͔Βͳ͍ ฏۉର਺໬౓ g(x)͕Θ͔Βͳ͍ͨΊະ஌ ͕ͩɺͲͷ f(x) Ͱ΋Ұఆ ૬ରతʹɺϞσϧ෼෍ 𝐟(𝐱)ͷ ༗ޮੑʹ͍ͭͯධՁͰ͖Δʁ

Slide 126

Slide 126 text

126 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ ~ +T T g x log f x 𝑑𝑥 l ฏۉର਺໬౓ l ର਺໬౓ ະ஌ 1 𝑛 b \]" W 𝐼(𝑥\) େ਺ͷ๏ଇʹΑΓɺ༩͑ΒΕͨ αϯϓϧσʔλͰۙࣅ ਅͷ໬౓ αϯϓϧσʔλ ͷ໬౓ 𝜃+ 𝜃678 ฏۉର਺໬౓ ର਺໬౓ ℓ = ) !/% ' log f x

Slide 127

Slide 127 text

127 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷਪఆ࠷໬ਪఆ l ࠷໬ਪఆ .-& l ର਺໬౓Λ༻͍ͨύϥϝʔλਪఆ๏ l ର਺໬౓Λ࠷େԽ͢Δύϥϝʔλ ࠷໬ਪఆ஋ 𝜃fgh ͕ۙࣅతʹ,-৘ใྔ Λ࠷΋࠷খԽ͢ΔύϥϝʔλʹͳΔͱਪఆͰ͖Δ l ର਺໬౓Λ࠷େԽ͢Δύϥϝʔλ ࠷໬ਪఆ஋ 𝜃fgh ͸ؔ਺ℓ(𝜃)͕𝜃Ͱඍ෼ ͯ͠ ij(k) ik = 0ͱͳΔ΋ͷΛਪఆ ਅͷ໬౓ αϯϓϧσʔλ ͷ໬౓ 𝜃+ 𝜃678 ฏۉର਺໬౓ ର਺໬౓ ℓ(𝜃) = ) !/% ' log f x|θ 𝜃1gh = max i ℓ(𝜃)

Slide 128

Slide 128 text

128 Ϟσϧͷਪఆɾબ୒ ࠷খೋ৐๏ͱ࠷໬ਪఆ l ྫɿճؼϞσϧͷ࠷໬ਪఆɻޡࠩΛਖ਼ن෼෍ͱԾఆ v 𝑦! = 𝑓 𝑥! = 𝑁(𝑤0 + 𝑤% 𝑥! , 𝜎") ℓ(𝜃) = ) !/% ' log 1 2𝜋𝜎" exp(− 𝑦! − (𝑤0 + 𝑤% 𝑥! ) " 2𝜎" ) = − 𝑛 2 log 2𝜋 − 𝑛 2 log 𝜎" − ) !/% ' ( 𝑦! − (𝑤0 + 𝑤% 𝑥! ) " 2𝜎" )

Slide 129

Slide 129 text

129 Ϟσϧͷਪఆɾબ୒ l ߏங͞Εͨؔ਺Ϟσϧ𝑓(ɾ)ͱ؍ଌ஋ͱͷؒͷೋ৐ޡࠩ ࢒ࠩ𝑒! ͷೋ ৐ ͕࠷খͱͳΔΑ͏ͳύϥϝʔλΛਪఆ l ֬཰ͷ֓೦͕ݱΕͳ͍ l ճؼ෼ੳͳͲʹ͓͍ͯɺղੳతʹύϥϝʔλಋग़͕Մೳ ࠷খೋ৐๏ͱ࠷໬ਪఆ 𝐸 = b \]" _ (𝑦\ − 𝑓(𝑥\))Z

Slide 130

Slide 130 text

130 Ϟσϧͷਪఆɾબ୒ ࠷খೋ৐๏ͱ࠷໬ਪఆ l ྫɿճؼϞσϧͷ࠷໬ਪఆɻޡࠩΛਖ਼ن෼෍ͱԾఆ v 𝑦! = 𝑓 𝑥! = 𝑁(𝑤0 + 𝑤% 𝑥! , 𝜎") ℓ(𝜃) = ) !/% ' log 1 2𝜋𝜎" exp(− 𝑦! − (𝑤0 + 𝑤% 𝑥! ) " 2𝜎" ) = − 𝑛 2 log 2𝜋 − 𝑛 2 log 𝜎" − ) !/% ' ( 𝑦! − (𝑤0 + 𝑤% 𝑥! ) " 2𝜎" ) ର਺໬౓ͷ࠷େԽ͸࠷খೋ৐๏ͱҰக 㱺 ࠷খೋ৐๏͸࠷໬ਪఆͷಛघྫͱҰக

Slide 131

Slide 131 text

131 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ 𝜃+ "3 "3*." ." ͲͷϞσϧΛ࢖͏ͱ ྑ͍ʁʁʁ

Slide 132

Slide 132 text

132 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ 𝜃+ "3 "3*." ." ͲͷϞσϧΛ࢖͏ͱ ྑ͍ʁʁʁ ର਺໬౓͕࠷΋େ͖͍ Ϟσϧ͕ྑ͍ͷͰ͸ʁ

Slide 133

Slide 133 text

133 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ଟ߲ࣜճؼͷର਺໬౓ from: https://elf-c.he.u-tokyo.ac.jp/courses/382 ࣍਺Λ૿΍͢ͱ ର਺໬౓΋্ঢ

Slide 134

Slide 134 text

134 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ଟ߲ࣜճؼͷର਺໬౓ from: https://elf-c.he.u-tokyo.ac.jp/courses/382 ࣍਺Λ૿΍͢ͱ ର਺໬౓΋্ঢ આ໌ม਺Λ૿΍͠·͘Ε͹ର਺໬౓ͷ؍఺ Ͱ΋ྑ͍ϞσϧΛ࡞ΕΔͷͰ͸ʁʁʁ

Slide 135

Slide 135 text

135 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ࠷΋ྑ͍ର਺໬౓Λ࣋ͭϞσϧΛબ୒ͨ͠৔߹ l աద߹ PWFSGJU ͷ໰୊ l ༩͑ΒΕͨσʔλ͸े෼આ໌Ͱ͖ͨͱͯ͠΋ɺ৽͍͠σʔλΛे෼ʹઆ໌Ͱ ͖Δͱ͸ݶΒͳ͍ from: https://datascience.foundation/sciencewhitepaper/u nderfitting-and-overfitting-in-machine-learning

Slide 136

Slide 136 text

136 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ࠷΋ྑ͍ର਺໬౓Λ࣋ͭϞσϧΛબ୒ͨ͠৔߹ l աద߹ PWFSGJU ͷ໰୊ l ༩͑ΒΕͨσʔλ͸े෼આ໌Ͱ͖ͨͱͯ͠΋ɺ৽͍͠σʔλΛे෼ʹઆ໌Ͱ ͖Δͱ͸ݶΒͳ͍ from: https://datascience.foundation/sciencewhitepaper/u nderfitting-and-overfitting-in-machine-learning ෳࡶ͗͢ͳ͍࠷దͳϞσϧΛ બ୒͢Δ͜ͱ͕ॏཁ

Slide 137

Slide 137 text

137 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ద੾ͳϞσϧબ୒ͷͨΊʹ l ର਺໬౓ʹΑͬͯਪఆ͞ΕΔ𝜃fgh ͱ ਅͷύϥϝʔλ 𝜃. ͷؒʹόΠΞε ͕ଘࡏ l ର਺໬౓ͱฏۉର਺໬౓ͷόΠΞε ΛධՁͯ͠ద੾ʹิਖ਼͢Δ͜ͱͰ ࠷దͳϞσϧΛબ୒͢Δඞཁ

Slide 138

Slide 138 text

138 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ద੾ͳϞσϧબ୒ͷͨΊʹ l 𝜃fgh ʹ͓͚Δର਺໬౓ͱฏۉର਺ ໬౓ͷࠩ 𝐷Λิਖ਼͍ͨ͠ 1 𝑛 ? '"* , log f x|𝜃-./ ͜ͷ෦෼Λ ิਖ਼͍ͨ͠ʂ ظ଴஋ΛͱΔ 𝔼0[𝐷] = t #$ $ g x [D]𝑑𝑥 ฏۉର਺໬౓ ର਺໬౓ 𝐷 = 1 𝑛 ? '"* , log f x|𝜃-./ − t #$ $ g x log f x|𝜃-./ 𝑑𝑥

Slide 139

Slide 139 text

139 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ l ద੾ͳϞσϧબ୒ͷͨΊʹ l 𝜃fgh ʹ͓͚Δର਺໬౓ͱฏۉର਺ ໬౓ͷࠩ 𝐷Λิਖ਼͍ͨ͠ 1 𝑛 ? '"* , log f x|𝜃-./ ͜ͷ෦෼Λ ิਖ਼͍ͨ͠ʂ ฏۉର਺໬౓ ର਺໬౓ ظ଴஋ΛͱΔ 𝔼0[𝐷] = t #$ $ g x [D]𝑑𝑥 ͜ͷิਖ਼෦෼͕ύϥϝʔλ਺Ͱ͋Δ 𝑝ʹۙࣅͰ͖Δ ิਖ਼ 𝐷 = 1 𝑛 ? '"* , log f x|𝜃-./ − t #$ $ g x log f x|𝜃-./ 𝑑𝑥

