基盤モデルのアーキテクチャを改造してみよう - 時系列基盤モデルのマルチモーダル拡張事例の紹介 -

by himura467

Slide 1

Slide 1 text

IJNVSB ج൫ϞσϧͷΞʔΩςΫνϟΛվ଄ͯ͠ΈΑ͏ ࣌ܥྻج൫ϞσϧͷϚϧνϞʔμϧ֦ுࣄྫͷ঺հ 1

Slide 2

Slide 2 text

ࣗݾ঺հ Self Introduction ઃָ࿕ਓ ژ౎େֶେֶӃ৘ใֶݚڀՊम࢜೥ɻ ػցֶशΛ༻͍ͨ࣌ܥྻ༧ଌʹؔ͢ΔݚڀΛ͍ͯ͠·͢ɻ Ұਓ฻Β͠ྺ̑೥໨ɻ ཛͰͱ͡Δܥͷྉཧ͕޷͖ɻ ํ޲Իஒɻ 2 Akito Shitara 5ZQF4DSJQU 1ZUIPO 1FSM Α͘ॻ͘ݴޠ Frequently used programming languages ;JH GitHub: @himura467 X: @himuhimu467

Slide 3

Slide 3 text

όઌͷ঺հ Basaki Introduction ϥΠϑΠζςοΫ 3 Life is Tech! தߴੜʹ ϓϩάϥϛϯάΛ ڭ͑ΔεΫʔϧͷ ϝϯλʔΛ͍ͯ͠·͢

Slide 4

Slide 4 text

೥Ն In the summer of 2024 5 4 ੜెͷ཭୤༧ଌΛ ߦ͍͍ͨ

Slide 5

Slide 5 text

೥Ն In the summer of 2024 5 5 ྲྀߦΓͷਂ૚ֶशʹ͸ σʔλ͕ඞཁෆՄܽ

Slide 6

Slide 6 text

೥Ն In the summer of 2024 5 6 ͕ɺσʔλ͕ͳ͍ ྲྀߦΓͷਂ૚ֶशʹ͸ σʔλ͕ඞཁෆՄܽ

Slide 7

Slide 7 text

೥Ն In the summer of 2024 5 7 Ͳ͏͠Α͏ʜ

Slide 8

Slide 8 text

ٹੈओݱΔ The savior has arrived 5 8 ࣌ܥྻج൫Ϟσϧ a:P

Slide 9

Slide 9 text

ͪͳΈʹ By the way 5 9 ʮ࣌ܥྻج൫Ϟσϧʯ ฉ͍ͨ͜ͱ͋Δํʔʁ a͸ʔ͍

Slide 10

Slide 10 text

5 10 ೔ຊͰ࠷΋ߴ͍ࢁ͸ ෋࢜ࢁ Common Crawl GitHub Wikipedia େن໛ݴޠϞσϧͷ֓ཁ About the concept of Large Language Models

Slide 11

Slide 11 text

5 11 ࣌ܥྻج൫Ϟσϧͷ֓ཁ About the concept of Time Series Foundation Model Earthquake Data Medical Data Temperature աڈͷגՁͷਪҠ ະདྷͷגՁ

Slide 12

Slide 12 text

5 12 ࣌ܥྻج൫Ϟσϧͷ֓ཁ About the concept of Time Series Foundation Model Earthquake Data Medical Data Temperature աڈͷגՁͷਪҠ ະདྷͷגՁ ৽نͷυϝΠϯʹରͯ͠΋΄Ͳ΄Ͳͷ༧ଌΛͯ͘͠ΕΔ

Slide 13

Slide 13 text

5 13 ࣌ܥྻج൫ϞσϧͳΒ σʔλ͕গͳͯ͘΋༧ଌՄೳ ࣌ܥྻج൫Ϟσϧͷ֓ཁ About the concept of Time Series Foundation Model

Slide 14

Slide 14 text

࣌ܥྻج൫ϞσϧΛࢼͯ͠ΈΔ Try Time Series Foundation Model 5 14 όΠτઌͷσʔλͰ ϑΝΠϯνϡʔχϯάͯ͠ ަࠩݕূͯ͠ΈΔ

Slide 15

Slide 15 text

ަࠩݕূͷ݁Ռ Cross-Validation Results 5 15 ϥϯμϜʹճ࣮ߦͨ݁͠ՌͰ͢ "3*."Ϟσϧͷύϥϝʔλ͸࠷దԽ͞Ε͍ͯ·ͤΜ σʔλ਺ͳͲॻ͖͖Ε͍ͯͳ͍৚͕݅ଞʹ΋ͨ͘͞Μ͋Γ·͢ 'JOFUVOFE 0SJHJOBM5JNFT'. "3*." .4& ."&

