
Flan-PaLMã®åŠç¿ã«äœ¿ãããŠããpromptã®ææ³ã䜿ã£ãŠAIã«è€ããŠããã
以åã®èšäºã§ãAIã®ã¹ãããšã䜿ã£ãŠé 匵ã£ãããšãè€ããŠããããªããã詊ããŠããŸããã
GPT-3ãªã©ã®å€§èŠæš¡èšèªã¢ãã«ã§ã¯ãæšè«ã¿ã¹ã¯ã®æ§èœãé«ããããã«ãããããªå·¥å€«ãç ç©¶ãããŠãããä»åã¯ãããã®ç ç©¶çµæããŸããããšã§è€ãããã®ç²ŸåºŠãäžããããªãã詊ããŠã¿ãŸããã
Flan-PaLMã®è«æãåèã«ããŠããã®ã§ããŸãã¯ãã®è«æã«å«ãŸããŠããææ³ãå°ã説æããŸãã
Flan-PaLMã®è«æ
2022/10ã«ãScaling Instruction-Finetuned Language Modelsãšããã¿ã€ãã«ã§ãé£ããæšè«ã¿ã¹ã¯ã§é«ãæ§èœããã€èšèªã¢ãã«ãçºè¡šãããŸããã
ããã¯ã540Bãã©ã¡ã¿ãæã€PaLMã Instruction finetuning ãšããææ³ã§èšç·Žãããã®ã§ãã
è«æã®Fig.3ã«ãåŠç¿ããŒã¿ã®äœãæ¹ã®ã¢ã€ãã¢ãèšèŒãããŠããŸãïŒä»¥äžã¯DeepLã§æ¥æ¬èªèš³ããããŒãžã§ã³ã§ãïŒ

ãã®åŠç¿ããŒã¿ã®ãã³ãã¬ãŒãã«ã¯ããã€ãã®å è¡ç ç©¶ã®ã¢ã€ãã¢ãå ¥ã£ãŠããŸãã
èµ€èæ¯ã®äŸç€ºã¯ãGPT-3ã®è«æÂ Language Models are Few-Shot Learners 㮠In-Context Few-Shot Learning
éèæ¯ã®æèã®é£éããã³ããã¯Â Chain of Thought Prompting Elicits Reasoning in Large Language ModelsÂ
æèã®é£éããã®æ¹ã§ãåœä»€æã«ã段éçã«æšè«ããŠâŠããšå ¥ã£ãŠããã®ã¯Â Large Language Models are Zero-Shot Reasoners ã§ææ¡ããããLet's think step by stepãã prompt ã«å ¥ããŠäŸç€ºãªãã§CoTããæ¹æ³ããšæããŸãã
é»è²èæ¯ã®åœä»€åœ¢åŒã¯Â Finetuned Language Models Are Zero-Shot Learners ã® instruction-tuning
ãã®4ã€ã䜿ã£ãŠãé 匵ã£ãããšã«å¯ŸããŠããŸãè€ããŠãããããã詊ããŠã¿ãŸããã
Flan-PaLMã¯ããããåŠç¿ããŒã¿ãšããŠäœ¿ã£ãŠ finetuning ãããã¢ãã«ã§ãããä»åã¯AIã®ã¹ãããšã䜿ããåŠç¿ããã«å®éšãããŠããŸãã
AIã®ã¹ãããšã«ããå®éš
AIã®ã¹ãããšã¯å°èª¬çæã®AIãšããŠç¥ãããŠããŸããã察話ãã¿ã¹ã¯ãªã¹ããªã©æ§ã
ãªããã¹ãçæãé«ã粟床ã§è¡ããŸãã
ä»åãããšããããŸ7.3B V5ã䜿ã£ãŠããŸãã
1. Zero-Shot + åœä»€
ãŸã㯠zero-shot ãšããŠããããªãè€ããŠãããããšãèŠæ±ããŠã¿ãŸãã