Slide 140

Slide 140 text

140 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ"*$ l ద੾ͳϞσϧબ୒ͷͨΊʹ l ࠷େର਺໬౓ όΠΞε 𝜃fgh ʹ͓͚Δฏۉର਺໬౓ 1 𝑛 ? '"* , log f x|𝜃-./ ∑!/% ' log f x|𝜃1gh − 𝔼x [𝐷] = 𝑛 ∫ +T T g x log f x|𝜃1gh 𝑑𝑥 ∑!/% ' log f x|𝜃1gh − 𝑝= n ∫ +T T g x log f x|𝜃1gh 𝑑𝑥 𝔼0[𝐷]

Slide 141

Slide 141 text

141 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ"*$ l ੺஑৘ใྔج४ "*$ l ฏۉର਺໬౓ͷ 𝜃fgh ʹ͓͚ΔਪఆྔͰ͋ΔҎԼͷࣜͷࠨลΛº͢Δ͜ͱͰ ౷ܭϞσϧͷࢦඪͱͳΔ l 𝑝͸ࣗ༝ύϥϝʔλ਺ l ͜ͷࢦඪ͕খ͍͞౷ܭϞσϧ͕ྑ͍Ϟσϧͱਪఆ͞ΕΔ l ৘ใྔج४ͷछ ∑!/% ' log f x|𝜃1gh − 𝑝= n ∫ +T T g x log f x|𝜃1gh 𝑑𝑥 AIC = −2 ) !/% ' log f x|𝜃1gh + 2𝑝

Slide 142

Slide 142 text

142 Ϟσϧͷਪఆɾબ୒ "*$ͷྫ l ଟ߲ࣜճؼͷ"*$ʹΑΔධՁ from: https://elf-c.he.u-tokyo.ac.jp/courses/382

Slide 143

Slide 143 text

143 Ϟσϧͷਪఆɾબ୒ "*$ͷྫ l ͲͷΑ͏ͳ#JOͷ਺Ͱදݱ͢Δͷ͕ྑ͍ͷ͔ʁ l ͲͷΑ͏ͳଟ߲෼෍ϞσϧʹϑΟοτͤ͞Δͷ͕ྑ͍ͷ͔ʁ

Slide 144

Slide 144 text

144 Ϟσϧͷਪఆɾબ୒ "*$ͷྫ

Slide 145

Slide 145 text

145 "*$ͷྫ Ϟσϧͷਪఆɾબ୒

Slide 146

Slide 146 text

146 Ϟσϧͷਪఆɾબ୒ from: https://elf-c.he.u-tokyo.ac.jp/courses/382 "*$ͷྫ

Slide 147

Slide 147 text

147 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબఆ"*$ l Ϟσϧͷઆ໌ม਺͕গͳ͍ͱɺ#JBT͕େ͖͘ͳΔ ֶशෆ଍ l Ϟσϧͷઆ໌ม਺͕ଟ͍ͱɺ#JBT͸খ͘͞ͳΔ͕ɺ7BSJBODF͕େ ͖͘ͳΓɺաֶशͱͳΔ l "*$ͳͲͷ৘ใྔج४͸ɺόΠΞεͱόϦΞϯεͷͭΛόϥϯε Α͘࠷খʹ͢Δʢظ଴༧ଌޡࠩͷ࠷খʣͷͨΊͷࢦඪ ༧ଌޡࠩ #JBT7BSJBODF

Slide 148

Slide 148 text

148 Ϟσϧͷਪఆɾબ୒ "*$ͷར༻ʹ͍ͭͯ l "*$ͷઈର஋ʹ͸ҙຯ͕ͳ͍͕ɺ"*$ͷࠩʹ͸ҙຯ͕͋Δ l "*$͸༧ଌ෼෍ΛධՁ͢Δ΋ͷͰɺઆ໌ม਺ͷޮՌͷݕূ΍ݕఆ ͳͲʹ༻͍Δ͜ͱ͸ద੾Ͱͳ͍ɻ l આ໌ม਺Λ૿΍͢͜ͱͰ"*$͕ৗʹ૿͑ଓ͚Δ৔߹ɺԾఆ͍ͯ͠ ΔϞσϧ଒͕ద͞ͳ͍Մೳੑ͕͋Δ

Slide 149

Slide 149 text

149 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબ୒ଞͷબ୒ํ๏ l ৘ใྔج४ *$ l Ϟσϧͷ౰ͯ͸·Γͷྑ͞ͱɺϞσϧͷෳࡶ͞ʹର͢ΔϖφϧςΟ͔Βߏ ੒͞ΕΔࢦඪͷҰൠܥ l ର਺໬౓ʹج͍ͮͨؔ਺ͱϖφϧςΟؔ਺𝐼(𝑛)ʹΑͬͯߏ੒ l ϖφϧςΟؔ਺ΛͲͷΑ͏ʹઃܭ͢Δ͔Ͱෳ਺ͷࢦඪ͕ଘࡏ l ৘ใྔج४͕খ͍͞Ϟσϧ͕ྑ͍Ϟσϧͷࢦඪ l "*$͸𝐼 𝑛 = 2 ͷϞσϧ IC = −2 ) !/% ' log f x|𝜃1gh + 𝐼(𝑛)𝑝

Slide 150

Slide 150 text

150 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબ୒ଞͷબ୒ํ๏ l ҳ୤౓৘ใྔج४ %*$ l ৘ใྔج४ͷछ l ର਺໬౓ͷࣄޙฏۉΛਪఆྔͱͯ͠ར༻͢ΔϕΠζϕʔεͷࢦඪ l ϖφϧςΟؔ਺͸ࣄલ෼෍ʹΑΔର਺໬౓ͱࣄޙฏۉͷର਺໬౓ͷࠩ 𝐷IC = −2 ~ log f x|𝜃y1 𝑑𝜃y1 + 𝐼 ・ 2 𝐼 ・ =2 ∫ log f x|𝜃y1 𝑑𝜃y1 − 2 ∫ log f x|𝜃 ∏(𝜃|𝑥) 𝑑𝜃 𝜃y1 = ∫ 𝜃 ∏(𝜃|𝑥) 𝑑𝜃

Slide 151

Slide 151 text

151 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબ୒ଞͷબ୒ํ๏ l ϕΠζ৘ใྔج४ #*$ l ৘ใྔج४ͷछ l ,-৘ใྔʹجͮ͘ͷͰ͸ͳ͘पล໬౓ʹج͍ͮͨ৘ใྔ l 𝑛 ≥ 8ͷ࣌ʹ"*$ΑΓ΋ϖφϧςΟ͕ڧ͘ͳΔɻͭ·Γɺ"*$ΑΓ΋ύϥ ϝʔλ਺͕খ͍͞ϞσϧΛબ୒͢Δ܏޲ 𝐵IC = −2 ) !/% ' log f x|𝜃1gh + 𝑝 log 𝑛

Slide 152

Slide 152 text

152 Ϟσϧͷਪఆɾબ୒ Ϟσϧͷબ୒ଞͷબ୒ํ๏ l .JOJNVN%FTDSJQUJPO-FOHUI .%- l σʔλΛϞσϧΛ༻͍ͯѹॖ͢Δࡍͷූ߸௕ͷ࠷খԽʹج͍ͮͨࢦඪ l ৘ใྔج४͸ਅͷ෼෍ͷڑ཭Λߟྀ͢Δ͕ɺ.%-͸σʔλѹॖʹ͍ͭͯ ߟ͑Δ l σʔλ΁ͷద߹ ର਺໬౓ ͱϞσϧίετͷͭͰߏ੒ l 𝐵͕ූ߸௕ʹ͋ͨΔ MDL = − ) !/% ' log f x|𝜃1gh + 𝐵𝑝

Slide 153

Slide 153 text

153 ਂ૚ֶशͷਐల

Slide 154

Slide 154 text

154 ਂ૚ֶशͷൃల "MQIB(P;FSP ը૾ੜ੒ 4UBCMF%JGGVTJPO ݴޠϞσϧ λϯύΫ࣭ߏ଄༧ଌ "MQIB'PME

Slide 155

Slide 155 text

155 ਂ૚ֶशͷൃల ҩྍը૾ จ๏ޡΓగਖ਼ ϩϘοτ ෆద੾౤ߘͷݕ஌

Slide 156

Slide 156 text

156 ਂ૚ֶशͷൃల *NBHF/FU$IBMMFOHF

Slide 157

Slide 157 text

157 ਂ૚ֶशͷൃల *NBHF/FU$IBMMFOHF

Slide 158

Slide 158 text

158 ਂ૚ֶशͷൃల "MFY/FU

Slide 159

Slide 159 text

159 ਂ૚ֶशͷൃల ਂ૚ֶश l ਂ૚ֶशਂ͍ߏ଄Λ࣋ͭχϡʔϥϧωοτϫʔΫ l χϡʔϥϧωοτϫʔΫ /FVSBM/FUXPSL ೴ͷߏ଄ɾػೳͷ໛฿͔ Βൃలͨ͠ϞσϧͰɺਓ޻χϡʔϩϯΛ૊Έ߹ΘͤͨωοτϫʔΫ l Ϟσϧ͸ؔ਺ͷू߹ ⼊⼒ 出⼒ 𝑓% 𝑓" 𝑓V ඍ෼Մೳͳؔ਺

