Upgrade to Pro — share decks privately, control downloads, hide ads and more …

AI研究者になる-学生時代の経験から最先端の研究まで- / 2023-09-27

Shoya Ishimaru
September 27, 2023
32

AI研究者になる-学生時代の経験から最先端の研究まで- / 2023-09-27

2023年9月27日に愛媛県立今治東中等教育学校で行った講演の資料です。

Shoya Ishimaru

September 27, 2023
Tweet

More Decks by Shoya Ishimaru

Transcript

  1. ੴؙᠳ໵ࣗݾ঺հ  • Ѫඤݝཱদࢁ੢த౳ڭҭֶߍଔۀ • େࡕ෎ཱେֶ޻ֶ෦ଔۀ ֶ࢜೥  • େࡕ෎ཱେֶେֶӃ޻ֶݚڀՊमྃ

    म࢜೥  • ΧΠβʔεϥ΢ςϧϯ޻Պେֶमྃ ത࢜೥  • %',*υΠπਓ޻஌ೳݚڀηϯλʔ ϙευΫ೥  • ΧΠβʔεϥ΢ςϧϯ޻Պେֶ δϡχΞڭत೥  • ೥݄͔Βେࡕެཱେֶ ಛ೚ڭत ߴߍੜͷࠒ X(PPHMF(MBTT X1I%)BU
  2. ਓ޻஌ೳ "SUJ fi DJBM*OUFMMJHFODF ͱ͸ԿͩΖ͏͔  νϟοτϘοτ͸"*ʁ ࣗಈӡసं͸"*ʁ ࣗಈυΞ͸"* ԿΛ΋ͬͯਓ޻஌ೳͱݺͿ͔͸ݚڀऀͷؒͰ΋ᐆດ

    ൚༻ɺಛԽɺFUD OBCDPTZTUFN ࠓ೔ͷߨԋͰ͸ʮ஌ੑΛײ͡ΒΕΔٕज़ʯΛਓ޻஌ೳͱ͠·͢ ֶͿɾߟ͑Δɾ఻͑Δೳྗ
  3. ͜͜਺೥Ͱ"*ݚڀ͕ٸ଎ʹൃలͨ͠ཧ༝͸ʁ  :FT /P  ೖྗ ग़ྗ ೖྗ ग़ྗ ೖྗ

    ग़ྗ ਓؒͷ஌ࣝΛϚχϡΞϧԽ ೴ͷਆܦωοτϫʔΫΛ࠶ݱ ར఺ ܽ఺ ࣮૷ͱཧղ͕ൺֱత؆୯ ϚχϡΞϧʹͳ͍໰୊͸ղ͚ͳ͍ ๲େͳֶशσʔλͱܭࢉࢿݯ͕ඞཁ ະ஌ͷ໰୊΋ֶशͯ͠ղ͚ΔΑ͏ʹ ͜ΕΒ͕ἧͬͨ
  4. ࠷৽ͷจॻੜ੒"*͸ԯΛ௒͑ΔύϥϝʔλΛ࣋ͭ  Zhao, Wayne Xin, et al. "A survey of

    large language models." arXiv preprint arXiv:2303.18223 (2023).
  5. ಛʹ͜͜਺ϲ݄Ͱจॻੜ੒"*͕ൃలͨ͠ཧ༝͸ʁ  • େن໛ݴޠϞσϧ -BSHF-BOHVBHF.PEFM ͷొ৔ • ݴޠϞσϧͱ͸ɺςΩετͷଓ͖Λ༧ଌ͢Δ֬཰Ϟσϧ ʮ೔ຊͷट౎͸ʯ ๺ژ

     ౦ژ  ژ౎  ӳࠃͷट౎͸ϩϯυϯͰݕࡧ͢Δͱ݅ ӳࠃͷट౎͸Ͱݕࡧ͢Δͱ ݅ P(ϩϯυϯ|ӳࠃ, ͷ, ट౎, ͸) = 8 21,700 ֬཰͕ΑΓߴ͍୯ޠΛ୳͢ ֬཰͸େྔͷจॻ ίʔύε ͔Βܭࢉ ୯ͳΔ ୯ޠ༧ଌػΛνϡʔχϯά͢Δͱ༷ʑͳ໰୊Λղ͚Δ͜ͱ͕෼͔ͬͨ Ԭ࡚, େن໛ݴޠϞσϧͷڻҟͱڴҖ, AIPγϯϙδ΢Ϝ੒Ռใࠂձ, 2023 https://speakerdeck.com/chokkan/20230327_riken_llm ΑΓൈਮ͠վม
  6. ಛʹ࿩୊ʹͳ͍ͬͯΔ$IBU(15ͱ͸Կʁ  • େن໛ݴޠϞσϧ(15 (FOFSBUJWF1SFUSBJOFE5SBOTGPSNFS Λ֦ு • ༩͑ΒΕͨࢦࣔ ϓϩϯϓτ ʹैͬͯԠ౴Λฦ͢Α͏ɺ

