INDEX
AIã®é²æ©ã¯æ¥é²ææ©ããã®ããããèµ°ãä¼æ¥ã®ä¸è§ããChatGPTããéçºããOpenAIã¨ãã¼ããã¼ã·ãããçµã¿ãAIæè¼ã®æ¤ç´¢ã¨ã³ã¸ã³ãBingããMicrosoft Edgeããéå¶ããç±³ãã¤ã¯ãã½ããã§ãããã¨ã«ééãã¯ããã¾ããã
ãããªå社ã¯2023å¹´7æã«ãããLongNetããRetNetãã¨ããã大è¦æ¨¡è¨èªã¢ãã«ããã®åºç¤ã¢ã¼ããã¯ãã£ã¨ãªã深層å¦ç¿ã¢ãã«ã«ã¤ãã¦ã®è«æãçºè¡¨ãã話é¡ãå¼ã³ã¾ãããæ¬è¨äºã§ã¯ããããä¸å¿ã«ãAIéçºç 究ã®æåç·ã«ã¤ãã¦ãªãã¼ããã¾ãã
2023å¹´7æ6æ¥ã«çºè¡¨ããããLongNetãã¯ãç¾å¨ã®æ·±å±¤å¦ç¿ã¢ãã«ã®ããã¡ã¯ãã¹ã¿ã³ãã¼ãã§ãããTransformerãã®çºå±åã®ä¸ã¤ã§ãã2017å¹´ã«Googleã®ç 究ãã¼ã ãã«ãã£ã¦çºè¡¨ãããè«æãAttention is All You Needãã«ã¦ææ¡ãããTransformerã
è«æã®ã¿ã¤ãã«ã§ã示ããã¦ããéãã深層å¦ç¿ã¢ãã«ã®æ§ç¯ã«ããã¦ãããã¾ã§çµã¿åããããã¦ããããªã«ã¬ã³ãããç³ã¿è¾¼ã¿ããããã¯ã¼ã¯ãæé¤ãããAttentionãæ©æ§ã®ã¿ã«åºã¥ããã¨ã§ã並åå¦çããã¬ã¼ãã³ã°ã«ãããæéã精度ã®å¤§å¹ åä¸ãéæããã®ãTransformerã§ãããä»ã§ã¯Googleã®ãBERTãããChatGPTã«ç¨ãããã¦ãããGPTããªã©ã代表çãªèªç¶è¨èªã¢ãã«ã®ãã¼ã¹ã¨ãªã£ã¦ãã¾ãã
èªç¶è¨èªå¦çã«ããã¦å¤§ããªææãéæãããTransformerãã§ãããå¦ç¿å ã¨ãªãæç« ã®é·ãï¼ã·ã¼ã±ã³ã¹é·ï¼ã«å¯¾ãã2次é¢æ°çã«è¨ç®éãå¢å ãããããå¦çã§ããããã¹ãã®é·ãã«éçããããã¨ãã課é¡ãåå¨ãã¾ãããã®è§£æ±ºã«ããã£ã¦ã2019å¹´çºè¡¨ã®Sparse Transformersãªã©ãè¨ç®ãå¹çåããææ³ãèæ¡ããã¦ãã¾ããã
ãLongNetãã¯ãã®ä¸ç¨®ã§ããããdilated attentionãã¨ããã·ã¼ã±ã³ã¹é·ã®é·ãã«åããã¦ææ°é¢æ°çã«æ³¨æã®ãã£ã¼ã«ããåæ£ããããã¨ã§ãä¸å³ã®éãå¤§å¹ ãªæ±ããã·ã¼ã±ã³ã¹é·ã®é£èºãéæããã¨ãããã¨ã§ãã
â»å¼ç¨å ï¼Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, Furu WeiãLongNet: Scaling Transformers to 1,000,000,000 TokensãâarXiv
ã覧ã®éããããã¾ã§ã®èªç¶è¨èªå¦çã¢ãã«ãæ±ãããã¼ã¯ã³æ°ï¼èªç¶è¨èªå¦çã«ãããããã¹ãã®æå°åä½ï¼ãï½æ°ç¾ä¸ä»¥å ã«åã¾ã£ã¦ããã®ã«å¯¾ãããLongNetããæ±ãããã¼ã¯ã³æ°ã¯10å以ä¸ã¨ãã¾ãã«æ¡å¤ãã§ãã