Slide 160

Slide 160 text

160 ਂ૚ֶशͷൃల ϑΟʔυϑΥϫʔυχϡʔϥϧωοτϫʔΫ Input Layer Hidden Layers Output Layer X = A[0] a[4] A[1] A[3] X Ŷ a[1] n a[2] 1 a[2] 2 a[2] 3 a[2] n a[3] 1 a[3] 2 a[3] 3 a[3] n A[2] A[4] 𝑓% ؔ਺ 𝑓% ͷྫ 𝐴[%] = 𝑓% 𝑥 = 𝑊𝑥 + 𝑏 ઢܗϞσϧ

Slide 161

Slide 161 text

161 ਂ૚ֶशͷൃల ϑΟʔυϑΥϫʔυχϡʔϥϧωοτϫʔΫ Input Layer Hidden Layers Output Layer X = A[0] a[4] A[1] A[3] X Ŷ a[1] n a[2] 1 a[2] 2 a[2] 3 a[2] n a[3] 1 a[3] 2 a[3] 3 a[3] n A[2] A[4] 𝑓% ؔ਺ 𝑓% ͷྫ 𝐴[%] = 𝑓% 𝑥 = 𝑊𝑥 + 𝑏 ΞϑΟϯม׵ 0.6 0.2 0.8 −0.3 −0.3 0.4 0.1 0.7 ॏΈΛσʔλʹ ߹͏Α͏ʹֶश

Slide 162

Slide 162 text

162 ਂ૚ֶशͷൃల ϑΟʔυϑΥϫʔυχϡʔϥϧωοτϫʔΫ Input Layer Hidden Layers Output Layer X = A[0] a[4] A[1] A[3] X Ŷ a[1] n a[2] 1 a[2] 2 a[2] 3 a[2] n a[3] 1 a[3] 2 a[3] 3 a[3] n A[2] A[4] 𝑓% 𝐴[%] = 𝑚𝑎𝑥{𝑥, 0} ׆ੑԽؔ਺ ɾ3F-6 𝐴[%] = 1/(1 + 𝑒+#) ɾ4JHNPJE

Slide 163

Slide 163 text

163 ਂ૚ֶशͷൃల ਂ૚Ϟσϧͷֶश ⼊⼒ 出⼒Ŷ 𝑓% 𝑓" 𝑓V l ଛࣦؔ਺ L 𝑊 = 4 !/% ' ℓ(𝑌! , 𝐹(𝑥! )) ೋ৐ଛࣦ NFBOTRVBSFEFSSPS ℓ 𝑦, 𝑦€ = (𝑦 − 𝑦€)" ަࠩΤϯτϩϐʔޡࠩ $SPTT&OUSPQZMPTT ℓ 𝑦, 𝑦€ = − 4 &/% • 𝑦& log 𝑦& € ͜ΕΛ࠷খԽ͢ΔΑ͏ʹֶश min l L 𝑊

Slide 164

Slide 164 text

164 ਂ૚ֶशͷൃల ਂ૚Ϟσϧͷֶश l ޡࠩٯ఻೻๏ l ύϥϝʔλߋ৽ͷख๏ l ଛࣦؔ਺ͷඍ෼ΛޙΖ ೖྗํ޲ ʹͲΜͲΜ࿈࠯཯Λ༻͍ͯ఻ൖͤ͞Δ 𝜕𝐿 𝜕𝑤% 𝑥 = 𝜕𝐿 𝜕𝑓V 𝜕𝑓V 𝜕𝑓" 𝜕𝑓" 𝜕𝑓% 𝜕𝑓% 𝜕𝑤% 𝑥 ⼊⼒ 𝑓% 𝑓" 𝑓V 出⼒Ŷ L 𝑊

Slide 165

Slide 165 text

165 ਂ૚ֶशͷൃల ਂ૚Ϟσϧͷֶश l ޯ഑߱Լ๏ l ଛࣦؔ਺͕࠷খͱͳΔΑ͏ʹ఻ൖ͞ΕͨޡࠩʹԠͯ͡ɺύϥϝʔλΛগ͠ ͣͭߋ৽͍ͯ͘͠ख๏ l ֤ύϥϝʔλͷଛࣦؔ਺ʹର͢Δภඍ෼ Λߦ͍ύϥϝʔλΛߋ৽ 𝑊* = 𝑊*+% − 𝜂 A 𝜕𝐿 𝜕𝑊 ‚/‚"'& From: https://axa.biopapyrus.jp/deep-learning/gradient_descent_method.htm

Slide 166

Slide 166 text

166 ਂ૚ֶशͷൃల ઢܗϞσϧͱͷҧ͍ a[4] a[4] X ઢܗϞσϧ ֶश ਂ૚ֶशϞσϧ ΧʔωϧϞσϧ ֶश ݻఆ ֶश ֶश

Slide 167

Slide 167 text

167 ਂ૚ֶशͷൃల ਂ૚ֶशͷಛ௃ l ສೳۙࣅೳྗ l ૚χϡʔϥϧωοτϫʔΫ͸े෼ͳ਺ͷύϥϝʔλ͕͋Ε͹ɺ࿈ଓؔ਺ Λ೚ҙͷਫ਼౓ͰۙࣅՄೳ l ཧ࿦తʹ͸σʔλ͕ແݶʹ͋Γɼૉࢠ਺͕ແݶʹ͋Δχϡʔϥϧωοτ ϫʔΫΛ༻͍Ε͹ͲΜͳ໰୊Ͱ΋ֶशՄೳ from: https://www.slideshare.net/trinmu/ss-226711018

Slide 168

Slide 168 text

168 ਂ૚ֶशͷൃల ਂ૚ֶशͷಛ௃ l ଟ૚ԽͷԸܙ l χϡʔϥϧωοτϫʔΫͷදݱྗ͸૚ ͷ਺ʹΑͬͯදݱྗ͕ࢦ਺తʹ޲্ l ਂ͍χϡʔϥϧωοτϫʔΫͷදݱೳ ྗΛઙ͍χϡʔϥϧωοτϫʔΫͰද ݱ͠Α͏ͱ͢Δͱେ͖ͳԣ෯ χϡʔ ϩϯ ͕ඞཁ Montufar, Guido F., et al. "On the number of linear regions of deep neural networks." Advances in neural information processing systems 27 (2014). Arora, R., Basu, A., Mianjy, P., & Mukherjee, A. (2016). Understanding deep neural networks with rectified linear units. arXiv preprint arXiv:1611.01491.

Slide 169

Slide 169 text

169 ਂ૚ֶशͷൃల ਂ૚ֶशͷಛ௃ l εέʔϦϯάଇ l 5SBOTGPSNFSͷੑೳ͸ύϥϝʔλ ਺ɺσʔληοταΠζɺܭࢉ༧ ࢉΛม਺ͱͨ͠γϯϓϧͳ΂͖৐ ଇʹै͏ l ͜ͷ๏ଇ͕ݴޠɾը૾ɾಈըͳͲ ͷ༷ʑͳλεΫʹద༻͞ΕΔ͜ͱ Λࣔ͢ from: Henighan, Tom, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun et al. "Scaling laws for autoregressive generative modeling." arXiv preprint arXiv:2010.14701 (2020).

Slide 170

Slide 170 text

170 ਂ૚ֶशͷൃల ਂ૚ֶशͷಛ௃ l 0WFSQBSBNFUFSJ[BUJPO l ଟ͘ͷେن໛ϞσϧͰͷֶशͰ͸σʔλ਺ ύϥϝʔλ਺ l ύϥϝʔλ਺Λ૿΍͠΋ɺ࣮ࡍͷֶशʹ͓͍ͯաֶश͸ੜ͡ͳ͍ l χϡʔϥϧωοτϫʔΫͰ͸൚Խޡࠩͱ܇࿅ޡࠩͷ͕ࠩখ͘͞ͳ ΔͨΊʁʢཧ࿦త෼ੳ͕ߦΘΕ͍ͯΔʣ Neyshabur, Behnam, et al. "The role of over- parametrization in generalization of neural networks." 7th International Conference on Learning Representations, ICLR 2019. 2019.