    
 ਓؒͷϑΟʔυόοΫͱڧԽֶश͕૊Έࠐ·Ε͍ͯΔ L Ouyang, J Wu, X Jiang, D Almeida, et. al. 2022. Training Language Models to Follow Instructions with Human Feedback. arXiv:2203.02155. (15ͷճ౴ ϓϩϯϓτ Ϣʔβʔ ཧ૝తͳճ౴ 0QFO"*ࣾͷ Ξϊςʔλʔ ֶश ޻෉ᶃ4VQFSWJTFE fi OFUVOJOH ޻෉ᶄ3FXBSENPEFM 0QFO"*ࣾͷ Ξϊςʔλʔ (15ͷճ౴ (15ͷճ౴ (15ͷճ౴ ճ౴ϥϯΩϯά ֶश (15ͷճ౴
  7. "*ݚڀऀͬͯͲΜͳ࢓ࣄʁ  • "*ΛΑΓݡ͘͢ΔͨΊͷΞϧΰϦζϜ ࢉ๏ ΛߟҊͨ͠Γ 
 "*ٕज़Λ༷ʑͳ෼໺ ҩྍɺڭҭɺ ʹԠ༻ͨ͠Γ͢Δ࢓ࣄ

    • ࢓ࣄͷ໨ඪ͸ɺݚڀ੒ՌΛ࿦จ΍੡඼ͱͯ͠ൃද͢Δ͜ͱ • ۈΊઌ͸େֶ ڭҭʹॏ఺ ͔Βاۀ ݚڀʹॏ఺ ·Ͱ༷ʑ • Ͳ͏͢Ε͹ͳΕΔʁത࢜߸ΛऔΔͷ͕Ұൠత ւ֎Ͱ͸ಛʹ  • େֶ ֶ࢜೥  େֶӃ म࢜೥ ത࢜೥ Ͱऔಘ͢Δ
  8. ݚڀ׆ಈ  • ݚڀςʔϚΛߟ͑Δ • ࿦จΛಡΜͰઌߦݚڀΛௐࠪ͢Δ • ࣮ݧ͢Δ ৘ใֶݚڀͷ৔߹͸γεςϜͷ࣮૷ͱධՁ 

    • ࿦จΛॻ͘ • ֶձͰݚڀ੒ՌΛ఻͑Δ • Ճ͑ͯɺݚڀࢦಋɾतۀɾֶձӡӦɾֶ಺༻຿ͳͲ
  9. ਎ମతߦಈ e.g., walking, standing, cycling, sleeping ೝ஌తߦಈ e.g., reading, writing,

    memorizing, talking ηϯαͱAIͷ૊Έ߹ΘͤͰਓͷߦಈ΍ঢ়ଶΛਪఆ͢Δ ৺ཧతঢ়ଶ
 e.g., interest, workload, con fi dence, fatigue
  10. ݚڀ঺հʮֶͿྗʯͷ֦ு  ڵຯ ཧղ౓ ೝ஌ෛՙ ΞΠτϥοΩϯά αʔϞάϥϑ ಡΈฦ͠ ඓ෦ද໘Թ౓ ஫ࢹ

    S. Ishimaru, et al. Cognitive State Measurement on Learning Materials by Utilizing Eye Tracker and Thermal Camera. Proc. ICDAR HDI 2017, pp. 32–36, 2017. S. Ishimaru, et al. Augmented Learning on Anticipating Textbooks with Eye Tracking. Positive Learning in the Age of Information (PLATO), pp. 387–398, 2018.
  11. ݚڀ঺հʮߟ͑Δྗʯͷ֦ு  S. Ishimaru, et al. The Wordometer 2.0: Estimating

    the Number of Words You Read in Real Life using Commercial EOG Glasses. Proc. UbiComp 2016 Adjunct, pp. 293–296, 2016. S. Ishimaru, et al. Reading Interventions: Tracking Reading State and Designing Interventions. Proc. UbiComp 2016 Adjunct, pp. 1759–1764, 2016. # 3 - Electrodes Horizontal axis: L - R [mV] Vertical axis: B - (L + R)/2 [mV] ؟ిҐ͔Βࢹઢํ޲Λਪఆ 100% 9:41 AM ػցֶशͰʮಡΜͩ୯ޠͷ਺ʯΛਪఆ Ϣʔβʔʹఏࣔ
  12. 52% 19% 0% 47% 75% 93% 0% 5% 6% Predicted

