Summilux 1.4/50 ASPH, Leica M10P @Chinkokuji Temple, Munakata, Japan
昨日のLLM(large language model 大規模言語モデル)議論ã®ç¶ããããå°ãæ¸ãã¦ã¿ããã¨æãã
DSåä¼ã®ã¹ãã«å®ç¾©å§å¡ä¼ã§ã¯IPAã¨ååãã2å¹´ã«ä¸åº¦ããã¼ã¿ãµã¤ã¨ã³ãã£ã¹ãã®ã¹ãã«æ¨æºãè¦ç´ããæ¹è¨çãçºè¡¨ãã¦ãã*1ãä»å¹´ã¯å¥ãããæ¹è¨å¹´ã ã£ãã®ã ãããã®æ¥ããããæ°ã¶æåã«è¯ã ããç»å ´ããChatGPTãç®ã®åã«ãã¤ã¤ãçæAIé åã«ããã¦ãã¼ã¿ÃAIãããã§ãã·ã§ãã«ï¼ãã¼ã¿ãµã¤ã¨ã³ãã£ã¹ã DSï¼ã®å ´åãæ±ããããã¹ãã«ã¯ã©ããããã®ããã¨ããè°è«ãéåã¨ããã
データサイエンティスト協会 10thシンポジウム スキル定義委員会発表資料(2023年10月20日)
çæAIã¯ãé³å£°ãé¡ç»åã®ãããªæ å ±ã®èå¥ãè¡ãAIã§ã¯ãªããå¦ç¿ããçµæã使ã£ã¦ãä¸ããæ å ±ã«è¿ããæ å ±ãç´¡ãã§ãå½¢ãè¨èãé³æ¥½ãªã©ãçã¿åºãAIã®ãã¨ã ãLLMã ãã§ãªããDeepfakeãçã¿åºããGAN*2ããæ¡æ£ã¢ãã«ã«åºã¥ãæ§ã ãªè¨èãã¤ã¡ã¼ã¸ã«ãã¦ãããmidjorney, DALL·E 3ãªã©ãããã«å½ããã
ã¡ãªã¿ã«éå以åï¼7å¹´åï¼ã®ããã°ã¨ã³ããªã§æ¸ããéããAIï¼artificial intelligence 人工ç¥è½ï¼ã¯ããã¾ã§ã¤ãã¢*3ã§ããæè¡ã§ã¯å®ç¾©ããã¦ããªããå¤æãä¼´ãæ å ±å¦çãèªååãããã®ã¯ãã¹ã¦AIã§ããã10å¹´è¿ãåã«æ¾å°¾è±å çã¨è©±ããè¨æ¶ããããããµã¼ã¢ã¹ã¿ããã§ãããã種ã®AIã¨è¨ãããæ©æ¢°å¦ç¿ï¼LLMãçã¿åºãã深層å¦ç¿ããã®ä¸ã¤ï¼ãèªç¶è¨èªå¦çæè¡ããªãã¦ãAIãªã®ã ãã¨ã¯ããå®å ¨ã«èªååãããã¨äººã¯AIã¨å¼ã°ãªããªãå¾åãããããã¹ã©ã«ä¹ãã¨ããªã¢ã«ã¿ã¤ã ã§å¨å²ã®ã¯ã«ãï¼è»ç¨®ã¿ã¤ããï¼ããã¼ã«ãèªèãã¦ãããã¨ã«é©ãããããã¨ã¦ãã¾ããªãAIã ã¨èª°ãè¨ããªããªãã ãããDeepLãé©ç°çãªç¿»è¨³åãèªãããèªå翻訳ãæ©æ¢°ç¿»è¨³ã¨å¼ã°ããã ãã§AIã¨ãã人ã¯é ãããå°ãªããªãã¨æãããã
話ãæ»ãããåããã¹ãã«å®ç¾©å§å¡ä¼ãã®çµè«ã¯ãããã®DSã«ã¨ã£ã¦çæAIã使ãããã©ããã¯ã¤ã·ã¥ã¼ã§ã¯ãªããä¼ç»ããåãä½ãã»å®è£ ããåããåããããã®ã«ãªããå¾æ¥ã®DSã¹ãã«ã«å ããAIå©æ´»ç¨ã¹ãã«ã¨ããã¹ããã®ãå¿ è¦ã«ãªããã¨ãããã®ã ã£ããã»ã¨ãã©ã ãã§ãã¤ãããã®ãçæAIã®ç¹å¾´ã ããã ã詳ããã¯çºè¡¨è³æãè¦ã¦é ããã°ã¨æãããä¸é¨æç²ãã¦ãããã