Slide 171

Slide 171 text

171 ࣌ܥྻσʔλͱ ਂ૚ֶश

Slide 172

Slide 172 text

172 ࣌ܥྻσʔλͱਂ૚ֶश 'FFE'PSXBSE//ͱ࣌ܥྻ l ௨ৗͷ'FFE'PSXBE //ʹ࣌ܥྻσʔλΛ ద༻͢Δ͜ͱ͸Մೳ͕ͩɺ࣌ؒͷྲྀΕΛଊ ͑Δ͜ͱ͕Ͱ͖ͳ͍ l ྫ͑͹ɺ༧ଌͷλεΫͰ͸ɺੲͷ෦෼͸ܰ ͘ѻͬͯɺ࠷ۙͷ෦෼͸ॏཁͳΑ͏ʹѻ͍ ͍ͨ l ࣌఺𝑡ͷֶशʹ͓͍ͯকདྷͷ஋𝑡 + 1Λ࢖Θ ͣɺաڈͷ஋͚ͩΛ༻ֶ͍ͯश͍ͨ͠ l ೖྗ௕͕ݻఆͰͳ͘ՄมͰ͋Δ͜ͱͷ΄͏ ͕ɺ࣌ܥྻΛೖྗͷࡍʹخ͍͠

Slide 173

Slide 173 text

173 ࣌ܥྻσʔλͱਂ૚ֶश 3FDVSSFOU/FVSBM/FUXPSL 3// ・・・ 𝑥% 𝑥" 𝑥V 𝑥W Feature Vecter l ӅΕঢ়ଶΛ࣋ͬͨχϡʔϥϧωοτϫʔΫ l ࣌ؒ৘ใͷ֓೦Λ࣋ͪɺ࣍ͷ࣌ؒʹӅΕঢ়ଶͷ৘ใΛ͓͘Δ RNN Block RNN Block RNN Block RNN Block

Slide 174

Slide 174 text

174 ࣌ܥྻσʔλͱਂ૚ֶश 3FDVSSFOU/FVSBM/FUXPSL 3// 𝑋 Feature Vecter RNN Block from: https://stanford.edu/~shervine/teaching/cs-230/cheatsheet- recurrent-neural-networks 𝑎o!p = 𝑔"(𝑊 qq𝑎o!-"p + 𝑊 qr𝑥o!p + 𝑏q) 𝑦o!p = 𝑔Z(𝑊 sq𝑎o!p + 𝑏s)

Slide 175

Slide 175 text

175 ࣌ܥྻσʔλͱਂ૚ֶश 3FDVSSFOU/FVSBM/FUXPSL 3// l ڧΈ l Մม௕ʹରԠ l ϞσϧαΠζ͕ೖྗ௕ʹΑͬͯมԽ͠ͳ͍ l աڈͷ৘ใΛߟྀͨ͠ܭࢉ͕Մೳ l ϞσϧͷॏΈ͕͕࣌ؒมԽͯ͠΋ڞ༗ l ऑΈ l ܭࢉͷ஗͞ l ݱࡏͷঢ়ଶʹରͯ͠কདྷͷೖྗΛߟྀͰ͖ͳ͍ l աڈͷ৘ใ͕࢒͍ͬͯͳ͍Մೳੑ

Slide 176

Slide 176 text

176 ࣌ܥྻσʔλͱਂ૚ֶश ଞͷ3//ܥ౷ͷϞσϧ l -POH4IPSU5FSN.FNPSZ -45. l ࣌ؒมԽ͢Δ௕ظهԱͷͨΊͷ ϝϞϦ𝑐o!pΛ࣋ͭ l ͭͷήʔτΛ࣋ͭ͜ͱͰաڈͷ৘ใͷ ๨٫΍ɺ৽͍͠ೖྗͷऔΓೖΕΛௐ੔ l Γ1 3FMFWBODF(BUF l Γ( 'PSHFU(BUF l Γ2: 6QEBUF(BUF l Γ3 0VUQVU(BUF 𝑐o!p = Γt ⋆ ̃ 𝑐o!p + Γu ⋆ 𝑐o!-"p ̃ 𝑐o!p = tanh(𝑊 v Γw ⋆ 𝑎o!-"p, 𝑥o!p + 𝑏v) 𝑎o!p = Γx ⋆ 𝑐o!p from: https://stanford.edu/~shervine/teaching/cs-230/cheatsheet- recurrent-neural-networks

Slide 177

Slide 177 text

177 ࣌ܥྻσʔλͱਂ૚ֶश ଞͷ3//ܥ౷ͷϞσϧ l (BUFE3FDVSSFOU6OJU (36 l -45.ΛҰൠԽͨ͠Ϟσϧ l -45.ΑΓ΋ܰྔͱͳ͓ͬͯΓɺͭͷ ήʔτͷΈͰߏ੒͠ܭࢉޮ཰ԽΛ޲্ l Γ1 3FMFWBODF(BUF l Γ2: 6QEBUF(BUF l -45.΋(36΋৽͍͠৘ใΛ֮͑ɺ աڈͷ৘ใΛద౓ʹ๨ΕΔػߏΛ උ͑ͨϞσϧ 𝑐o!p = Γt ⋆ ̃ 𝑐o!p + (1 − Γt) ⋆ 𝑐o!-"p ̃ 𝑐o!p = tanh(𝑊 v Γw ⋆ 𝑎o!-"p, 𝑥o!p + 𝑏v) 𝑎o!p = 𝑐o!p from: https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-recurrent-neural-networks

Slide 178

Slide 178 text

178 ࣌ܥྻσʔλͱਂ૚ֶश 3FDVSSFOU/FVSBM/FUXPSL 3// l ڧΈ l Մม௕ʹରԠ l ϞσϧαΠζ͕ೖྗ௕ʹΑͬͯมԽ͠ͳ͍ l աڈͷ৘ใΛߟྀͨ͠ܭࢉ͕Մೳ l ϞσϧͷॏΈ͕͕࣌ؒมԽͯ͠΋ڞ༗ l ऑΈ l ܭࢉͷ஗͞ l ݱࡏͷঢ়ଶʹରͯ͠কདྷͷೖྗΛߟྀͰ͖ͳ͍ l աڈͷ৘ใ͕࢒͍ͬͯͳ͍Մೳੑ

Slide 179

Slide 179 text

179 ࣌ܥྻσʔλͱਂ૚ֶश ૒ํ޲3// #JEJSFDUJPOBM3// from: https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-recurrent-neural-networks l ܥྻͷ࠷ॳ͔Βͷॱ఻ൖʹՃ͑ͯɺܥྻ ͷ࠷ޙ͔Βॱʹٯ޲͖ͷ఻ൖɺ྆ํ޲ʹ ߦ͏ωοτϫʔΫ l ࣌ܥྻσʔλʹద༻͞ΕΔ͜ͱ͸ଟ͘ͳ ͍͕ɺݴޠγʔΫΤϯεͳͲʹద༻͞Ε Δ͜ͱ͕ଟ͍

Slide 180

Slide 180 text

180 ࣌ܥྻσʔλͱਂ૚ֶश 3FDVSSFOU/FVSBM/FUXPSL 3// l ڧΈ l Մม௕ʹରԠ l ϞσϧαΠζ͕ೖྗ௕ʹΑͬͯมԽ͠ͳ͍ l աڈͷ৘ใΛߟྀͨ͠ܭࢉ͕Մೳ l ϞσϧͷॏΈ͕͕࣌ؒมԽͯ͠΋ڞ༗ l ऑΈ l ܭࢉͷ஗͞ l ݱࡏͷঢ়ଶʹରͯ͠কདྷͷೖྗΛߟྀͰ͖ͳ͍ l աڈͷ৘ใ͕࢒͍ͬͯͳ͍Մೳੑ

Slide 181

Slide 181 text

181 ࣌ܥྻσʔλͱਂ૚ֶश ޯ഑ফࣦ 7BOJTIJOHHSBEJFOU ・・・ 𝑥% 𝑥" 𝑥V 𝑥W Feature Vecter RNN Block RNN Block RNN Block RNN Block 𝜕𝐿 𝜕𝑤% 𝑥% = 𝜕𝐿 𝜕𝑅𝑁𝑁W … 𝜕𝑅𝑁𝑁% 𝜕𝑤% 𝑥% ޡࠩٯ఻೻ʹඞཁͳޯ഑͕ඇৗʹ খ͘͞ͳΓɺֶश੍͕ޚͰ͖ͳ͘ ͳΔ໰୊

Slide 182

Slide 182 text

182 ࣌ܥྻσʔλͱਂ૚ֶश 3//ͷར༻ํ๏ l 0OFUP.BOZܕ l ࣌ܥྻੜ੒ͳͲ l .BOZUP0OFܕ l ࣌ܥྻ෼ྨ

Slide 183

Slide 183 text

183 ࣌ܥྻσʔλͱਂ૚ֶश 3//ͷར༻ํ๏ l .BOZUP.BOZܕ l ҟৗ஋ݕ஌ͳͲ l &ODPEFS%FDPEFSܕ l ࣌ܥྻ༧ଌͳͲ

Slide 184

Slide 184 text

184 ࣌ܥྻσʔλͱਂ૚ֶश &ODPEFS%FDPEFS 4FRVFODFUP4FRVFODF l ௕ظ༧ଌɾੜ੒Λߦ͏ͨΊͷϑϨʔϜϫʔΫ l ೖྗܥྻΛ&ODPEFSͱݺ͹ΕΔ3//CBTFEΛ༻͍ͯϕΫτϧʹม׵ &ODPEF ͠ɺ%FDPEFSͱݺ͹ΕΔผͷ3//CBTFEΛ༻͍ͯλʔ ήοτܥྻʹۙ͘ͳΔΑ͏ʹग़ྗ͢ΔϞσϧ