2023å¹´7æ18æ¥ã«çºè¡¨ããããRetNetï¼Retentive Networkï¼ãã¯ããTransformerãã®å¾ç¶ã¢ãã«ã¨ãã¦ææ¡ããã深層å¦ç¿ã¢ãã«ã§ããè«æãRetentive Network: A Successor to Transformer for Large Language Modelsãã§ã¯ãGPUã¡ã¢ãªã®æ¶è²»ãã¹ã«ã¼ããããã¬ã¤ãã³ã·ã¼ã¨ãã£ãææ¨ã§ç²¾åº¦ãç ç²ã«ãããã¨ãªããTransformerããä¸åããã¨ã示ããã¦ããããLongNetãã¨åæ§ã«ãæ±ããã·ã¼ã±ã³ã¹é·ã®æ¡å¤§ãå¦çå¹çã®åä¸ãå®ç¾ãããã¨ãäºæ³ããã¾ãã
ãã¦ããLongNetãããRetNetãã«ãã大è¦æ¨¡è¨èªå¦çã®çºå±ã¯ä½ãæã ã«ããããã®ã§ããããï¼
ãã®ã¤ã¡ã¼ã¸ã¨ãã¦æããããã®ããã³ã¼ãã¹ï¼æ§é åãããèªç¶è¨èªã®å¤§è¦æ¨¡ãã¼ã¿ãã¼ã¹ï¼ãã¤ã³ã¿ã¼ãããå ¨ä½ãä¸åº¦ã«å¦ç¿ã§ããããã«ãªãã¨ãããã¨ã§ããããã«ããããã£ãããããAIã質åã«çããããã«ä½¿ããæ å ±ã®éã¯å¤§ããé«ã¾ããè³ææ°ååãä¸ç¬ã§åãè¾¼ã¾ãããªã©ç®çã«åãããAIã®ãã¥ã¼ãã³ã°ã容æã«ãªãã¾ãããã®ç¯å²ã¯è¨èªå¦çã«çã¾ãããç»åå¦çãªã©ä»ã®åéã«ãåã¶ãã¨ãäºæ³ããã人工ç¥è½å ¨ä½ã®é©æ°ãæ¨ãé²ãããã¨ãäºæ³ãããã®ã§ãã
ãã¡ãããAIãç 究ããä¼æ¥ã¯ãã¤ã¯ãã½ãã社ã ãã§ã¯ããã¾ããã大è¦æ¨¡è¨èªã¢ãã«ã深層å¦ç¿ã¢ãã«ã®çºå±ã¯æ¥ã é²ãã§ãã¾ãã
ãã¨ãã°ã2023å¹´7æ18æ¥ã«çºè¡¨ãããMetaã®ãLlama2ãã¯ãGPT-3.5ã®3æ1æ¥æç¹ã«å¹æµããæ§è½ãæã¡ãªããä¸é¨ãé¤ãã¦ç 究/åç¨ã許諾ããã¦ãããªã¼ãã³ã½ã¼ã¹æ§ã®é«ããé åã§ãã
ã¾ããTransformerããã¼ã¹ã¨ããGoogleã®ãBardãã2023å¹´3æ22æ¥ã«ä¸è¬å ¬éããã対å¿è¨èªãåºããã¨ã¨ãã«ï¼æ¥æ¬èªã¯ãã§ã«å¯¾å¿æ¸ã¿ï¼ãGoogle ã¬ã³ãºã¨é£æºãç»åãåç §ã§ããããã«ãªããé³å£°ã§ã®å¯¾å¿ãéå§ãããªã©ãã¢ãããã¼ããéãã¦ãã¾ãã
2023å¹´åé ãããChatGPTãããã£ããã«ã«ããã«çãä¸ãã£ããçæAIãã¼ã ããããããAIç²ããæãã¦ããæ¹ãããããããã¾ãããããããæ¥ã æ°ããªé²å±ãå ±ãããããã®ç¶æ³ãã¾ã ã¾ã ç®ãé¢ããªãã¨ãããã§ãããã
2023å¹´7ææ«æç¹ã«ããããAIéçºã®ææ°ãã¥ã¼ã¹ããã¤ã¯ãã½ããã®å¤§è¦æ¨¡è¨èªã¢ãã«ãLongNetããä¸å¿ã«åãä¸ãã¦ãã¾ããã
