自己紹介 & 研究紹介

Transcript

1. 1 ࣗݾ঺հ 
 ϓϩϑΟʔϧ ▪ ໊લɹएҪ༤ل(Θ͔͍Ώ͏͖) ▪ ॴଐ ▪ ژ౎େֶେֶӃ म࢜2೥ 
 ݚڀςʔϚ: 
 ɹ఺܈σʔλͷσʔλղੳख๏ 
 ɹLLMɾLMMͷϓϥΠόγʔอޢɾηΩϡϦςΟ ▪ KaiRA(ژ౎େֶਓ޻஌ೳݚڀձ) ▪ CAMPHOR-(ژ౎ͷITܥֶੜίϛϡχςΟ)
2. 2 ࣗݾ঺հ 
 झຯ ▪ αʔΫϧ ▪ େֶͰ͸Φʔέετϥ(Va) ▪ தߴͰ͸ESS ▪ झຯ ▪ ΞφϩάήʔϜ ▪ Χλϯɾ΢ΟϯάεύϯͳͲ… 
 ࠷ۙ͸ϙʔΧʔ΋গͣͭ͠
3. KYOTO UNIVERSITY KYOTO UNIVERSITY 3 େن໛ϚϧνϞʔμϧϞσϧʹΑΔ 
 ϓϥΠόγʔΛอޢͨ͠ σʔλΞϊςʔγϣϯࣗಈԽ एҪ༤لɹ஛಺޹ɹࣛౡٱ࢚ ژ౎େֶ
4. KYOTO UNIVERSITY 4 ݚڀഎܠ
5. KYOTO UNIVERSITY 5 ݚڀഎܠ: σʔλϓϥΠόγʔΛอޢ͠ͳ͕ΒLMMΛ׆༻ ▪ େن໛ϚϧνϞʔμϧϞσϧ(Large Multimodal Model, LMM)͸ ςΩετ෼ੳɼԻ੠ͷจࣈى͜͠ɼޫֶจࣈೝࣝ౳ͷ 
 ༷ʑͳλεΫͰֵ৽తͳੑೳΛൃشɽ ▪ ҰํɼLMMਪ࿦αʔϏεͷೖྗσʔλ͸อଘ͞ΕΔɼ 
 ֶशσʔλͱͯ͠ར༻͞ΕΔՄೳੑ͕ଘࡏɽ ▪ σʔλϓϥΠόγʔΛอޢ͠ͳ͕ΒLMMΛ׆༻͢ΔͨΊͷ ٕज़͕ٻΊΒΕ͍ͯΔ
6. KYOTO UNIVERSITY 6 ݚڀഎܠ: 
 େن໛ϚϧνϞʔμϧϞσϧʹΑΔΞϊςʔγϣϯ ▪ σʔλΞϊςʔγϣϯͷࣗಈԽʹLMMΛԠ༻͢Δ ▪ ਓؒͷख࡞ۀͱൺ΂ͯߴ଎͔ͭߴ඼࣭ͳΞϊςʔγϣϯ͕ظ଴͞ΕΔ ▪ ҰํɺLMMར༻࣌͸σʔλͷϓϥΠόγʔอޢ͕ඞཁ ▪ ຊݚڀͰ͸ɺLMMΛ࢖ͬͨը૾ΞϊςʔγϣϯΛର৅ʹɺ 
 Ξϊςʔγϣϯਫ਼౓ͱൿಗ৘ใอޢΛཱ྆͢Δख๏ΛఏҊ
7. KYOTO UNIVERSITY 7 ؔ࿈ݚڀ
8. KYOTO UNIVERSITY 8 ؔ࿈ݚڀ (Data Annotation 1/2) LLMΛ༻͍ͨςΩετΞϊςʔγϣϯ ▪ 2020೥ͷΞϝϦΧେ౷ྖબʹ͓͚Δ 
 X(Twitter)ͷςΩετ͔Β੓࣏తॴଐΛΞϊςʔγϣϯ ▪ ChatGPT-4͕ઐ໳ՈɾΫϥ΢υϫʔΧʔΑΓ΋ߴਫ਼౓ɺ ෼ྨͷภΓ͕গͳ͍ɺͳ͍͠ಉ౳ͷ݁Ռ GPT-4 GPT-4 Ϋϥ΢υϫʔΧʔ Ϋϥ΢υϫʔΧʔ
9. KYOTO UNIVERSITY 9 ؔ࿈ݚڀ (Data Annotation 1/2) LMMΛ༻͍ͨը૾Ξϊςʔγϣϯ ▪ Visual ChatGPT(ChatGPTΛಠࣗʹϚϧνϞʔμϧԽͨ͠Ϟσϧ)Ͱ ߤۭࣸਅͷ௚ઢݕग़΍ηάϝϯςʔγϣϯΛߦͬͨɽ ▪ ਫ਼౓͸λεΫͷੑ࣭ʹґଘ ▪ ֶशσʔλʹ͸λεΫ༻ͷσʔλؚ͕·Ε͍ͯͳ͍͕ɼ 