Slide 185

Slide 185 text

185 ࣌ܥྻσʔλͱਂ૚ֶश %FDPEFS΁ͷೖྗ l ੜ੒͞Εͨ΋ͷΛͦͷ··ೖྗ͢Δͱɺޡ͕ࠩͲΜͲΜੵΈॏͳΓ ֶश͕ෆ҆ఆɾऩଋ͕஗͘ͳΔͱ͍͏໰୊

Slide 186

Slide 186 text

186 ࣌ܥྻσʔλͱਂ૚ֶश %FDPEFS΁ͷೖྗ l ਖ਼ղΛͦͷ··ೖྗ͢Δख๏ l ֶश͕҆ఆɾऩଋ͕଎͘ͳΔ͕ɺ܇࿅࣌ͱςετ࣌ͷ෼෍͕ҟͳΔͱ ͍͏໰୊΋ 5FBDIFS'PSDJOH

Slide 187

Slide 187 text

187 ࣌ܥྻσʔλͱਂ૚ֶश %FDPEFS΁ͷೖྗ l ֬཰ʹԠͯ͡ੜ੒σʔλ͔ڭࢣσʔλͷೖྗΛมԽͤ͞Δख๏ 4DIFEVMFE4BNQMJOH from: Bengio, Samy, et al. "Scheduled sampling for sequence prediction with recurrent neural networks." Advances in neural information processing systems 28 (2015).

Slide 188

Slide 188 text

188 ࣌ܥྻσʔλͱਂ૚ֶश %FDPEFS΁ͷೖྗ l ("/Λ༻͍ͯੜ੒σʔλͱڭࢣσʔλͷೖྗʹΑΔੜ੒ͷࠩΛখ͞ ͘͢Δ 1SPGFTTPS4BNQMJOH from: Lamb, Alex M., et al. "Professor forcing: A new algorithm for training recurrent networks." A neural information processing systems 29 (2016).

Slide 189

Slide 189 text

189 ࣌ܥྻσʔλͱਂ૚ֶश $POWPMVUJPOBM/FVSBM/FUXPSL $// l ը૾ͳͲΛଊ͑ΔͨΊʹ༻͍ΒΕΔ/FVSBM/FUXPSL l $POWPMVUJPOBM͸৞ΈࠐΈͱݺ͹Εɺબ୒͞ΕͨϑΟϧλʔαΠζʹ Ԡͯ͡ը૾৘ใΛѹॖ͍ͯ͘͠

Slide 190

Slide 190 text

190 ࣌ܥྻσʔλͱਂ૚ֶश $POWPMVUJPOBM/FVSBM/FUXPSL $// l ը૾ͳͲΛଊ͑ΔͨΊʹ༻͍ΒΕΔ/FVSBM/FUXPSL l $POWPMVUJPOBM͸৞ΈࠐΈͱݺ͹Εɺબ୒͞ΕͨϑΟϧλʔαΠζʹ Ԡͯ͡ը૾৘ใΛѹॖ͍ͯ͘͠ ྖҬΛ৞ΈࠐΜͰѹॖ͢Δख๏Λ ࣌ܥྻσʔλʹԠ༻Ͱ͖ͳ͍͔ʁ

Slide 191

Slide 191 text

191 ࣌ܥྻσʔλͱਂ૚ֶश %JMBUFE$BVTBM$POWPMVUJPO from: https://deepmind.com/blog/article/wavenet-generative-model-raw-audio

Slide 192

Slide 192 text

192 ࣌ܥྻσʔλͱਂ૚ֶश %JMBUFE$BVTBM$POWPMVUJPO from: Oord, A. V. D., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., ... & Kavukcuoglu, K. (2016). Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499. l $//Λ༻͍ͨ࣌ؒ೾ܗ Ի੠ੜ੒ ʹର͢Δ֬཰తੜ੒ϞσϧͰ͋Δ 8BWF/FUͰఏҊ͞Εͨߏ੒ཁૉͷͭɻ l ೖྗ૚͔ΒॱʹͦΕͧΕɺɺɺݸͣͭεΩοϓ͠ͳ͕Β৞Έࠐ ΈΛܭࢉɻ৞ΈࠐΈ૚ͷਂ͞ʹԠͯ͡ɺೖྗͰ͖Δܥྻ௕͕ࢦ਺తʹ ૿Ճ͍ͯ͘͠ɻ l શ݁߹//ͷΑ͏ʹύϥϝʔλ͕ଟ͘ͳ͘ɺ3//ͷΑ͏ʹճؼతͳ઀ ଓ͕ແ͍ͷͰɺֶश͕3//ΑΓ଎͘ͳ͍ͬͯΔ l ௕ظͷܥྻΛೖྗͱͯ͠΋ޯ഑ফࣦ͕ੜ͡ͳ͍

Slide 193

Slide 193 text

193 ࣌ܥྻσʔλͱਂ૚ֶश 5SBOTGPSNFS l l"UUFOUJPOJT"MM:PVOFFEzͰఏҊ l ओʹࣗવݴޠॲཧ΍ը૾ॲཧͳͲʹ͓͍ͯߴ͍ਫ਼౓Λୡ੒͍ͯ͠ΔϞ σϧͷҰछ BERT ViT from: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.

Slide 194

Slide 194 text

194 ࣌ܥྻσʔλͱਂ૚ֶश 5SBOTGPSNFS l 4FMG"UUFOUJPOͱ'FFE'PSXBSEͰओʹߏ੒ l &ODPEFS%FDPEFSϞσϧ l $//ͱൺ΂େҬతͳؔ܎Λଊ͑Δ͜ͱ͕Մೳ

Slide 195

Slide 195 text

195 ࣌ܥྻσʔλͱਂ૚ֶश 5SBOTGPSNFS l 4FMG"UUFOUJPOͱ'FFE'PSXBSEͰओʹߏ੒ l &ODPEFS%FDPEFSϞσϧ l $//ͱൺ΂େҬతͳؔ܎Λଊ͑Δ͜ͱ͕Մೳ l 1PTJUJPOBM&ODPEJOHͱ͍͏֤ೖྗͷ Ґஔ৘ใΛϕΫτϧͰදݱ

Slide 196

Slide 196 text

196 ࣌ܥྻσʔλͱਂ૚ֶश 1PTJUJPOBM&ODPEJOH l 5SBOTGPSNFSͷಛੑ্ɺೖྗܥྻͷҐஔ৘ใ͕ࣦΘΕͯ͠· ͏ͨΊҐஔΛ໌ࣔతʹࣔͨ͠ಛ௃ྔΛೖྗʹՃ͑Δ l 4JOVTPJEBMQPTJUJPOBMFNCFEEJOHT l ೖྗͷ֤UPLFOͷҐஔΛTJOؔ਺ͱDPTؔ਺ʹΑΔϕΫτϧͰදݱ l ֶशແ͠ʹ૬ରతͳҐஔ৘ใΛදݱ

Slide 197

Slide 197 text

197 ࣌ܥྻσʔλͱਂ૚ֶश 1PTJUJPOBM&ODPEJOH from: https://kazemnejad.com/blog/transformer_architecture_positional_encoding/

Slide 198

Slide 198 text

198 ࣌ܥྻσʔλͱਂ૚ֶश 1PTJUJPOBM&ODPEJOH l 5JNF4UBNQ QPTJUJPOBMFNCFEEJOH l ࣌ܥྻσʔλΛೖྗͷࡍʹ೥݄೔࣌ؒͳ Ͳͷ৘ใ͕͋Ε͹ར༻Ͱ͖Δ&NCFEEJOH ͷҰछ l *OGPSNFS΍"VUPGPSNFSͳͲͷ࠷৽ͷ 5SBOTGPSNFSϕʔεͷ࣌ܥྻ༧ଌϞσϧ ʹ༻͍ΒΕ͍ͯΔ

Slide 199

Slide 199 text

199 ࣌ܥྻσʔλͱਂ૚ֶश 5SBOTGPSNFS Transformer Block l 4FMG"UUFOUJPOͱ'FFE'PSXBSEͰओʹߏ੒ l &ODPEFS%FDPEFSϞσϧ l $//ͱൺ΂େҬతͳؔ܎Λଊ͑Δ͜ͱ͕Մೳ l 1PTJUJPOBM&ODPEJOHͱ͍͏֤ೖྗͷ Ґஔ৘ใΛϕΫτϧͰදݱ l &ODPEFS %FDPEFSͱ΋ʹ .VMUJIFBEBUUFOUJPOͱ'FFE'PSXBSEͰ ߏ੒͞Εͨ5SBOTGPSNFS#MPDLͷੵΈॏͶ

