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

AI倫理について

 AI倫理について

AI倫理を巡る各種議論や実際の事例について国際基督教大学で講義した際のスライドです

IKEDA Yasuhiro

May 29, 2024
Tweet

Other Decks in Technology

Transcript

  1. ࣗݾ঺հ ஑ా ହ߂ʢIKEDA Yasuhiroʣ • 2010೥ɺNTTݚڀॴʹೖࣾɻωοτϫʔΫͷ҆શͳӡ༻ʹ޲͚ͯɺػցֶशΛ׆༻ͨ͠ ҟৗݕ஌ΞϧΰϦζϜΛ։ൃɻޙʹʮAIҟৗ༧ஹݕ஌ιϦϡʔγϣϯ ˏDeAnoSʯ ͱͯ͠NTTΞυόϯεςΫϊϩδΑΓ੡඼Խ •

    2019೥ɺPKSHA TechnologyʹࢀըɻΤϯδχΞϦϯάϚωʔδϟʔͱͯ͠ɺઌਐతͳ ٕज़ݕ౼͔Βࣄۀಋೖʹ޲͚ͨ։ൃ·ͰΛखֻ͚ΔɻΫϨδοτΧʔυෆਖ਼ݕ஌Ξϧΰ ϦζϜͷ։ൃ΍σΟʔϓϥʔχϯάΛ༻͍ͨΦʔσΟΤϯε֦ுɺ޿ࠂίϯόʔδϣϯ ༧ଌϞσϧߏங౳ͷ࣮੷͋Γ • 2021೥ɺσʔλαΠΤϯεʹΑΔࣾձ՝୊ղܾΛՃ଎͢Δ΂͘߹ಉձࣾpratyaΛઃཱ 2
  2. "*ࢢ৔ن໛ͷਪҠ • ૯຿লͷௐࠪʹΑΔͱɺੈքͷAIࢢ৔ن໛͸2022೥࣌఺Ͱ໿18ஹԁͱͳ͓ͬͯΓ 2030೥·Ͱ͸Ճ଎౓తʹ੒௕͢Δͱ͞Ε͍ͯΔ • ੜ੒AIͷ୆಄ʹΑͬͯɺࢢ৔ͷ੒௕͸͞ΒͳΔՃ଎͕ݟࠐ·Ε͍ͯΔ Ø 2023೥1݄ʹϦϦʔε͞ΕͨAzure OpenAIͷಋೖࣾ਺͸1೥Ͱ2,300ࣾʹ 4

    Artificial intelligence (AI) market size worldwide in 2021 with a forecast until 2030 https://www.statista.com/statistics/1365145/artificial-intelligence-market-size/ 日本マイクロソフトが生成AIサービスの利用社数を公開し続ける理由 https://japan.zdnet.com/article/35213126/
  3. +PCJO "OOB .BSDFMMP*FODB BOE&GGZ7BZFOB5IFHMPCBMMBOETDBQFPG"* FUIJDTHVJEFMJOFT /BUVSF.BDIJOF*OUFMMJHFODF    •

    աڈ5೥ʹ౉Δɺ͋ΒΏΔࠃͷاۀɾݚڀॴɾ࣏ࣗମͷൃදͨ͠AIྙཧΨΠυϥΠϯͷ಺༰Λௐࠪ • සग़͢ΔΩʔϫʔυΛ෼ੳͨ݁͠Ռʮಁ໌ੑʢઆ໌ੑɺղऍੑʣʯʮਖ਼ٛɾެฏੑʯʮඇ༗֐ੑʯ ʮ੹೚ੑʯʮϓϥΠόγʔʯͷ5߲໨͕ɺա൒਺ͷΨΠυϥΠϯ಺Ͱग़ݱ 7 https://www.nature.com/articles/s42256-019-0088-2
  4. ಁ໌ੑʢઆ໌ੑɾղऍੑʣ 8 • AIͷ൑அʹ͍ͭͯઆ໌ɺղऍͰ͖Δ͜ͱ ü AIʹΑΔة֐ͷ཈੍ ü AIʹΑΔޮՌͷ૿෯ ü ๏తͳཧ༝