    class low middle high Actual class high middle low BDDVSBDZ S. Ishimaru and K. Kise.“Quantifying the Mental State on the Basis of Physical and Social Activities”. Proc. UbiComp '15 Adjunct, pp. 1217–1220, 2015. ։ൃͷ͖͔͚ͬᶃ ࣗ਎ͷෆௐ͔ΒίϯσΟγϣϯͱηϯασʔλͷؒʹ૬ؔΛൃݟ
  13. ৺Թܭ͸೔ʑͷߦಈྔͷมԽ͔Β৺ͷঢ়ଶΛਪఆ͢Δ  ೖྗ ग़ྗ ؾ෼ ػ 
 ց 
 ֶ

    
 श ਎ମߦಈྔ ೝ஌ߦಈྔ ࣾձߦಈྔ ਭ຾ ಛ௃நग़ ׆ྗ iPhone, AppleWatch Fitbit JINS MEME Twitter, Facebook RescueTime ਪఆ
  14. ݚڀ঺հʮ఻͑Δྗʯͷ֦ு  ൃ࿩ɾᰐ͖ɾস͍Λߴਫ਼౓Ͱਪఆ 
 'TDPSF   Τϯήʔδ౓߹͍ ൃ࿩ɾॻهɾ಺৬ Λߴਫ਼౓Ͱਪఆ

    
 ϢʔβʔඇґଘͷֶशͰ"DD 'TDPSF Chen, et al. Quantitative Evaluation System for Online Meetings Based on Multimodal Microbehavior Analysis. Sensors and Materials 34 (8), pp. 3017–3027, 2022. Watanabe, et al. EnGauge: Engagement Gauge of Meeting Participants Estimated by Facial Expression and Deep Neural Network IEEE Access, 2023 (Ealy Access).
  15. %JTDVTTJPO+PDLFZ %+  #(.ͳ͠ #(.ݻఆ #(.ಈత ఏҊγεςϜ ൃ࿩ྔʹԠ֤ͯࣗ͡ͷ1$ͷ#(.͕มΘΔ ΦϯϥΠϯϛʔςΟϯάγεςϜΛ։ൃ ಛఆͷࢀՃऀ͕஻Γ͗͢ΔͷΛ཈ࢭ

    Ͱ͖Δ͜ͱΛ࣮ݧͰ໌Β͔ʹͨ͠ H. Suzawa et al.Supporting Smooth Interruption in a Video Conference by Dynamically Changing Background Music Depending on the Amount of Utterance. UbiComp '22 Adjunct
  16. ։ൃͨ͠ΞϓϦαʔϏε  νίΫΠΠϫέϩϘ ໿ଋͷ࣌ؒʹ஗Εͨͱ͖ͷ ݴ͍༁Λݕࡧ͢ΔΞϓϦ 3FTU$BTU ࣍ʹτΠϨʹߦͩ͘Ζ͏ ࣌ؒΛ༧ใ͢ΔΞϓϦ ,PUPEBNB ৸๥Λޙչͨ͠πΠʔτΛूΊͯ

    ਂ໷ʹͻͨ͢Β౤ߘ͢Δ໎࿭#PU ਓͷߦಈ΍ੈͷதͷৗࣝΛม͑Δ΋ͷ͕޷͖ 
 ୭΋͕ϚΠφεͩͱࢥ͍ͬͯΔ΋ͷΛϓϥεʹͰ͖ͳ͍͔
  17. "*ݚڀऀʹͳΔͨΊʹ໾ཱͬͨ͜ͱ͸ʁ தֶʙߴߍ  • தֶ೥ੜͷͱ͖ʹେֶͷΦʔϓϯΩϟϯύεʹߦͬͨ • डݧ͸ઌͰ΋ʮେֶʯʮݚڀʯͷΠϝʔδΛ͔ͭΊͨ • ՝֎׆ಈʹࢀՃͨ͠ αΠΤϯεΩϟϯϓɺੜ෺ֶΦϦϯϐοΫFUD

     • ڵຯ΍ਐ࿏ͷબ୒ࢶ͕૿͑ͨ۩ମԽͨ͠ • ྭ·͠ڝ͍߹͑Δ༑ਓͱஂମઓͷडݧษڧΛͨ͠ • ୈҰࢤ๬ߍʹ͸ߦ͚ͳ͔͕ͬͨɺʮసΜͰ΋ͨͩͰ͸ى͖ ͳ͍ʯੑ֨΍ݚڀʹඞཁͳجૅӳޠྗͳͲ͕਎ʹ͍ͭͨ ·ͱΊ