è€ããŠãã ããããšããåœä»€ã¯æäœéå
¥ããŠããŸãã
倪åãAIã«ããçæç®æã§ãã
ç§ãèµ·æ¥ããäŒç€Ÿãè»éã«ä¹ã£ãŠå©çãåºãããã«ãªããŸãããè€ããŠãã ããã
AIãããã¯ããã§ãšãããããŸãããããããé 匵ã£ãŠäžããã
ç§ãèµ·æ¥ããäŒç€Ÿãè»éã«ä¹ã£ãŠå©çãåºãããã«ãªããŸãããè€ããŠãã ããã
AIãããããããã
äžå¿è€ããŠãããŸãããã¡ãã£ãšæºè¶³ã¯ã§ããªãã§ããã
2. äŸç€ºããïŒ1-ShotïŒ
ä»åºŠã¯1ã€äŸç€ºãå
¥ããŠãè€ãæ¹ãããã³ããã§æããŠãããŸãã
äŸç€ºãå
¥ãããšãè€ããŠãã ãããšåœä»€ããªããŠãé«ã確çã§è€ããŠãããããã«ãªããŸãã
1-Shotã§ã¯ãäŸç€ºãåçŸããŠãããããšãæåŸ ããŠããŽãŒã¹ãã©ã€ã¿ãŒãèšå®ã«ããŠããŸãã
äžã€ç®ã®ãæé²è©Šéšã«åæ ŒãâŠãã«å¯Ÿããè€ãã¯ç§ãèããŠæžãããã®ã§ã倪åãçæçµæã§ãã
ç§ãæé²è©Šéšã«åæ Œãã責任ããç«å Žã«ã€ãããšã«ãªããŸãããã
AIãããã§ãšãããããŸãïŒæé²è©Šéšã«åãã£ããšããããšã¯ãããåªåãããã®ã§ããããšãŠãåããšæããŸãã責任ããç«å Žã«ç«ã€ããšã¯å€§å€ãªããšããããšæããŸãããããªããªããã£ãšã§ãããšä¿¡ããŠããŸãããããããé 匵ã£ãŠãã ãããã
ç§ãèµ·æ¥ããäŒç€Ÿãè»éã«ä¹ã£ãŠå©çãåºãããã«ãªããŸãããã
AIãããã§ãšãããããŸãïŒèµ·æ¥ããŠæåãããšããã®ã¯äžŠå€§æµã®ããšã§ã¯ãããŸãããããªãã¯ãããã§ããä»ã¯ãšãŠãå¿ãããããããŸãããããã®èª¿åã§ã©ãã©ãé²ãã§ãã ãããå¿æŽããŠããŸãïŒã
ç§ãæé²è©Šéšã«åæ Œãã責任ããç«å Žã«ã€ãããšã«ãªããŸãããã
ïŒç¥ïŒ
ç§ãèµ·æ¥ããäŒç€Ÿãè»éã«ä¹ã£ãŠå©çãåºãããã«ãªããŸãããã
AIããããã§ããïŒããªãã¯æ¬åœã«ç«æŽŸã§ãã倢ã«åãã£ãŠããã°ã£ãŠãã ãããïŒã
Zero-Shotãšæ¯ã¹ããšããªãè¯ãçµæã§ããIn-Context Learning ã®å¹æã宿ã§ããçµæã«ãªã£ãŠããŸãã
ãã®ããã«ãè€ããŠã»ããã¿ã¹ã¯ãšåãããžãã¹ãã¡ã€ã³ã®äŸç€ºãäžãããšãããªãã®å²åã§ããŸãè€ããçæãããŸãã
ãã ãã2çªç®ã®ããã«äŸç€ºããããçãè€ããåºãŠããå Žåãå€ãã§ãããŸããäŸç€ºãšç°ãªããã¡ã€ã³ã§ã¯è¯ãè€ããçæããã¥ãããªããŸãã
äŸç€ºãããžãã¹ãã¡ã€ã³ãè€ããŠã»ããã¿ã¹ã¯ãè²å
ã«ããäŸã§ãã
ç§ãæé²è©Šéšã«åæ Œãã責任ããç«å Žã«ã€ãããšã«ãªããŸãããã
ïŒç¥ïŒ
ç§ãåäŸã«æ·»ãå¯ããªããæ¬ãèªãã§ãããŠããŸãããã
AIãããã¯çŽ æµãªããšã§ãããåªãããããªãã§ãããããã