2023å¹´7æ27æ¥ã«ã¯ãã¤ã¯ãã½ããããæ¥æ¬æ¿åºã¸ChatGPTæè¡ãæä¾ããããã¨ãå ±ãããã¾ãããæ©æ¢°å¦ç¿ã¸ã®ãã¼ã¿å©ç¨ã«ãããèªç±åº¦ã®é«ããããæ©æ¢°å¦ç¿ãã©ãã¤ã¹ãã¨ç§°ããããã¨ã®ããæ¥æ¬ã«å¯¾ãããAIéçºä¼æ¥ã®æ³¨ç®åº¦ã¯å°ãªããªãã¨èãããã¾ããæã èªèº«ããã®ç°å¢ãæ´»ç¨ã§ãããããAIéçºã»æ´»ç¨ã®æåç·ã«ç®ãåãã¦ããã¾ãããã
ï¼å®®ç°ææºï¼
Âã»Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, Furu WeiãLongNet: Scaling Transformers to 1,000,000,000 TokensãâarXiv ã»Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, Furu WeiãRetentive Network: A Successor to Transformer for Large Language ModelsãâarXiv ã»Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia PolosukhinãAttention Is All You NeedãâarXiv ã»å±±ä¸è£æ¯ ï¼Seamlessï¼ã10åãã¼ã¯ã³ãå¦çã§ãããã¤ã¯ãã½ããã®è¨èªçæAIãLongNetããåã®MRIããAIé³å£°åæãªã©5ã¤ã®éè¦è«æã解説ï¼çæAIã¦ã£ã¼ã¯ãªã¼ï¼ãâTECHNOEDGE ã»Introducing Llama 2âMetaAI ã»Sparse Attentionã«ã¤ãã¦åããããã解説ï¼âAGIRobots ã»å¤§è¦æ¨¡è¨èªã¢ãã«âNRI ã»Advancing AI for humanity ã»Ignacio de GregorioãMicrosoft Just Showed us the Future of ChatGPT with LongNetãâMedium ã»ãµã¤ãã¦ã±ã³ã¸ï¼ITmediaãChatGPTï¼3.5ï¼ã«å¹æµãããLlama 2ãããã¼ã«ã«PCã§åããã¦ã¿ããâAIï¼ by ITmedia NEWS ã»Google Bard ã»æ¨½äº ç§äººãGoogleã®AIãã£ãããBardããããã¹ãããã« ï½æ¯è¼ç大è¦æ¨¡ãªã¢ãããã¼ããå®æ½ãâçªã®æ ã»Microsoftãæ¥æ¬æ¿åºã«ChatGPTæè¡æä¾ãçå¼æ¡ã«æ´»ç¨âæ¥æ¬çµæ¸æ°è ã»ãæ¥æ¬ã¯æ©æ¢°å¦ç¿ãã©ãã¤ã¹ãããã®çç±ã¯èä½æ¨©æ³ã«ããâAIï¼ by ITmedia NEWS
ã¡ã«ãã¬ç»é²ããã¦ããã ãã¨ãè¨äºãã¤ãã³ããªã©ã®ææ°æ å ±ããå±ããããã¾ãã
30ç§ã§ç解ï¼ã¤ã³ãã©ã°ã©ãã£ãã¯ãåç»ã§è§£èª¬ï¼ãã©ãã¼ãã¦ã1æ¥1è¨äºãã¤ã³ããããããï¼