データサイエンティスト協会 10thシンポジウム スキル定義委員会発表資料(2023年10月20日)
ï¼ãªããããã«æ±ããããAIå©æ´»ç¨åã®ã¡ã¤ã³ã§ãããä¼ç»ããåãä½ãã»å®è£ ããåãã¯ç¸å½ã«ãã£ã¼ãã§ãããèæ¯ç解ã®ä¸ããã¸ãã¹èª²é¡ããããã¯å¤¢ãå½¢ã«è½ã¨ãã¦ããåãæ±ããããããã¨ãã°ã¤ã¼ãã³ã¿ãããªã¯ã¤ã«ãã§é¢ç½ãåµæ¥è ããã£ã¦ãã¦ãè²ã 夢ãèªã£ãã¨ãã«ãããå½¢ã«ããã¤ã¡ã¼ã¸ãæã¦ããã¨ãããã¨ã ãã»ã¨ãã©ã®äººã«ãªãarchitectçãªã¹ãã«ã ãããããããäºæ¥ããµã¼ãã¹ãå½¢ã«ããèã§ãããã¨ã¯æããã ããã®ä¸ã§ãå½ç¶ã®ãã¨ãªãããåã¢ãã«ãfine-tuningãå«ãã¦å©æ´»ç¨ã§ããon-goingã§ç¶æ³ãã©ãã©ãã¨å¤ããä¸ãæè¡ãå«çãæ¨é²èª²é¡ãç¾å ´ã§è§£æ±ºãã¦ããåãå¿ é ã¨è¨ãããå ·ä½çã«ã¯ãä¸ã®ãªã³ã¯å ãåã³ãæ¬ç¨¿ã®æå¾ã«ãªã³ã¯ã®ããã¹ãã«ãã§ãã¯ãªã¹ããã¿ã¹ã¯ãªã¹ãããåç §ï¼
-
ãããã£ã¦ãã»ã¨ãã©ã®äººã«ã¨ã£ã¦ãLLM*4ã使ãããã©ããã¯ã¤ã·ã¥ã¼ã§ã¯ãªããPromptingã¨å¼ã°ããChatGPTã¸ã®æ示ã®åºãæ¹ã«ã¤ãã¦ã®ç¹éè¨äºããã³ã³ããã«ä¸¦ã¶éèªã§ãããçµã¾ãã¦ãããããã種ã®è¶çªã¨ãè¨ãããè¨èã§ããã°ã誰ã ã£ã¦ãããªãã«ä½¿ããã«æ±ºã¾ã£ã¦ããããã ããã®ä¸ã使ãæ¹èªä½ãChatGPTã«èãã°ãããªãã«æãã¦ãããã*5
ææ§ã«ããæ示ãã§ããªã人ã¨ãã¡ããã¨æ示ã§ãã人ã®éãããã¾ããã¦ãããã§åããã¦ãããçµå±äººã«ãã®ãé ¼ãã¨ãã¨åãåé¡ãããã§ãçã¾ãã¦ããã¨ãããã¨ã ã空æ°ãèªãã§è¡åããããã¨ãããããªæ¥æ¬çãªæ¯ãèãã¨ãããè¡åè¦ç¯ã¨ããã®è¨èªåã«ããæ示åºãï¼promptingï¼ã¯ç¸å½ã«ä¹é¢ãã¦ãããLLMã®ç»å ´ã«ãããã¤ãã«è¨èªåè½åããã¾ãé«ããªãã¨æãããæ¹ã ã®è¨èã«ããåãé«ã¾ãæãæ¥ã¦ããã¨ããè¦æ¹ãã§ããã
-