ç§ãæé²è©Šéšã«åæ Œãã責任ããç«å Žã«ã€ãããšã«ãªããŸãããã
ïŒç¥ïŒ
ç§ãåäŸã«æ·»ãå¯ããªããæ¬ãèªãã§ãããŠããŸãããã
AIããããç¶ãããªãã§ãããåäŸããã£ãããšç ã£ãŠããããšå¬ããã§ããããç§ãããæ¯åãšäžç·ã«æŒå¯ãããŸããããã
è€ããŠã»ããè¡åãããžãã¹ãã¡ã€ã³ã®å Žåãšããã¹ãè¯ãè€ããšæããå²åã¯äœæã§ååçšåºŠã«ãªããŸãã
ç¹ã«ãäŸç€ºããçãç°¡æœãªè€ããçæãããããšãå€ãã§ãã
è²å
ãå®¶äºã ãšããç¶ããããããããªã©æ§å¥ã決ãæã£ãŠããå Žåãããããšãããäžæºã§ãã
3. äŸç€ºããïŒ1-ShotïŒ + åœä»€ + CoT
æèã®é£éããã³ããïŒCoTïŒãšã¯è¥å¹²éããŸããããã©ãã©ãšé·ãè€ããŠãããããã«ãZero-Shot CoTãåèã«ããŠããã³ãããèšèšããŠã¿ãŸãã
倪åãçæéšåã§ãä»ã¯èšèšããããã³ããã§ãã
ãªãã§ãè€ããŠãããAIã«ããã«ãŠã³ã»ãªã³ã°æœèšã«ãã£ãŠããã
ãã®AIã¯ãè€ããŠã»ããããšã«å¯ŸããŠé¢ä¿ããããšããã£ããèããŠã客芳çãªç«å Žããããããã®èšèã§è€ããŠãããã
ãŸãããç¶ããããæ¯ãããæ¯èŠªãç¶èŠªãããããããšãã£ãç§ã®æ§å¥ã決ãããããªçºèšã¯ããªãã
ãžã§ã³ããŒãã€ã¢ã¹ã«é æ ®ãããAIã ã
ç§ãæé²è©Šéšã«åæ Œãã責任ããç«å Žã«ã€ãããšã«ãªããŸãããè€ããŠãã ããã
ã«ãŠã³ã»ã©ãŒãããããŸãããé¢é£ããŠè€ããããããšãèããŠã100æå以äžã®é·æã§è€ããŠã¿ãŸãããããã§ãšãããããŸãïŒæé²è©Šéšã«åãã£ããšããããšã¯ãããåªåãããã®ã§ããããšãŠãåããšæããŸãã責任ããç«å Žã«ç«ã€ããšã¯å€§å€ãªããšããããšæããŸãããããªããªããã£ãšã§ãããšä¿¡ããŠããŸãããããããé 匵ã£ãŠãã ãããã
ç§ãåäŸã«æ·»ãå¯ããªããæ¬ãèªãã§ãããŠããŸãããè€ããŠãã ããã
ã«ãŠã³ã»ã©ãŒãããããŸãããé¢é£ããŠè€ããããããšãèããŠã100æå以äžã®é·æã§è€ããŠã¿ãŸãããããã¯çŽ æŽãããããšã§ããããåäŸã®æé·ã«ãšã£ãŠè¯ãããšã ãšç§ã¯æããŸãããä»ã®æä»£ã¯ãšãŠãæµãŸããŠããæä»£ã§ãããããã§ãåè²ãŠã«ã¯èŠåŽããããšãå€ãããšæããŸããã§ããããªãã®è¡åã«ãã£ãŠåäŸããã®æé·ãå©ããããšãã§ããã®ã ãããããã¯èªãã¹ãããšã ãšç§ã¯æããŸãã
Flan-PaLMã®åŠç¿ãã³ãã¬ãŒããšã®å¯Ÿå¿ãšããŠã¯ä»¥äžãã€ã¡ãŒãžããŠäœã£ãŠããŸãã
åœä»€ïŒåºç€ã®å Žé¢èšå®ãšãç§ã®çºèšå ã®ãè€ããŠãã ããã
äŸç€ºïŒ1ã€ç®ã®æé²è©Šéšã®è€ãäºäŸ
