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 -

YAPC::Fukuoka 2025 における LT の資料です

Transcript

1. IJNVSB ج൫ϞσϧͷΞʔΩςΫνϟΛվ଄ͯ͠ΈΑ͏
   
   ࣌ܥྻج൫ϞσϧͷϚϧνϞʔμϧ֦ுࣄྫͷ঺հ
    1
2. ࣗݾ঺հ Self Introduction ઃָ
   ࿕ਓ ژ౎େֶେֶӃ
   ৘ใֶݚڀՊ
   म࢜೥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
3. όઌͷ঺հ Basaki Introduction ϥΠϑΠζςοΫ 3 Life is Tech! தߴੜʹ
   ϓϩάϥϛϯάΛ
   ڭ͑ΔεΫʔϧͷ
   ϝϯλʔΛ͍ͯ͠·͢
4. ೥Ն In the summer of 2024 5 4 ੜెͷ཭୤༧ଌΛ
   ߦ͍͍ͨ
5. ೥Ն In the summer of 2024 5 5 ྲྀߦΓͷਂ૚ֶशʹ͸
   σʔλ͕ඞཁෆՄܽ
6. ೥Ն In the summer of 2024 5 6 ͕ɺσʔλ͕ͳ͍ ྲྀߦΓͷਂ૚ֶशʹ͸
   σʔλ͕ඞཁෆՄܽ
7. ೥Ն In the summer of 2024 5 7 Ͳ͏͠Α͏ʜ
8. ٹੈओݱΔ The savior has arrived 5 8 ࣌ܥྻج൫Ϟσϧ a:P
9. ͪͳΈʹ By the way 5 9 ʮ࣌ܥྻج൫Ϟσϧʯ
   ฉ͍ͨ͜ͱ͋Δํʔʁ a͸ʔ͍
10. 5 10 ೔ຊͰ࠷΋ߴ͍ࢁ͸ ෋࢜ࢁ Common Crawl GitHub Wikipedia େن໛ݴޠϞσϧͷ֓ཁ About the concept of Large Language Models
11. 5 11 ࣌ܥྻج൫Ϟσϧͷ֓ཁ About the concept of Time Series Foundation Model Earthquake Data Medical Data Temperature աڈͷגՁͷਪҠ ະདྷͷגՁ
12. 5 12 ࣌ܥྻج൫Ϟσϧͷ֓ཁ About the concept of Time Series Foundation Model Earthquake Data Medical Data Temperature աڈͷגՁͷਪҠ ະདྷͷגՁ ৽نͷυϝΠϯʹରͯ͠΋΄Ͳ΄Ͳͷ༧ଌΛͯ͘͠ΕΔ
13. 5 13 ࣌ܥྻج൫ϞσϧͳΒ
    σʔλ͕গͳͯ͘΋༧ଌՄೳ ࣌ܥྻج൫Ϟσϧͷ֓ཁ About the concept of Time Series Foundation Model
14. ࣌ܥྻج൫ϞσϧΛࢼͯ͠ΈΔ Try Time Series Foundation Model 5 14 όΠτઌͷσʔλͰ
    ϑΝΠϯνϡʔχϯάͯ͠
    ަࠩݕূͯ͠ΈΔ
15. ަࠩݕূͷ݁Ռ Cross-Validation Results 5 15 ϥϯμϜʹ
    
    ճ࣮ߦͨ݁͠ՌͰ͢
    "3*."
    Ϟσϧͷύϥϝʔλ͸࠷దԽ͞Ε͍ͯ·ͤΜ
    σʔλ਺ͳͲॻ͖͖Ε͍ͯͳ͍৚͕݅ଞʹ΋ͨ͘͞Μ͋Γ·͢ 'JOFUVOFE 0SJHJOBM
    5JNFT'. "3*." .4&    ."&   
16. 5 16 ैདྷख๏Λ্ճΔਫ਼౓Λ
    ग़͢͜ͱ͕Ͱ͖ͨ ަࠩݕূͷ݁Ռ Cross-Validation Results
17. 5 17 ͜͜·Ͱ͕લ࠲Ͱ͢
18. ͔͜͜Β͕ຊ୊Ͱ͢ The main discussion begins 5 18 Ұ౓ຊདྷͷ໨తΛࢥ͍ग़͢
19. ຊདྷͷ໨త The original purpose 5 19 ੜెͷ཭୤༧ଌΛ
    ߦ͍͍ͨ
20. 5 20 ੜెͷ཭୤༧ଌΛ
    ߦ͍͍ͨ ཭୤Λ๷ࢭ͍ͨ͠ ຊདྷͷ໨త The original purpose
21. ࣮ݧ࣌ʹσʔλ͔ΒಘΒΕͨࣔࠦ Insights gleaned from data 5 21 ग़੮܏޲͕Ұ౓མͪ࢝ΊΔͱ
    ໭Βͳ͍͕ͪ ܽ੮͍ͯ͠ΔؒʹίϛϡχςΟ͕ৢ੒͞Εͯ͠·͍ɺૄ֎ײΛײͯ͡͠·ͬͨΓʜʁ
22. ग़੮܏޲ͷมԽ
    Πϝʔδ Attendance Trend Graph 5 22 ग़੮೔਺      ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄
23. 5 23 ग़੮೔਺      ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ख஗Ε ग़੮܏޲ͷมԽ
    Πϝʔδ Attendance Trend Graph
24. 5 24 ग़੮೔਺      ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ݄ ͜͜ͰΞϥʔτΛग़ͯ͠΄͍͠ ग़੮܏޲ͷมԽ
    Πϝʔδ Attendance Trend Graph
25. ࣌ܥྻ༧ଌͷݶք Limitations of Time Series Forecasting 5 25 ͍͘Βਫ਼౓্͕͕ͬͯ΋
    ࣌ܥྻͷΈΛઆ໌ม਺ͱͨ͠
    ༧ଌͰ͸ݪཧతʹ࣮ݱෆՄೳ
26. ࣌ܥྻج൫Ϟσϧͷ֦ு Extending Time-Series Foundation Models 26
27. ࣌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
28. 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
29. 5 29 ϥϯμϜʹ
    
    ճ࣮ߦͨ݁͠ՌͰ͢
    "3*."
    Ϟσϧͷύϥϝʔλ͸࠷దԽ͞Ε͍ͯ·ͤΜ
    σʔλ਺ͳͲॻ͖͖Ε͍ͯͳ͍৚͕݅ଞʹ΋ͨ͘͞Μ͋Γ·͢ .VMUJNPEBM 'JOFUVOFE 0SJHJOBM
    5JNFT'. "3*." .4&     ."&     ަࠩݕূͷ݁Ռ Cross-Validation Results
30. 5 30 ఔʑʹ΍Ε͍ͯͦ͏ ަࠩݕূͷ݁Ռ Cross-Validation Results .VMUJNPEBM
    ͸ج൫ϞσϧࣗମͷύϥϝʔλΛౚ݁ͨ͠ঢ়ଶͰਫ਼౓ͷ޲্͕ݟΒΕ͍ͯΔ
31. 5 31 ࣌ܥྻج൫Ϟσϧͷັྗ The Appeal of Time-Series Foundation Models ͍ͭઌ݄຤ʹ"NB[POͷ࣌ܥྻج൫Ϟσϧ͕ ϝδϟʔΞοϓσʔτΛܴ͑ΔͳͲ
    ೔ਐ݄าͰਐԽ͍ͯ͠ΔݚڀྖҬ
32. 5 32 ʒ Conclusion l#FU
    ࣌ܥྻج൫Ϟσϧz
    ͯ͠Έ·ͤΜ͔ʁ