    ü AIͱͷ৴པؔ܎ߏங • AIͷಁ໌ੑ޲্ʹ޲͚ͯɺԼه͕ॏཁͱߟ͑ΒΕ͍ͯΔ ü AIγεςϜ։ൃʹ͓͚Δ৘ใͷ։ࣔ ü ιʔείʔυɺσʔλͷ։ࣔ ü ؂ࠪੑ
  5. ਖ਼ٛɾެฏੑ 9 • ʮਖ਼ٛʯͱ͸… ü ެฏੑɺόΠΞεɾࠩผͷ཈ࢭɺଟ༷ੑɺΠϯΫϧʔδϣϯɺAIͷܾఆʹର͢Δҟٞਃཱͯ͠ɾ ٹࡁાஔɺAI΍σʔλɾAIͷԸܙ΁ͷΞΫηεʹର͢Δެฏੑɺ࿑ಇࢢ৔ɾຽओओٛɾࣾձ໰୊ ΁ͷߩݙɺֶशσʔλͷଟ༷ੑ୲อɺetc. • ʮਖ਼ٛʯΛ୲อ͢ΔͨΊʹ͸ҎԼ͕ॏཁͰ͋Δͱߟ͑ΒΕ͍ͯΔ

    ü ن֨ͷࡦఆ΍໌ࣔతͳنൣͷϞσϧ૊ΈࠐΈ ü ಁ໌ੑΛ୲อ͠ɺݖར΍੍ݶʹ͍ͭͯެ։͢Δ ü ςετɾ؂ࢹɾ؂ࠪͷపఈ ü ๏ʹΑΔن੍ͷൃలɾڧԽɺҟٞਃཱͯ͠ͷݖརɾٹࡁાஔͷ֬อ ü ֤छ੍౓ͷΞοϓσʔτ ü ੓࣏తΞΫγϣϯɺֶࡍతɾଟ༷ͳ࿑ಇ؀ڥɺΠϯΫϧʔγϒͳࣾձɺརӹͷ෼഑ɺetc.
  6. .FISBCJ /JOBSFI FUBM"TVSWFZPOCJBTBOEGBJSOFTTJONBDIJOF MFBSOJOH "$.$PNQVUJOH4VSWFZT $463    10

    • طଘݚڀΛௐࠪ͠ɺػցֶशʹ͓͚Δެฏੑʹ͍֤ͭͯछఆٛΛ෼ྨ ü 23छͷόΠΞεʢྺ࢙తόΠΞεɺਓޱόΠΞεɺetc.ʣ ü 6छͷࠩผʢ௚઀తɾඇ௚઀తࠩผɺ੍౓ʹΑΔࠩผɺetc.ʣ ü 10छͷެฏੑʢ౳ՁΦοζɺػձۉ౳ɺແҙࣝʹΑΔެฏɺetc.ʣ • "ެฏͳػցֶश"ͷख๏ΛʮલॲཧʯʮதॲཧʯʮޙॲཧʯͰ෼ྨ ü લॲཧɿֶशσʔλʹ಺ࡏ͢ΔࠩผΛࣄલʹऔΓআ͘ ü தॲཧɿطଘͷֶशϞσϧʹରͯ͠ɺֶशͷաఔʹ͓͍ͯࠩผ͕ى͜Βͳ͍Α͏ ໨తม਺΍੍໿৚݅ΛՃ͑Δ ü ޙॲཧɿ௨ৗ௨ΓʹϞσϧΛֶशͨ͠ޙͰɺ༧ଌ݁ՌΛެฏੑ͕ຬͨ͞ΕΔΑ͏ิਖ਼
  7. ެฏੑͷτϨʔυΦϑͱྙཧֶ 11 • ެฏੑͷई౓ʹ͸ʮଐੑؒͰͷ֨ࠩʢImpact Disparityʣͷճආʯͱ ʮଐੑͷΑ͏ͳηϯγςΟϒͳ৘ใΛ൑அʹ༻͍ΔʢTreatment Disparityʣ͜ͱͷճආʯ ͷ2छ͕ଘࡏ* ü લऀͷྫɿ༩৴ͷείΞʹ͍ͭͯɺੑผؒͰείΞͷ෼෍͕౳͘͠ͳΔΑ͏ʹ͢Δ

    ʢެฏੑʹ͓͚Δ”Demographic Parity”ʣ ü ޙऀͷྫɿ༩৴Λ͓͜ͳ͏ϞσϧͷΠϯϓοτʹੑผͷ৘ใΛ༻͍ͳ͍ ʢެฏੑʹ͓͚Δ”Fairness Through Unawareness”ʣ • ྙཧֶͷ؍఺ͰɺલऀΛʮ݁Ռʯʹண໨ͨ͠ޭརओٛతΞϓϩʔνɺޙऀΛʮҙਤʯʹ ண໨ͨٛ͠຿࿦తΞϓϩʔνͱղऍ͢Δݚڀ΋͋Δ** ** Mougan, Carlos, and Joshua Brand. "Kantian Deontology Meets AI Alignment: Towards Morally Robust Fairness Metrics." arXiv preprint arXiv:2311.05227 (2023). https://arxiv.org/abs/2311.05227 * Zafar, Muhammad Bilal, et al. "Fairness constraints: Mechanisms for fair classification." Artificial intelligence and statistics. PMLR, 2017. https://proceedings.mlr.press/v54/zafar17a.html
  8. $IBU(15ͷٕज़ཁૉɿ5SBOTGPSNFS • 2017೥ʹGoogle BrainͷݚڀऀΒʹΑΔ࿦จ”Attention Is All You Need”ͰఏҊ͞Εͨख๏ɻ຋༁౳ ͷλεΫʹ͓͍ͯେ෯ͳਫ਼౓޲্Λ࣮ݱ ü