Slide 200

Slide 200 text

200 5SBOTGPSNFS#MPDL ࣌ܥྻσʔλͱਂ૚ֶश • &ODPEFSɼ%FDPEFSͱͱ΋ʹ5SBOTGPSNFS #MPDLͷੵΈॏͶͰߏ੒ • 5SBOTGPSNFS#MPDLͷߏ੒ .VMUJ)FBE4FMGBUUFOUJPO 3FTJEVBM$POOFDUJPO ࢒ࠩ઀ଓ -BZFS/PSNBMJ[BUJPO 1PTJUJPOXJTF'FFE'PSXBSE %SPQPVU Transformer Block

Slide 201

Slide 201 text

201 4FMGBUUFOUJPO ࣌ܥྻσʔλͱਂ૚ֶश ⼊⼒系列の 潜在表現 系列⻑ × 次元数 Key : K Query: Q 𝑊y 𝑊z Value: V Attention Map : M 系列⻑ × 系列⻑ 𝑊{ 𝑊xt! Output 𝑠𝑜𝑓𝑡𝑚𝑎𝑥( 𝑄𝐾# 𝑑 )

Slide 202

Slide 202 text

202 4FMGBUUFOUJPO ࣌ܥྻσʔλͱਂ૚ֶश Value: V Attention Map : M 系列⻑ × 系列⻑ 𝑊xt! Output 𝑠𝑜𝑓𝑡𝑚𝑎𝑥( 𝑄𝐾# 𝑑 ) ֤࣌఺ͷಛ௃͕ଞͷ࣌఺Λߟྀ ͯ͠࠶ߏ੒͞ΕΔ

Slide 203

Slide 203 text

203 .VMUJ)FBE4FMGBUUFOUJPO ࣌ܥྻσʔλͱਂ૚ֶश • ઌఔͷ4FMGBUUFOUJPOͷܭࢉΛෳ਺ฒߦ ࣮ͯ͠ߦ • ෳ਺ͷϕΫτϧʹ੾Γ෼͚ͯܭࢉ͢Δ͜ͱ Ͱೖྗؒͷଟ༷ͳྨࣅੑΛൃݟ͠ɼΑΓ ଟ༷ͳදݱྗΛ֫ಘ

Slide 204

Slide 204 text

204 5SBOTGPSNFS#MPDL ࣌ܥྻσʔλͱਂ૚ֶश • &ODPEFSɼ%FDPEFSͱͱ΋ʹ5SBOTGPSNFS #MPDLͷੵΈॏͶͰߏ੒ • 5SBOTGPSNFS#MPDLͷߏ੒ .VMUJ)FBE4FMGBUUFOUJPO 3FTJEVBM$POOFDUJPO ࢒ࠩ઀ଓ -BZFS/PSNBMJ[BUJPO 1PTJUJPOXJTF'FFE'PSXBSE %SPQPVU Transformer Block

Slide 205

Slide 205 text

205 3FTJEVBM$POOFDUJPO ࢒ࠩ઀ଓ ࣌ܥྻσʔλͱਂ૚ֶश Residual𝑂𝑢𝑡𝑝𝑢𝑡 = 𝑥 + 𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛(𝑥) l ೖྗ𝑥Λ.VMUJ)FBE4FMG"UUFOUJPOͷग़ྗʹ Ճ͑Δ͜ͱͰɺֶशͷ҆ఆԽΛਤΔ l 3FT/FUͳͲͷϞσϧͰఏҊ͞Εͨख๏Ͱɺ ޯ഑ফࣦɾരൃ໰୊ΛճආͰ͖Δ

Slide 206

Slide 206 text

206 5SBOTGPSNFS#MPDL ࣌ܥྻσʔλͱਂ૚ֶश • &ODPEFSɼ%FDPEFSͱͱ΋ʹ5SBOTGPSNFS #MPDLͷੵΈॏͶͰߏ੒ • 5SBOTGPSNFS#MPDLͷߏ੒ .VMUJ)FBE4FMGBUUFOUJPO 3FTJEVBM$POOFDUJPO ࢒ࠩ઀ଓ -BZFS/PSNBMJ[BUJPO 1PTJUJPOXJTF'FFE'PSXBSE %SPQPVU Transformer Block

Slide 207

Slide 207 text

207 -BZFS/PSNBMJ[BUJPO ࣌ܥྻσʔλͱਂ૚ֶश 𝐿𝑎𝑦𝑒𝑟𝑁𝑜𝑟𝑚 𝑥 = 𝛾 𝑉𝑎𝑟 𝑥 ∗ 𝑥 − 𝜇 𝑥 + 𝛽 l ֤࣍ݩͷग़ྗΛฏۉ෼ࢄʹਖ਼نԽ͢Δख๏ l ޯ഑ফࣦɾޯ഑രൃΛ཈͑Δ໾ׂ

Slide 208

Slide 208 text

208 5SBOTGPSNFSº *OGMVFO[BGPSFDBTUJOH ࣌ܥྻσʔλͱਂ૚ֶश Wu, N., Green, B., Ben, X., & O'Banion, S. (2020). Deep transformer models for time series forecasting: The influenza prevalence case. arXiv preprint arXiv:2001.08317.

Slide 209

Slide 209 text

209 ਂ૚ֶश WT౷ܭతϞσϧ ࣌ܥྻσʔλͱਂ૚ֶश Wu, N., Green, B., Ben, X., & O'Banion, S. (2020). Deep transformer models for time series forecasting: The influenza prevalence case. arXiv preprint arXiv:2001.08317. .σʔληοτͷTUFQBIFBE GPSFDBTUJOHͷਫ਼౓ݕূ

Slide 210

Slide 210 text

210 ࣌ܥྻλεΫͱ ۙ೥ͷݚڀ

Slide 211

Slide 211 text

211 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ࣌ܥྻ༧ଌ ? ݱࡏ কདྷ ֶश

Slide 212

Slide 212 text

212 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ࣌ܥྻ༧ଌ ݱࡏ কདྷ ֶश

Slide 213

Slide 213 text

213 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ /#FBUT l 'VMMZ$POOFDUFEͰߏ੒͞Εͨϒ ϩοΫͷੵΈ্͛ʹΑΔ࣌ܥྻ༧ଌ Ϟσϧ l ϒϩοΫؒͰ͸࢒ࠩػߏʹΑΔ઀ଓ ͱɺϒϩοΫ಺෦Ͱ͸ઢܗ݁߹Ͱߏ ੒͞Ε͓ͯΓɺղऍੑͷߴ͍Ϟσϧ ͱͳ͍ͬͯΔʢͲͷ෦෼Ͱ൓Ԡͯ͠ ͍Δ͔͕෼͔Δʣ from: Oreshkin, Boris N., et al. "N-BEATS: Neural basis expansion analysis for interpretable time series forecasting." International Conference on Learning Representations. 2019.

Slide 214

Slide 214 text

214 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ঢ়ଶۭؒϞσϧ º ਂ૚ֶश4 l εςοϓҎ্ͷඇৗʹ௕͍γʔέϯεʹରԠͰ͖Δঢ়ଶۭؒϞσ ϧΛ׆༻ͨ͠ख๏ΛఏҊ l ঢ়ଶۭؒϞσϧΛ3FDVSSFOUͱ$POWPMVUJPOBMදݱʹஔ͖׵͑ l )JQQPߦྻΛ3//ʹ૊ΈࠐΉ͜ͱͰ௕ظهԱੑΛ֫ಘ from: Gu, Albert, Karan Goel, and Christopher Re. "Efficiently Modeling Long Sequences with Structured State Spaces." International Conference on Learning Representations. 2021.