Slide 16

Slide 16 text

5 16 ैདྷख๏Λ্ճΔਫ਼౓Λ ग़͢͜ͱ͕Ͱ͖ͨ ަࠩݕূͷ݁Ռ Cross-Validation Results

Slide 17

Slide 17 text

5 17 ͜͜·Ͱ͕લ࠲Ͱ͢

Slide 18

Slide 18 text

͔͜͜Β͕ຊ୊Ͱ͢ The main discussion begins 5 18 Ұ౓ຊདྷͷ໨తΛࢥ͍ग़͢

Slide 19

Slide 19 text

ຊདྷͷ໨త The original purpose 5 19 ੜెͷ཭୤༧ଌΛ ߦ͍͍ͨ

Slide 20

Slide 20 text

5 20 ੜెͷ཭୤༧ଌΛ ߦ͍͍ͨ ཭୤Λ๷ࢭ͍ͨ͠ ຊདྷͷ໨త The original purpose

Slide 21

Slide 21 text

࣮ݧ࣌ʹσʔλ͔ΒಘΒΕͨࣔࠦ Insights gleaned from data 5 21 ग़੮܏޲͕Ұ౓མͪ࢝ΊΔͱ ໭Βͳ͍͕ͪ ܽ੮͍ͯ͠ΔؒʹίϛϡχςΟ͕ৢ੒͞Εͯ͠·͍ɺૄ֎ײΛײͯ͡͠·ͬͨΓʜʁ

Slide 22

Slide 22 text

ग़੮܏޲ͷมԽ Πϝʔδ Attendance Trend Graph 5 22 ग़੮೔਺ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄

Slide 23

Slide 23 text

5 23 ग़੮೔਺ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ख஗Ε ग़੮܏޲ͷมԽ Πϝʔδ Attendance Trend Graph

Slide 24

Slide 24 text

5 24 ग़੮೔਺ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ͜͜ͰΞϥʔτΛग़ͯ͠΄͍͠ ग़੮܏޲ͷมԽ Πϝʔδ Attendance Trend Graph

Slide 25

Slide 25 text

࣌ܥྻ༧ଌͷݶք Limitations of Time Series Forecasting 5 25 ͍͘Βਫ਼౓্͕͕ͬͯ΋ ࣌ܥྻͷΈΛઆ໌ม਺ͱͨ͠ ༧ଌͰ͸ݪཧతʹ࣮ݱෆՄೳ

Slide 26

Slide 26 text

࣌ܥྻج൫Ϟσϧͷ֦ு Extending Time-Series Foundation Models 26

Slide 27

Slide 27 text

࣌ܥྻج൫Ϟσϧ 5JNFT'. ͷߏ଄ The architecture of TimesFM 5 27 ࣌ܥྻ ݄ ݄ Residual Block Vector MSE: Loss Function ͜͜Λֶश͢Δ Stacked Transformer Ref: Abhimanyu Das, Weihao Kong, Rajat Sen, Yichen Zhou. A decoder-only foundation model for time-series forecasting. ICML 2024

Slide 28

Slide 28 text

ϚϧνϞʔμϧ֦ுͯ͠ΈΔ Try multi-modal extension 5 28 ʮࠓ݄͸ؤுͬͨʂʯ ʮਐḿඍົ͔΋ʜʯ ςΩετܥྻ ࣌ܥྻ ݄ ݄ Residual Block Vector Fusion Module MSE: Loss Function ͜͜Λֶश͢Δ Vector Text Encoder Stacked Transformer https://github.com/himura467/multimodal-timesfm

Slide 29

Slide 29 text

5 29 ϥϯμϜʹճ࣮ߦͨ݁͠ՌͰ͢ "3*."Ϟσϧͷύϥϝʔλ͸࠷దԽ͞Ε͍ͯ·ͤΜ σʔλ਺ͳͲॻ͖͖Ε͍ͯͳ͍৚͕݅ଞʹ΋ͨ͘͞Μ͋Γ·͢ .VMUJNPEBM 'JOFUVOFE 0SJHJOBM5JNFT'. "3*." .4& ."& ަࠩݕূͷ݁Ռ Cross-Validation Results

Slide 30

Slide 30 text

5 30 ఔʑʹ΍Ε͍ͯͦ͏ ަࠩݕূͷ݁Ռ Cross-Validation Results .VMUJNPEBM͸ج൫ϞσϧࣗମͷύϥϝʔλΛౚ݁ͨ͠ঢ়ଶͰਫ਼౓ͷ޲্͕ݟΒΕ͍ͯΔ

Slide 31

Slide 31 text

5 31 ࣌ܥྻج൫Ϟσϧͷັྗ The Appeal of Time-Series Foundation Models ͍ͭઌ݄຤ʹ"NB[POͷ࣌ܥྻج൫Ϟσϧ͕ ϝδϟʔΞοϓσʔτΛܴ͑ΔͳͲ ೔ਐ݄าͰਐԽ͍ͯ͠ΔݚڀྖҬ

Slide 32

Slide 32 text

5 32 ʒ Conclusion l#FU࣌ܥྻج൫Ϟσϧz ͯ͠Έ·ͤΜ͔ʁ