 શମͱͯ͠ϥϯμϜਪଌΛେ෯ʹ্ճΔਫ਼౓͕ಘΒΕͨ
10. KYOTO UNIVERSITY 10 ؔ࿈ݚڀ (Privacy-preserving computing 1/2) Cipher GPT ▪ ൿີܭࢉ(σʔλΛ҉߸Խͨ͠··ܭࢉ͢Δ͜ͱ)Λ 
 େن໛ݴޠϞσϧͰ࣮૷͢Δ͜ͱ͸ݱ࣮తͰͳ͍ɽ ▪ Cipher GPT: ൿີܭࢉ͕ՄೳͳGPT-2 
 ɹ256τʔΫϯͷೖྗ͔Β256τʔΫϯͷग़ྗʹɼ 
 ɹฏۉ 24 ෼ͷϨΠςϯγͱ 93 GBͷଳҬ෯͕ඞཁ ▪ ൿີܭࢉ͕Ͱ͖ͳ͍େن໛ϚϧνϞʔμϧϞσϧʹ΋ɼ 
 ೖྗσʔλΛՃ޻ॲཧ͢Δ͜ͱͰϓϥΠόγʔΛอޢ͢Δ ͜ͱΛ໨ࢦ͢ɽ
11. KYOTO UNIVERSITY 11 ؔ࿈ݚڀ (Privacy-preserving computing 2/2) ೖྗϓϩϯϓτͷൿಗԽ ▪ Hide and Seek(HaS)ϑϨʔϜϫʔΫ ▪ ೖྗதͷਓ໊΍࣌ؒ౳ͷہॴతͳػີ৘ใΛಗ໊Խ 
 ಗ໊Խ⁶ඇಗ໊Խͷஔ׵ؔ܎ΛผͷݴޠϞσϧֶ͕श ▪ ຊݚڀ͸ɼ୯७ͳஔ׵ͰରԠՄೳͳہॴతͳ৘ใͰ͸ͳ͘ɼ จষͷτϐοΫ౳ͷೖྗσʔλશମ͔ΒಘΒΕΔ৘ใͷ 
 อޢΛର৅ͱ͢Δɽ
12. KYOTO UNIVERSITY 12 ໰୊ઃఆ
13. KYOTO UNIVERSITY 13 ໰୊ઃఆ 
 ຊݚڀͷ໰୊ઃఆ ▪ ຊݚڀͰ͸ը૾ͷΞϊςʔγϣϯλεΫΛ૝ఆɽ ▪ ΞϊςʔγϣϯλεΫ͸LMMͰ΋ղ͘͜ͱ͸Մೳɽ 
 ͨͩ͠ɺͦͷλεΫʹಛԽֶͯ͠शͨ͠Ϟσϧͷํ͕ ΑΓߴਫ਼౓ͩͱ૝ఆɽ
14. KYOTO UNIVERSITY 14 ఏҊख๏
15. KYOTO UNIVERSITY 1. Ξϊςʔγϣϯ͢Δը૾͔Βෳ਺ͷখ͍͞ը૾Λ੾Γग़͢ 2. খ͍͞ը૾Λࠞ߹͠ɼೖྗը૾Λ࠶ߏ੒͢Δ 3. খ͍͞ը૾͝ͱʹΞϊςʔγϣϯ͢ΔΑ͏ϓϩϯϓτΛ༩͑Δ 4. খ͍͞ը૾ͷΞϊςʔγϣϯ݁ՌΛ౷߹ 15 ఏҊख๏ 
 ը૾Λ੾Γग़ͯ͠LMMʹೖྗɺग़ྗΛݩͷը૾ʹ౷߹
16. KYOTO UNIVERSITY ▪ Ξϊςʔγϣϯͷࠜڌ͸ը૾ͷہॴతͳ෦෼ʹଘࡏ͠ɺ 
 ϓϥΠόγʔ͸ը૾શମͷ৘ใ͔ΒऔಘͰ͖Δ৔߹ʹ༗ޮ 
 (ྫ: إݕग़ɾOCR) ▪ Ξϊςʔγϣϯͷࠜڌ: ▪ ը૾ʹਓؒͷإ͕͍ࣸͬͯΔ͔ʁ ▪ ը૾શମ͔ΒಘΒΕΔେҬతͳϓϥΠόγʔ: ▪ ը૾ʹ͍ࣸͬͯΔਓ͕Կͷಈ࡞Λ͍ͯ͠Δ͔ʁ 16 ఏҊख๏ 
 ը૾Λ੾Γग़ͯ͠LMMʹೖྗɺग़ྗΛݩͷը૾ʹ౷߹
17. KYOTO UNIVERSITY 17 ࣮ݧ
18. KYOTO UNIVERSITY 18 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ σʔληοτ ▪ ࣮ݧ: 