Slide 215

Slide 215 text

215 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ਂ૚ֶश º ෼෍ग़ྗ from: Salinas, David, et al. "DeepAR: Probabilistic forecasting with autoregressive recurrent networks." International Journal of Forecasting 36.3 (2020): 1181-1191. l %FFQ"3 l ֬཰తͳ༧ଌΛ࣮ݱ͢ΔͨΊʹෛͷೋ߲໬౓ؔ਺ͱϞϯςΧϧϩαϯϓ Ϧϯάͷग़ྗܗࣜͷಋೖ Training Prediction

Slide 216

Slide 216 text

216 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ਂ૚ֶश º ෼෍ग़ྗ from: Guen, Vincent Le, and Nicolas Thome. "Probabilistic time series forecasting with structured shape and temporal diversity." arXiv preprint arXiv:2010.07349 (2020). l 453*1& l Ψ΢ε෼෍ͳͲͷҰൠతͳ ෼෍Ͱ͸ͳ͘ɺඇఆৗͳ࣌ ܥྻʹ΋ରԠͨ࣌ؒ͠తɺ ܗঢ়త؍఺Ͱॊೈͳ෼෍༧ ଌ͕Մೳͳ&ODPEFS %FDPEFSϞσϧͷఏҊ

Slide 217

Slide 217 text

217 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ਂ૚ֶश º ෼෍ग़ྗ from: de Bézenac, E., Rangapuram, S. S., Benidis, K., Bohlke-Schneider, M., Kurle, R., Stella, L., ... & Januschowski, T. (2020). Normalizing kalman filters for multivariate time series analysis. Advances in Neural Information Processing Systems, 33, 2995-3007. l /PSNBMJ[JOH,BMNBO'JMJUFST l ઢܗΨ΢εঢ়ଶۭؒϞσϧΛ/PSNBMJ[JOHGMPXͰิڧͯ͠ଟมྔ࣌ܥྻ ͷϞσϦϯάΛ࣮ݱ͢Δ l 3FBM/71ͷΞʔΩςΫνϟΛར༻

Slide 218

Slide 218 text

218 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ਂ૚ֶश º ෼෍ग़ྗ from: https://lilianweng.github.io/posts/2018-10-13-flow-models/ l /PSNBMJ[JOH'MPX l ୯७ͳ֬཰ม਺𝑧. ʹରͯ͠ඇઢܗม׵𝑓\ ΛॏͶΔ͜ͱʹΑͬͯɺॊೈͳ෼ ෍𝑝y(𝑍y)Λ֫ಘ͢ΔͨΊͷख๏

Slide 219

Slide 219 text

219 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ෆ౳ִؒ࣌ܥྻ from: Chen, Ricky TQ, et al. "Neural ordinary differential equations." Advances in neural information processing systems 31 (2018). l 0%&3// l &ODPEFS%FDPEFSϞσϧʹ͓͚Δσίʔμͷજࡏม਺ग़ྗΛ0%&/FUʹ ஔ͖׵͑Δ͜ͱͰෆ౳ִؒ࣌ܥྻʹରԠ

Slide 220

Slide 220 text

220 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ෆ౳ִؒ࣌ܥྻ from: Chen, Ricky TQ, et al. "Neural ordinary differential equations." Advances in neural information processing systems 31 (2018). l /FVSBM0SEJOBSZ%JGGFSFOUJBM&RVBUJPO l 3FT/FUͱৗඋ෼ํఔࣜʹྨࣅ఺ʹண໨͠ɺৗඍ෼ํఔࣜͷղ๏Λχϡʔ ϥϧωοτͷදݱʹ༻͍Δख๏ l ͜ΕΛ༻͍Δ͜ͱͰϝϞϦޮ཰ͷߴ͍ ࣌ؒ࿈ଓͳϞσϧΛߏங͢Δ͜ͱ͕Մೳ ͱͳΔ l જࡏදݱ͸0%&4PMWFSʹΑͬͯܭࢉՄೳ

Slide 221

Slide 221 text

221 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ 5SBOTGPSNFSº ࣌ܥྻ༧ଌ l *OGPSNFS l 5SBOTGPSNFS͕௕ظܥྻ༧ଌʹ͸ؤ݈Ͱͳ͍͜ͱ΍ɺܭࢉޮ཰͕ѱ͍͜ ͱΛղܾ͢ΔͨΊʹ4FMGBUUFOUJPOͷܭࢉޮ཰޲্ͷςΫχοΫΛఏҊ from: Zhou, Haoyi, et al. "Informer: Beyond efficient transformer for long sequence time-series forecasting." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. No. 12. 2021. Method Complexity per Layer Convolutional 𝑂 𝐾 ~ 𝐷0 ~ 𝐿 Recurrent 𝑂 𝐿 ~ 𝐷0 Self-attention (Transformer) 𝑂 𝐿0 ~ 𝐷 K: the length of filter D: dimensionality of space L: input length N: Number of layers Computational Complexity 𝑶 𝑵×(𝑳𝟐 ‰ 𝑫)

Slide 222

Slide 222 text

222 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ 5SBOTGPSNFSº ࣌ܥྻ༧ଌ l *OGPSNFS ̎ͭͷ$PNQMFYJUZ࡟ݮख๏ΛఏҊ l 1SPC4QBSTFॏཁ౓ͷߴ͍"UUFOUJPO.BQͷΈΛར༻ 𝑂 𝐿Z ‹ 𝐷 ˠ 𝑂 𝐿 log 𝐿 ‹ 𝐷 ʹ࡟ݮ l 4FMGBUUFOUJPO%JTUJMMJOH ηϧϑΞςϯγϣϯ૚Λग़Δ౓ ܥྻͷ௕͕͞൒෼ʹͳΔΑ͏ৠཹ 𝑂 N ‹ ⋯ ˠ 𝑂 2 − 𝜖 ‹ ⋯ ʹ࡟ݮ

Slide 223

Slide 223 text

223 1SPC4QBSTF ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ ⼊⼒系列の 潜在表現 系列⻑(L) × 次元数(D) Key : K Query: • 𝑄 𝑊y 𝑊z Value: V Attention Map : M 系列⻑ × 系列⻑ 𝑊{ 𝑊xt! Output 𝑠𝑜𝑓𝑡𝑚𝑎𝑥( • 𝑄𝐾# 𝐷 ) 上位u件のQuery のみを利⽤

Slide 224

Slide 224 text

224 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ 5SBOTGPSNFSº ࣌ܥྻ༧ଌ l "VUPGPSNFS l 5SBOTGPSNFS಺෦ʹ࣌ܥྻ෼ղͷػߏΛඋ͑ͨϞσϧ from: Wu, Haixu, et al. "Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting." Advances in Neural Information Processing Systems 34 (2021): 22419-22430.

Slide 225

Slide 225 text

225 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ Ґஔ৘ใ º ࣌ܥྻ༧ଌ l %JGGVTJPO$POWPMVUJPOBM3// l ଟมྔ࣌ܥྻͷؔ܎Λάϥϑͱͯ͠ଊ͑ͯ ಓ࿏ͷಓͳͲ ࣌ܥྻ༧ଌΛ ߦ͏ϞσϧΛఏҊ l (SBQI/FVSBM/FUXPSLͱ3//ͷ૊Έ߹Θͤ from: Li, Yaguang, et al. "Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting." International Conference on Learning Representations. 2018.

Slide 226

Slide 226 text

226 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ (SBQI/FVSBM/FUXPSL l άϥϑσʔλͷֶशͷͨΊʹχϡʔϥϧωοτϫʔΫʹద༻ͨ͠Ϟσϧ l άϥϑ͸ϊʔυͱΤοδͰߏ੒͞ΕΔදݱํ๏Ͱɺϊʔυ΍Τοδͷද ݱΛྡ઀ϊʔυͷ৘ใΛ༻͍ͯߋ৽͍ͯ͘͠

Slide 227

Slide 227 text

227 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ (SBQI/FVSBM/FUXPSL l छྨͷ(// l (SBQI$POWPMVUJPOBM//$//ͷ৞Έ͜Έͷ֓೦Λάϥϑʹద༻

Slide 228

Slide 228 text

228 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ (SBQI/FVSBM/FUXPSL l (SBQI$POWPMVUJPOBM// l $//ͷ৞Έ͜Έͷ֓೦Λάϥϑʹద༻

Slide 229

Slide 229 text

229 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ༧ଌ (SBQI/FVSBM/FUXPSL l (SBQI"UUFOUJPO// l ϊʔυͷྡ઀ʹ͍ͭͯͷॏΈΛ"UUFOUJPOͰදݱ

Slide 230

Slide 230 text

230 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ෼ྨ ࣌ܥྻ෼ྨ from: https://medium.com/@hassanismailfawaz/deep-learning-for-time- series-classification-a-brief-overview-73b58767ed0f ?

Slide 231

Slide 231 text

231 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ෼ྨ 4IBQFMFU l Ϋϥε෼ྨʹ༗ޮͳ೾ܗύλʔϯɻ෼ྨʹ༗ޮͳಛ௃͸࣌ܥྻશମͰ͸ ͳ͘গ਺ͷ෦෼࣌ܥྻʹදΕΔͱ͍͏ΞΠσΞʹجͮ͘ɻ l 4IBQFMFUΛൃݟޙɺߴ଎ʹਪ࿦Ͱ͖Δ from: Grabocka, Josif, et al. "Learning time-series shapelets." Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 2014.

Slide 232

Slide 232 text

232 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ෼ྨ *ODFQUJPO5JNF l ը૾෼ྨʹ༻͍ΒΕΔϑϨʔϜϫʔΫͰ͋Δ*ODFQUJPOΛ࣌ܥྻʹԠ༻ l ࣌ܥྻΛ#PUUMFOFDLͰѹॖ͠ɺ$POWPMVUJPOBMΧʔωϧͰֶश from: Ismail Fawaz, Hassan, et al. "Inceptiontime: Finding alexnet for time series classification." Data Mining and Knowledge Discovery 34.6 (2020): 1936-1962.