人éã®ç¥æ§ã¯æ¦ããè¨èã®ãã³ããªã³ã°è½åããã¸ã¥ã¢ã«åè½åãä½ããå½¢ã«ããæã®åããã®è¨èãã¤ã¡ã¼ã¸ãçã¿åºãã使ããã ãã®ä¸ã®ä¸ãå¿ã®ä¸ã®æ·±ãç解åï¼ç¥è¦ï¼ã«åºã¥ãã¦ããã6å¹´ååããã¼ãã¼ãã»ãã¸ãã¹ã»ã¬ãã¥ã¼(DHBR)ã®ç¥æ§ã®æ ¸å¿ã«ã¤ãã¦ã®è«èã§ã¾ã¨ããéããæ å ±ã¯å¤ããå ¥ã£ã¦ãã¦ãå¦çãããåºã¦ãããããã®inputã¨outputãã¤ãªãé¨åãããããæ å ±å¦çã ãããã俯ç°ããã¨ãç¥æ§ã¨ã¯ãªã«ããã¨ããåãã«ã¤ãã¦ã®çãã¯ããªãã·ã³ãã«ã§ãæ å ±å¦çã«ãããã¤ã³ãããã¨ã¢ã¦ããããï¼I/Oï¼ãã¤ãªãåã¨è¨ãæãããã¨ãã§ããããã®ããã«å®ç¾©ãã¦ãã人ã¯èªå以å¤ãã¾ãè¦ããã¨ããªãããããã¯ãã©ããã«è¦ãã¨å¦å®ãé£ãã ããã
ãã®ä¸èº«ãè¦ãã°ãæè¦ã®ãããªå¤é¨æ å ±ãä½æã§ãããã®ã¸ã®å¤æããã¯ãã¾ãã対象ãã³ã³ããã¹ããè«çã®ç解ã¨è¨ãã¹ãããªãé«åº¦ãªæ å ±å¦çãé£ãªããã¾ã並è¡ãã¦ç¸å½ã®ã¬ãã«ã¾ã§èµ·ããããããã¯ç·åãã¦ãç¥è¦ãã¨ããã¹ããã®ã§ãã*6ããããæ å ±å¦çãããªãã¡ãç¥æ§ã¨ããã¹ããã®ã®å¤§åãå ãããå®éã®ã¨ãããLLMãèªåèµ°è¡ã®åæã¨ãªãcomputer visionãå«ãç¾å¨ã®äººå·¥ç¥è½ã®ãããã®å¤§åã¯ãç¥è¦ãé¨åã®å¼·ãã«ãããã®ã ã
å®å®
å人ãããããã®äººæè²æãèããã令å5å¹´3æ24æ¥
文部科学省 今後の教育課程、学習指導及び学習評価等の在り方に関する有識者検討会(第3回)
話ãæ»ãã¨ã人éã®å ´åããã®ãç¥è¦ãã¯ç¥çä½é¨ã人çä½é¨ãæç´¢ä½é¨ã®ï¼ã¤ã«ãã£ã¦çã¿åºããããã¨ããã®ããã®æã®è«èã®ããªãããã¤ã¾ãã è°è«ã®è骨ã ã£ãããã詳ãããã®è°è«ã«è§¦ãããããæ¹ã«ã¯ãç´æ¥ãã·ã³ã»ããã³ãã®è©²å½ç®æããããã¯ãªãªã¸ãã«ã§ããDHBRã®è«èãèªãã§é ããã°ã¨æãã
ãã®è¨èã«ãããã¨ããé¨åã¯ç¥è¦ã®ã¢ã¦ããããè¡åã«å½ãããè¨èã«ããä¸ã§ãåãããµãã£ã¨æããä½ãããã種ã®ã·ã³ãã«åããããªã«ãå®ä½åãããã¾ãã¯çåãè¨èã«ãããã¨ããã§ããã®ã ããããããªã«ãå¾ä½ã®ãããªããã¨ããã¾ãè¨èã§ã¤ãããã¨ããç¬éãããã ããã
æã ãç¥æ§ãæããè¡åããããã¯ç¥çã ã¨èããè¡åã®ä¸ã¤ã«ãç¥è¦ãããé©åãªè¨èãä¸ãããã¦ããªããªã«ãã«è¨èãä¸ããè¡åãããããã¾ãã«ããã ãå°èª¬å®¶ã¯å¸¸ã«ããã§åè² ãã¦ããããã ããå¸äºã®åãã®å¨ãã®é¢ã ã次ã ã«è©¦ã¿ã¦ããããã¨ã¢ãããã¾ã£ãããã¨ãã£ãè¨èãã·ã§ããããã®æå³ã人ã«ä¼ãããã¨ã¯å¯è½ã ããããç¸å½ã«é£ããã¯ãã ããããã®è¨èãçã¿åºãã人ãã¡ã¯å大ã ã