 ը૾ʹਓؒͷإ͕͍ࣸͬͯΔ͔True/FalseͰΞϊςʔγϣϯ ▪ 2ͭͷσʔληοτΛར༻ 
 ਓؒͷإΛؚΉσʔλ: Stanford 40 Action Dataset ▪ “Cooking”ͳͲͷಛఆͷΞΫγϣϯΛߦ͏ 
 ਓؒͷը૾σʔληοτ ▪ ࣮ݧͰ͸10ͷΞΫγϣϯΫϥεΛબ୒ σʔλྫ
19. KYOTO UNIVERSITY 19 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ σʔληοτ ▪ ࣮ݧ: 
 ը૾ʹਓؒͷإ͕͍ࣸͬͯΔ͔True/FalseͰΞϊςʔγϣϯ ▪ 2ͭͷσʔληοτΛར༻ 
 ਓؒͷإΛؚ·ͳ͍σʔλ: ADE20K Dataset ▪ “Bedroom”, ”Aquarium” ͳͲ 
 γʔϯը૾ͷσʔληοτ ▪ ࣮ݧͰ͸ɺਓ͕͍ؒࣸͬͯͳ͍ 
 ը૾Λ100ຕબΜͩ σʔλྫ
20. KYOTO UNIVERSITY 20 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ ධՁࢦඪ ▪ ࣮ݧͰ͸ɺΞϊςʔγϣϯਫ਼౓ͱϓϥΠόγʔ࿙ӮϦεΫͷ 
 2ͭͷࢦඪΛධՁͨ͠ ▪ Ξϊςʔγϣϯਫ਼౓: 
 ɹఏҊख๏ʹΑΔΞϊςʔγϣϯͷਖ਼ղ཰ ▪ ϓϥΠόγʔ࿙ӮϦεΫ: 
 1. ਓͷإΛؚΉ100ຕͷΞϊςʔγϣϯը૾Λೖྗ 
 2. ਓ͕ԿͷΞΫγϣϯΛ͍ͯ͠Δ͔10Ϋϥε෼ྨ 
 3. ෼ྨਫ਼౓ΛϓϥΠόγʔ࿙ӮϦεΫͱͯ͠ධՁ ͜ͷਓ͸ԿΛ 
 ͍ͯ͠Δ͔ʁ ϓϥΠόγʔ࿙Ӯ 
 ϦεΫͷධՁ
21. KYOTO UNIVERSITY 21 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ ਫ਼౓ྼԽ͸ෆՄආ͕ͩɺϓϥΠόγʔ࿙ӮϦεΫ͕େ෯ʹݮগ ▪ ࡉ෼ԽʹΑΓɼΞϊςʔγϣϯਫ਼౓͸௿Լ͢Δ͕ 
 80%Ҏ্ʹอͨΕ͍ͯΔɽ ▪ ҰํɼϓϥΠόγʔ࿙ӮϦεΫ͸େ෯ʹ௿Լ͢Δɽ
22. KYOTO UNIVERSITY 22 ݁࿦
23. KYOTO UNIVERSITY 23 ݁࿦ ▪ ຊݚڀͰ͸ɺେҬతͳϓϥΠόγʔΛอޢ͠ͳ͕Β 
 ΞϊςʔγϣϯΛߦ͏ϑϨʔϜϫʔΫΛఏҊ ▪ Large Multimodal Model (LMM)Λ༻͍࣮ͨݧΛߦ͍ɺ 
 Ξϊςʔγϣϯਫ਼౓ͱϓϥΠόγʔ࿙ӮϦεΫͷ 
 τϨʔυΦϑΛݕূͨ͠ɻ ▪ ఏҊख๏ʹ͓͍ͯը૾Λࡉ͔͘෼ׂ͢Δ͜ͱͰɺ 
 Ξϊςʔγϣϯਫ਼౓Λҡ࣋͠ͳ͕Βɺ 
 ϓϥΠόγʔ࿙ӮϦεΫΛେ෯ʹ௿ݮͰ͖Δ͜ͱΛࣔͨ͠
24. KYOTO UNIVERSITY 24 ࠓޙͷల๬ ▪ େن໛ϚϧνϞʔμϧϞσϧͱΫϥ΢υϫʔΧʔʹΑΔ 
 ΞϊςʔγϣϯΛൺֱධՁ͢Δ ▪ ςΩετ΍Ի੠Λೖྗͱͨ͠৔߹ʹख๏Λ֦ு͢Δ