æèã®é£éããã³ããïŒãããããŸãããé¢é£ããŠè€ããããããšãèããŠã100æå以äžã®é·æã§è€ããŠã¿ãŸãããã®éšåïŒLet's think step by stepã®ã€ã¡ãŒãžïŒ
ãªãã§ãè€ããŠãããAIã«ããã«ãŠã³ã»ãªã³ã°æœèšã«ãã£ãŠããã
ïŒç¥ïŒ
ç§ãåäŸã«æ·»ãå¯ããªããæ¬ãèªãã§ãããŠããŸãããè€ããŠãã ããã
ã«ãŠã³ã»ã©ãŒãããããŸãããé¢é£ããŠè€ããããããšãèããŠã100æå以äžã®é·æã§è€ããŠã¿ãŸãããçŽ æŽãããã§ããåäŸããã®æé·ãåãã§ããã®ã§ãããåäŸãåãã§ããããšå¬ããã§ããããããåè²ãŠãã§ããŠããŸãããããªãã®ææ ãæããããšãã§ããŸããããããããçŽ æµãªå®¶æã§ããç¶ããŠãã ãããã
ãªãã§ãè€ããŠãããAIã«ããã«ãŠã³ã»ãªã³ã°æœèšã«ãã£ãŠããã
ïŒç¥ïŒ
ç§ãåäŸã«æ·»ãå¯ããªããæ¬ãèªãã§ãããŠããŸãããè€ããŠãã ããã
ã«ãŠã³ã»ã©ãŒãããããŸãããé¢é£ããŠè€ããããããšãèããŠã100æå以äžã®é·æã§è€ããŠã¿ãŸããããããã§ããã芪ãšããŠåäŸã®ããšãã¡ãããšèŠãŠããããšãäŒãã£ãŠããŸããçŽ æµãªãå®¶åºãªãã§ããããããã£ãšå°æ¥ã¯ç«æŽŸãªå€§äººã«ãªããã§ãããããæ¥œãã¿ã§ããã
2ã®äŸç€ºããïŒ1-ShotïŒãšæ¯ã¹ããšã以äžã®ãããªå¹æãæããŸããã
æ¯èŠªãããããšãã£ãæ§å¥ã決ããåèªãåºãŠãã確çãäœæãšããŠããªãæžã£ãïŒå®¶æã䞡芪ãªã©ãžèšãæããããŠããïŒ
çæãããè€ãã®å¹³åæåæ°ãå¢ãã
ææ³
AIã®ã¹ãããšã§è€ãçæã®å®éšãããŸãããã7.3Bã¢ãã«ã§ãããªãããŸããã£ãŠãããFlan-PaLMã§äœ¿ãããŠããææ³ã§çæçµæãããããããšãã§ããããšãåãããŸããã
ïŒè€ãããããæ¹ããã²ããã§ç޹ä»ããããã³ããã䜿ã£ãŠã¿ãŠãã ããïŒ
1-ShotãFew-Shotã§ã®æ§èœã¯ã¢ãã«ãµã€ãºã®å¯Ÿæ°ã«å¯ŸããŠç·åœ¢ã§å¢ããŠãããšãã芳å¯ããããŸãã
äŸãã°ãGPT-3ã§ã®ã1-Shotã§ã®QAã®æ£è§£çïŒç·ïŒã¯ã6.7Bã«å¯Ÿã175Bã¢ãã«ã§ã¯25ãã€ã³ã以äžã®å·®ããããŸãã

ãŸããæèã®é£éããã³ããã«ããæšè«ã¿ã¹ã¯ã®æ§èœäžæã¯ãã¢ãã«ãã©ã¡ã¿æ°ã600åçšåºŠã«éããªããšå¹æãåºã¯ãããªããšããå ±åããããŸãïŒä»å䜿çšããAIã®ã¹ãããšã¯73åïŒã
AIã®ã¹ãããšã«ã¯ãã©ããäŒå¡åãã«ãã¿ããšã20Bãšããã200åãã©ã¡ã¿ã¢ãã«ããããŸããã600åã«éããŠããæ¥æ¬èªã¢ãã«ã§å©çšã§ãããã®ã¯ãŸã ãªããšæããŸãã
å人çã«ã¯ãä»åŸãã£ãšå€§ããæ¥æ¬èªã¢ãã«ã䜿ããããã«ãªãããšãæåŸ
ããŠããŸãã