æ並ã¿ãªè¨èã§ãã話ããªã人ã¨ããã®äººãªãã®è¨èã§è©±ãã人ã®éãããã¾ãã¯ãã®äººãªãã®ç¥è¦ã®æç¡ã«ãããã¨ã¯æããã ããå®ã¯ãã®ä¸ã§ããã®æãã¦ããä½ããä½ããã®è¡¨ç¾ã¨ãã¦å½¢ã«ããåã®å·®ã«ããããã¨ããã®çµé¨å¤ãç¥è¦ãã®ãã®ã®ã¬ãã«ã«ç¸å½ã®å½±é¿ãä¸ãã¦ãããã¨ã«æ°ã¥ãã¦ãã人ã¯å°ãªãã
å ã®è«èã§ã¯ãç¥è¦ã磨ãä¸ã§å¤§åãªãã¨ã®ä¸ã¤ã¨ãã¦ãè¨èã«ãªã£ã¦ããªããã®ãä¸ã®ä¸ã®å¤§åã§ãããã¨ãèªãããã¨ãããã¦æãããã¨ãè¨èã§ãçµµã§ãããããã¢ã¦ãããããããã¨ã¨æ¸ããã®ã ããã¾ãã«ãããä»æ±ãããã¦ããã¨æãã®ã¯æ°ã®ããã§ã¯ãªãã ãã*7ã
ps. å年以ä¸ãã»ã¼æ¯é±ã®ç±çãªè°è«ã¨ãåã¹ãã«å®ç¾©å§å¡ã®ããªãã®æéæä¸ãçµã¦çã¾ããã¹ãã«ãã§ãã¯ãªã¹ã ver.5ã¨ã¿ã¹ã¯ãªã¹ã ver.4ã¯ãã¡ã
www.datascientist.or.jp
ãªããããããDSåä¼10thã·ã³ãã¸ã¦ã ã§çºè¡¨ããæ(2023.10.20)ã®åç»ã¯ãã¡ãã
www.youtube.com
*1:ä»åã®ã¹ãã«ãã§ãã¯ãªã¹ãã¯version5ãã¿ã¹ã¯ãªã¹ãã¯version4ã«ãªãã
*2:Generative adversarial networkãæµå¯¾ççæãããã¯ã¼ã¯
*3:ä¸è¬çã«ã¯ç念ãWikipediaã«ããã°non-physical, timeless, absolute, and unchangeable essences of all things ï¼è¨³ï¼éç©ççã§ãæéãè¶ è¶ãã絶対çã§ãä¸å¤ãªä¸ç©ã®æ¬è³ªï¼ Theory of forms - Wikipedia
*4:ChatGPTã®ãããªLLMã¢ããªãå«ã
*5:ç¾å¨ãpromptingã®æã«ãã使ãã度åãã®å·®ãããã®ã¯äºå®ã ããããããã®å¿ è¦æ§ã¯ç¸å½ã«å¼±ã¾ã£ã¦ããã¨æ¨å®ãããã
*6:ãæè¦ããå¥ãåã£ããèªç¥ãã¨ããè¨èãããããå®éã«ã¯æè¦èªä½ãããªãé«åº¦ãªæ å ±å¦çã§ããããããå«ãã¦ãç¥è¦ãã¨ãã¦èããã¹ãã ã¨ããã®ãåã®èãã ãä½ãããã®ä¸ãç©ççãªä¸çãã«è²åããé¦ããè触ããåå¨ããªããè³å ã§ã®æ å ±ã®ã¤ã¡ã¼ã¸ã§ãã質æã«ãããªãã®ã ããã
*7:ãã®æ稿ã®ä¸ã«ç¾ããè¨èãããã£ã¼ãããããã£ã¦çã¿åºããã¦ãããã®ã§ãããåèªèº«ãä¸ã¤ä¸ã¤ã«èªåèªèº«ï¼ããã¦ä»²éãã¡ï¼ã®å¿ã®å身ã¨æããããæããã