Slide 233

Slide 233 text

233 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ෼ྨ 30$,&5 l ཚ਺഑ྻͱ࣌ܥྻσʔλΛ͔͚߹Θͤͨಛ௃ྔΛେྔʹੜ੒͠ɺੜ੒͞ Εͨܥྻͷಛ௃ʢ࠷େ஋ͱ಺ੵ͕ਖ਼Ͱ͋ͬͨλΠϜ΢Οϯυ΢ͷׂ߹ʣ Λ༻͍Δ͜ͱͰ࣌ܥྻ෼ྨΛֶश from: Dempster, Angus, François Petitjean, and Geoffrey I. Webb. "ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels." Data Mining and Knowledge Discovery 34.5 (2020): 1454-1495. 畳み込み Time Window 1.23 2.34 Time Window分の特徴を⽣成 ・・・ 特徴量を作成 最⼤: 2.34 正の割合: 0.21 学習

Slide 234

Slide 234 text

234 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ෼ྨ 5-PTT l ଛࣦؔ਺Λڑ཭ֶशʹ͢Δ͜ͱͰڭࢣͳ͠Ͱ࣌ܥྻͷಛ௃ྔΛֶश l Ϟσϧࣗମ͸%JMBUFE$//Λར༻ Franceschi, Jean-Yves, Aymeric Dieuleveut, and Martin Jaggi. "Unsupervised scalable representation learning for multivariate time series." Advances in neural information processing systems 32 (2019).

Slide 235

Slide 235 text

235 ࣌ܥྻλεΫͱۙ೥ͷݚڀ࣌ܥྻ෼ྨ ڑ཭ֶश .FUSJD-FBSOJOH l ࣅͨαϯϓϧಉ࢜ͷಛ௃ྔ͸͚ۙͮɺࣅ͍ͯͳ ͍ಛ௃ྔͷڑ཭͸ԕ͚͟Δͱ͍͏ૢ࡞Λଛࣦؔ ਺Λ௨ֶͯ͡श͢Δख๏ l ୅දతͳڑ཭ֶश༻ͷଛࣦؔ਺ͱͯ͠ɺ5SJQMFU -PTT΍$POUSBTUJWF-PTTͳͲ͕ଘࡏ

Slide 236

Slide 236 text

236 ࣌ܥྻλεΫͱۙ೥ͷݚڀҟৗݕ஌ ҟৗݕ஌ มԽݕ஌ From: https://neptune.ai/blog/anomaly-detection-in-time-series ҟৗʂ ҟৗʂ

Slide 237

Slide 237 text

237 ࣌ܥྻλεΫͱۙ೥ͷݚڀҟৗݕ஌ -PDBM0VUMJFS'BDUPS -0' l ࣗ਎ͷ఺͔Βۙ๣Lݸͷ఺ͱ͍͔ʹີ͔Ͱ͋Δ͔ࣔ͢ہॴີ౓ -PDBM EFOTJUZ Λ༻͍ͨ֎Ε஋ͷݕग़

Slide 238

Slide 238 text

238 ࣌ܥྻλεΫͱۙ೥ͷݚڀҟৗݕ஌ &ODPEFS%FDPEFSʹΑΔѹॖ from: Li, Shu, et al. "Fair Outlier Detection Based on Adversarial Representation Learning." Symmetry 14.2 (2022): 347.

Slide 239

Slide 239 text

239 ࣌ܥྻλεΫͱۙ೥ͷݚڀҟৗݕ஌ *TPMBUJPO'PSFTU from: https://qiita.com/tchih11/items/d76a106e742eb8d92fb4

Slide 240

Slide 240 text

240 ࣌ܥྻλεΫͱۙ೥ͷݚڀҟৗݕ஌ .BUSJY1SPGJMF from: https://towardsdatascience.com/introduction-to-matrix-profiles-5568f3375d90

Slide 241

Slide 241 text

241 ࣌ܥྻλεΫͱۙ೥ͷݚڀҟৗݕ஌ .BUSJY1SPGJMF

Slide 242

Slide 242 text

242 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά ϞσϦϯά モデル

Slide 243

Slide 243 text

243 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά 8FCσʔλͷϞσϦϯά from: Murayama, Taichi, Yasuko Matsubara, and Sakurai Yasushi. "Mining Reaction and Diffusion Dynamics in Social Activities." arXiv preprint arXiv:2208.04846 (2022). l ֦ࢄ൓ԠํఔࣜͱχϡϥʔϧωοτϫʔΫͷ૊Έ߹Θͤ

Slide 244

Slide 244 text

244 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά 4/4ؒͷϑϩʔ l ଟมྔ)BXLFTաఔΛ༻͍ͨ4/4ؒͷྲྀΕͷՄࢹԽ from: Zannettou, Savvas, et al. "The web centipede: understanding how web communities influence each other through the lens of mainstream and alternative news sources." Proceedings of the 2017 internet measurement conference. 2017.

Slide 245

Slide 245 text

245 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά 1PJOU1SPDFTT ఺աఔ l ΠϕϯτσʔλͷϞσϦϯάʹ༻͍ΒΕΔ౷ܭతϞσϧ l Ұఆظؒͷۭؒɾ࣌ؒɾΠϕϯτͷੑ࣭ΛϞσϦϯά l 1PJTTPOաఔ l ۚ༥΍஍਒ͳͲͷ֬཰తʹൃੜ͢Δࣄ৅ͷϞσϦϯάʹ༻͍ΒΕΔ l )BXLFTաఔ l ࣗ෼ͷաڈͷΠϕϯτʹґଘ ͠ TFMGFYDJUJOH ͳੑ࣭Λ࣋ͪɺ 4/4ͷόʔετͳͲΛϞσϦϯά

Slide 246

Slide 246 text

246 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά ਓؒͷߦಈʹ͍ͭͯͷϞσϦϯά l ਓؒͷߦಈɺӡಈɺਭ຾ɺ৯ࣄͳͲΛه࿥͢Δ.PCJMFIFBMUI BQQMJDBUJPOΛ׆༻ͨ͠ߦಈͷϞσϦϯά l (BVTT෼෍ͷࠞ߹ʹΑΔ*OUFOTJUZͰਓؒͷߦಈ܏޲΍)BXLFT1SPDFTT Ͱपظతͳ૬ޓґଘͷؔ܎Λଊ͑Δ from: Kurashima, Takeshi, Tim Althoff, and Jure Leskovec. "Modeling interdependent and periodic real-world action sequences." Proceedings of the 2018 world wide web conference. 2018.

Slide 247

Slide 247 text

247 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά ਓؒͷձ࿩ʹ͍ͭͯͷϞσϦϯά l ʮਓͷౖΕΔஉʯʹ͓͚Δܶதձ࿩ͷϞσϦϯά from: Guo, Fangjian, et al. "The bayesian echo chamber: Modeling social influence via linguistic accommodation." Artificial Intelligence and Statistics. PMLR, 2015.

Slide 248

Slide 248 text

248 ࣌ܥྻλεΫͱۙ೥ͷݚڀϞσϦϯά ϑΣΠΫχϡʔεʹ͍ͭͯͷϞσϦϯά l ϑΣΠΫχϡʔεͷ֦ࢄͰ͸ɺ֦ࢄλΠϛϯάͱमਖ਼λΠϛϯάͷͭͷ ظ͕ؒଘࡏͦ͠ΕͧΕόʔετݱ৅͕ੜ͡Δ͜ͱΛ)BXLFT1SPDFTTͰ ϞσϦϯά from: Murayama, Taichi, et al. "Modeling the spread of fake news on Twitter." Plos one 16.4 (2021): e0250419.

Slide 249

Slide 249 text

249 લ൒෦෼ͰओͳϕʔεͱͳͬͯΔॻ੶ l ΢ΥϧλʔΤϯμʔεʮ࣮ূͷͨΊͷܭྔ࣌ܥྻ෼ੳʯ l ๺઒ݯ࢛࿠ ʮ࣌ܥྻղੳೖ໳ʯ l ԭຊཽٛ ʮܦࡁɾϑΝΠφϯεσʔλͷܭྔ࣌ܥྻ෼ੳʯ ࢀߟॻ

Slide 250

Slide 250 text

250 ͦͷଞ l খ੢ఃଇɼ๺઒ݯ࢛࿠ʮ৘ใྔن४ʯ l അ৔ޱ ొ தଜ ࿨ߊʮ৽͍͠৴߸ॲཧͷڭՊॻʯ l Ҫख ߶ ʮೖ໳ ػցֶशʹΑΔҟৗݕ஌ʕ3ʹΑΔ࣮ફΨΠυʯ l ۙߐ ਸ޺ ໺ଜ ढ़Ұʮ఺աఔͷ࣌ܥྻղੳʯ l "JMFFO/JFMTFOʮ࣮ફ ࣌ܥྻղੳ ʕ౷ܭͱػցֶशʹΑΔ༧ ଌʯ l ਺ཧख๏ᶝ ࣌ܥྻղੳ ౦ژେֶʮ਺ཧɾσʔλαΠΤϯεڭ ҭϓϩάϥϜʯ ࢀߟॻ