    GPTʢGenerative Pre-trained TransformerʣͰ͸Decoder෦෼ͷΈΛ૊Έ߹Θͤͯ࢖༻ • 2020೥ʹɺOpenAIͷݚڀऀΒʹΑΔ࿦จ”Scaling Laws for Neural Language Models”Ͱɺ Transformerͷਫ਼౓͕ϞσϧαΠζʗσʔληοταΠζʗܭࢉྔ͕૿͑Δʹ࿈Εͯ΂͖৐ଇʹ ैͬͯ޲্͠ɺͦͷ্ݶʹ͍ͭͯ͸؍ଌ͞Εͳ͔ͬͨ͜ͱ͕ࣔ͞Εͨ →ϦιʔεΛ౤͡Δ΄Ͳਫ਼౓্͕͕Γଓ͚ΔՄೳੑΛࣔࠦ 14 左図:Attention Is All You Need https://arxiv.org/abs/1706.03762 右図:Scaling Laws for Neural Language Models https://arxiv.org/abs/2001.08361
  9. $IBU(15ͷٕज़ཁૉɿ3-)' • ChatGPTͰ͸Ϟσϧͷେن໛ԽʹՃ͑ͯɺਓखʹΑΔϥϕϦϯάΛ׆༻ͨ͠ڧԽֶश ʢRLHFɿReinforcement Learning from Human FeedbackʣʹΑͬͯɺΑΓਓؒʹͱͬͯ޷·͍͠ ର࿩Λ͓͜ͳ͏Ϟσϧ΁ͱνϡʔχϯά Ø

    Step 1ɿਓखͰ࡞੒͞Εͨʮ޷·͍͠ճ౴ʯΛਖ਼ղσʔλͱͨ͠ڭࢣ͋ΓֶशʹΑΔϑΝΠϯνϡʔχϯά Ø Step 2ɿೖྗจʹର͢Δෳ਺ͷճ౴ީิʹରͯ͠ɺਓखʹΑΔϥϯΫ෇͚Λ࣮ࢪɻͦͷσʔλΛ༻͍ͯɺೖྗจ ʹର͢Δճ౴จͷϥϯΫΛ༧ଌ͢ΔReward ModelʢRMʣΛֶश Ø Step 3ɿRMʹΑΔධՁ͕࠷େͱͳΔΑ͏ͳग़ྗΛ͓͜ͳ͏ϞσϧΛڧԽֶशͰ࡞੒ 15 Training language models to follow instructions with human feedback https://arxiv.org/pdf/2203.02155
  10. ը૾ੜ੒"*ͷٕज़ཁૉʢҰྫʣ • Stable Diffusionͷߏ੒ཁૉ Ø Variational AutoencoderɿσΟʔϓϥʔχϯάʹΑΔσʔλѹॖͷٕज़ɻը૾σʔλͷѹॖʹ࢖༻ Ø Diffusion ModelɿϊΠζ͔Βը૾Λੜ੒͢ΔϓϩηεΛֶश͢ΔϞσϧɻֶशσʔλʹରͯ͠෇༩ͨ͠ϊΠζΛ

    আڈ͢ΔΑ͏ͳϞσϧΛֶशͤ͞Δ Ø Denoising U-NetɿDiffusion Modelʹ͓͍ͯϊΠζআڈʹ༻͍ΒΕΔσΟʔϓϥʔχϯάͷϞσϧ Ø CLIPɿը૾ͱςΩετΛಉ͡જࡏۭؒʹຒΊࠐΉϞσϧɻςΩετͷຒΊࠐΈʹ͸Transformer͕༻͍ΒΕ͍ͯΔ 16 High-Resolution Image Synthesis with Latent Diffusion Models https://arxiv.org/pdf/2112.10752 "a photograph of an astronaut riding a horse" Wikipedia : Stable Diffusion https://ja.wikipedia.org/wiki/Stable_Diffusion
  11. 0QFO"* l(155FDIOJDBM3FQPSUzBS9JW QSFQSJOUBS9JW   • GPT-4ͷొ৔ʹΑͬͯAI͕༷ʑͳ࣭໰ʹ౴͑ΒΕΔΑ͏ʹͳͬͨ͜ͱͰɺྙཧతͳ໰୊ʹؔ͢Δ஫໨ ͕ߴ·Δ • OpenAI͕GPT-4ͷϦϦʔεͱಉ࣌ʹެ։ͨ͠Technical

    ReportͰ͸ɺର࿩ܕAI๊͕͑Δةݥੑͱͦͷ ରॲʹ͍ͭͯղઆɻةݥੑͷྫͱͯ͠ҎԼΛڍ͍͛ͯͨ Ø ϋϧγωʔγϣϯʢݬ֮ɺͰͬͪ͋͛ʣɺϓϥΠόγʔɺ༗֐ͳ಺༰ɺ৘ใૢ࡞ɺةݥͳ૑ൃతߦಈɺetc. 17 GPT-4 Technical Report https://arxiv.org/abs/2303.08774
  12. .PUPLJ 'BCJP 7BMEFNBS1JOIP /FUP BOE7JDUPS3PESJHVFT.PSFIVNBOUIBOIVNBO .FBTVSJOH$IBU(15 QPMJUJDBMCJBT 1VCMJD$IPJDF  

     • ChatGPTͷ੓࣏తͳόΠΞεʹ͍ͭͯධՁͨ͠࿦จɻChatGPTʹରͯ͠ຽओౘʗڞ࿨ౘͦΕͧΕͷ ཱ৔ΛऔΒ্ͤͨͰͷճ౴ͱɺ੓࣏తཱ৔Λࢦࣔ͠ͳ͍σϑΥϧτͷճ౴Λൺֱͨ͠ͱ͜Ζɺ ChatGPT͸ຽओౘدΓʢࠨ೿دΓʣͷόΠΞε͕͋Δ͜ͱ͕֬ೝ͞Εͨ • ถࠃͱಉ༷ʹ੓࣏తʹೋۃԽ͍ͯ͠ΔϒϥδϧɾӳࠃΛର৅ͱͨ͠ධՁͰ΋ɺ͍ͣΕ΋ࠨ೿دΓͷ ݁Ռ͕ಘΒΕͨ • ֶशͷݩͱͳͬͨ΢Σϒσʔλࣗମ΍ɺֶश࣌ʹߦΘΕͨΫϦʔχϯά΍৘ใ௥Ճʹ͓͚ΔόΠΞε ͕Өڹ͍ͯ͠ΔՄೳੑ͕͋Δ 18 https://arxiv.org/abs/2303.08774
  13. 4PNFQBMMJ (PXUIBNJ FUBM%JGGVTJPOBSUPSEJHJUBMGPSHFSZ JOWFTUJHBUJOHEBUB SFQMJDBUJPOJOEJGGVTJPONPEFMT 1SPDFFEJOHTPGUIF*&&&$7'$POGFSFODFPO $PNQVUFS7JTJPOBOE1BUUFSO3FDPHOJUJPO • Diffusion Modelֶ͕शσʔλΛͦͷ··࠶ݱͨ͠σʔλΛग़ྗͯ͠͠·͏Մೳੑʹ͍ͭͯௐࠪ

    • ެ։σʔληοτLAIONΛ༻͍ͯධՁɻը૾ͷΩϟϓγϣϯΛؚΜͩϓϩϯϓτΛStable Diffusionʹ ༩͑Δ͜ͱͰɺର৅ͷը૾ʹࠅࣅͨ͠ग़ྗΛಘΔ͜ͱʹ੒ޭͨ͠ 19 https://arxiv.org/pdf/2212.03860
  14. ෆద੾ͳσʔλΛ༻͍ͨ"*ֶशͷ໰୊ 27 • ը૾ੜ੒AIͷֶश༻σʔληοτʮLAION 5Bʯʹ਺ඦຕͷࣇಐͷੑతٮ଴ը૾ؚ͕·Ε͍ͯͨ ͜ͱ͕ถελϯϑΥʔυେͷݚڀʹΑΓൃ֮ • LAION 5BΛֶशʹར༻͍ͯ͠ΔStability AIࣾ͸ϞσϧͷΞοϓσʔτ΍ϓϩϯϓτͷϑΟϧλϦϯά

    ʹΑͬͯ໰୊ʹରԠɻҰํɺಉ༷ʹLAION 5BΛར༻͍ͯ͠ΔMidjourney͔Βͷίϝϯτ͸ແ͔ͬͨ Stable Diffusion 1.5 Was Trained On Illegal Child Sexual Abuse Material, Stanford Study Says https://www.forbes.com/sites/alexandralevine/2023/12/20/stable-diffusion-child-sexual-abuse-material-stanford-internet-observatory/